3.1. Scientific Production
In the Scopus database, 1535 documents are found, in which 3672 keywords are employed by 4893 authors. The authors of single-authored documents are 79; the authors of multi-authored documents are 4814; the coefficient, showing co-authors per documents, is 4.08; and the collaboration index is 3.46.
In the scientific database Web of Science, a smaller number of documents (434) is identified, written by 1810 authors as the authors of single-authored documents are 27 and the authors of multi-authored documents are 1783. The coefficient, presenting co-authors per documents, is 4.72, and the collaboration index is 3.38.
It seems that a big part of the indexed documents in Scopus and Web of Science are prepared in collaboration among several authors as the countries of the corresponding authors with the biggest contribution according to Scopus are: USA, China, India, Japan, and Germany (
Figure 2). The parameter Multiple Countries Publication (MCP) measures the international collaboration as at least one co-author should be from another country. The parameter Single Country Publication (SCP) indicates that the corresponding author and his co-authors are from one country.
The annual scientific production in the period from 2011 to 2020 is presented in
Figure 3: for 2011, 35 documents are indexed in Scopus; for 2012, 58; 2013, 63; 2014, 65; 2015, 60; 2016, 81; 2017, 120; 2018, 223; 2019, 384; 2020, 446. The observed tendency is described with an increasing curve that can be explained with enhanced interest to this topic by researchers. No jumps causing turning points are visible on the curves. Contrariwise, the curve smoothly raises as the calculated annual growth rate is 32.68%. The annual scientific production indexed in Web of Science is also described with a smoothly increasing curve as the found documents for 2011 year are 8; for 2012 and 2013, 16; 2014, 15; 2015, 16; 2016, 33; 2017, 44; 2018, 54; 2019, 107; 2020, 121.
The USA, China, UK, India, Canada, Switzerland, Italy, Japan, Germany, and the Netherlands are among the most cited countries. The maximal value of the average citations per documents, indexed in Scopus, is 8.5, obtained for 2017 (
Figure 4).
3.2. Characteristics of Sources
Receiving information about the sources of published papers and their impact gives an opportunity for detailed investigation into the topics of research interest. The intention is concrete papers and their discussed problems to be further explored for outlining some important key research points and main scientific themes. For this purpose, the results from the bibliographic analysis are used to point out the most influential sources. According to the Scopus database, the papers are published mainly in different conference proceedings, journals and books; the twenty most relevant sources are presented on
Figure 5. After the source grouping, it is evident that the most relevant proceedings are from the following scientific forums:
Industrial Electronics Conference,
IEEE International Symposium on Industrial Electronics,
International Symposium on Low Power Electronics and Design,
Annual Conference of the IEEE Industrial Electronics Society,
SPIE conferences, organized by the International Society for Optical Engineering,
and the most preferred journals and books are:
IEICE Transactions on Information and Systems,
Lecture Notes in Electrical Engineering,
IEEE Access,
Sensors,
Lecture Notes in Computer Science,
IEEE Journal of Solid-State Circuits,
Scientific Reports.
According to the bibliographic data extracted from Web of Science, the top sources are the following scientific journals: IEEE Access, Scientific Reports, Journal of Instrumentation, Sensors, ACS Applied Materials & Interfaces, Applied Sciences-Basel, Electronics, Elektronika ir Elektrotechnika, IEEE Consumer Electronics Magazine, IEEE Journal of Solid-State Circuits. The obtained results show that the conference proceedings include a smaller collection of the scientific papers in comparison to the journal contributions.
Bradford’s law is depicted on
Figure 6 (according to Scopus). It presents the number of sources in the first zone (so called core sources), which have the biggest contribution for papers publication in the explored scientific area. According to this law, the journals are grouped in three zones, as the number n of the published papers related to the examined topic forms a geometric series: 1:n:n
2:n
3. In the case of the query “
machine learning and
electronics”, the first zone includes 49 sources, mainly journals and conference proceedings, as the number of all sources is 821. The curve, which describes Bradford’s law and is constructed according to the Web of Science database, has a similar form; in the first zone, there are 57 sources from 347.
