By combining AAL with subject monitoring, it becomes possible to create personalized and proactive care solutions, promoting independent living while offering a safety net for vulnerable individuals. Subject monitoring can be conducted in both outdoor and indoor environments. Outdoor environments expose frail people to various risks, such as falls, extreme temperatures, and potential wandering or confusion in individuals with early symptoms of dementia. In this context, AAL systems aim to provide support to frail individuals by facilitating route checking, anomalous behavior recognition, motion activity evaluation, and other relevant functionalities (Fernandes et al., 2020; Lee, 2021). Since wearables are the only devices that can be employed outdoors, they acquire a fundamental importance. Smartwatches and smart wristbands are the most commonly used devices, while Inertial Measurement Units (IMUs) are typically the sensors chosen (the same used in indoor scenarios) (Bhargava et al., 2017; Iadarola et al., 2022). In literature, other solutions have been proposed that address alternative or complementary approaches (Sokullu et al., 2019; Kenfack Ngankam et al., 2020; Rupasinghe and Maduranga, 2022): for example, Global Navigation Satellite System (GNSS) sensors for outdoor localization (Junior et al., 2023); instrumented insoles (commercial or customized) as an aid for gait detection and consequently for pointing out unsteady walking or falls (Cristiani et al., 2014; Sivakumar et al., 2018) and sensors mounted on the body of a walker as a low-cost solution for people with limited mobility (Ding et al., 2022). Conversely, indoor scenarios present frail individuals with risks closely associated with their living spaces.

Notably, the indoor environment has been identified as a significant contributing element to falls (Lee, 2021), which are attributed to factors such as uneven or slippery floor surfaces (including carpets and mats), tripping obstacles, inadequate lighting, poorly designed or maintained stairs without handrails and unsuitable furniture. These criticalities increase the likelihood of tripping, falling, or slipping for frail individuals. Additional hazards arise from the absence of safety or preventive devices, such as night lights and grab rails (Carter et al., 1997; Lee, 2021). Indeed, the requirements of monitored subjects can vary significantly across different indoor scenarios. In private homes, where individuals live alone or with a caregiver, the primary focus of monitoring is on preventing domestic accidents and delivering essential healthcare services. On the other hand, in retirement residences where multiple individuals share common spaces, subject monitoring systems are primarily designed to facilitate group activities and controlled physical exercises (Nastac et al., 2019; Cicirelli et al., 2021).

Adapting the monitoring approach to suit the specific needs and dynamics of each indoor setting is crucial. This ensures that monitored subjects receive personalized support and care tailored to their circumstances. Additionally, it is important to recognize that AAL systems cannot remain static, as people’s needs and habits evolve over time, along with the parameters that need to be observed. Consequently, data analysis methodologies must account for the evolving nature of these systems, allowing for the possibility of dynamically weighting or customizing certain parameters over others (Cicirelli et al., 2021). By embracing flexibility and adaptability, AAL systems can continue to provide effective and relevant support to individuals in various indoor environments.

The selection of appropriate sensors considers multiple factors, including the specific objectives of the AAL system, sensor cost, intrusiveness, user acceptability, and privacy concerns. However, more complex sensor networks, comprising environmental sensors, object sensors, cameras, and wearable sensors constitute the foundation of indoor AAL. The living facilities may be equipped with an array of interconnected sensors and actuators, enabling remote control and capable of detecting various environmental parameters such as door openings and room brightness. These sensors are strategically deployed to monitor the daily activities of individuals, ensuring security and safety. The selection of appropriate sensors considers multiple factors, including the specific objectives of the AAL system, sensor cost, intrusiveness, user acceptability, and privacy concerns. Communication protocols, such as ZigBee, Bluetooth, ZWave, USB, Ethernet, among others, are utilized to interconnect sensors, actuators, and smart devices throughout the environment (Tazari et al., 2011; Lloret et al., 2015; Babangida et al., 2022). Typically, raw or pre-processed data from sensors are transmitted to a collection center, either local or remote, where they undergo integration and analysis using robust algorithms (Plentz and De Pieri, 2018). A thorough and reliable data analysis becomes crucial in indoor scenarios equipped with automatic dangerous situation detection or capable of requesting help triggering alarms to third parties.

