Integration of AI in Self-Powered IoT Sensor Systems

Abstract

The acceleration of digitalization has caused an increase in demand for autonomous devices. In this paper, the technologies of artificial intelligence (AI), and especially machine learning (ML), integrated into applications that use self-powered Internet of Things (IoT) sensors are analyzed. The study addresses the issue of the lack of a standardized classification of IoT domains and the uneven distribution of AI integration in these domains. The systematic bibliometric analysis of the scientific literature between 1 January 2020 and 30 April 2025, using the Web of Science database, outlines the seven main areas of IoT sensor usage: smart cities, wearable devices, industrial IoT, smart homes, environmental monitoring, healthcare IoT, and smart mobility. The thematic searches highlight the consistent number of articles in the health sector and the underrepresentation of other areas, such as agriculture. The study identifies that the most commonly used sensors are the accelerometer, electrocardiogram, humidity sensor, motion sensor, and temperature sensor, and analyzes the performance of AI models in self-powered systems, identifying accuracies that can reach up to 99.92% in medical and industrial applications. The conclusions drawn from these results underscore the need for an interdisciplinary approach and detailed exploration of ML algorithms to be adapted to the hardware infrastructures of autonomous sensors. The paper proposes future research directions to expand AI’s applicability in developing systems that integrate self-powered IoT sensors. The paper lays the groundwork for future projects in this field, serving as a reference for researchers who wish to explore these areas.

1. Introduction

Sensors initially used as measuring elements have become components of intelligent systems in the contemporary world through an Internet of Things (IoT) approach. Thus, IoT sensors provide real-time data in various physical environments, which is made possible by combining artificial intelligence (AI) technologies with other modern data acquisition and processing technologies. IoT sensors are processed through AI algorithms to make automated decisions for future predictions or to optimize various processes across many fields. This review explores how the convergence between IoT sensors and modern AI technologies contributes to transforming the surrounding world through an intelligent digital network.

The technological evolution of recent years has taken place under the simultaneous development of self-powered sensors, AI algorithms, and IoT networks. However, these technologies have been treated in the specialized literature as separate components. The present paper aims to analyze the synergistic collaboration between these components. This study provides an overview of the current state of research and trends at a moment of technological transition that impacts various fields of activity.

Self-powered IoT systems that use energy harvesting technologies represent a challenge in the context of AI due to constraints in energy, space, processing, connectivity, and communication. In the case of minimalist hardware systems that have self-powered sensors, AI algorithms based on ML require computational resources that are usually insufficient at the level of available hardware. Under these conditions, devices in this category have the ML component hosted in the cloud. In order to access cloud resources, the devices require Internet connectivity. Instability of energy sources, such as variation in vibration, illumination, or movement, affects the reliability of the continuous execution of these algorithms. For these reasons, the field is insufficiently explored, given that the majority of research addresses individual elements. For example, some works focus on optimizing energy consumption, while others address the performance of AI models, without analyzing the necessary trade-off between the two. On the other hand, the lack of integration standards and the difficulties in testing under real conditions delay the maturity of this technological intersection. Therefore, a systematic literature review addressing these components simultaneously will allow readers to understand current limitations and identify viable directions for further development.

The present paper has as its main scope to carry out a bibliometric and thematic analysis of the way in which self-powered IoT sensors interact with AI technologies, especially machine learning ones. The paper does not limit itself exclusively to hardware components or to AI algorithms in isolation. The paper highlights trends in the literature, focusing on hardware–software integration challenges together with application directions within IoT contexts. The paper addresses both technological aspects, mentioning types of sensors, energy sources, and ML models, as well as the functional framework in which these are implemented. The purpose of such an approach is to provide the reader with an overview of the maturity and viability of these solutions in real-world applications.

By analyzing the specialized literature, it was found that domains group sensors, and within each domain, there is a taxonomy of sensors based on the area in which they are used and the specific type of sensor. According to Rozsa et al. [1], the classification includes the industrial domain, smart cities, and healthcare. The same paper states that the industrial field comprises the agricultural, logistics, and plant floor domains, while smart cities include transport, buildings, and environment, and healthcare includes monitoring and management. Javaid et al. [2] take a different approach to classifying sensors, emphasizing a classification based on their size and purpose. Thus, sensors are classified according to Industry 4.0 into nanosensors, microsensors, nuclear sensors, and passive sensors.

In the past, sensors represented measuring devices. With the evolution of technology, they have become active components of an innovative ecosystem that, in most cases, is connected to an infrastructure that allows real-time data transmission in the context of automation. Therefore, sensors are mostly treated from the perspective of IoT applications as measuring devices with advanced capabilities. According to Rehman et al. [3], IoT applications are divided into seven domains: transportation, smart homes, markets, agriculture, healthcare, and smart cities.

Naresh and Lee [4] provide a rigorous classification of sensors based on most of the relevant criteria for the automation field. Thus, sensors are classified based on the type of physical quantity measured, dividing them into acoustic, electrical, optical, biological, magnetic, mechanical, chemical, thermal, and radiation sensors. Furthermore, in the same research [4], sensors are grouped into active and passive based on their operating principle. Also, depending on the physical contact typology, sensors are categorized into contact and non-contact sensors. Sensors can also be classified into analog and digital sensors based on signal output. Also, in the research by Naresh and Lee [4], sensors are organized based on signal conversion into biological, physical, and chemical sensors. Additionally, sensors can also be grouped based on application, specificity, and comparability. In the context of IoT, the classification based on application is considered the reference. Thus, Naresh and Lee [4] identify the following industries as domains in which sensors operate: agriculture, automotive, engineering, domestic appliance, information and telecommunication, health and medicine, marine, defense, space, transportation, energy, and manufacturing.

Santos et al. [5] divide the specific domains of smart cities into energy, education, traffic, transport, city services, infrastructure, health, home, and environment. Still, upon closer analysis, it can be observed that these elements are standalone domains.

Analyzing the matters from the specialized literature, it is found that most sensors are classified based on their field of applicability. Still, this approach is disparate, lacking a standard classification to which most works can refer. Moreover, as already mentioned, most modern sensors are of the IoT type. Under these conditions, the issue arises regarding the level of research on AI technologies for IoT sensors in each field of activity. This research aims to answer the following research questions (RQs).

• What IoT domains are not explored in the context of AI technologies, especially those based on machine learning (ML)?
• What future research directions should focus on IoT sensors, AI technologies, and the integration of ML in applications that use these sensors?
• What field should be the primary focus of interest for researchers to accelerate the evolution of AI techniques applied in combination with self-powered IoT sensors?

To answer these questions, the authors have the following contributions:

• Identifying IoT domains in the literature through a Web of Science (WoS) analysis to propose standardizing sensor classification domains with IoT technology.
• Distinguishing the reference sensors for each IoT domain and the level of interest of researchers in these types of sensors over the past five years, namely 2020–2025.
• Exploring how many works in the specialized literature address each type of sensor in explicit relation to IoT for the previously mentioned period.
• Adding restrictions to sensor analysis by incorporating AI constraints to outline how research on IoT sensors has been reported about AI technologies in the last five years.
• Narrowing down to target a specific direction of AI that focuses on ML algorithms. For this, the number of articles generated for identified sensors in combination with IoT and ML is searched in the specialized literature.

The research analyzes the papers published in the specialized literature between 1 January 2020 and 30 April 2025 through a search in WoS. Each section also included a detailed analysis of the specialized literature beyond the statistical component, whose role is to identify researchers’ interest in a particular area. The direct consequence is identifying areas that require further research in the future, outlining a series of directions for researchers interested in the field of IoT sensors concerning AI technologies.

The paper is structured into nine sections. Section 2 presents the methodology, and Section 3 synthesizes the specialized literature regarding IoT domains related to sensors specific to each field. Section 4 presents the classification of sensors, whereas Section 5 introduces the classification of self-powered IoT sensors. Section 6 analyzes the self-powered IoT sensors in the AI field, and Section 7 details ML implications in the self-powered sensors IoT domain. Section 8 discusses and identifies the answers to the research questions, the limitations of the study, and the future development directions of the identified areas. Section 9 is associated with the conclusions of the paper.

2. Methodology

To answer the RQs, this paper analyzes specialized works through a detailed literature review and subsequently conducts a systematic analysis based on quantitative and qualitative research methods. Beyond the specialized literature review, the paper employs WoS analysis to identify IoT domains. Figure 1 presents the overall workflow of this study. For this, a search is conducted in this database corresponding to the period between 1 January 2020 and 30 April 2025, using the query TS = (“IoT sensors”) AND PY = (2020–2025), with the primary goal of identifying the most frequent domains associated with IoT sensors. The authors classified the fields into smart cities, wearable devices, industrial IoT, smart homes, environmental monitoring, healthcare IoT, and smart mobility.

Subsequently, the sensors were established according to the measured input, and articles specifically addressing these types of sensors in the context of IoT were identified. The literature review on the link between IoT sensors and AI was conducted based on the query TS = (“self-powered sensing” AND (“artificial intelligence” OR “AI”)) AND PY = (2020–2025). The models in which IoT sensors are addressed in the scientific literature concerning AI were evaluated at this stage. The thematic and bibliometric analysis conducted using the VOSviewer 1.6.20 software highlighted an analysis of the co-occurrence of keywords extracted from the selected articles. This allowed for identifying dominant thematic clusters to highlight the connections between the concepts of self-powered sensors and AI technology. The results are implemented in a detailed framework where the studies are compared regarding performance indices.

Future research directions are identified along with recommendations for technological development in the field of self-powered IoT sensors, and ML, through additional filtering that includes only works that explicitly address ML algorithms (TS = (“self-powered sensing” AND (“machine learning” OR “ML”)) AND PY = (2020–2025)). Thus, the publications were analyzed using the algorithms and the reported performance metrics.

A schematic of the WoS filtering strategy used to select and categorize the papers is presented in Figure 2.

Figure 2 presents the workflow diagram illustrating the WoS filtering strategy. The methodology includes three sequential queries to extract articles on IoT sensors, followed by filters related to AI and ML. This approach enables a targeted bibliometric and thematic analysis of research trends between 2020 and 2025.

The search strategy is based on the combinations of key terms presented in Figure 2. These terms were extracted by the authors of this paper from the specialized literature and included in the queries performed in WoS. These terms were combined using the logical operators AND and OR to generate detailed searches in the specialized literature regarding self-powered sensors. The authors opted for the WoS database because it has a broad coverage of indexed sources and offers the possibility of extracting metadata for bibliometric analysis. Methodological limitations include the following:

• The possible exclusion of works indexed only in alternative databases such as Scopus or IEEE Xplore;
• The volatility of citation numbers for recent work;
• The fact that the analysis is based exclusively on articles published in English;
• The inclusion of review articles in the analysis.

