Healthiness and Safety of Smart Environments through Edge Intelligence and Internet of Things Technologies

Abstract

Smart environments exploit rising technologies like Internet of Things (IoT) and edge intelligence (EI) to achieve unseen effectiveness and efficiency in every tasks, including air sanitization. The latter represents a key preventative measure–made even more evident by the COVID-19 pandemic–to significantly reduce disease transmission and create healthier and safer indoor spaces, for the sake of its occupants. Therefore, in this paper, we present an IoT-based system aimed at the continuous monitoring of the air quality and, through EI techniques, at the proactively activation of ozone lamps, while ensuring safety in sanitization. Indeed, these devices ensure extreme effectiveness in killing viruses and bacteria but, due to ozone toxicity, they must be properly controlled with advanced technologies for preventing occupants from dangerous exposition as well as for ensuring system reliability, operational efficiency, and regulatory compliance.

1. Introduction

Pathogenic microorganisms are commonly found in the air of crowded environments such as schools, museums, and workplaces [1]. In these settings, a simple sneeze can quickly spread colds or seasonal flu, leading to significant social costs. Even more serious are airborne diseases like whooping cough, measles, acute lung infections, and tuberculosis, which cause millions of deaths worldwide each year [2]. Improving air quality can, therefore, provide significant benefits to the individual and the community: this was clearly highlighted during the SARS-CoV-2 [3] (coronavirus) pandemic, which prompted extensive action from health authorities.

Biological contaminants are difficult to control due to the small size of microorganisms, which can escape even the most efficient filters. While sterilization is hard to achieve in normal environments, disinfection–maximally reducing pathogenic germs through physical or chemical methods–is a more attainable goal. Installing germicidal ultraviolet-C (UV-C) lamps in ventilation ducts is a highly effective method for achieving high levels of air disinfection without exposing occupants to intense radiation [4]. Indeed, when UV-C rays hit oxygen (O₂) in the air, ozone (O₃) is created, and this has very high disinfectant properties and can penetrate places where UV-C rays cannot reach, providing greater coverage. In scenarios like schools, offices, and public places, the so-called process of “ozonation” helps maintain low microbiological contamination, being able to theoretically achieve up to 99.99% air disinfection with proper levels and duration of irradiation. Therefore, interest in UV-C lamp air disinfection has recently surged, driven by the increasing resistance of germs to chemical disinfectants and greater awareness of the risks posed by pathogens in crowded environments. However, when designing an air disinfection system with UV-C lamps, it is essential to consider key aspects related to the safety and effectiveness of the treatment: indeed, given the ozone toxicity and the UV-C lamps aging (thousands of hours of work with proper temperature and humidity levels), these systems must be properly controlled with advanced technologies for preventing occupants from dangerous exposition while ensuring proper sanitization standards. Therefore, accurate presence monitoring systems, ubiquitous deployments of heterogeneous environmental sensors, adequate notifications systems, and dynamic actuators’ operations are some of the functionalities that make the Internet of Things (IoT) [5] a key enabler for the healthiness and safety of smart environments (i.e., a physical space provided by a network of interconnected devices that collect data and use it to provide intelligent services to users) [6].

In this article, we propose an IoT-based system that, through a continuous monitoring of the air quality within an indoor scenario (e.g., office, classroom, mall, museum), enables its sanitization by exploiting ozone lamps. By integrating different hardware and software technologies and distributing computation and storage tasks across the local and remote elements, the proposed system fully automates the operation of UV-C lamps in full safety: such approach advances the state-of-the-art solutions, which are, in most cases, exclusively focused on air quality monitoring and dependent on human operators for managing sanification actions (with obvious difficulties and costs if the target deployment is of a large scale). In particular, edge intelligence (EI) [7] techniques and the Amazon Web Services (AWS) [8] infrastructure are exploited to streamline the management of on-premise IoT devices and to distribute the processing of their data between the edge and the cloud, thus maximizing the effectiveness of UV-C lamps and their reliability, safety, and lifetime. Indeed, the analysis, design, and implementation phases of such hybrid edge–cloud systems are detailed by presenting both hardware and software components and their contribution to the implementation of the different system functionalities (monitoring, warning and notification, storage, etc.).