In the field of AAL, Human Activity Recognition (HAR) has emerged as a valuable tool with multifaceted utility. Within the AAL domain, HAR presents a range of solutions aimed at enhancing the quality of life of frail individuals (elderly and/or disabled people) and maintaining improved health and independence. Additionally, it provides also support to caregivers and medical professionals. HAR has garnered substantial interest as a prominent field of study in recent times. HAR methodologies are devised with the objective of autonomously detecting and classifying individuals’ routine activities within defined contexts. Depending on the task and the employed technologies, two main methodologies are commonly adopted. The first approach relies on a threshold analysis method, which can suffice for triggering alerts when detecting dangerous events (Zdravevski et al., 2017; Al Machot et al., 2018a,b; Cicirelli et al., 2021). The second and more recent approach (see Figure 1, fourth generation), employs Artificial Intelligence (AI) solutions such as Machine Learning (ML) and Deep Learning (DL) algorithms for HAR (Aggarwal and Ryoo, 2011; Wang J. et al., 2019).

Among the various possible applications, human activities can be classified into four distinct groups based on the involvement of various body parts (Jegham et al., 2020; Minh Dang et al., 2020):

Detailed framework of the HAR process will be discussed in the upcoming section. Finally, the aim of this work is to provide an overview of recent literature of AAL systems focusing on HAR from a technological perspective, tackling emerging evidences, challenges and future directions (Stavropoulos et al., 2020).

Papers have been revised by searching published works in the last 10 years in the following databases: IEEE Xplore, PubMed, Scopus. A series of keywords have been used, alone or in combination: ambient assisted living, ambient assisted living technology, healthy ageing, human activity recognition, ambient sensors, wearable sensors, wearable technology, activity monitoring, machine learning, deep learning, frail person.

The rest of the paper is organized as follows: Section 2 describes the Monitoring Solutions, focusing on the main components of an AAL system and detailing wearable and non-wearable sensors. Section 3 presents the processing chain, outlining the most relevant methodologies for data acquisition, processing and analysis. Section 4 proposes a critical discussion that also addresses future directions of this research field.

Wearable devices typically consist of small sensors that can be integrated into cloths, rings, shirts, watches or other garments and devices. Such sensors gather body and context information to be locally processed or directly transmitted, mainly wirelessly through appropriate communication protocols, to a central processing unit of an AAL system (Kumari et al., 2017).

In the last few years, most wearable devices have been miniaturized and have optimized their power consumption (Cicirelli et al., 2021). Wearable devices, especially fitness trackers, wristbands and smartwatches (Andò et al., 2016; Lutze, 2018; Xie et al., 2020), have various built-in/integrated sensors such as accelerometers, gyroscopes and orientation sensors. Smartphones represent an additional solution for their characteristics of cost effectiveness and high number of embedded sensors (Ramanujam et al., 2021). Moreover, smartphones’ embedded sensors can be used alone or in combination with other wearable technology to evaluate posture and activities and to prevent falls, together with biological and behavioral monitoring (Haghi et al., 2019; Badgujar and Pillai, 2020; Nouredanesh et al., 2020). In this review, since users take their smartphone with them almost everywhere, even though they are not always in direct contact with the body, they are considered wearable devices at the same level as wristwatches, rings, glasses and necklaces, as opposed to environmental sensors and cameras. Apps for recording the device’s sensors data can be run on all commercial operating systems (Android or iOS) and they can be combined with commercial smartwatches, self-developed smart bands or devices like Shimmer nodes (Sanchez-Comas et al., 2020). However, not all applications provide an integration with smartphones; instead, many studies considered custom-developed solutions of electronic components. A significant number of works developed technologies in the laboratory, whereas fewer studies used commercial devices. Inertial sensors are the most common wearable elements used for HAR in AAL; in some cases, accelerometers (Tay et al., 2023), gyroscopes, magnetometers, temperature and object sensors may be applied, mainly worn on the waist or the hip (Pierleoni et al., 2019; Sarabia-Jácome et al., 2020).

Identification of the user’s position can be obtained even with passive RFID sensors or Bluetooth Low Energy (BLE) technology, that paved the way to the Bluetooth Smart for wearable devices (Ciabattoni et al., 2019; Paolini et al., 2019; Bilbao-Jayo et al., 2021). Also, wearables usually commercialized for fitness purposes have been evaluated for elderly wellbeing in the AAL domain (Piwek et al., 2016; Seneviratne et al., 2017; Cedillo et al., 2018; Surendran et al., 2018; Stavropoulos et al., 2020).