To synthesize the interaction of self-powered sensors within the context of IoT domains, Figure 3 proposes a conceptual framework for integrating self-powered sensors into IoT applications that incorporate AI components. In this model, self-powered sensors act as autonomous data generators. They collect energy from the environment and transmit multimodal data to a cloud service. AI modules interpret these streams using ML algorithms. These algorithms extract patterns and perform classifications or predictions based on data that has been previously pre-processed. The extracted information is applied in IoT domains to make context-based decisions regarding autonomous operation.

The framework presented in Figure 3 highlights the feedback loop in which domain-specific requirements influence the selection of the AI model. Additionally, hardware constraints influence the performance of the AI component in sensor optimization.

The paper includes all articles from 2020 to 2025 that integrate IoT sensors and AI components, regardless of whether they utilize TinyML, on-device learning, or self-powered edge intelligence. Regardless of the hardware configuration in which AI components are integrated with self-powered sensors, the works in the literature are extracted within the bibliometric analysis. A total of 230 scientific articles were extracted from WoS and included in this analysis.

In order to ensure a structured approach, the selected works were classified according to three major axes of analysis:

• The AI method used (regression, classification, neural networks, classical machine learning, TinyML, etc.);
• The energy source of the sensors (triboelectric, piezoelectric, biochemical, thermal, etc.);
• The IoT application domain (healthcare, smart cities, industry, agriculture, environment, etc.).

Through these three axes established by the authors in this analysis, it is highlighted how AI methods integrate differently depending on the energy constraints.

3. Domain Identification in the Literature Using IoT Sensors

For a rigorous analysis of the specialized literature, WoS searches were conducted for 2020–2025, searching for all review articles focusing on IoT sensors. From the analysis of the specialized literature, seven reference domains were identified in which IoT sensors are used. The seven domains are smart cities, wearable devices, industrial IoT, smart homes, environmental monitoring, healthcare IoT, and smart mobility.

Regarding the field of smart cities, most papers discuss the use of IoT sensors in traffic monitoring, urban infrastructure, the development of parking at the urban level and in airports, and the transformation of urban services through modern technologies specific to AI and big data. Refs. [6,7] analyze how smart cities provide better air quality, improving the population’s health. It monitors early disease detection and personalized treatments in real-time through IoT sensors connected to mobile applications. Bhattacharya et al. [8] study the integration of IoT sensors in monitoring traffic, drainage, and urban mobility by integrating AI technologies, specifically deep learning (DL), a component of ML. Following these studies, Hornik et al. [9] collect data from IoT sensors, focusing on mobility and infrastructure. Also at the level of AI technology, Gheorghe and Soica [10] detail the transformation of urban traffic from the perspective of IoT and AI, and analyzes optimization methods regarding the operation of traffic lights and the response to various emergencies. Refs. [11,12] conduct systematic analyses in the specialized literature regarding smart parking systems for identifying free spaces using cameras that provide data to ML algorithms. Similarly, Koroniotis et al. [13] examine a perspective on smart airports by applying IoT sensors for the automation, security, and real-time monitoring of airport traffic. Mohammadzadeh et al. [14] analyze the impact of smart cities on public health using IoT sensors and mobile applications to improve access to medical services. Also, in an AI approach, Koumetio Tekouabou et al. [15] and Rosca et al. [16] discuss modeling urban forms to address challenges related to health, security, mobility, and other aspects that affect the population’s quality of life.

Wearable devices have generated many articles in the literature that analyze the role of self-powered wearable sensors. Health monitoring applications, such as fall detection and arterial pulse, step count monitoring, and heart rate monitoring, have facilitated the study of IoT sensors concerning AI technologies, simultaneously emphasizing energy autonomy [17]. In this context, Xi et al. [6] analyze self-powered wearable sensors and TENG technologies for medical applications and human–machine interfaces. In continuation of this research, Peng and Du [18] present a detailed analysis of TENGs in health monitoring applications, specifically targeting fall detection and pulse monitoring. The discussion on energy conversion in TENG systems for wearable devices is identified by Peng and Du [18]. It mentions the conversion technologies and their performance. Another reference work for wearable devices concerning IoT is the research by Aslam et al. [19], which analyzes the applications of wearable sensors in health, identifying a series of gaps, including energy efficiency [20].

The Industrial IoT field investigates predictive maintenance (PdM) approaches, cybersecurity for industrial infrastructure, and modern technologies such as ML and IoT sensors to ensure process safety. Cacciuttolo et al. [21] use a communication protocol to transmit IoT-specific wireless data, Long Range Wide Area Network (LoRaWAN), and IoT sensors to monitor underground mines. Another direction in the industrial sector is outlined by Tuptuk et al. [22], who review cybersecurity for industrial water and wastewater systems. ML technologies in industrial PdM are discussed by Polverino et al. [23], with a focus on algorithms and Key Performance Indicators (KPIs) for mechanical equipment. Finally, refs. [24,25] study the integration of IoT and distributed systems in the manufacturing industry. This paper analyzes the impact on productivity and equipment monitoring.

Smart homes is a field that uses IoT sensors to recognize human activities, monitor thermal energy efficiency at the household level, increase comfort and safety, and integrate at the smart community level. Shaharuddin et al. [26] propose a conceptual framework for using IoT sensors in fire prevention. Additionally, Mishu et al. [27] explore human activity recognition for using ambient sensors to improve local autonomy and Tomat et al. [28] discuss thermal comfort in smart homes through crowdsensing and simulations with dynamic models.

Studies on air, water, and overall environmental monitoring are intensively researched in the field of environmental monitoring. These studies predominantly integrate IoT sensors and AI technologies for pollution detection, prevention of ecological disasters, anticipation of certain behaviors, and simulation of situations generated by different contexts. In this regard, most applications are in the agricultural sector, where work by Jafar et al. [29] uses AI and IoT for plant disease detection or by Bandara et al. [30], which monitors urban water quality through the Global Navigation Satellite System (GNSS) and IoT technologies [31]. Alahmad et al. [32] detail the integration of technologies in precision agriculture to reduce agricultural losses and optimize resources, using wireless sensors and communication networks for monitoring as reference elements. Shahid et al. [33] focus on air quality monitoring using IoT networks in the Middle East. This analysis utilizes nanomaterial-based sensors for pollutant detection and integrates ML technologies to forecast pollution sources. Ahmadzadeh et al. [34] review wireless sensor networks concerning the use of agricultural systems to study the importance of communication protocols concerning mobile technologies aimed at optimizing smart agriculture. Also at the farm level, Huang and Khabusi [35] propose an advanced architecture for IoT networks in agriculture, integrating wireless sensors to reduce water waste and improve irrigation systems. Also, at the prediction and agricultural yield levels, Tangorra et al. [36] explore the use of IoT sensors to collect large volumes of data that can later be analyzed for various trends through AI tools.

The Healthcare IoT field is managed to acquire medical data to prevent and remotely treat patients and use modern technologies such as Blockchain and AI in innovative health systems. Xi et al. [6] examine how smart cities improve medical services by integrating IoT sensors for health monitoring and personalized treatments. Ejaz et al. [37] also present the possibility of remote monitoring and processing of medical data in the medical sector. Still, it emphasizes the security of data integrated at the level of electronic medical records. A systematic review of IoT applications used in intensive care units is presented by Arefin et al. [38]. It analyzes different medical sensors and data visualization platforms to provide an overview of current challenges and future perspectives of the Internet of Medical Things (IoMT). Another direction in the health field focused on analyzing the possibility of combating the COVID-19 pandemic through integrating the IoT-AI tandem. Kollu et al. [39] investigate early detection through patient monitoring and the automation of diagnostic processes. It mentions a ML model called Long Short-Term Memory (LSTM) applied to analyze respiratory and cough sounds. Also, in the medical field of obstetrics and gynecology, medical sensors help monitor pregnancy, an aspect investigated by Pavan et al. [40]. The study improves maternal and fetal health by integrating medical systems to reduce mortality through early anomaly detection mechanisms. Mohammadzadeh et al. [14] discuss the potential of smart cities to transform medical services by integrating these modern technologies.

The smart mobility sector overlaps with smart cities, as most works address the optimization of urban and airport traffic. However, the works targeting this field explicitly focus on combining IoT sensors and AI for transportation analysis. For this reason, the authors of the present work have treated the field as separate from smart city, which focuses more on the quality of human life and less on algorithms targeting traffic. Thus, Bhattacharya et al. [8] analyze the integration of IoT sensors and AI algorithms into an architecture aimed at the electric vehicle charging infrastructure. The paper identifies research directions for the experimental validation of the proposed infrastructures in future studies. Gheorghe and Soica [10] investigate the transformation of urban mobility by integrating AI, IoT, and predictive analytics in traffic management. The paper identifies a reduction in delays of up to 30% when AI models are integrated with IoT sensors at the traffic level. Refs. [12,41] review smart parking systems, discussing the challenge of detecting parking spaces under various conditions of size, brightness, and obstacles and proposing techniques based on video devices mounted on a vehicle’s windshield (dashcam) and neural networks (NNs) for identifying available spaces. The paper by Koroniotis et al. [13] addresses smart airports and IoT services implemented through automation techniques.

The analysis of the specialized literature shows that the fields overlap because they examine common elements that target the quality of human life. In Figure 4, the way in which different technological domains related to IoT intersect and influence each other is suggested. At the center of all these domains, smart cities are identified as the main element of interconnection, as all the different domains contribute to the development of this field. Smart cities employ technologies from all other domains through the integration of wearable sensors for medical monitoring systems, the integration of sensors that influence traffic, the integration of sensors that affect air quality and soil quality, which reflects in the health of the food consumed by the population, etc. Therefore, all these elements contribute to improving the quality of life for citizens. Wearable devices focus on the use of wearable sensors to monitor individual health. These devices operate in the healthcare IoT and smart cities sectors because the collected data is integrated into urban systems that facilitate access to personalized services [42]. The industrial IoT sector aims to optimize industrial processes, which eases the work of the employed staff. Thus, from this perspective, the citizens’ quality of life is also influenced by an indirect action in the area of smart cities. Additionally, optimizing industrial processes through IoT sensors influences the field of environmental monitoring by monitoring environmental parameters affected by industrial activity. The smart homes domain focuses on creating smart houses that facilitate human activity by creating a pleasant ambient environment, replacing time-consuming processes, optimizing resource usage, and ensuring enhanced comfort and advanced safety. This is made possible by the IoT-AI components that operate in the smart homes area, which also directly influence Smart Cities, as smart homes are part of larger communities. Environmental monitoring uses AI and IoT to detect pollution and prevent ecological disasters, among other elements influencing smart cities and industrial IoT. Healthcare IoT collects medical data and collaborates with wearable devices and smart cities. Additionally, smart mobility, which focuses exclusively on optimizing urban traffic and managing transportation infrastructure, influences smart cities, industrial IoT, and environmental monitoring by reducing or exacerbating gas emissions. Figure 4 highlights the interdisciplinarity of these fields in an IoT-AI approach. At the center of all these domains lies smart cities as a central element, as they interact with all other domains to provide a global connectivity experience that influences the quality of human life. This graphic representation shows the importance of interdisciplinary cooperation among researchers for integrating IoT technologies in various fields to solve global problems that directly influence the quality of human life.