The rest of this article is organized as follows: in Section 2, background concepts about smart environments, IoT, and EI are provided with a brief analysis of the related work. In Section 3, the application scenario is analyzed, its requirements elicited, and the resulting design choices illustrated, with benefits, limitations, and future work discussed in Section 4. Section 5 concludes the article with final remarks.

2. Background and Related Work

2.1. Smart Environment

Smart environments represent the combination of computational technologies, advanced sensing, and communication paradigms, facilitating real-time monitoring, control, and enhancement of diverse environmental parameters. Typically supported by IoT technologies, these environments are generally defined by the widespread presence of interconnected devices and sensors that gather data, allowing smart systems to react automatically to the evolving conditions within the space [9]. The notion of smart environments spans multiple domains including smart buildings, homes, workspace, hospitals, museums, and many others that emphasize improving occupants’ comfort, system efficiency, safety, and sustainability.

For example, in smart homes and workplaces, the environment can automatically adjust lighting, temperature, and security settings to suit the occupants’ needs, thereby creating a more comfortable and safe atmosphere. In smart museums, interactive displays and augmented reality initiatives can captivate visitors in innovative manners, tailoring tours to align with individual interests and offering profound insights into the exhibited collections. These technologically enriched environments significantly improve the visitor experience. In smart cities, integrated transportation frameworks can optimize vehicular flow and lessen traffic congestion by leveraging real-time data to manage public transport systems and traffic signals. This approach fosters a more sustainable urban ecosystem that prioritizes the welfare of its inhabitants. Moreover, the recent advancements in machine learning algorithms have enabled smart environments to adapt these changing conditions dynamically. This predictive capability enhances the overall responsiveness and efficiency of energy consumption by optimizing heating ventilation and air conditioning (HVAC) systems.

2.2. Edge Intelligence

EI is an emerging paradigm that effectively integrates edge and cloud computing within a continuum [7], being particularly effective for smart environments where challenges related to privacy, scalability, and economic viability are of utmost importance. If edge computing primarily focuses on processing and storage at or near the data source, regardless of its specific nature (traditional computers, smartphones, edge servers, etc.), EI aims to establish a more comprehensive computing ecosystem, where IoT devices are in the spotlight and provided with ad hoc, optimized AI models consistent with their resources’ availability [10]. Being cloud-independent but cloud-interoperable, EI enables a collaborative framework with flexible workload distribution based on real-time conditions, performance requirements, and resource availability.

Such a new paradigm is key because a large amount of data are produced in smart environments such as smart homes, commercial buildings, smart hospitals, etc. The ability to process these data in real time is crucial for sustaining the responsiveness and operational efficiency of these systems. EI fulfills this requirement by facilitating data processing in proximity to the data source—at the network’s edge—rather than depending exclusively on centralized cloud servers [11]. This computational approach significantly reduces latency, thereby facilitating intelligent systems to react more promptly to varying environmental conditions, which is especially vital in domains where immediate decision-making is paramount, such as in healthcare and safety systems. Furthermore, EI ensures privacy, as it is not necessary to send the locally processed data to the cloud, which not only abides by privacy regulations but also minimizes the risk of data breaches. Privacy concern is especially related to smart environments where personal and sensitive information is mostly at stake, such as in smart homes and healthcare systems [12]. Another beneficial aspect of EI is its scalability. Smart environments are able to manage a large number of sensor devices without over burdening the underlying network. The scalability makes smart environments more efficient in terms of robustness and economics because limited volumes of data are transmitted to the cloud.

2.3. Related Work

Smart environment monitoring is a well-explored topic in the literature, with many works focused on different goals and technologies. A comparison with the current state-of-the-art is given in

Table 1. Summary of related work on smart environment monitoring systems.