The majority of HAR systems in AAL are dedicated to the identification and management of falls, as confirmed by literature (Dhiman and Vishwakarma, 2019). In such systems, wearables are one of the key elements due to their mobility, portability, cost and availability and several studies have been conducted, mostly using a single wearable device (Wang X. et al., 2020). Even in this context, inertial sensors represent a large percentage of the research, whereas only a minority deployed other solutions (Bourke et al., 2007). Although Shimmer nodes, smartphones and smart watches often contain sensors like magnetometers, such elements were not normally used to detect falls; indeed, the only sensors embedded in smartphones used for this purpose are accelerometers and gyroscopes (Shi et al., 2016; Islam et al., 2017; Medrano et al., 2017; Chen et al., 2018).

Even combining multiple sensors into a single framework can provide valuable data for meaningful and complex predictions, thus achieving a more versatile, robust and trustworthy wearable system for HAR purposes. Moreover, commercial tools are widely used, such as Samsung Galaxy Gear Live (Faye et al., 2017), Microsoft Band 2 (Garcia-Ceja et al., 2018) and Intel Basis Peak (Kang and Larkin, 2017), as well as other alternatives like Empatica E3 (Clapés et al., 2018), Fitbit (Kang and Larkin, 2017), and Google Glass (Shewell et al., 2017; Clapés et al., 2018).

It is worth noticing that most of the time, wearable technology alone would be sufficient to assess activity recognition in indoor environments and AAL systems. This is an important advantage, combined with their low cost, portability and unobtrusiveness. However, a hybrid approach combining wearable and non-wearable sensors often overcomes possible drawbacks due to users not wearing the device correctly, thus leading to low quality signals, not being comfortable in putting on a wearable that could excessively stand out, forgetting to charge it or even to wear it continuously.

Non-wearable sensor solutions for HAR encompass devices or systems that can detect and analyze human activities without needing to be directly attached to the body. These solutions play a crucial role in the functionality and effectiveness of AAL systems. Operating in a passive manner, these sensors autonomously monitor room occupants without the need for manual intervention. This eliminates the need for users to carry additional devices during their daily activities.

One example of a non-wearable sensor solution is radio-frequency-based systems, as demonstrated by (Fan et al., 2020). These systems utilize the analysis of radiofrequency (RF) signal reflections to monitor various activities performed by individuals. In this study, Radio Frequency sensors were embedded directly in the floor to capture the everyday activities of residents (Fan et al., 2020). As part of radiofrequency-based systems, radar and ultra-broadband technologies, as well as automotive-derived solutions, also emerged as interesting approaches for human activity recognition (Guendel et al., 2021; Senigagliesi et al., 2022).

Furthermore, contemporary monitoring and behavior analysis tasks can be facilitated by diverse image-based technologies.

Nowadays, low-cost cameras have emerged as viable options for monitoring individuals’ daily activities, ensuring their well-being (Geman and Costin, 2014; Cebanov et al., 2019; Minh Dang et al., 2020; Ryselis et al., 2020; Bouchabou et al., 2021; Chen et al., 2021; Raeis et al., 2021). These devices enable continuous monitoring of individuals without requiring their active involvement. These include RGB, Depth, and RGB-D cameras, as well as IR (infrared) array sensors, known as thermal cameras.

RGB cameras, being widely available and affordable, provide information about the shape, color, and texture of the scene (Zerrouki et al., 2018). However, they have some drawbacks such as a limited field of view, complex calibration procedures, sensitivity to environmental variations (e.g., lighting conditions, type of illumination and cluttered background) and privacy concerns. To address this latter issue, depth cameras offer distance information from the sensor to elements in the scene, capturing detailed spatial information while maintaining heightened privacy protection (Jayaraj et al., 2019; Beddiar et al., 2020). Depth sensors also exhibit superior resilience to variations in illumination, color, and texture compared to RGB devices. However, noisy measurements can occasionally affect accurate object or subject identification, necessitating data processing and refinement. In recent years, low-cost devices integrating RGB and depth sensors, such as Kinect and Intel RealSense systems, have been employed as environmental sensors in AAL systems. Another alternative, IR array cameras, measure thermal energy emitted by the human body or other objects (Mashiyama et al., 2015; Spasova et al., 2016; Karayaneva et al., 2023). These low-resolution IR arrays offer advantages such as privacy preservation, low power consumption, insensitivity to ambient lighting variations, operation in complete darkness, fast response time, easy deployment, and straightforward image processing.