It is noticeable that the number of sources with open access among the most relevant ones remains low; according to Scopus, only IEICE Transactions on Information and Systems and Journal of Physics: Conference Series have open access. According to Web of Science: Sensors, ACS Applied Materials & Interfaces, Applied Sciences-Basel, Electronics, and Elektronika ir Elektrotechnika are open access.
The source importance could be evaluated through different parameters; one of the most popular is the Hirsch index (h-index). The h-index is a measure, showing the number of published papers h as each paper must be cited at least h times. The sources with the highest impact concerning h-index are summarized in
Table 1 according to Scopus and Web of Science. The results are gathered separately, searching in Scopus and Web of Science. Then, the sources are combined and arranged in the direction of higher to lower h-index.
3.4. Thematic Evolution
The scientific knowledge dynamics and knowledge evolution are grasped through the generation of some thematic maps, which are created after applying the clustering algorithm Louvain regarding the most utilized 500 authors’ keywords. Each thematic map consists of four quadrants as each quadrant shows one or more main themes/clusters. The first quadrant includes so-called “motor” clusters, the second quadrant collects “highly developed and insolated” themes, the third quadrant “emerging or declining” clusters and the fourth quadrant “basic and transversal” themes. Additionally, a thematic map is characterized by two main parameters: (1) centrality, which is related to the theme importance and (2) density, which concerns the theme development. Each cluster on the thematic map is presented through a bubble, whose name is a word with the highest occurrence in this cluster. The bigger bubbles are characterized by a higher frequency of word occurrence. The themes evolution can be studied as the examined period (2011–2020 year) is divided to time slices. Here, four time-slices are formed, taking into account the curve form of the annual scientific production (see
Figure 3): time slice 1 reflects on the period 2011–2014, time slice 2 summarizes the 2015–2016 years, time slice 3 includes 2017–2018 years and time slice 4 describes 2019–2020 years.
In the first time slice, presented on
Figure 8, the “motor” cluster is machine learning, which includes several other important keywords such as:
classification,
fault diagnosis,
computer vision,
pattern recognition,
feature extraction,
human-computer interaction,
condition monitoring,
defect prediction,
consumer electronics.
Neural networks and
genetic algorithms are also classified as “motor” themes. In the “emerging or declining” quadrant is the theme
sentiment analysis. The “basic and transversal” themes are:
support vector machines (other valuable keywords in this cluster are
wireless sensor networks and
anomaly detection),
accelerometer and
data mining. It can be noticed that a big part of the cluster names are keywords typical for machine and deep learning scientific fields, data science and computational linguistics (sentiment analysis). Only the
accelerometer (electronic device used for example in vehicle control system) cluster pertains to the electronics’ domain.
This time slice characterizes the period from 2011 to 2014 year, when the “motor” clusters are machine learning, neural networks and genetic algorithms. The machine learning cluster is presented with the biggest bubble in comparison to the other clusters on the map, which indicates the biggest research interest and contribution to this topic. Support vector machine is a machine learning algorithm and is separated in a different cluster, placed into the “basic and transversal” themes. It means that especially this algorithm is so popular in the first observed time slice among researchers, and it is utilized in different scientific works. The cluster accelerometer is not placed in the center that is related to its not-so-big importance for research according to the parameter centrality.
In time slice 2 (
Figure 9), the “motor” theme is related to
neural network and the “highly developed and insolated” themes are connected to the keywords
low-power design and
artificial intelligence. The theme
support vector machine is seen as a “emerging or declining” cluster. In this case, it is a “declining” cluster because in the next time slices it does not appear as a separate cluster as this keyword is included in other clusters. In the fourth quadrant, “basic and transversal” themes include
Internet of Things (IoT), industry 4.0, machine learning and
classification. In time slice 1, the keyword
classification was part of the cluster
machine learning, while here it forms a separate cluster. The closest cluster to the center of the coordinate system is
artificial intelligence that could be explained with the theme “importance and progressive development” in this time slice.
