The Artificial Intelligence of Things (AIoT) integrates Artificial Intelligence (AI) and the Internet of Things (IoT) to create smart systems capable of environmental sensing and real-time interaction. Low-cost sensors (LCSs) offer scalable and continuous air quality monitoring, but their accuracy is often limited due to inherent biases, even after factory calibration. Real-world calibration is therefore essential for reliable measurements. This study presents an AIoT-based solution called the Real-Time Particulate Matter Sensing and Calibration (RPM-SC) System, which uses customized PMS7003 sensors equipped with externally controlled pulse-width modulation (PWM) fans to improve airflow and measurement consistency. The sensors are integrated into a oneM2M-compliant IoT platform for standardized data collection and communication. To enhance measurement accuracy, we propose a novel calibration model: the Trans-Convolutional Fusion Memory-Aware Network (TCM-Net). This model combines Transformer networks, Temporal Convolutional Networks (TCNs), and a Memory-Aware Module to capture temporal dependencies and correct sensor bias effectively. TCM-Net was trained on real-world datasets from RPM-SC systems co-located with a certified PM reference instrument. It achieved a root mean square error (RMSE) of 0.325 μg/m3and a coefficient of determination (R²) of 0.999, outperforming conventional models. The results confirm the effectiveness of the proposed AIoT architecture and calibration model in providing accurate and reliable PM2.5data from low-cost sensors.

Accurate fine particulate matter (PM2.5) estimation is essential for air quality monitoring, but high-cost sensors have limited spatial coverage. To address this, we introduced the concept of virtual sensors to estimate PM2.5 levels in unmonitored areas by leveraging data from nearby physical sensors. To support this study, we developed a real-time PM monitoring system integrating long range (LoRa)-based low-cost sensors with a oneM2M-compliant Internet of Things (IoT) platform, enabling data collection. To enhance the accuracy of PM measurements, we applied a mechanistic model that adjusts PM values using environmental factors like humidity and temperature. Using this calibrated data, we propose the Geo-Temporal TransGraph Attention Network (GTA-Net), a deep learning model that estimates PM2.5 levels in unmonitored regions. A distance-based graph captures spatial dependencies, while transformer-based temporal encoding improves predictive accuracy. GTA-Net was evaluated against LSTM, GRU, CNN+LSTM, and Vanilla Transformer models, and the 10-fold cross-validation results showed that it achieved on average an RMSE of 0.679 ± 0.526 μg/m and an R2 close to 1.000, performing better than all baseline models. A performance comparison based on different PM embedding sizes revealed that 1-hour PM embeddings yield the most accurate estimations. This research highlights the potential of virtual sensors and spatiotemporal modeling to enhance air quality monitoring.

AIoT-based smart healthcare system that utilize smart devices can provide personalized, proactive care to patients, reducing the reliance on limited medical service capacities and promoting everyday health management. However, the lack of compatibility in health data formats and services across devices from different manufacturers hinders the growth of AIoT-based healthcare services. To address this issue, we propose a unified system for the collection of raw sensor data from smart devices (i.e., smartphones, smartwatches) of various manufacturers as well as health survey data. The data collection application part of our system is available on Android, Wear OS and iOS, and stores data on a oneM2M-based IoT platform via REST APIs. Using the system, we collected data from 300 participants over the course of 60 days. After collecting the data, we analyzed the dataset and then proposed a method to archive health-related lifelog in a standardized manner following the Fast Healthcare Interoperability Resources (FHIR) protocol. With our proposed method, we envision cohort monitoring for smart healthcare services. We expect that our system will improve data compatibility across different devices and facilitate the growth of AIoT-based healthcare services.