These correlations illustrate possible future research directions, as the lack of interdependencies in areas highlights potential future connections. For example, the interaction between wearable devices and smart homes could automatically adjust the internal environment based on the user’s physical or emotional state. Thus, a smartwatch that detects a high-stress level could trigger changes in temperature and lighting or play relaxing music from the home’s audio system. Another connection that could be explored is between healthcare IoT and smart mobility through intelligent emergency systems that analyze the patient’s vital data and transmit it in real-time to autonomous ambulances. These correlations can be inspected by adjusting the route based on the clinical condition and traffic. Figure 4 should be analyzed more in the sectors with no correlations between domains, as these indicate future synergies that can bring innovations in human quality of life. This is only possible through integrating IoT sensors in collaboration with these fields using AI technologies.

The challenges associated with integrating sensors in the identified fields of activity vary depending on the specific application. For smart cities, the need for implementation over an extensive area is identified, as well as the necessity of network-level connectivity. Exposure to weather conditions, such as dust or rain, imposes a certain level of sensor protection. Ensuring interoperability and compatibility between subsystems, including traffic, parking, lighting, and pollution, is also crucial.

In the field of wearable devices, integration features include the necessity for sensors to be miniature, very slim, lightweight, not uncomfortable to wear, and biocompatible. They are mounted directly on the body, skin, clothing, or accessories, and must ensure both comfort and safety for the user. The biggest challenge of wearable devices is the balance between performance, autonomy, comfort, and the safety they must ensure for the user.

In the industrial field, the integration specifics refer to the harsh conditions in which they must operate. Among these conditions are vibrations, dust, extreme temperatures, humidity, and other harsh weather conditions. These sensors are mounted on mobile or hard-to-access equipment, which requires predictive maintenance and continuous monitoring of the equipment. Additionally, the compatibility criteria for integration with Supervisory Control and Data Acquisition (SCADA), Enterprise Resource Planning (ERP), or Manufacturing Execution System (MES) systems must always be analyzed, as well as the risk of cyberattacks that necessitate certain layers of security.

In the field of smart homes, integration features refer to increasing comfort, safety, energy efficiency, and design. Energy autonomy is a constant concern in smart home applications, as the goal is to minimize maintenance.

For environmental monitoring, sensors require installation in isolated, hard-to-reach environments such as streams, rivers, forests, fields, mountains, and caves, which necessitates access to extreme climatic areas, making self-powered sensors a requirement. The need for remote communication integration is a characteristic of these sensors. Another class of challenges for these sensors concerns calibration, durability, longevity, and minimal maintenance.

In healthcare IoT, the authors identified integration features as the precision of the sensors, along with the safety they must ensure for patients, the confidentiality of operations, and their reliability. Sensors must be biocompatible, non-invasive, or minimally invasive. The data requires real-time and secure transmission in accordance with the General Data Protection Regulation (GDPR) or the Health Insurance Portability and Accountability Act (HIPAA) regulations. Another concern is energy autonomy, especially when these devices are wearable or implantable.

For smart mobility, integrated sensors at the level of road infrastructure, as well as vehicles, require portable or real-time monitoring, which faces the issue of latency that needs to be minimized as much as possible.

4. IoT Sensors Identification Based on Measured Input

IoT sensors are classified into various types and used in the 7 previously identified fields, depending on the nature of the measured input parameter.

Table 1. Types of IoT sensors.

Sensor Type	References
Accelerometer	[43,44,45,46,47,48,49,50,51,52,53,54,55]
Air quality sensor	[56,57,58,59,60]
Blood pressure sensor	[61]
Chemical sensor	[62,63,64]
Color sensor	[65]
Contact sensor	[66]
Current sensor	[67,68,69]
Electrochemical sensor	[63]
Electrocardiogram	[70,71,72,73,74,75,76,77,78,79,80,81,82,83,84]
Electroencephalogram	[85]
Flow meter	[86]
Gas sensor	[63,64,87,88,89,90]
Gyroscope	[43,44,51,53,54]
Humidity sensor	[91,92,93,94,95,96,97,98,99,100]
Infrared sensor	[101]
Level sensor	[102]
Light sensor	[103,104,105]
Load cell	[106]
Motion sensor	[43,49,64,66,107,108,109,110,111,112]
Passive infrared	[101,113]
Pressure sensor	[61,95,114,115]
Proximity sensor	[110]
Rain sensor	[116]
Smoke sensor	[117]
Sound sensor	[118]
Strain gauge	[46]
Temperature sensor	[46,92,111,119,120,121,122,123,124,125,126,127,128,129,130,131,132,133,134,135]
Tilt sensor	[95]
Ultrasonic sensor	[48,106,136,137,138]
Vibration sensor	[109,139,140,141]
Water quality sensor	[142]

summarizes a series of articles from the literature by running the WoS query TS = ((“IoT sensors”)) AND PY = (2020–2025), which yielded 1739 results. From these articles, those materials that have a specific type of sensor as their central element were selected. Table 1 shows that the most commonly used sensors are the accelerometer, electrocardiogram, humidity sensor, motion sensor, and temperature sensor.

From the totality of materials synthesized in Table 1, the following analyzes the first five types of sensors that have provided the highest number of articles in the specialized literature.

The temperature sensor is the most studied, due to its degree of use in all IoT fields. From the earliest days of human life, temperature sensors have been considered essential monitoring equipment [120,143]. In this regard, there is extensive research on real-time body temperature monitoring with automatic alerts sent to emergency teams [121]. These sensors were considered essential even during the COVID-19 pandemic [123]. One of the most bizarre applications identified in the literature was the research [133] that uses IoT and DL to identify Egyptian cobra bites in real-time, achieving an accuracy of 90.9%. Another sector where temperature sensors are used is the agricultural field. The approaches in this field are varied, ranging from real-time fire prediction [126], continuing with the optimization of decisions regarding crops [135], and ending with the study of energy savings to extend the battery life of IoT devices by over 75% [125]. Aquaculture is a distinct branch of agriculture that also uses temperature sensors [128] to monitor water quality and prevent diseases. The transportation sector also uses temperature sensors [122], analyzing the integration of IoT and composite materials for real-time monitoring of railway infrastructure health [46]. Most of these studies also integrate AI components [134], data security [129], modern software tools [130], or blockchain [111].

The second most popular sensor in this analysis is the electrocardiogram (ECG) sensor. Arrhythmia disorders are the leading causes of global mortality, which is why IoT devices measure ECG, and the data provided by this sensor is trained by ML algorithms for classification, aiming at patient prevention and monitoring [83]. John et al. [70] present a technique for evaluating the quality of ECG signals in IoT sensors based on the k-nearest neighbors (KNN) classification algorithm. Modern medical facilities are made possible by digital tools that integrate digital twins and AI [71]. The detection of heartbeats based on the data provided by these sensors is examined by John et al. [72]. In the pharmaceutical industry, the integration of digital twin technologies alongside AI components strategically aligns the operational needs of companies with data processing requirements, enabling them to generate the best metrics associated with AI components [144]. The study by Fantozzi et al. [145] demonstrates the practical integration of digital twin technologies in a production line for packaging biological drugs. The model simulates reality and identifies bottlenecks. In this way, the packaging process is optimized without disrupting actual operations. The accuracy of the virtual model is over 90% correlation with reality. Overall, the two studies [144,145] demonstrate the capabilities of digital twin technologies in the industry through the integration of AI and IoT technologies in relation to sensors.

Portable IoT sensors monitor health by detecting obstructive apnea [73] and by integrating Convolutional Neural Network (CNN) models for multimodal analysis for the same purpose [79]. The CNN model presented in the research by Chang et al. [74] detects atrial fibrillation from ECG, using pruning and quantization, with 91.1× compression and 91.7% accuracy. The one-dimensional convolutional neural network (1D CNN) for classifying heartbeats from ECG signals obtained with wearable IoT sensors has been studied by Xiaolin et al. [75].

At the same time, John et al. [76] compare AI classifiers for detecting the integrity of ECG data from IoT sensors, achieving 99.47% accuracy. Early diagnosis through arrhythmia classification from ECG is achieved with the help of ML algorithms, reaching 99.92% accuracy [80]. These algorithms operate in real-time by acquiring data from the electrocardiogram sensor and processing it at the algorithm level, achieving remarkable performance metrics, such as 99.75% accuracy [81]. Abdul Razak et al. [78] mention that human errors cause 80% of road accidents. In this context, ECG monitoring can improve traffic safety by detecting conditions harmful to drivers. Sivapalan et al. [82] developed an NN for real-time ECG anomaly detection, with the primary objective of reducing energy consumption in wearable IoT devices. The same major concerns aimed at improving the efficiency of these devices are addressed by Xiaolin et al. [77]. Ali et al. [84] present an IoT ECG module for predicting vascular age and the impact of smoking on ECG traces, and the results are validated through ML models.

Analyzing the specialized literature, the accelerometer is one of the most popular sensors. It is identified by Toupas et al. [43] as helping recognize human activities at the level of smart buildings. From a logistical standpoint, Arz Von Straussenburg et al. [49] use accelerometers for real-time monitoring of office occupancy. Tang et al. [146] continuously monitor the condition of motor bearing faults to prevent various accidents. With wireless signal processing, IoT sensors in the accelerometer category maximize energy efficiency by integrating AI methods using Edge and Cloud technologies [44]. In Industry 4.0, accelerometers are used for monitoring bridges [45], examining the integration of composite materials at the railway level [46], or are utilized in medical applications to improve the prediction of alcohol intoxication [47,147]. Shahbazi and Byun [53] use these sensors at an industrial level to increase data production in defect detection through ML technologies. In the paper by Ghafoor [48], accelerometers identify potholes, ensuring high road safety. Also, Ghafoor [48] mentions that this objective is achieved by integrating the Extreme Gradient Boosting (XGBoost) algorithm into the analysis of sensor results. Rosca et al. [51] propose an urban traffic monitoring system. Park et al. [52] use accelerometers to monitor the neck, thereby predicting potential strokes through ML algorithms. Durairaj et al. [55] monitor cows’ health using accelerometers and AI techniques that achieve an accuracy of 92.45% in predicting lameness.