It seems that during these two years (2015 and 2016), the popularity of machine learning for solving a wide variety of problems in electronics increases as it is converted from “motor” cluster to “basic and transversal”. Algorithms for classification are separated in different clusters that draw their importance from the investigated domain. Internet of Things and Industry 4.0 are formed as different clusters as the first one is more developed, taking into account the parameter Density. The “motor” cluster in this time slice is neural network, which is connected to its possibility for deciding deep learning tasks. The bubble is bigger than this in time slice 1, which shows a higher number of scientific papers that utilize neural network apparatus.
In the third time slice (
Figure 10) the “motor” theme is
classification, which includes keywords
feature extraction, fault detection and
big data. The second quadrant “highly developed and insolated” themes consist of three clusters:
condition monitoring, security and
power electronics, as the
optimization cluster is placed on the boundary between the second and third quadrant. In the third quadrant, “emerging or declining” clusters are grouped in
artificial neural networks, recurrent neural networks and
organic electronics themes. The fourth quadrant “basic and transversal” theme includes
neural network, IoT, machine learning and
neuromorphic computing.
In this time period (2017–2018), the cluster machine learning continues to be popular in scientific society and is placed in the “basic and transversal” quadrant. The cluster classification is removed from the “basic and transversal” themes, becoming the “motor” topic. The opposite can be said for neural networks, whose cluster is moved from “motor” themes to the “basic and transversal” quadrant. Obviously, the research attention is focused on big data and how the classification problems could be decided.
The “motor” clusters in time slice 4 (
Figure 11) are formed around the keywords
Internet of Things, neural networks and
artificial neural networks. The
wireless sensor networks,
electronics and
sentiment analysis are classified as “highly developed and insolated” themes. The “emerging or declining” clusters are topics related to
low power and
neuromorphic. The fourth quadrant includes “basic and transversal” themes:
power electronics, IoT, machine learning and
reinforcement learning. It can be observed that in this time slice, more sector names are keywords, typical for the electronics domain. It means that the mutual connection among keywords from the scientific areas of machine learning and electronics is strengthened in comparison to the previous time slices.
In this period (2019–2020), the machine learning cluster remains its place in the “basic and transversal” quadrant, continuing to be in the center of the scientific interest. The classification cluster is absorbed by the machine learning cluster, stopping to present a separate research topic. A new cluster, reinforcement learning, is formed, which is detached from the machine learning cluster. Neural networks are again among the “motor” themes, confirming their importance in scientific development.
If we consider the time slice position of the theme machine learning on the thematic maps and investigate it, its changeable mode is visible. In time slice 1, machine learning is a “motor” theme; in the next three time slices, it is classified into the “basic and transversal” cluster. Its position in the fourth quadrant is also changeable. The motion is visible from a more central to more peripheral place and again back to a more central position. This cluster content is also changeable as some merges and divergences with other themes are noted.
The cluster neural networks is also characterized by its dynamics. In the first time slice, although it is placed among the “motor” clusters, it is presented with a very small point that indicates smaller scientific production devoted to this research topic. The interest in the neural networks increases in the next time slices, as during periods 2 and 4, they form bigger “motor” clusters. In time slice 3, two clusters related to neural networks emerge: artificial neural networks and recurrent neural networks, which are placed in “emerging or declining” themes.
While the clusters low-power design and Internet of Things emerge in time slice 2, the cluster power electronics is formed in time slice 3 and the new cluster electronics occurs in time slice 4.
The thematic evolution map (
Figure 12) shows the complex process and metamorphoses of a starting field of science to its merging or splitting with other scientific areas. Additionally, new emerging scientific sections are depicted.
3.5. Analysis of Selected Papers
For further analysis, the chosen journals are among the most relevant and with the highest-impact sources; another important factor for selection is their open access. Open access is the preferred format for further exploration because it enables research ideas to be widely spread among scientists regardless of their geolocation. More and more authors prefer to publish in open access journals for wide spreading of their novel and innovative solutions. Such an approach is extremely suitable for idea dissemination in emerging scientific areas, as is the case of the utilization of machine learning in electronics. Thus, the following open source journals are taken into account for investigation as the total number of reviewed scientific papers is 44:
Journal of Physics: Conference Series (8 papers are discussed),
MDPI Sensors (15),
MDPI Applied Sciences (5),
MDPI Electronics (6),
Elektronika ir Elektrotechnika (7),
Nature Communications (3).