All these devices suffer from subject occlusion, which occurs when certain body parts of the subject are hidden or obscured by other objects or body parts within the room, leading to incomplete or inaccurate tracking of the subject’s movements. To overcome the occlusion limitation, a practical solution is to employ a multiple camera setup that covers various areas of the room from different viewpoints. By using multiple cameras, the chances of occlusion can be reduced, as different cameras capture different perspectives of the scene. This approach allows for a more comprehensive view of the subject’s activities and improves the accuracy of tracking. However, it is important to note that using multiple cameras increases HAR systems’ costs and requires synchronization among them to ensure proper coordination and alignment of the captured data. Synchronizing the cameras enables the seamless integration of the captured images or depth data, allowing for a more complete understanding of the subject’s movements and activities.

In the context of monitoring human activities, sensors can be also embedded in everyday objects within the environment. Contrary to wearable sensors, which directly measure human activities, these sensors enable the detection of movements and activities through the usage of specific objects, providing valuable insights into the daily lives of individuals (Bassoli et al., 2017; Rafferty et al., 2017). For instance, sensors can be integrated into furniture items such as carpets, beds, fridges, and more, allowing for unobtrusive monitoring of daily living activities. Presence statistics of users in different spaces can be gathered by monitoring the sensors embedded in furniture. Power meters can be employed to track appliance usage (Bianchi et al., 2019), such as monitoring TV sets, while smart pill box devices assist in checking medication intake (Keum et al., 2019). Roland et al. proposed the installation of an accelerometer attached to a smart drinking cup to efficiently identify the user’s drinking movement (Roland et al., 2018). Bassoli et al. installed sensors directly on the furniture of the subject’s house for HAR. Pressure pads are used to monitor bed or chair occupancy, while sensors inside the fridge provide indirect information about feeding habits (Marenzi et al., 2012, 2013a,b; Bassoli et al., 2017). Chaccour et al. developed a smart carpet with piezoresistive pressure sensors to detect falls of the inhabitant (Chaccour et al., 2015; Singh et al., 2020).

Everyday object sensors offer a less invasive and privacy-friendly alternative to cameras, as they are designed to specifically recognize human activities related to the intended use of the object. By focusing on that, these sensors can provide valuable insights and functionality while minimizing potential privacy concerns. Everyday object sensors can detect interactions with household appliances or devices, such as opening a refrigerator, without capturing or storing detailed visual information of individuals. Yue et al. proposed an RF-based system that accurately monitors sleep postures overnight in the user’s own house. By analyzing RF reflections and distinguishing them from other signals, the system can identify different sleeping postures such as supine, left side, right side, and prone (Yue et al., 2020). This approach can help alleviate privacy concerns while still enabling the development of innovative and convenient technologies that enhance user experiences in a responsible and respectful manner.

Sensors and devices are characterized by outputs that are either punctual values (i.e., pixels in an image) or temporal series (i.e., position, acceleration). HAR applications could rely on raw data or need further pre-processing steps to enhance the signal quality (i.e., noise reduction, data normalization, segmentation) and/or to obtain derived data. For instance, from data captured by RGB-D cameras it is possible to estimate the position over time of specific points on the human body, often corresponding to anatomical points of repère or joints with respect to a specific coordinate system typically intrinsic to the device. This data processing stage is called skeleton tracking. Moreover, joint angles can be obtained starting from body joint positions (Jegham et al., 2020). Notably, in the context of AAL, it is common practice to store and analyze the skeleton tracking data instead of RGB or depth data to prioritize privacy preservation (Gasparrini et al., 2014; Tu et al., 2018; Li and Sun, 2021; Srimath et al., 2021). The pre-processing stage aims at noise reduction, data normalization and segmentation.

The feature extraction procedure consists in the definition of a set of parameters able to discriminate the activities to be classified. Based on the given data nature and characteristics, the features can be divided into several categories: time-domain, frequency-domain and kinematic features (Dhiman and Vishwakarma, 2019; Sousa Lima et al., 2019; Ahad, 2020, 2021). The time-domain features are usually defined to describe the data amplitude variation and distribution over time (for instance mean, variance and kurtosis). On the other hand, the frequency-domain features show the distribution of signal energy (i.e., Fast Fourier Transform, entropy and power spectral density). Kinematic features include all the characteristics of the subject’s movements, acceleration and posture (joints positions and angles). The kinematic features describe geometric relations between body joints (Müller et al., 2005; Guerra et al., 2020).