The fourth most studied sensor is the motion sensor. These sensors have the primary objective of optimizing energy consumption [43,107], being used both at the household level [66] and at the industrial level [108], in office buildings [49]. In the literature, several works correlate these sensors with the elderly, helping to reduce consumption [64,112]. IoT sensors have gained significant interest in the literature due to their capabilities in recognizing human activities through various ML architectures.

The fifth most used sensor is the humidity sensor. These are used as reference elements in the agricultural field [93]. Power supply remains a constant issue in the literature. Magneto-mechano-electric generators power soil moisture sensors through interaction with a magnetic flux concentrator (MFC) [91]. Personal protective equipment (PPE) includes IoT sensors for monitoring masks, assessing humidity, and temperature [92]. Fe-Al-Si-based thermoelectric (FAST) materials, which are cost-efficient and non-toxic, are used for autonomous energy generation for IoT sensors using residual heat [94]. The Telecommunication Tower Stability Profile (TTSP) system proposed by Fahim et al. [95] monitors the stability of telecommunications towers through optical and wireless sensors by analyzing environmental and structural factors, including humidity sensors. The retail industry uses ML to predict demands and manage inventory, analyzing seasonal patterns, environmental conditions detected through temperature and humidity sensors, and promotional effects [96]. Rönnberg et al. [97] explore the use of sonification and visualization to reduce disturbances caused by construction transport by integrating them into the sensor suite and humidity type. In the agricultural field, early detection of water stress in lettuce is achieved through data acquisition from moisture sensors [98], provided that continuous functionality is ensured, which can be achieved by implementing models for identifying defects in IoT agricultural sensors [99].

The classification of IoT sensors presented in this paper is based on a systematic approach related to the measured physical parameter. The authors mention that in applications, sensors have multiple functions, which makes it possible to classify them into multiple categories. A concrete example is smart environmental sensors that simultaneously measure temperature, humidity, and specific air quality parameters. This multi-attribute approach influences the integration of AI algorithms because these algorithms need to simultaneously handle data of different natures. These data can be numerical, they can be signals, they can be images, they can be voice, and so on [148]. From an AI perspective, multifunctional sensors generate multimodal datasets that require preprocessing techniques, data standardization, tests on how data influences model performance metrics, data fusion, feature selection, as well as customized analysis regarding data transformation in relation to improving performance metrics. Consequently, the classification proposed in this paper is a conceptual guide based on the authors’ expertise.

5. Self-Powered IoT Sensors Identification and Applications

One of the biggest challenges of IoT sensors concerns the power source. To prevent this problem, autonomous IoT sensors have been produced, which generate their energy in addition to their primary function of collecting and transmitting data in IoT networks. This category of sensors does not require external power sources, such as traditional batteries or direct electrical power. Instead, autonomous IoT sensors capture energy from the ambient environment and transform it into a source that helps them perform basic functions. The autonomous IoT sensor collects data from the surrounding environment and transmits it to the central system via the network, operating without batteries or other conventional energy sources. These sensors are self-powered by generating energy using elements from the surrounding environment. Ambient energy harvesting technologies classify autonomous IoT sensors into three categories of sensors [149]:

• Sensors that utilize mechanical energy, such as movement, vibrations, or pressure. This category includes piezoelectric and triboelectric generators. Piezoelectric sensors generate electrical energy by applying a mechanical force (pressure, vibrations). This category of sensors is discussed in a multitude of specialized papers. For example, Kuang et al. [150] describe an instantaneous sensing system based on a piezoelectric generator, providing a solution for energy harvesting and immediate data transmission. Tang et al. [151] propose a self-powered accelerometer based on TENG, including a V-Q-a theoretical model applicable to piezoelectric sensors. Tang et al. [151] make a valuable contribution by providing the mathematical foundation for the operation of self-powered sensors. Another application is presented by Liao et al. [152], where a piezoelectric transducer integrated into a composite structure for self-sensing is analyzed. Sensors that integrate energy harvesting with measurement functionality are described by Haghayegh et al. [153]. This paper presents a sensor based on a piezoelectric cantilever, which collects energy from vibrations and measures acceleration autonomously. The system uses a circuit designed for real-time multiplexing to achieve this objective. Zhao et al. [154] present a piezoelectric device that converts motion into electrical signals, with the advantage of a simple structure and high instantaneous power density, making stationary and flexible sensors operate efficiently. Besides purely piezoelectric solutions, some studies explore hybrid combinations [155]. For example, Gao et al. [156] propose a flexible epidermal sensor based on the hybridization of piezoelectric and triboelectric effects. The device is designed to be biocompatible with the skin. Another important aspect of piezoelectric sensors is their stability under variable conditions, as demonstrated by Shi et al. [157], which presents a flexible piezoelectric sensor with self-calibration for temperature variations. The sensor contains a system that automatically adjusts the output based on temperature. The piezoelectric energy harvester’s (PEH) performance is detailed by Paranjape et al. [158], demonstrating that these devices can be used as an autonomous power source for wireless sensors. Tian et al. [159] introduce a hybrid piezoelectric–triboelectric sensor created to recognize motor tics. It can detect involuntary movements of the human body. Due to their ability to generate energy using the surrounding environment and provide signals using this self-generated energy source, these piezoelectric sensors become stand-alone devices in the IoT ecosystem. TENG sensors represent one of the most studied technologies for developing self-powered IoT systems. They can convert ambient mechanical energy (vibrations, movement, friction) into electrical energy. The applications in which these sensors are used are diverse, ranging from designing a self-powered platform based on discharge induced by visible light [160] to medical monitoring applications that measure physiological parameters such as pulse and respiration [161,162]. A fibrous strain sensor combined with TENG is described by Han et al. [163], which aims to be integrated into clothing or sports equipment to collect energy, allowing the sensor to provide real-time data on the user’s physical activity. Deng et al. [164] describe a system consisting of two TENGs mounted in the sole for recognizing human activities. By analyzing the signals generated by footsteps, this system identifies types of activities (walking, running, climbing stairs, etc.), with applications in lifestyle monitoring. Bio-inspired sensors from natural structures are suitable for applications in smart cities and robotics [165]. The paper by Tang et al. [166] addresses the use of TENGs in augmented reality (AR) and virtual reality (VR) devices. Additionally, a self-powered solid–liquid TENG mercury sensor designed for water quality monitoring is presented by Zhang et al. [167]. This sensor utilizes the principles of TENGs to detect the presence of toxic pollutants. Li et al. [168] propose a self-powered sensing system based on the impedance effect, where the sensor signal is independent of the output variations in the TENG. The integration of TENGs with emerging technologies such as AI, big data, and cloud computing is explored by Liu et al. [169]. Kovalska et al. [170] present a TENG sensor designed for applications in industrial environments. The material used allows for the maintenance of performance even in extreme conditions, which extends the range of applications for these sensors. Domingos et al. [171] introduce a self-powered IoT node based on a wind-driven TENG for environmental monitoring. In contrast, Liang et al. [172] present a self-powered capacitive system based on an Electric Double Layer (EDL). This sensor does not require an external power source and can measure variations in pressure and humidity. In the paper by Huang et al. [173], TENG is explored as a multifunctional sensor capable of capturing both physical (pressure, vibrations) and chemical (gases, ions) stimuli. This type of sensor can be used in complex applications such as industrial monitoring and civil protection. Zhang et al. [174] describe a triboelectric pressure sensor manufactured using 3D-printing technology, proposed for robot, smart prosthetics, and sports equipment applications. Rui et al. [175] propose self-powered TENG insoles for measuring plantar pressure. These allow for monitoring body weight distribution during walking and are used in medical and sports applications [176].
• BFCs represent the second category of sensors. Sensors that use thermal energy by converting temperature differences into electrical energy. These sensors use thermoelectric materials as technology. Thermoelectric sensors are used in refs. [177,178,179,180]. Van Toan et al. [177] present a self-powered sensing system capable of storing energy and powering IoT devices with Bluetooth. In the paper by Tian et al. [178], a flexible thermoelectric generator (TEG) based on Bi₂Te₃ is analyzed and integrated into a wireless sensor for monitoring temperature in smart buildings. Guan et al. [179] propose a hybrid multifunctional sensor that combines foam graphene and ionic hydrogels to measure temperature and deformation simultaneously. Zhang et al. [180] introduce a thermoelectric material based on PEDOT/MWCNT, used in a flexible textile sensor for temperature and motion detection. Sun et al. [181] present a glucose detection sensor based on a zinc-air battery with advanced catalysts (SA-Ir/NC), offering high sensitivity and a stable signal without needing an external power source. In the research by Mao et al. [182], an epidermal self-powered biosensor is described as measuring lactate and glucose levels in sweat using an enzyme-activated biochemical cell (EBFC), making it ideal for real-time metabolic monitoring. Jin et al. [183] propose a self-powered multi-channel sensor for detecting food toxins (ochratoxin A and patulin), built on a dual photovoltaic platform, which offers precise detection independent of light conditions. These studies highlight the potential of electrochemical and photovoltaic systems in developing autonomous devices for food safety and health.
• HEGs are sensors that use solar energy through miniature photovoltaic panels that convert light into energy. In this category, classic photovoltaic and self-powered electrochemical sensors based on Photofuel Cell (PFC)-type photovoltaic cells stand out. Qian et al. [184] present an actuator with sensor functionality, based on the photo-thermoelectric effect. It integrates a thermoelectric generator into a flexible device powered by near-infrared light. This demonstrates how solar energy can be directly converted into motion and an electric signal. In the paper by Yang et al. [185], a polyelectrolytic hydrogel is proposed for mechanical–electrical conversion and self-sensing, inspired by natural energy generation processes through salinity gradients and having applications in biomedicine. A self-powered ratiometric sensor for estrogen built on a two-channel PFC eliminates the influence of light fluctuations on measurements, providing detection of the hormone 17-beta-estradiol [186]. This is an example of integrating solar energy harvesting and biosensors for health monitoring. Sun et al. [181] explore a self-powered glucose sensor that uses a zinc–air battery with catalysts. The sensor detects glucose levels in biological fluids. A self-powered epidermal biosensor for lactate and glucose in sweat uses a biochemical cell that extracts energy from organic compounds, allowing continuous monitoring of metabolic parameters without recharging or periodic battery replacement [182,187]. Jin et al. [183] propose a portable self-powered sensor for food toxins that simultaneously detect toxins such as ochratoxin A and patulin. Kim et al. [188] introduce SolarSense, a self-powered gesture recognition system using solar panels. It offers over 97% accuracy in motion recognition and demonstrates how solar energy can be integrated into human–machine interaction.

To conduct a comparative analysis between the two types of sensors, battery-powered and self-powered sensors, the Efento Wireless Indoor Air Quality Logger and EnOcean Easyfit ETHS sensors were considered.