In the Journal of Physics: Conference Series, 25 documents are found to satisfy the query “machine learning and electronics”, from which eight are selected for examination because of their strong relevance to the researched topic.
In the first one, Arul et al. report an intelligent system for fire detection as early as possible to avoid bigger damage [
16]. CCTV cameras are used for visual information gathering. The system performs further processing for the generation of a notification message. The machine learning technique is applied to compare the currently grasped visual pictures of the place with the pre-defined picture set of the same place when there is a fire.
Wei and Jia present how machine learning is used for the identification of an electromyographic signal (sEMG) for capturing the position of a robotic arm based on a muscle sensor [
17]. The accuracy of six algorithms is compared: Latent Dirichlet Allocation (LDA), Decision Tree (DT), Backpropagation Neural Networks (BNN), Support Vector Machines (SVM), Random Forest (RF) and K-Nearest Neighbors (KNN) for recognition of four arm gestures. The chosen algorithm for experimentation is SVM.
Cheng et al. propose a method for the detection of defects in power equipment through the Faster region convolution neural network [
18]. The method is characterized by high accuracy for defect detection for purposes of power inspection.
Lasbahani and Taoussi present a new approach for big data securing in real time through applying an unsupervised machine learning algorithm [
19]. The anomaly detection is revealed during the big data processing.
An energy management system with the capability to control its building microgrids is presented in [
20]; it generates alarm messages when equipment failure occurs. An improved SVM algorithm is used for prediction of the fault equipment based on microgrid data. The accuracy of the created model is verified through an example.
Deep machine learning is applied for the realization of an improved voice recognition system [
21]. The proposed method for end-to-end speech recognition leads to simplifying the traditional complex process of speech recognition.
Machine learning is used for sentiment analysis as an improved feature selection method is applied [
22]. The authors Danyang and Huimin have proven the effectiveness of this method at emotion classification through experiments.
An application of machine learning in the implementation of a social robot and its facial and expression recognition systems is presented in [
23]. The experimental results prove a good accuracy of recognition and good human–robot interaction.
From the journal Sensors, the topics of 15 documents are examined below.
Nam et al. discuss contemporary technologies for touchscreens and the application of machine and deep learning for different purposes, such as: user identification through touchscreen, detection of different gestures, improvement the accuracy of the touch location recognition and correct recognition of multiple input sources [
24].
An application of machine and deep learning algorithms (convolutional neural networks (CNN), kNN, SVM) for the identification of product quality is presented in [
25]. The proposed learning architecture is able to determine welding defects according to extracted information from infrared pictures.
Signoretti et al. present a new algorithm called Tiny Anomaly Compressor for data compression in the infrastructure of Internet of Things [
26]. It is based on machine learning techniques for intelligent tiny devices that can analyze local data in real time.
Wearable devices for monitoring a person’s health when different working activities are performed, anomaly detection in their health state and possibilities for accidents reduction are discussed in [
27]. The working environment is becoming safer through a combination of the capabilities of supervised and unsupervised machine learning and suitable IoT infrastructure.
Morales-Molina et al. propose an approach for IoT cyber-attack detection and protection framework driven by artificial intelligence methods [
28]. Unsupervised techniques are utilized for the selection of the main characteristics of low power and Lossy Network and Dense Neural Network algorithm for precisely deciding a classification task.
Novac et al. present a framework for deep neural networks training and deployment on microcontrollers called MicroAI [
29]. It is compared with embedded inference engines concerning the main characteristics, such as the power consumption and memory utilization.
Sensor-based gloves with included microphone, gyroscope and accelerometers are described by Cerro et al. to facilitate the realization of correct connections among parts in a production line [
30]. This intelligent system improves the products quality and decreases its cost.
The problem of noise pollution in smart cities and its impact on human health is discussed in [
31] as well as a solution in the form of a mobile embedded system that is based on a machine learning approach and geo-sensing application. Such a system gives more information about the current noise situation to the authorities.