In the case of images, usually global and local features are computed. The first ones describe the image frames as a whole, providing different types of information (spatial, temporal, frequency) (Ke et al., 2013). Local features extract information around a set of interest points or describe a selected image region, through techniques like histograms of oriented gradients (Aly and Sayed, 2019).

After feature extraction, the relevant ones are selected to achieve dimensionality reduction by finding the smallest subset of features which efficiently defines the data for the given problem (Blum and Langley, 1997; Chandrashekar and Sahin, 2014; Jindal and Kumar, 2017; Ayesha et al., 2020). It can be accomplished using different methods, such as filter, wrapper, embedded, and the more recent hybrid approach (Blum and Langley, 1997; Minh Dang et al., 2020; Zebari et al., 2020). Filter methods measure the relevance of features using statistical standards for evaluating a subset, they process data before the classification occurs and are independent from the latter. The features are ordered according to the ranking of importance (computed with suitable score metrics) and those below a certain threshold are removed. Among the different algorithms, the most used are: ReliefF, statistical techniques such as Principal Component Analysis, Independent Component Analysis, Neighborhood Component Analysis and Correlation Based filter (Suto et al., 2016; Alzahrani et al., 2019; Siddiqi and Alsirhani, 2022). Wrapper method selects the optimal features subset evaluating alternative sets by running the classification algorithm on the training data. It employs the classifier estimated accuracy as its metric (Bhavan and Aggarwal, 2018; Zebari et al., 2020). The most used iterative algorithms are the Recursive Feature Elimination with Support Vector Machine, the Sequential Feature Selection algorithm and the Genetic Algorithm (Liu and Shao, 2013; Guerra et al., 2022). Compared to the filter method, the wrapper method achieves better performance and higher accuracy, nevertheless it increases computing complexity due to the need to recall the learning process for each feature set considered (Jindal and Kumar, 2017; Zebari et al., 2020). In the embedded method, as the name suggests, the selection occurs within the learning algorithm. The most common are the tree-based algorithms like, for example, the Random Forest and the Decision Tree. Embedded methods can be used in multiclass and regression problems and compared to a wrapper method, it is computationally more effective while retaining similar performance (Negin et al., 2013). Finally, the hybrid approach combines filter and wrapper methods to achieve the benefits of both. Usually, the filter technique is first applied to reduce the search space and then, a wrapper model is used to acquire the best subset (Peng et al., 2010).

Dataset construction concerns the process that divides the acquired data into training, validation and test sets. Generally, a set of data is required to train the classification model and a set of validation data is used to evaluate the performance of the model during training epochs for fine tuning the hyperparameters and to estimate if the model does not overfit, i.e., when a statistical model fits exactly against its training data at the expenses of its generalization abilities. Finally, test data, different from those involved in the training set, are used to evaluate the performance of the model (Bouchabou et al., 2021). The data contained in the training, validation and test could be described by labels also called classes. As will be stressed in the next Section, this operation is of crucial importance for the classification algorithms. In HAR tasks, the classes represent the type of activity to be recognized (i.e., walking, sitting, lying down, and so on) (Sathya and Abraham, 2013; Sindhu Meena and Suriya, 2020).

In HAR, three methods have been used to divide the data into training, validation and test set. In the first one, called cross-subject, the subjects are divided into two groups. The data of the first group are used for the training phase, whereas those of the second one are involved in the validation and test phase (Khan and Ghani, 2021). The cross-subject method aims at guiding the learning process of the AI model so that it becomes as robust as possible, in order to adapt it to the heterogeneity of the subjects. The second splitting method is characterized by randomly dividing the whole dataset based on a percentage such as 70-30%, 80-20%, and so on. The larger portion is fed for training the model where the other portion is kept for validation and test (Ahad, 2021). This is the most used splitting criteria in the general problems of classification algorithms and have been reported in HAR with success. Alternatively, when data are acquired by multiple cameras with different points of view, a cross-view method can be used. In this case the data coming from one or more cameras are used for the training phase and those of the remaining ones for the validation and test phase (Wang L. et al., 2020).

The most critical step of HAR systems relies on classification algorithms. In the literature, two main categories can be identified, namely threshold-based and AI methods. Furthermore, AI algorithms can be divided into ML and DL techniques. In the following, these methods are analyzed, highlighting the main advantages and issues in their application for HAR data analysis.