Paper	Year	Monitored Air Component	Hardware	Real Time	Kind of Local/Remote Computation	Level of Intelligence/Automation	Application Scenarios
[13]	2020	Aerosol, VOC, CO, CO2, temperature, and humidity	Pollutant detection sensors, microcontroller, and an LTE modem	Yes	Remote	Alert system and user notifications	General indoor environment
[14]	2018	CO2, CO, SO2, NO2, O3, Cl2, temperature, and humidity	Microcontroller for gas sensor nodes, wireless sensor nodes, gateways, and an IoT server	Yes	Local and remote	Gateway data processing	Residential indoor environments
[15]	2018	Temperature, humidity, pressure, particulate material, TVOC, and eCO2	Sensor nodes, Raspberry Pi 3, and Launchpad CC2650 kit	Yes	Local and remote	Data collection and analysis	Public indoor environments
[16]	2021	Temperature, humidity, eCO2, and TVOC	Air quality sensors BME680 and CCS811, along with ESP32 microcontrollers	Yes	Remote	Alert system and user notifications	Residential indoor environments
[17]	2019	CO2, temperature, humidity, CO, formaldehyde, and VOCs	ZigBee wireless sensor network	Yes	Remote	Alert system	Healthcare indoor environments
[18]	2019	CO2	ESP8266 Thing Dev (Sparkfun) microcontroller and an MHZ-19 CO2 sensor	Yes	Remote	Moderate level of automation	Public and residential indoor environments
[19]	2020	Temperature, humidity, and PM2.5 levels	Dust sensor GP2Y1014AU, DHT11 sensor, Arduino Leonardo board	Yes	Remote	alert system and user notifications	Healthcare indoor environments
[20]	2019	Temperature, humidity, PM2.5, CO2, and formaldehyde	STM32 MCU module, RS485, and NB-IoT.	Yes	Local and remote	No automation	General indoor environments
[21]	2020	VOCs, CO2, PM2.5, PM10, temperature, humidity	HUZZAH32 microcontroller board	Yes	Remote	Recommendation system	Public indoor and outdoor environments
[22]	2016	CO and HCHO gases	PC, Marvell board, sensors, and mobile devices	Yes	Local and remote	Alert system and user notifications	General indoor and outdoor environments
[23]	2020	Toxic gases like NOx, CO2, benzene, and smoke	NodeMCU ESP32, MQ-135, and DHT-11 sensors	Yes	Remote	Alert system	Industrial areas
[24]	2017	Ozone (O3)	Arduino BT IoT Prototyping Board	Yes	Local	Generate warnings when levels exceed a predetermined threshold value	Specific indoor environment
[25]	2019	O3, NO2, temperature, and relative humidity	Sensors with testbeds	Semi real	Local	Comparing and evaluating calibration methods	Indoor and outdoor environments
[26]	2019	CO2, CO, and CH4 gases	Arduino Uno, ESP8266, gas sensors, and Android devices	Yes	Remote	Prediction and alert notification	Outdoor air monitoring in smart cities
Our	2024	O3, temperature, humidity	DHT-11 and KH-NPOD-100 sensors, RaspberryPi, UV-C lamp	Yes	Amazon AWS and IoT technology for both cloud and edge computing	Autonomous UV-C lamp (de)activation, anomaly detection	Public indoor environments

, which shows, for each related work, the monitored air component, hardware devices used, an indication of whether they are implemented in a real-time environment or used just for simulations and test-beds, an explanation of the type of computation and level of intelligence, and the application scenario.