An electronic nose in support of a forest employee who has to recognize forest pathogens such as Pythium and Phytophthora is discussed in [
32]. This application includes sensors and the data collected by them are classified through the SVM algorithm.
Rodríguez-Rodríguez et al. collect data from wearable sensors and predict the blood glucose level with the aim of finding the optimal therapy for each patient [
33]. The accent of their work is how to create accurate short-term predictive models, personally driven, with minimal data volumes.
Rezaei et al. propose a core–sheath fiber strain sensor suitable for smart textile clothes for grasping the movement and for monitoring the persons’ activities [
34]. The received sensor signals are processed with a machine learning algorithm.
Peng and Li present a literature review regarding sensor-based radar systems with applications for short-range localization and life tracking [
35]. The improved radar systems with integrated machine learning algorithms for data processing are also examined.
A new sensor for monitoring the blood pressure, based on machine learning algorithms, is introduced in [
36]. Machine learning supports the recognition of values from the shape of the pulse wave. The distance change between the moving surface and the fixed electrode is measured as the change in capacity. This approach is suitable for non-invasive blood pressure monitoring, diagnosis, and treatment.
An embedded smartphone app for stand-alone usage with the possibility for evaluation of the balance and for the prevention of the fall risk of elderly people is discussed in [
37]. The evaluation function is performed by a machine learning algorithm. Additionally, inertial sensors for the smartphone are used for monitoring the balance and risk of falling.
The problem for a cheaper approach to spatial tracking of unknown signals in ad hoc wireless sensor networks is solved in [
38]. For this purpose, machine learning algorithms are used, which work in two phases related to modeling and tracking of the spatial signal. The sensors are presented as neurons as the signal identification is possible after calling a part of these neurons.
From the journal MDPI Applied Sciences, five scientific papers are selected for further examination.
Teo et al. use machine learning algorithms to predict the correct location of the touch on the screen of mobile devices such as tablets and smart phones [
39]. These devices are constructed with precise transducers (accelerometers and microphones) and produced by signals at touch processed through the Random Forest algorithm. The created predictive model is characterized by high accuracy.
Byeon et al. analyzed 10 postures from a large volume of images with the aim of supporting elderly people when they move alone at home and to prevent some dangerous situations [
40]. The recognition of postures is performed through applying convolutional neural networks.
Dineva et al. propose a new methodology for multi-label classification at the diagnosis of multiple faults occurring in electrical machines and drive systems [
41]. In addition, the authors evaluate the severity of faults in noisy environments. The created multi-classification models are compared and validated through experiments.
Li et al. is interested in the realization of flexible system with more flexible devices (such as thin-film transistors) for usage in sensors, wearable devices and energy suppliers, including in the computation parts [
42]. Therefore, they propose a new circuit architecture for the purposes of an image processing accelerator based on thin-film transistors. Image pre-processing and classification is performed in real time with good parameters regarding the energy consumption and image classification speed.
A region-based convolutional neural network (R-CNN) is used for dent localization on a car surface that could be formed during the manufacturing process [
43]. The accuracy of the dent detection is very high, which contributes to producing vehicles with high quality, because of the improved inspection process.
Six papers from the journal MDPI Electronics are chosen as more relevant, and then they are reviewed.
The first one treats security problems as in this work an intrusion detection system for attack identification in a wireless network is proposed [
44]. The system works in two stages as the created machine learning model conducts classification tasks concerning the traffic records at each stage. The model is validated and is characterized with high accuracy.
A kinetic energy harvester model for horse monitoring is presented in [
45]. Accelerometer data are processed through machine learning techniques in order to identify early symptoms of health problems. For continuous data-sensing during training and sleeping sessions, different wireless technologies are applied as the key observed factor is optimal energy consumption.
Mursi et al. show one work from the area of crypto analysis as they present a method, based on neural networks, for attacking XOR Arbiter Physical unclonable functions (XPUF) (it is a hardware security primitive) [
46]. XPUFs are used for the generation of a cryptographic key or for device authentication. The experimental results prove that this new method is capable of breaking the 64-bit 9-XPUF for 10 min.