An important aspect related to HAR for frail people is the time needed by the AAL system to detect possible dangerous situations. This time is strictly related to the computational complexity of the adopted algorithms and to the computing power of the devices which perform the processing chain.

Threshold-based methods are the lowest computing power demanding techniques, since they are based on simple comparisons with a fixed value. Even in the case of threshold computed at runtime and/or applied to statistical indicators, the computational complexity can be easily managed by standard microcontrollers and does not have a critical impact on processing times.

AI methods are characterized by a higher computational complexity than threshold-based techniques. In particular, DL methods feature the highest computational complexity, not only for their training, which is run off-line and may require multiple CPU and/or GPU, but also for their implementation as a classifier once all the weights are determined. Therefore, their processing is typically performed by a centralized unit for both wearable and non-wearable devices. However, recently some researchers have designed DL algorithms on low power devices suitable for wearable applications.

Recent and very complex solutions exploited ensemble learning also with DL algorithms (Kumar and Suresh, 2023) enhancing the classification performance of single techniques.

The increasing popularity of AI solutions has led to the development of various software tools, libraries, and frameworks that facilitate the implementation of these algorithms on devices with limited resources. An example of such a tool is TensorFlow Lite for microcontroller³ (TFLM), an open-source library designed to enable the implementation of AI methods (i.e., ANNs, LSTM, CNN and so on) on a wide variety of MCUs and Digital Signal Processors (DSPs). TFLM allows the execution of pre-trained algorithms developed using TensorFlow for on-device inference. Another prominent solution is X-Cube-AI, a software tool developed by STMicroelectronics.⁴ It offers a comprehensive environment for generating and optimizing AI algorithms developed using popular ML and AI frameworks such as TensorFlow,⁵ Keras,⁶ or PyTorch.⁷ X-Cube-AI is tailored for deployment on the STM32 family of MCUs, empowering developers to leverage familiar frameworks and simplify the integration of AI capabilities into their applications. In addition to these tools, there are cloud-based platforms like Edge Impulse, which provides a flexible environment for the development of AI models. Edge Impulse⁸ supports various embedded platforms, including MCUs and smartphones, enabling developers to create AI models that cater to diverse hardware constraints. NanoEdge AI Studio⁹ is another valuable tool that supports both learning and inference directly inside the MCUs. Notably, it offers the advantage of automatically selecting the most suitable machine learning libraries based on the provided data (Shumba et al., 2022).

Inference needs to be performed under real-time constraint, especially when a potentially dangerous condition needs to be identified. This means that the classification should be strictly performed prior to a fixed temporal deadline, which is defined by the acquisition time window. In other words, the system acquires a window of data and its pre-processing and classification should be performed before the end of acquisition of the following time window. The computational complexity of the algorithms and the computing units included in the system play a critical role in the real-time compliance (Avram and Pop, 2023; de la Cal et al., 2023; Gonçalves et al., 2023; Saliba et al., 2023; Zeng et al., 2023). Delayed or sluggish processing can hinder the effectiveness of AAL systems in providing timely assistance, which is crucial for ensuring the safety and well-being of individuals. Efficient algorithms and optimized implementations are necessary to overcome these constraints and enable real-time processing on resource-limited platforms. The computational unit controls the sampling and acquisition of data: usually this is performed by Commercial-Off-The-Shelf (COTS) low power and low cost microcontrollers, supporting interfaces and data transfer protocols (SPI, I2C…). When high speed, flexibility and control over the elements of the architecture are required, Field Programmable Gate Arrays (FPGAs) and Application-Specific Integrated Circuits (ASICs) can be used, with increased costs and higher time for development and/or production. The computational unit is equipped with initial signal conditioning and processing algorithms together with specific classification methods that perform data analysis for local decision making, real-time response and forwarding processed data to successive layers. Characteristics that can influence all these choices include power consumption, computational and storage capacity, complexity and results accuracy of the algorithms, privacy concerns and latency requirements. Pre-trained models using computationally efficient algorithms may be used for anomaly detection and the results can produce warnings or propose a course of action. Alternatively, a fraction of an ANN can perform partial on-device data processing, to forward only intermediate data, thus ensuring also the privacy of the user (Zhang et al., 2023). Lastly, after the elaboration and analysis of data, results, inferences, or alarms are packaged and forwarded to the upper layers for further processing or management (Shumba et al., 2022).