Recent advancements in the monitoring of indoor air quality (IAQ) have capitalized on the IoT to significantly improve the collection, analysis, and interaction with real-time data. Numerous studies underscore the integration of IoT technology with various sensors, facilitating continuous monitoring and instantaneous updates concerning indoor air quality parameters. For example, the Smart-Air device incorporates sensors for pollutants such as aerosols, volatile organic compounds (VOCs), carbon monoxide (CO), and carbon dioxide (CO₂) alongside LTE technology for the transmission of real-time data and cloud-based analysis, thereby demonstrating practical applications in environments such as Hanyang University [13]. Likewise, modular IoT platforms, which include sensor nodes and gateways, accentuate real-time data collection and adaptability, as evidenced by implementations at Qatar University [14]. Other systems employ custom hardware and the Constrained Application Protocol (CoAP) for device communication, thereby accommodating a variety of indoor environments [15]. Research frequently emphasizes the integration of sensors and the design of systems, utilizing an array of sensors such as BME680 (BOSCH, Gerlingen, Germany) and CCS811 (ScioSense, Eindhoven, The Netherlands) to monitor parameters including CO₂, VOCs, particulate matter (PM2.5), and temperature [16,19]. These systems employ a range of communication interfaces, including RS485, LoRa, and WiFi, thereby showcasing diverse methodologies for data transmission and system design [20]. The automation and real-time functionalities of these systems are also of considerable significance. Systems equipped with alert mechanisms notify users via SMS when air quality declines [19], while cloud platforms are capable of dispatching notifications when pollutant concentrations surpass predetermined thresholds [22]. Specialized systems designed for monitoring specific pollutants, such as ozone in proximity to photocopy machines, exemplify targeted automation in designated environments [24]. Data storage and analysis represent crucial elements within these monitoring systems. The prevalent utilization of cloud computing platforms and local databases for data storage and analysis is noteworthy. For instance, systems that utilize InfluxDB (2.7.6 version) and those integrated with Amazon Web Services (AWS) cloud illustrate the efficacy of cloud-based solutions for extensive data management [16,22]. Furthermore, the calibration of low-cost sensors within IoT networks has been addressed, contributing to enhanced data interpretation and system accuracy [25]. Practical applications and case studies are often incorporated to highlight the tangible impact of these systems in real-world scenarios. For instance, the successful deployment of the Smart-Air system at Hanyang University [13] and the operational efficacy of modular systems at Qatar University [14] serve to illustrate practical implementations. Additional studies have emphasized the application of these systems within industrial contexts [23] and the development of mobile applications aimed at pollution management [26]. This synthesis encapsulates the diverse methodologies and innovations in IAQ monitoring, thereby underscoring the pivotal role of IoT technology in enhancing real-time data collection, analysis, and user engagement across a multitude of indoor environments. Unlike other works that emphasize air quality monitoring and alerting systems, our approach focuses on the monitoring of ozone with an automated management of UV-C lamps to increase safety and efficiency. Moreover, with respect to these related works, the proposed system focuses on sanitization rather than air quality or comfort by exclusively monitoring the ozone level and, for maintenance reasons of UV-C lamps, the environmental conditions (i.e, humidity and temperature). As for many other work, IoT technologies are widely used but, as an innovative element, EI techniques are exploited to achieve a higher level of operational efficiency, automation, and safety, as discussed in the following sections.

3. System Development

In this section, we discuss the system’s development, following the analysis, design, and implementation phases. The resulting system architecture realizing the smart environment is shown in Figure 1. In particular, it can be immediately noticed that the edge components of the smart environment are wrapped in the green box and the cloud services in the gray one. Details about system functionalities, design choices, and implementation technologies follow.

3.1. Analysis Phase

In the analysis phase, the main functionalities of the system are elicited, with the goal of driving the subsequent design choices and implementation technologies. In detail, the smart environments should expose the following functionalities:

• Warning and Notification: Warnings about the operational status of UV-C lamps and ozone emission are critical to prevent accidental exposure. Therefore, to properly notify people that the area is undergoing ozonation, notifications should be sent through their smartphones. Similarly, alerts should be sent to the administrator when environmental conditions are unsuitable for a correct lamp operation or if it is nearing the end of its life or has failed, ensuring timely replacement and maintaining system effectiveness.
• Video surveillance: Once the treatment is started, it is important to avoid the presence of people in the room since ozone, if inhaled in significant quantities, can be harmful. Therefore, a video recognition system should be designed to make sure, in primis, people are not detected in the room while ozonation is taking place, and, hence, no pets.
• Operational Efficiency and Automation: Lamps’ activation and deactivation can be dynamically performed according to the manufacturer’s instructions so as to ensure continuous safe operation without manual intervention. For example, the system’s operation can be optimized by adjusting settings in response to changing conditions: ozone dose can be the enforced or the exposure time extended according to current sensor data, or, conversely, ozonification interrupted if unsafe conditions are detected, or UV-C lamps and ozone generators shut down when they are not needed, thus saving energy. Finally, once the ozonation process is complete, ventilation of the room should be initiated for the time needed to convert any remaining ozone back to oxygen.
• Reliability and Maintenance: UV-C lamps have a limited operational lifespan and may degrade over time, being sensitive to temperature, humidity, and pollutants. These can reduce the efficiency of the lamp, accelerating the degradation of the bulb or reducing the effectiveness of the lamp. Figure 2 presents the reliability of a typical UV-C lamp [27] over the operating time. In particular, the blue line shows that at 500 h and with an ambient temperature of 35 °C, 50% of the population will emit at least 45 mW (L75 means that the LED produced 75% of the initial UV-C output of 60 mW) while only 10% of devices (i.e., B10) emit less than 36 mW of their initial output, which is the L60B10 value. Hence, by monitoring environmental conditions, system administrators can ensure timely replacement and maintain system effectiveness so as to balance lifetime and effectiveness. Likewise, in the direction of predictive maintenance, trends in IoT data can be analyzed to predict when the system’s components require maintenance or are likely to fail.
• Logging and Storage: For regulatory compliance and performance analysis, it is important to keep a record of all the operations. In particular, many industries have rigid rules about ozonification, thus log data ensure that the system remains compliant with health and safety regulations.