Table 2. Technical specification of the sensors.

Features	Efento Wireless Indoor Air Quality Logger with Battery	EnOcean Easyfit ETHS Self-Powered
Measured Parameters	Temperature, Humidity	Temperature, Humidity
Measurement Interval	Configurable: 1 min–10 days	Periodic + event-based (significant change detection)
Wireless Technology	Bluetooth Low Energy (BLE)	EnOcean wireless (868 MHz EU/902 MHz US)
Power Supply	Battery-powered (up to 5 years autonomy)	Solar-powered, optional coin cell for backup
Cloud/Remote Monitoring	With Efento Gateway + Efento Cloud	Possible via EnOcean-based systems

presents the comparative analysis based on their technical specifications.

Table 2 shows that both sensors are wireless and can be compared by monitoring indoor environmental conditions. The Efento Wireless Indoor Air Quality Logger sensor has the advantages of internal memory where data can be stored locally, the ability to be easily integrated into mobile applications, high measurement precision, and a wide temperature range. The EnOcean Easyfit ETHS sensor has the significant advantage of not requiring batteries, as it is solar-powered, which translates to low maintenance. This is ideal for automation at the level of smart buildings, as it has small dimensions and is easy to install. The most significant advantage of the EnOcean Easyfit ETHS sensor that sets it apart from the Efento Wireless Indoor Air Quality Logger is its self-powering capability. On the other hand, the technical documentation of the Efento Wireless Indoor Air Quality Logger sensor mentions a battery life of up to 5 years, which minimizes the advantages of self-powered sensors.

Figure 5 depicts the operational flow of a battery-powered IoT sensor compared to a self-powered one. In Figure 5, the specific stages of the operational process of a battery-powered IoT sensor are presented. In the case of this type of sensor, constant battery monitoring is necessary, which requires the inclusion of a maintenance stage that interrupts the operational flow. Therefore, this process is prone to interruptions and entails a certain maintenance cost, which means low reliability in isolated or hard-to-reach environments. The self-powered sensors in Figure 5 have a separate energy harvesting process, and the system runs in a continuous autonomous flow. Basically, the process is uninterrupted due to the energy self-supply capabilities, eliminating the need for an external source. Thus, long-lasting autonomous operation is ensured, which is ideal for hard-to-reach environments or dense sensor networks. Figure 5a,b highlight the operational superiority of self-powered sensors in modern IoT applications. Futhermore, the two diagrams outline the elimination of battery dependence, not only as an optimization process but also as an extension of the functionality of these sensors in isolated or complex environments.

In

Table 3. Comparative analysis of battery-powered and self-powered IoT sensors.

Feature	Battery-Powered IoT Sensor	Self-Powered IoT Sensor
Energy Source	Internal battery	Harvested from the environment (motion, light, etc.)
Maintenance Required	Periodic battery replacement	No maintenance needed
Operational Continuity	Prone to interruptions due to battery depletion	Continuous and autonomous operation
Long-Term Cost	Higher (battery replacement, labor)	Lower (no recurring energy cost)
Deployment Suitability	Limited in remote or inaccessible areas	Ideal for remote, embedded, or high-density systems
Sustainability	Low (battery waste, higher energy demand)	High (eco-friendly and energy efficient)

, the authors conduct a direct comparative analysis of battery-powered sensors and self-powered sensors, highlighting the advantages of self-powered sensors.

Table 3 highlights the advantages of self-powered sensors compared to traditional battery-powered ones. Battery-powered sensors constantly consume resources, both materially and humanly, from a maintenance perspective. Self-powered ones eliminate these needs and function autonomously. From the standpoint of operational continuity, powered sensors represent a superior solution due to their long-term operation, especially in hard-to-reach environments or dense networks, such as in agriculture, urban infrastructure, ecological monitoring, or the medical field. From the perspective of sustainability, self-powered sensors reduce electronic waste, resulting in a decrease in the energy footprint. Thus, these sensors align with current directions of sustainable development [189].

Battery-powered sensors require intelligent energy management strategies, which also apply to hybrid systems that combine traditional batteries with supercapacitors. In applications where energy comes from renewable sources, rule-based adaptive strategies are studied to maintain the availability of wireless sensor nodes when energy levels temporarily drop [190]. In this way, the continuation of monitoring and data transmission is ensured. The implementation of intelligent battery management systems improves the way energy is delivered and distributed to sensor components [191].

Regarding battery degradation and lifespan, the intelligent policy adapted to the operating conditions of the sensors reduces the negative impact of degradation, maintaining the quality of service provided by them. Compared to classical discharge methods, charge and discharge cycles extend the lifespan of batteries, thereby preventing premature failures [192]. The use of modern charging techniques, such as adaptive or pulsed charging, reduces battery wear [191].

Wireless solutions for battery management are the most convenient solutions for sensor implementation due to reduced maintenance complexity, energy efficiency, increased long-term reliability, and user convenience [193].

The efficiency with which self-powered IoT sensors harvest energy is closely correlated with environmental conditions [194]. These conditions contribute to the performance and autonomy of the systems. For example, in the case of TENGs sensors, the conversion efficiency decreases in environments with high humidity or very low temperatures. This is due to the density of triboelectric charge [195]. Regarding TENGs sensors, they depend on the temperature difference between the ambient environment and the heat source. In this way, under constant temperature conditions, efficiency decreases. In the case of applications taking place in indoor environments, solar sensors generate insufficient energy in the absence of direct light [196]. Studies [178,180,197] highlight these limitations, where hybrid power supply strategies or AI algorithms for self-adaptive operation concerning available energy address these issues. Consequently, the way these systems are designed must relate to computational techniques associated with realistic operating scenarios in which environmental conditions are explicitly mentioned in the design stage.

6. Self-Powered IoT Sensors in the AI Field

The development of self-powered sensing technologies has been supported by integrating advanced systems equipped with AI components. To identify progress in this sector, a WoS search was conducted that integrated the keywords self-power sensing, AI, and the time constraint (TS = (“self-powered sensing” AND (“artificial intelligence” OR “AI”)) AND PY = (2020–2025)). The paper’s technical analysis is presented in the following sections. Subsequently, a systematic study of researchers’ interests and the evolution of research using data generated by WoS is conducted.

6.1. Detailed Literature Analysis of Self-Power IoT Sensors in the AI Field

The search yielded 52 results, which were analyzed by grouping them according to the field in which the elements of novelty are presented. A detailed analysis of the identified AI components was also conducted. The results are summarized in

Table 4. Self-powered sensors integrated with AI by domain.

Domain	Description	Reference
Environmental Monitoring	Review of TENGs for IoT	[198]
	Underwater pressure sensing with TENG	[199]
	Advances in multimodal mechanoluminescent sensors	[200]
Healthcare IoT	Microneedle patches for biosensing	[201]
	Self-powered plantar stress sensing insole	[202]
	Bimode temperature-pressure self-powered sensor	[203]
	Self-powered leg system with AI	[204]
	Ultrasensitive self-powered mechanoluminescence smart skin	[205]
	Piezoelectric polymer nanofibers for pressure sensors	[206]
	Smart textile TENGs	[207]
	Performance of flexible TENGs	[208]
Industrial IoT	Self-powered sensing in robotics and AI	[209]
	Smart bearing with self-diagnosis	[210]
	Hybrid vibration energy harvester	[211]
	Review of ML-assisted TENG sensors	[212]
	Self-powered sensing and intelligent systems	[213]
	On-device predictive maintenance system	[214]
	3D printed conductive hydrogel-based energy harvesting device	[215]
	Advanced 3D printing-based TENG for mechanical energy harvesting and self-powered sensing	[216]
Smart Cities	Self-powered sensors in civil infrastructure	[217]
	Wind energy harvesting with AI	[197]
	ML for TENG-based sensing systems	[178]
	Intelligent piezoelectric AIoT node	[218]
	Self-powered sensing in rail transit	[219]
	Origami-inspired generator for IoT	[220]
Smart Homes	Wood-based triboelectric sensor for smart homes	[221]
	Cellulosic TENG for sustainable sensing	[222]
	Self-powered speech recognition system	[223]
	Film and textile-based single electrode TENGS	[224]
Smart Mobility	Hybrid backpack energy converter with AI	[225]
	Self-powered gyroscope sensor based on TENG	[226]
Wearable Devices	Dual-modal self-powered sensor for writing	[188]
	Thermoelectric fabric for respiration monitoring	[227]
	TENG applications in biomedical and AI	[228]
	Self-powered tactile sensing review	[229]
	CL@rGO hydrogel for self-powered sensing	[230]
	Multilayer hydrogel for self-powered sensing	[231]
	Multi-functional T-TENG with UV protection	[232]
	Organic thermoelectrics for sensing	[233]
	Triboelectric paper for tactile sensing	[234]
	Facial recognition mask using TENG	[235]
	Stretchable CCF fiber for sensing	[236]
	Graphene textile TENG for motion monitoring	[237]
	Intelligent E-skin with multimodal perception	[238]
	High-performance flexible piezoelectric sensor array	[239]
	Core-sheath piezoelectric sensor based on steel wire and PVDF microfibrillar bundle	[240]
	Fiber/fabric-based piezoelectric nanogenerators and TENGs	[241]
	Flexible electronic skins based on the composite of ALTFM	[242]
	Improved ultrathin stretchable TENG with postcharging material	[243]
	Water repellent fabric-based TENG for harvesting human mechanical energies	[244]

Note: ALTFM—Aligned liquid metal/thermoplastic polyurethane fibrous mat; DNN—Deep neural network; EMG—Electromagnetic Generator; GRU—Gated Recurrent Unit; RB-TENG—Rolling Ball Triboelectric Nanogenerator; PVDF—Polyvinylidene fluoride; RF—Random Forest.

, and the papers are discussed in detail to identify areas that need further investigation.

The analysis of these works shows that the combination of self-power sensing concerning AI primarily activates in smart cities, wearable devices, healthcare IoT, and industrial IoT through autonomous sensors that process and interpret data intelligently [178,245]. The most commonly used sensors are TENGs, which collect ambient mechanical energy and convert it into electrical energy. This is why many studies investigate TENG technology combined with AI algorithms, such as LSTM, RF, DL, or Recurrent Neural Network (RNN) [204,214,235]. For example, the paper by He et al. [235] is dedicated to an innovative mask for facial recognition using an LSTM model. The results achieved an accuracy of 99.87%, highlighting the enhanced capabilities of combining technologies in human–machine interaction and virtual reality.