A platform for distributed parallel processing of image streams in real time, which utilizes a deep learning model for inference, is discussed in [
47]. This platform manages different services in an efficient way, distributes resources and ensures large-scale image stream processing.
Jose et al. discuss a strategy for the implementation of an artificial neural network algorithm on microcontrollers that is suitable for IoT and Machine-to-Machine (M2M) applications [
48].
Du et al. propose a method for feature extraction and selection for the optimization and classification of faults at a two-level three-phase pulse-width modulating rectifier [
49]. Such a fault diagnosis approach based on open-circuit measurement can prevent unwished shutdown and can decrease the number of failures.
From the journal Elektronika ir Elektrotechnika, seven papers are selected and reviewed.
Sovilj-Nikic et al. use a regression tree algorithm for the creation of a predictive model about phone durations in the Serbian language [
50]. The model has been developed taking into account several features that are obtained from a speech database in the Serbian language. The results from comparison of the evaluated model with models for other languages show similar findings.
A method for Alzheimer’s disease diagnosis through a machine learning classification algorithm is proposed in [
51]. The recorded EEG signals are bases for feature extraction and their further analysis. The received results prove that this method is a suitable alternative to well-known approaches for diagnosis.
Aoad presents a design of a microchip antenna with spiral shape, which is modeled through machine learning with the aim for its achieved accuracy to be high [
52]. Three machine learning models are created, and they are compared through data taken from simulation and from measurement. The findings show that the developed small antenna operates in a wide high frequency diapason.
Du reports a new method for fault diagnosis of mine hoist that includes machine learning techniques such as SVM and genetic algorithm [
53]. The proposed method is evaluated through experimental data as the detection rate of fault diagnosis is increased.
Teodorovic and Struharik present a hardware accelerator and novel algorithm for sparse decision tree induction with possible usage in embedded and edge applications, where there are several constraints concerning memory usage, storage capacity, and limited bandwidth [
54]. The proposed approach is validated experimentally through standard UCI machine learning datasets and the results show a significant reduction in memory and storage usage as the accuracy remains the same.
Gradolewski et al. report a safety system for a collaborative robot that has the ability to prevent human health damage [
55]. It is developed upon techniques from distributed computing, machine and deep learning, computer vision and sensing, robot motion control. The authors plan to improve the proposed system as the person at the workplace will be tracked all the time and the system will react only when entering the robot operational zone.
A usage of the Q-learning algorithm for control management of solar-powered environmental wireless sensor network nodes is presented in [
56]. With the aim to improve the computer performance, machine learning techniques are applied.
Three papers from the Nature Communications journal are considered as relevant and they are reviewed.
Loke et al. discuss the realization of wearable devices for monitoring the physiological states of human body integrated through digital fiber in a shirt [
57]. Some devices are used such as temperature sensors and memory devices as the collected data are proceeded through the neural network algorithm. This approach allows measured and stored data in a fiber to be used from algorithms to make suitable inferences in the area of the body temperature control or autonomous drag delivery.
Maurya et al. present a process for 3D printing of piezoresistive graphene sensors, which are integrated in car tire with the aim to measure the voltage waveforms during car motion [
58]. In addition, a machine learning algorithm is created to evaluate the tire pressure. The authors believe that this work is an important step in the construction of smart tires for autonomous cars.
An architecture of multi-resistive synapses typical for neuromorphic computing is presented by Boybat et al. as it is designed to operate in a wide dynamic range and with high accuracy [
59]. The authors achieved very good results after experimentation with memory devices used for unsupervised learning.
In summary, the findings from the performed bibliometric analysis and literature review in this section are presented through a developed conceptual framework (
Figure 13). It outlines the current research areas that benefit in parallel from scientific achievements in machine learning and electronics in order to different intelligent hardware-based solutions to be created, improved or optimized. Several supervised and unsupervised machine learning techniques, algorithms for deep learning and reinforcement learning as well as evolutionary and genetic algorithms are used for the development of new methods that are capable of improving existing events, systems and processes or to propose novel solutions.