3.2. Design Phase

In this phase, the design of the system with the selected AWS components is discussed. Among the several cloud platforms (Microsoft, Google, Bosch, etc.) offering services for managing IoT devices across the edge–cloud continuum [28], we opted for AWS due to its rich ecosystem of services designed to ensure scalability, high-availability, and flexibility in the seamless integration between local and remote computing [8,29]. Indeed, despite its steep learning curve due to its wide range of services, AWS stacks up against competitors owing to its maturity in global infrastructure and convenient pricing structure [30,31]. For example, AWS can be preferred both to Microsoft Azure (which generally comes with higher costs and Microsoft-centric environments) and Google Cloud Platform (whose market share and third-party ecosystem are limited, being often used as part of a multi-cloud strategy), just to name a few. Thus, in order to support the functionalities listed in Section 3.1, we designed our system according to the following AWS architectural components:

• AWS IoT Greengrass enables real-time operations and data analysis directly on IoT devices, without the need for a constant connection to the cloud. In particular, AWS IoT Greengrass runs AWS Lambda functions locally, synchronizes data with the cloud, and manages secure communication with other devices. In our smart environment, AWS IoT Greengrass processes sensor data such as temperature, humidity, and ozone and sends them to a PostgreSQL database for storage.
• Amazon S3 is a scalable storage service that allows for the storage and retrieval of data through an Internet connection, at any time and from anywhere. In our scenario, Amazon S3 stores images and videos captured by the camera connected to the RaspberryPi to realize the video surveillance functionality.
• AWS IoT Core enables the secure connections of IoT devices to the AWS cloud, manages data, and interacts with other AWS services. In particular, in our smart environment, it is exploited to manage communication between the Raspberry Pi and AWS services, using the MQTT protocol to transfer the collected data.
• AWS Lambda is a serverless computing service that automatically executes code in response to events and manages the necessary resources. In our scenario, the functions process the sensor data and the video arriving through the IoT Core, triggering other operations within the system. Below we present snippets of functions used for processing and storing sensor data (notifying in the case of humidity, temperature, or ozone values over the established thresholds) and for image processing (notifying in the case of people or pets detected in the video).

• AWS SageMaker is a fully managed service for building, training, and deploying machine learning models at scale over the data collected from the sensors or images, using machine learning models to make predictions or classifications.
• Amazon Rekognition is an image and video analysis service that uses artificial intelligence to recognize objects, faces, scenes, text, and activities. With respect to our system, it is aimed at detecting the presence of people inside the room where the ozonation occurs, triggering a notification through Amazon SNS service if necessary.
• Amazon DynamoDB is a fully managed NoSQL database service designed for applications that require high performance and scalability. Amazon DynamoDB stores data related to events or analysis results in a NoSQL database, providing a scalable solution for data management.
• Amazon SNS is a messaging service for sending push-notifications to mobile devices, email, SMS, or forward messages to services like SQS and Lambda. In our scenario, events such as the presence of people where ozonation is occurring or the (de)activation of an ozonizer in adequate environmental conditions act as triggers for the SNS notification system.
• Alarm (CloudWatch) is back-office service aimed at the monitoring and management of the deployed infrastructure. In particular, it allows for notifications based on system metrics, such as CPU, memory usage, and anomalies during the execution of services.