The healthcare IoT field stands out with innovations such as self-powered footwear that monitors the plantar condition of the foot, making it possible to identify gait anomalies using AI tools. This system assists in medical diagnosis and patient recovery [202]. Additionally, autonomous smart prosthetics are integrated using TENG sensors, which use DL models to detect lower limb movements in data analysis. Thus, the paper by Kong et al. [204] achieved an accuracy of over 99%.

Data privacy is the largest issue facing the medical domain, especially concerning the self-powered sensors. Strict safeguards pertaining to GDPR and HIPAA regulations apply to their use. Unauthorized access is prevented by local edge computing, secure communication protocols, multifactor authentication, end-to-end encryption technologies, and other custom-made software layers. Medical data is stored on certified servers or in cloud environments that adhere to international data protection standards. Protection technologies, frequent audits, certifications, studies on large patient samples, assessments by specialized compliance assurance committees, and data processing transparency are required to guarantee the clinical compliance of self-powered IoT sensors in medical applications. Integrating self-powered sensors in the medical field is made possible by guaranteeing all of these security and privacy measures.

In the field of smart cities, AI components predict weather conditions, such as wind direction and speed, using a self-powered wind generator based on an EMG with an RB-TENG and a GRU algorithm. This algorithm represents a simplified version of LSTM. The results led to an accuracy of 96% for speed and 99.04% for wind direction [197].

In the industrial IoT field, AI components are used to detect industrial machine failures early. The on-device PdM system uses self-powered sensors and TinyML technology, achieving over 99% accuracy with a 66.8% energy savings compared to traditional sensors [214]. Other research explores the lubrication of bearings using DL techniques [210].

The field of wearable devices introduces systems for smart textiles that also use TENG sensors. These systems identify body movement to monitor physical activities and emotional states. In this context, ML algorithms analyze the signals from these sensors and interpret them for sports or medical purposes [233,237]. Xiong et al. [237] use a graphene textile to distinguish human movements, demonstrating superior results.

Besides TENG, there are also hybrid energy sources that incorporate piezoelectricity combined with EMG in mobile applications. These applications target self-powered backpacks that use AI algorithms to recognize users’ movements to optimize energy consumption, generating up to 4.5 mW of average power during running [225].

Optimizing energy consumption, increasing spectral efficiency, adapting transmission to the variable conditions of the communication channel, autonomy, and scalability are dictated by the modulation techniques used by IoT sensors. New modulation schemes have been introduced in the recent literature. Such a modulation scheme is present in the system based on reconfigurable intelligent surfaces [246]. It transmits information and energy harvesting with superior performance in terms of spectral efficiency and energy reliability. Additionally, the scheme of the paper [247] uses the initial condition indexing of chaotic sequences. They are used in short-range wireless communications, having low complexity.

6.2. Bibliometric and Systematic Thematic Analysis for Self-Powered Sensors and AI Components

The temporal evolution of publications demonstrates, through the obtained results, a growing international interest in self-powered sensing technologies integrated with AI components. In 2020, the search yielded two publications, reaching 16 publications in this field by 2024. The first four months of 2025 led to seven publications, which is estimated to increase the number of publications in the field. The results are presented in Figure 6, where a constant rise in publications can be observed.

Regarding the type of publications, the distribution shows 37 research papers and 16 review papers, according to Figure 7. The fact that most publications are original articles, namely 69.81%%, indicates an intense experimental research activity. Additionally, 30.19% of review articles provide a synthetic perspective on progress.

The publications are classified into various WoS categories, most in multidisciplinary material science, specifically 39. The following fields are nanotechnology and applied physics, with 25 papers each, and physical chemistry, with 23 papers. The last two positions correspond to multidisciplinary chemistry with 14 articles and electrical and electronics engineering with 6, according to Figure 8. Thus, raw materials occupy the top position in the categories of powered sensors, indicating potential for development or a necessity for growth in other fields.

The geographical distribution by country places China first, with 46 studies, followed by the USA with 9, and Singapore with 6. Figure 9 shows the low interest of other countries in these top technologies for the evolution of autonomous devices.

The main publications that disseminated these results were Elsevier, with 19 articles, followed by Wiley, with 17, and the American Chemical Society, with 6. Paradoxically, IEEE, a publication dedicated to advancements in the electrical field, recorded only 3 publications in the mentioned period, as shown in Figure 10.

To provide an analysis of the scientific impact of the reviewed works, two graphs were constructed to show the influence of each work in the scientific community through the papers that cite them and the recent interest in the field through the number of views in the sample. These graphs show the qualitative dimension of the works in the specialized literature. The two indicators assess scientific relevance along with the attention given to short-term research on publications from the 2020–2025 period. Figure 11 shows the distribution of the number of cited references, highlighting the works with a high academic impact. Figure 11 depicts the number of users over a period of 180 days, reflecting the readers’ interest in relation to the works in the analyzed literature. The two metrics provide a qualitative dimension to the analysis.

In Figure 11, the maximum identified value is 310 for the number of citations of the references. This value indicates the most cited reference in the list. The minimum identified value is 27. This value reflects references that have a lower impact or are more recently published. Most of the values are concentrated between 40 and 80 citations, and the authors of the present work interpret this value as a balanced distribution of reference influence.

In Figure 12, the number of accesses for each article in the last 80 days is represented. The maximum value identified is 194, indicating that 194 people accessed the article within the previous 180 days. The minimum value identified is 0, which means that the paper has not been accessed at all. Most of the values are concentrated between 5 and 50 accesses. The authors interpret this balanced distribution of the frequency of use. Values below 5 are found in a small group of only six articles, indicating the readers’ interest in the works in the field.

The publications predominantly address areas of material science research, with 41 articles. Subsequently, science and technology recorded 28 articles, followed by physics with 26, chemistry with 25, engineering with 11, instrumentation with 4, and energy fuels with 2. Analyzing these values presented in Figure 13, it is observed that the predominant area is raw materials, followed by the area of physical–chemical engineering.

The value analysis of the keyword clusters obtained using the VOSViewer 1.6.20 tool represents a way to identify the main themes in the publication collection based on the frequency of keyword occurrences. This analysis highlights the connections between key concepts and identifies the dominant trends in the researched field, as seen in Figure 14. In this analysis, 316 keywords were used, of which the 12 keywords that appeared at least five times were selected. These keywords were subsequently grouped into three distinct clusters, each highlighting a subdomain or a main research direction. This way, a systematic picture of the thematic structure of the scientific literature on self-powered sensors integrated into AI applications was obtained.

The first cluster (colored in red) contains six elements (energy harvesting, sensor, system, transparent, performance, triboelectric nanogenerator) and focuses on the sensors’ energy and performance. This cluster is associated with collecting ambient mechanical energy through self-powered sensors, including TENG sensors in this category. Within the cluster, performance elements regarding the functionality of these devices without an external power supply are included. The term transparent suggests an interest in transparent or flexible materials used by these devices, which are usually in the wearable or innovative category.

The second cluster (colored in green) contains five elements (artificial intelligence, energy, nanogenerators, self-powered sensing, triboelectric nanogenerator) and is associated with the combination of AI and TENG. The cluster highlights the importance of modern technologies, such as AI, in self-powered sensor technologies. The term energy demonstrates the correlation between AI tools and TENG to provide intelligent interpretations of human movements, vibrations, or other ambient stimuli.

The third cluster (colored in blue) is exclusively dedicated to sensors, having a single element represented by the self-power sensors technology (sensors).

This analysis, conducted using the VOSViewer tool, synthesizes the keywords of the 52 results provided based on the initial search. The study identifies self-powered energy as the central point in cluster 1, from which the importance of AI in interpreting data generated by sensors in the second cluster derives.

The analysis of these research results highlights the progress in self-powering sensing integrated with AI and underscores the countries that show increased interest in this direction. These results are discussed in detail in the dedicated discussion section.

7. The Use of ML Models in Self-Powered Sensor Systems

ML technologies have been applied in recent years in a multitude of fields, including technologies aimed at self-powered sensors. These algorithms are used in intelligent systems for monitoring, recognition, diagnosis, prediction, or identification tasks. ML algorithms have been used in applications for these tasks. Still, the performance of the proposed model in the research is evaluated based on specific indicators of the type of algorithm used. A benchmark indicator for all these algorithms is accuracy, a widely used metric to express the proportion of correct predictions relative to the total predictions made. Although accuracy is insufficient to evaluate a model’s performance fully, it provides an overview of the quality of the developed model.

In the specialized literature, 61 papers were published between 2020 and 2025 that address ML models in self-powered sensor systems.

Table 5. ML in self-powered sensors.

ML Algorithm Used	Application Integrating Self-Powered Sensor	Accuracy (%)	Reference
Two-layer LSTM model	Self-sensing FRM in VR and HMI	99.87	[235]
ML algorithm	Self-powered bionic hand with multi-functional actuation and sensing	96.80	[248]
RF	Human activity recognition using TENG-based HAR	88	[249]
RF, DNN	On-device PdM of machinery	99	[214]
NN	Speed monitoring of rotating machinery using TENG	>90	[250]
RF, LR, DT	Sitting posture monitoring vest	96.60; 95.50; 94.3	[251]
Automated ML	Defect diagnosis of rolling bearings using TENG	99.48	[252]
CNN	Workplace activity monitoring using tactile tribo/piezo sensor	>98	[253]
ML techniques	Human motion monitoring using hybrid energy harvester	85.6–100	[254]
ML	Gait recognition and rehabilitation monitoring system	96.70	[255]
NN	Identity recognition via gait analysis	100	[256]
DL model	Monitoring vehicle collisions and human gestures using BC-PVA-BCZT aerogels	100	[257]
K-Means, SVM	Wrist posture recognition using PS-TENG	95.18	[258]
ML algorithms	Speech recognition and organ monitoring using 3D-printed wearable triboelectric sensors	99	[259]
RF	HAR	93	[260]
Optimized ML algorithm	Motion sensing and smart sports applications	99	[261]
ML	Coaxial fibers for wearable strain sensing and triboelectric fabric	95	[236]

Note: BC—Bacterial cellulose; BCZT—BaCaZrTiO; DT—Decision Tree; HAR—Human activity recognition; HMI—Human Machine Interface; LR—Logistic Regression; PS—Polyacrylonitrile/sodium; PVA—Poly(vinyl alcohol); SVM—Support vector machine.

presents a comparative analysis of the models used in the reviewed works, the authors’ accuracy, and the research’s primary objective.

He et al. [235] use the two-layer LSTM model for facial and emotion recognition in the VR environment. The model achieved an accuracy of 99.87%, demonstrating its capabilities for use in psychological therapy. Similarly, Weng et al. [248] employ an autonomous bionic sensor and an ML algorithm to classify human gestures, achieving an accuracy of 96.8%. This sensor contributed to the development of a multifunctional robotic hand. For monitoring human activities, Harris et al. [249] use RF, which achieves an accuracy of 88% in classifying different types of activities, such as walking, running, or jumping. Another baseline study is by Chen et al. [214], where the RF algorithm and the Deep Learning Network (DLN) are tested in PdM systems, achieving an accuracy of 99% and demonstrating a 66.8% reduction in energy consumption through integrating self-powered sensors.