3.3. Implementation Phase

The specific hardware and software technologies selected to realize the smart environment are listed below. A summary of the overall development process, from analysis to implementation, is shown in

Table 2. Summary of development process from analysis to implementation.

Functionalities	Design Choices	Implementation
Video Surveillance	AWS Rekognition, AWS IoT Greengrass	Video analysis with AWS Rekognition, video capture with Raspberry Pi, TensorFlow Lite and OpenCV for people recognition
Warning and Notification	AWS SNS	Email notification
Operation Efficiency and Automation	AWS Lambda	Process automation with AWS Lambda functions
Reliability and Maintenance	AWS IoT Greengrass, AWS CloudWatch	On-Premise Monitoring with AWS IoT Greengrass, alarms, and predictive maintenance with AWS CloudWatch
Logging and Storage	AWS S3, AWS DynamoDB	AWS S3 for file and log storage, AWS DynamoDB for structured data, PostgreSQL for relational data

.

• A RaspberryPi4 microcomputer acts as a central node for collecting data from a group of sensors and a camera. In particular, it collects data from the temperature, humidity, and ozone sensors, while the camera captures images or videos for detecting the presence of people. This is possible by combining TensorFlow Lite and OpenCV, so as to capture videos from the camera and apply machine learning models to process each frame in real time.
• Sensors such as DHT11 (for humidity and temperature) and KH-NPOD-100 DGOzone (for ozone).
• CRJ O3-UV-500 Ozonizer is a sanitizing machine equipped with UV-C lamps that generate ozone and a fan device; see Figure 3.
  It is automatically activated based on the results of analyses performed by AWS services or by the RaspberryPi.
• Node-Red, an open-source framework, commonly used for industrial automation and IoT projects, aimed at connecting hardware devices, APIs, and AWS online services through a flow-based, block-oriented interface. In particular, we configured the environment by using Node-RED to collect, store, and display data locally with a database, as shown in Figure 4.
• PostgreSQL to locally store (in a relational way) the data collected from the sensors for analysis and reporting; see Figure 5.
• MQTT (Message Queuing Telemetry Transport) as the messaging protocol owing to its efficient management of large-scale IoT deployments, low-bandwidth requirement, energy efficiency, and robustness in unreliable networks. These features make MQTT an ideal candidate to interconnect the smart environments’ components deployed at edges and on the cloud. In particular, we opted for MQTT version 5, a keep-alive duration of 30 s, and a maximum packet size of 149,504 bytes.

3.4. Results Analysis

The prototype of our system described above was tested in order to assess its correct operation and to perform a preliminary performance evaluation. In over 100 tests, the actuation commands on the UV-C lamp (activation and deactivation) were never lost; we report in

Table 3. Performance evaluation.

Sensor Data Processing		Image Processing
Operation	Time Range (ms)	Operation	Time Range (ms)
Parsing SQS message	1–5	S3 metadata retrieval	10–20
DynamoDB write	10–20	Recognition API call	500–1000
Anomaly check	1–2	Label analysis	1–5
SNS notification (if needed)	20–40	SNS notification (if needed)	20–40
Total (w/o notification)	12–27	Total (w/o notification)	511–1025
Total (with notification)	32–67	Total (with notification)	531–1065

the execution time of the two main lambda functions of sensor data processing and video processing.