In the industrial field, Zhang et al. [250] propose a neural model for monitoring the speed of rotating machines, achieving an accuracy of 90%. The paper demonstrates early defect detection. In the paper by Dong et al. [252], an ML model combined with a signal decomposition method allowed for diagnosing bearing defects, achieving an accuracy of 99.48%.

In the field of health, refs. [251,255] use TENG sensors together with RF, LR, and DT methods, achieving accuracies of 96.6%, 95.5%, and 94.3% for monitoring the human body’s position. This type of application prevents problems associated with a sedentary lifestyle. Wei at al. [255] introduce a textile system for recognizing pathological gait, achieving an accuracy of 96.7% in the medical recovery process. Some studies explore applications related to sports, and, thus, the well-being of the human body, such as refs. [255,262], where the hybrid piezoelectric–electromagnetic system, together with ML techniques, monitors human movements and achieves an accuracy ranging from 85.6% to 100%. Huang et al. [256] present a gait recognition system based on a piezoelectric sensor with precise identification of three individuals, achieving an accuracy of 100%. Refs. [257,258] detect collisions and recognize postures using a CNN model, achieving an accuracy of 100%. In contrast, in the second work, K-Means and SVM are used, achieving an accuracy of 95.18% in recognizing 12 wrist postures. Also in the medical field, Tong et al. [259] present 3D printed triboelectric sensors, along with ML algorithms for organ monitoring and silent speech recognition, achieving an accuracy of 99%. Demircan et al. [260] introduce the RF algorithm with an accuracy of 93% in recognizing human activities, a system that integrates TENG sensors with mobile applications. Zu et al. [261], through a customizable textile sensor, use ML algorithms that achieve 99% accuracy in recognizing five body movements, while Wang et al. [236] employ a coaxial wire resistant to 15,000 cycles for gait and sign language recognition, achieving 95% and 100% accuracy, respectively. These algorithms have led to the development of complex applications that allow the integration of self-powered sensors. In the specialized literature, RF, LSTM, and CNN algorithms have been identified, and they have achieved accuracies of over 95%, allowing for the optimization of results from the hardware equipment. These applications range from medical monitoring, staff assistance, and helping people with disabilities to predictive diagnostics of machines, demonstrating the extensive use of fields for self-powered sensors.

ML algorithms are evaluated through specific metrics, with accuracy being one of the most commonly used indicators in the comparative analysis of algorithms. The actual performance of self-powered sensor systems is influenced by the hardware constraints of the environment. Additional metrics, such as random-access memory (RAM) and flash memory consumption, inference time, optimality, completeness, and energy consumption per operation, can be evaluated in ML algorithms as future research directions. RF and SVM algorithms require more memory than optimized models, such as compact TinyML, MobileNet, or simple regressions. Studies [214,261] mention energy savings of up to 66.8% when opting for lightweight models adapted to the hardware. From this approach, future research directions include incorporating computational efficiency metrics alongside standard metrics such as accuracy, precision, recall, and F1-score, to facilitate the real-world implementation of self-powered IoT sensors in various applications.

These results demonstrate the need to increase the research volume in the field by exploring new ML algorithms that could provide superior accuracy compared to those currently integrated. Additionally, expanding research will lead to new concepts and contribute to the technological advancement of self-powered sensors within applications that integrate ML algorithms.

8. Discussion

This paper stands out from other syntheses by addressing the interdisciplinary combination of self-powered IoT sensors and AI algorithms, particularly ML ones. Unlike other works, this synthesis highlights the collaboration between these modern technologies. Most synthesis papers address either IoT technology, AI applications in Edge systems, the collaboration between energy harvesting technologies, or exclusively energy consumption optimization techniques. Thus, the paper is an original contribution through its integrative and bibliometric structure on the intersection of the three domains, identifying the connections between the key components of this technological ecosystem.

This research was guided by three RQs, which were answered based on the bibliometric and thematic analysis conducted from 1 January 2020 to 30 April 2025. The data were extracted using the WoS platform, and the study addressed the RQs to outline future directions for exploring self-powered IoT sensors integrated into AI technologies.

The synthesis presents an overview of the current state of the art. The studies predominantly address the issue of TENG-type self-powered sensors and the manner in which they are integrated into applications utilizing ML algorithms such as LSTM, RF, or CNN. Most applications that integrate this combination of technologies are in the fields of healthcare, smart cities, and industrial settings. The fact that most of the works are in the area of human activity recognition, predictive maintenance, real-time monitoring, and intervention with ML models whose accuracy exceeds 95%, demonstrates the ability of these algorithms to contribute to the evolution of the fields in which they operate. However, the authors encourage researchers to explore other ML algorithms, which should be intensively studied in other fields where they are underdeveloped.

The first WoS search aimed to identify the areas of use for IoT sensors, identifying 1739 articles. These articles extracted data regarding the seven regions of applicability—smart cities, wearable devices, industrial IoT, smart homes, environmental monitoring, healthcare IoT, and smart mobility. As part of this analysis, the authors also ranked the most commonly used types of sensors: temperature sensors in over 20 papers, electrocardiogram sensors in 15 papers, and accelerometers in 13 papers.

The second search focused on integrating AI technologies in self-powered sensors, yielding 52 articles with an increasing distribution from 1 January 2020 to 30 April 2025, reflecting the steadily growing scientific interest in this field. However, the number of articles for such a vast field is minimal, which leads to the idea of the necessity for the maturation of the concept of self-powered sensors integrated with AI. Of the 52 articles, 70% were contribution articles, and 30% were reviews. The WoS scientific domain predominantly identifies these articles as multidisciplinary materials (39 articles), followed by nanotechnology (25), applied physics (25), chemistry–physics (23), and electrical and electronic engineering, with only 6 works. These values are at least surprising given the fact that sensors mostly have an electrical component and are integrated into systems associated with the field of electrical and electronic engineering, which reinforces the idea of the need for the concept itself to mature and for a much larger volume of scientific contributions in the field to be generated. Regarding geographical distribution, China has the highest contributions, accounting for approximately 88.4% of the total works, and the top publisher for disseminating the results is Elsevier, which has 19 articles.

The VosViewer analysis identifies the semantic connections between key concepts, such as self-powered energy and AI technologies. This highlights the interdisciplinary nature of these works, emphasizing the connection between TENG and AI, which demonstrates that modern technologies cannot function independently, that is, in isolation, but require an interdisciplinary approach. This analysis prioritizes future research directions, which indicate the need for new flexible and biocompatible materials that can be integrated into self-powered systems as well as the second research direction addresses the necessity of ML or DL algorithms to optimize the interpretation of data generated by sensors. A systematic perspective is provided on collaboration between AI algorithms and self-powered sensors by grouping keywords into clusters. This analysis has the following practical implications:

• For researchers, it serves as a basis for identifying gaps in the specialized literature and outlining future research directions.
• For developers, it is a way to plan projects based on dominant trends, such as replacing traditional sensors with TENG sensors, which integrate very easily into wearable or IoT applications.
• It provides an overview of the research priorities to which funders should allocate additional funding resources to encourage technological advancement.
• In the industrial area, it facilitates the technological transfer of research results into commercial products by integrating powered sensors in medical, industrial, or urban fields, identified as reference points in this work.

The VosViewer analysis is a methodology through which information was extracted from existing textual datasets in research disseminated in various publications, and the identified clusters highlighted three primary directions:

• Harvesting self-powered energy through the identification of new materials.
• Integrating AI algorithms for interpreting and optimizing sensor-generated data.
• Self-powered sensors should be widely integrated into applications across various fields.

The third search focused on the number of articles that exclusively address ML algorithms in the context of self-powered sensors. The result provided 61 articles with remarkable accuracy, most having values over 90%. Many of these articles operated in the medical field for psychological therapy, arrhythmia classification, medical rehabilitation, assisting users with body posture, or movement monitoring. Beyond this achievement regarding the enhanced performance of ML algorithms, the research highlighted technologies capable of saving energy by up to 66.8% and operating under extreme conditions. AI models, especially CNN, LSTM, and RF, are intensively studied in low-energy consumption applications.

The very low research values in the literature show many fields but also unexplored ML models. The annual increase in publications shows an emerging trend. Still, their volume is far too small compared to the number of ML models that need to be explored and the integration possibilities of these battery-free sensors into a vast field of applications.

The study’s limitations concern the single data source, WoS, which, although includes most of the literature research, may exclude some Scopus-type research or local databases. Another limitation of the study concerns the research analysis and the quality of ML models relating to accuracy, although the models are evaluated based on multiple performance indicators.

RQ1 analyzes the areas not sufficiently explored in the context of AI technologies, particularly through the ML component. The analysis of keyword co-occurrence and the distribution of articles in specific domains led to identifying seven areas, four of which are considered dominant: smart cities, wearable devices, healthcare IoT, and industrial IoT. In contrast, agriculture, education, environmental monitoring, and smart homes are insufficiently explored in the context of AI. The lack of research indicates opportunities to integrate these sensors into prediction, monitoring, identification, or optimization tasks for crop production or disease prevention at the crop level.

RQ2 aims at research directions that should integrate IoT sensors with AI, particularly ML. Based on the values extracted from the analysis, five primary directions have been identified:

• Standardization of IoT sensor taxonomy to approach a universal classification framework, based on which comparisons between works can be made.
• Develop optimized ML algorithms for the hardware of self-powered sensors by integrating less-explored models in the literature.
• Addressing end-to-end systems that allow for AI integration, using data interpretation, and covering the entire chain—capture, local processing, and automated decision-making. To integrate such a component, interdisciplinary collaboration between hardware engineers, AI researchers, and software developers is necessary.
• Creating interconnected platforms in the IoT field, as many works in the specialized literature address the domains in isolation. One direction is to develop a holistic solution that considers the effects of a decision in one domain on others, for example, the automated adjustment of the home environment based on data provided by a wearable device.
• Comparative studies between self-powering technologies are needed because, although TENG dominates the literature, comparing performances between TENG, piezoelectric, BFC, HEX, or hybrid sources is limited. Future research should evaluate these technologies under real usage conditions, particularly to assess them in the long term.

The compatibility with existing technical standards is highlighted in

Table 6. Comparative analysis between emerging IoT communication technologies and existing standardized technologies.

Criterion	Emerging Technologies	Established Technologies
Standardization	0	1
Interoperability	0	1
Hardware compatibility	0	1
Support for existing protocols	0	1
Market acceptance	0	1
Energy efficiency	1	0
Innovation potential	1	0
Suitability for battery-free operation	1	0

. Self-powered sensors based on TENG, reconfigurable intelligent surface (RIS)-indexed modulation [263], and chaotic schemes such as Differential Chaos Shift Keying (DCSK) [247] are limited by current technical standards. Additionally, the level of market acceptance is low compared to established protocols such as long range (LoRa) [264], ZigBee [265], BLE [266], Narrowband Internet of things (NB-IoT) [267], etc. New technologies are not yet standardized, which necessitates a series of modified hardware and software architectures. For example, emerging solutions require the development of dedicated circuits for energy harvesting and management in the case of TENG. The lack of standardization makes it more difficult for new technologies to be widely adopted. Therefore, future research directions should include a stage of analysis of compatibility with communication protocols, platforms, market openness, consumer reluctance, and existing infrastructure. Additionally, proposals for gradual integration strategies are necessary to facilitate the transition without compromising the functionality of existing systems. In Table 6, the value of the upper solution was marked with 1 and the value of the lower solution with 0.

Analyzing Table 6, it is observed that emerging technologies are superior in terms of energy efficiency, being able to function without a battery. In contrast, established technologies are superior in terms of compatibility, standardization, integration into current infrastructure, market openness, and the existence of applications that integrate them.

RQ3 analyzes which field should be prioritized for accelerating AI technologies combined with powered IoT sensors, and the bibliometric analysis shows that the healthcare IoT sector is the most active in this area. That is precisely why the fields that should be intensively explored are those in the agricultural industry due to their importance in ensuring food for the population and animals. The impact in this regard would be twofold: it would increase resources and provide an approach to optimizing food production. Therefore, the authors believe that agricultural research integrating AI and powered sensors should be a priority.

The bibliometric analysis conducted in this paper demonstrates a geographical dispersion of publications that includes locations across multiple continents. This geographical diversity reflects the global concern for the development of self-powered IoT sensors integrated with AI algorithms. The performance of these systems is influenced by local factors such as humidity, extreme temperatures, solar exposure, or other weather conditions, justifying the necessary studies to validate their applicability in various climatic and industrial contexts. In this way, the interest of researchers in these technologies is also justified, even if they are in different regions (China, the United States, Europe, Southeast Asia, etc.).

The fields of applicability of self-powered IoT sensors have been identified based on their frequency of occurrence in the scientific literature. According to the bibliometric analysis, the most studied fields from the perspective of integrating AI technologies with self-powered sensors have been identified. The authors acknowledge that this classification is influenced by the uneven distribution of scientific interest. Additionally, fields such as agriculture are underrepresented in the implementation of these technologies, an aspect that the authors highlighted in their paper [268]. Moreover, the authors of the present paper have experience in integrating IoT sensors in the agricultural field [269]. The authors will contribute to future research that directly addresses these underrepresented areas, especially in the agricultural field, and encourage other authors to take the same initiatives in their research.

The list of future research directions that emerge from the analyses conducted in this paper is as follows:

• Standardization of the taxonomy of self-powered IoT sensors;
• The integration of less commonly addressed ML algorithms in self-powered system applications;
• Development of end-to-end systems that cover the entire functional chain from capture—local processing—automatic decision-making;
• Creating interconnected IoT platforms between applications to ensure compatibility between infrastructures;
• Conducting comparative studies between self-powering technologies (TENG, piezo, BFC, HEG) tested under real conditions;
• Optimizing AI architectures for energy-limited hardware when Cloud infrastructure is not a viable option;
• Expanding field studies in varied climatic and geographical conditions (tropical, polar, desert, harsh industrial;
• Exploration of emerging subfields such as Federated Learning for distributed training without raw data transfer, neuromorphic computing for event-based inference in ultra-low-power systems, and implementations on microcontrollers powered by renewable sources.

In addition to the trends identified in the bibliometric analysis, there are subdomains that redefine, from an architectural point of view, self-powered IoT networks. Among these are the following:

• Federated learning, which trains ML models in a distributed manner without transferring raw data;
• AI model compression makes it possible to reduce memory space and energy consumption for local inference;
• Implementations on microcontrollers powered exclusively from recoverable energy sources demonstrate the feasibility of applications in real-world environments.
• Neuromorphic computing is a research direction that processes event-based AI components in ultra-low-power systems.

These directions are underrepresented in the analyzed literature at this stage and may represent future development paths for AI components in self-powered IoT nodes.

AI algorithms integrated into self-powered IoT systems highlight a series of specific technical challenges. These are rarely addressed directly in the specialized literature. One of these challenges is the duty cycling of AI processing. This processing refers to running algorithms exclusively at times when the sensor has enough stored energy to perform such operations. To accomplish such a task, synchronization between data collection and energy availability is necessary, as well as for the actual execution of inference in cases where the model is not resident in the cloud. Another issue concerns the trade-off between inference accuracy and energy consumption. This refers to the fact that complex AI models provide results with very good evaluation metrics, but they cannot operate efficiently under fluctuating energy conditions. Another issue concerns the temporal inconsistency of data caused by intermittent power sources, as real-time sequential interpretation is performed. These challenges require solutions such as intelligent buffering, adaptive model selection, anticipatory prediction of available energy, energy consumption streamlining and optimization, or AI models tolerant to missing data. These represent research directions that researchers should consider as benchmarks in this field.

The open issues and future research directions aimed at ensuring the maturation of the field primarily concern the lack of standardization of the interfaces between self-powered sensors, AI modules, and IoT networks. The real-time implementation of AI algorithms on intermittently powered nodes requires studies addressing buffering, data loss, latency, and real-time implications. The study of AI models under variable energy conditions requires new paradigms, such as continuous learning or incremental learning, directly on self-powered IoT nodes. These paradigms will improve the adaptability of systems to environmental variations and sensor wear without the need for complete retraining. In this context, future research should address the integration of these directions into the proposed architectures.

9. Conclusions

The analysis of the works from the specialized literature presented in this study provides an overview of the evolution of research in IoT sensors, emphasizing self-powered sensors and their integration into applications that use AI technologies. Based on the obtained data, the paper identifies the main IoT domains where sensors are used, classifies the types of sensors concerning these domains, and explores the level of academic interest in them concerning AI and ML technologies. The applications of self-powered sensors are discussed in detail, and the AI tools are integrated into data collection, processing, transmission, and interpretation. The study’s results outline future research directions.

One of the study’s most important findings is that the field of IoT cannot be approached without integrating an AI component. Thus, the sensors used in modern applications perform extended functions. In addition to their measuring function, they have become standalone devices that integrate active elements of the connected digital ecosystem, allowing them to provide real-time data and make automated decisions. Decision-making is possible due to integrating AI components, mostly ML algorithms. The literature analysis reveals disparities within the IoT field regarding the degree of adaptation of AI technologies. Most works focus on the healthcare domain, where CNN and LSTM are addressed, leaving other ML algorithms underexplored. Fields such as agriculture have provided limited papers in the specialized literature. This difference indicates a need to expand research on the applicability of AI tools in all IoT domains where self-powered sensors are used.

Another point of interest in discussions related to the energy autonomy of these IoT sensors highlights the capability of TENG, BFCs, and HEG sensors to be integrated into standalone applications. The paper highlighted significant gaps regarding the durability aspects of these sensors under extreme conditions, outlining future research directions:

• Comparative studies between different energy harvesting technologies by varying the environment in which they perform measurements;
• Evaluating the impact of power variations on the accuracy of ML models;
• The development of hybrid power systems that ensure energy redundancy in operation.

The studies in the specialized literature that integrate AI models demonstrate their superior capabilities in transforming data to obtain helpful information necessary for automated systems. The most explored models are RF, LSTM, CNN, and SVM. These prove their efficiency in searching for human activity recognition, anomaly detection, industrial fault prediction, image classification, medical monitoring, etc. The performance analysis of these models is mainly based on the accuracy indicator, which raises significant issues related to the need for expanded research on performance indicators that can be used in a comparative analysis of these models.

Another research direction in the context of AI-IoT models includes the following:

• Multicriteria comparative analysis that, in addition to accuracy, includes energy consumption, speed, complexity, etc.;
• Exploring optimization methods for AI models to function in tandem with the embedded hardware of self-powered sensors;
• Comparison between edge computing versus cloud computing implementations in IoT-AI applications that include self-powered sensors.

Based on a detailed analysis of the existing research in the literature, the authors derive the following future research directions:

• Standardization of IoT sensor taxonomy, as the lack of a universally accepted classification leads to difficulties in comparing research and guiding development strategies. Thus, the scientific community should establish a common classification framework for self-powered sensors.
• Expanding AI research into underdeveloped fields, such as agriculture, light industry, education, and so on, future research projects should explore customized methods for integrating AI into these domains.
• Optimizing AI models for the hardware of self-powered sensors, especially for reduced computational resources. Thus, future research should focus on compressing ML models, using dedicated edge AI frameworks, and energy-aware ML modeling for IoT applications.
• AI integration in the capture, processing, and decision-making chain, as most current research focuses on data capture or processing. Thus, a future direction should be the approach of end-to-end systems that integrate AI into the entire flow of data capture, local data processing, and automated decision-making. To achieve such a goal, interdisciplinary collaboration between engineers, AI researchers, and material designers is necessary.
• Exploring synergies between IoT domains requires creating holistic solutions that consider the impact of a decision in one domain on others. This involves the development of IoT integration platforms capable of interconnecting heterogeneous sensors in communication between applications.
• Studying interoperability in IoT-AI networks by securing the data transmission method in architectures for distributed data processing, authentication and encryption mechanisms adapted to self-powered sensors, and communication standardization in IoT ecosystems.

The results of bibliometric and thematic analyses demonstrate the difficulties that new trends in the integration of self-powered sensors into AI algorithms face from the perspective of theoretical and applied approaches. From a theoretical perspective, the integration of AI components into ultra-low power consumption systems requires the development of new adaptive inference models that operate under strict hardware constraints when using local models. In the case of cloud-hosted models, IoT sensors require cloud-compatible infrastructure. The lack of standardization in the field of self-powered sensors creates a discontinuity between laboratory innovations and industrial scaling. From a practical perspective, the need for studies targeting the following technological strategies is highlighted:

• The development of ML algorithms compatible with low-power microcontrollers;
• The implementation of modular platforms where self-powered sensors can be integrated into existing IoT networks;
• The stability of common performance measurement standards for self-powered devices.

In conclusion, the present research demonstrates that the evolution of self-powered IoT sensors is closely linked to the advancement of AI technologies, specifically ML components. To accelerate the potential of technological evolution, the authors found the necessity of a multidisciplinary approach that requires academic–industrial collaborations, governmental support, and sustained investments in applied research.