It can be immediately noticed that, in both cases, a relevant but bearable delay is introduced by the notification system (in the cases in which we need to notify the administrator) while, specifically for the image processing, calling the AWS Recognition API introduces between 0.5 and 1 s of delay. The latter can be significantly reduced in the case of local execution, where only the label analysis on the video frame is needed and it takes between 280 ms and 400 ms (but with the possibility of exclusively identifying people), with almost identical values for the SNS notification time. Obviously, advanced (and more expensive) devices provided with specific AI-components such as Google Coral Board and Nvidia Jetson Nano are able to support powerful ML models, and they can more than halve the image processing time with respect to a RaspberriPi4, which is intended for general-purpose computing tasks. However, in our case, a response time of approximately 1 s is acceptable and not critical for the system’s operation. Conversely, it is worth noting that, regardless of the specific edge device, the bandwidth consumption in the case of local execution is limited to a few KBs for the notification, thus avoiding GBs of video to be forwarded on the cloud for processing, with subsequent money saving. For the sensor data processing, instead, the PostgreSQL operations are also practically instantaneous, and the anomaly checks between 10 and 15 ms. Finally, it should be considered that the Raspberry Pi 4 uses about 3.1 watts while reading sensor data and up to 5.5 watts at maximum load during video analysis: in similar conditions, the power consumption of a cloud server is in the order of kWatts.

4. Benefits, Limitations, and Future Work

Regarding the certified quality of ozonation treatments for indoor environments, the proposed IoT-based system has several additional benefits. First, the automation degree provided by the design choices and implementation technologies allow for higher effectiveness, also avoiding the need of a constant (and error-prone) human intervention. Then, by resorting to the AWS infrastructure, the system’s resources are automatically optimized and secured, with obvious advantages in terms of capital expenditures, time-to-market, and scalability. Indeed, if an entirely locally built serverless platform would offer full control over the infrastructure as well as lower latency for lightweight local workloads and no vendor-lock issues, the fully-managed AWS Lambda solution offers a compelling choice to prioritize agility, operational efficiency, and interoperability. Indeed, it totally abstracts away operational overheads, reduces upfront investments (in hardware, maintenance, and power, etc.), and introduces a limited delay, since cold starts have been notably mitigated through recent AWS innovations like Provisioned Concurrency. Last but not least, the choice of a hybrid edge–cloud setting dictated by the EI paradigm provides significant flexibility by allowing systems to be dynamically deployed case-by-case across the continuum based on real-time conditions, resource availability, and expected performance requirements. With respect to our scenario, this introduces a two-fold strength: on the one hand, the local analysis of sensor data and video pre-processing improves privacy, responsiveness, and bandwidth efficiency; the cloud, on the other data, brings relevant resources for a data store and intensive data tasks such as predictive maintenance.

Some limitations, however, are detectable in the proposed system, and they pave the way for interesting future research directions. Indeed, at least in its current status, the system is not suitable for providing its functionalities in outdoor environments. While the outlined IoT infrastructure might remain mostly untouched, additional hardware components need to be considered for fostering the air purification [32]. In order to evaluate the system in additional deployments, instead, scalability tests are needed. Resorting to simulations for preliminary assessments of the performance in different settings is, however, a well-established methodology [33]. Even then, a key activity like the devices’ calibration is still performed manually and resorts to operators’ expertise, albeit its automation would bring relevant advantages: some contributions in such direction are already available [25], whereas it can be still considered an open challenge. Finally, additional machine learning models should be added to enable, over the gathered data and other contextual information, air quality forecast or predictive maintenance functionalities. Indeed, the latter are a key task in industrial contexts [34] in order to enforce the overall system reliability by reducing the downtime; but, as for the devices’ calibration, the required human intervention is not negligible. In summary, although it is in the prototype stage, the solution implemented by the proposed IoT system is very promising, and none of the aforementioned limitations represents an actual drawback but rather a hint for possible future work.

5. Conclusions

By splitting oxygen molecules to form ozone, it is possible to effectively achieve water treatment, sterilization, and air purification. The importance of these treatments, especially the latter, arose prominently in recent years due to the COVID emergence, and it is still in the spotlight to mitigate the risks of future disease transmission. In this article, we deal with the healthiness and safety of indoor smart environments: in particular, we describe an IoT-based system that continuously monitors ozone concentration and environmental conditions and, if any parameter deviates from the desired range, it adjusts operations automatically. Leveraging on EI techniques for distributing intelligence among the edge–cloud continuum, the proposed system ensures efficient, safe, and reliable operations by the ozonation system, maximizing its effectiveness while minimizing costs, human intervention, and risks for occupants.