The Internet of Things (IoT) now provide a considerable advantage in terms of intelligent devices. For example, sensor data of medical equipment could be used to extract meaningful insights to enable a range of tasks to be performed intelligently at the device location, such as blood pressure (BP) prediction several minutes ahead. In this article, we propose a computing device called Medical Edge, designed to convert conventional medical equipment into IoT-enabled devices. We developed an interworking proxy that allows Medical Edge to be interfaced with medical devices, such as patient monitors (e.g., GE CARESCAPE B650), to smoothly collect physiological data and upload them to an IoT server platform. We used a oneM2M standard-based server platform to provide access to these data in a standardized manner. To demonstrate a promising application of our proposed Medical Edges, we performed a study on BP estimation based on photoplethysmography (PPG) signal only. We propose a hybrid neural network architecture and apply it with a publicly available data set called multiparameter intelligent monitoring in intensive care II (MIMIC II). The model consists of five 1-D convolutional neural networks (CNNs), three Bi-directional long short-term memory networks, and four fully connected layers. The mean absolute error (MAE) and standard deviation (STD) of the proposed model were, respectively, 0.95 and 1.44 millimetres of mercury (mmHg) for diastolic BP (DBP), and 1.38 and 2.13 mmHg for systolic BP (SBP). The experimental results are in full compliance with the international standards of the Association for the Advancement of Medical Instrumentation (AAMI) and the British Hypertension Society (BHS).

In recent years, particulate matter (PM) having a diameter smaller than 2.5 μm has become a significant issue due to its severe impact on human health. With the advent of IoT-enabling technologies, a ubiquitous IoT sensing infrastructure is now used to constantly monitor aspects of our surrounding environment, such as ambient air pollution. In this article, we introduce a PM-sensing system composed of off-the-shelf LoRa-based wireless hardware boards and low-cost PM sensors. By leveraging software platforms that are compliant with an IoT standard called oneM2M, PM data sets can be collected and accessed in a standardized manner, i.e., via oneM2M-defined representational state transfer application programmable interfaces. Also, for reliable PM monitoring, a short-term (i.e., within 2 h) PM forecasting method based on autoregressive integrated moving average and vector autoregressive moving average (VARMA) models is proposed and evaluated with a 30-day PM data set collected from 15 LoRa-based PM sensor nodes installed at a university campus. The experimental results show that the overall root-mean square error and correlation coefficient of the VARMA models integrated with hierarchical clustering are improved by 7.77% and 3.7%, respectively, compared with the single node-based forecast model.

Machine learning has played a significant role in building intelligent systems in the history of data science. In the recent paradigm where objects in the world will be connected with each other, commonly referred to as the Internet of Things (IoT), people begin to consider the challenges and opportunities to utilize the huge data sets generated, also referred to as Big data. One of the active research topics in dealing with the IoT's big data is the practical feasibility of algorithms used in classical machine learning but also in a newly emerging branch, called deep learning. In this article, we demonstrate a quantitative analysis comparing performance between classical machine learning and deep learning algorithms with a human movement direction detecting application based on analog pyroelectric infrared (PIR) sensor signals. The sensing data acquisition and retrieval system is implemented with the open-source IoT software platforms based on the oneM2M standard. With the analog PIR data sets collected from 30 subjects, we perform experimental studies comparing classical machine learning and deep learning algorithms in terms of economic feasibility, scalability, generality, and real-time detection performance. The results show that classical machine learning shows better performance in real-time detection (i.e., with the sensing values within the first 0.5 s). In contrast, our simple deep learning model achieves about 90% accuracy for detecting moving directions even with the data sets from only three subjects and a single PIR sensor. Moreover, it could be applied to a larger number of subjects without updates.

Since the Internet-of-Things (IoT) has been introduced, it is considered as one of the emerging technologies providing great opportunities to many vertical industries. One of the major IoT application areas that gets significant attention is smart city. Since it is unrealistic to expect full convergence toward a single IoT platform in the near future, it is mandatory to enable interworking between different platforms based on multiple standards, coexisting in the emerging smart cities. In this paper, we take the example of two global IoT standards, FIWARE and oneM2M, which are actively used in many smart city projects, and analyze them to show the feasibility of IoT platforms interworking. Based on the analysis, we design and implement a novel IoT interworking architecture providing a semantic driven integration framework suitable for smart city. The core idea behind our approach is to introduce interworking proxies that: 1) conduct a static mapping of sensor information between IoT platforms and 2) perform semantic interoperability using semantically annotated resources via a semantic interworking proxy that dynamically discovers new kinds of information and adapts itself to enable automatic translation of semantic data between given source and target IoT platforms while it is running. We present the system based on these proxies and evaluate it in Santander smart city. The results demonstrate that it is able to discover and manage IoT sensors connected to both oneM2M and FIWARE. It appears that the semantic approach provides the flexibility and dynamic adaptivity needed for fast growing and rapidly changing urban environments.

The oneM2M global initiative has been working to standardize horizontal machine-to-machine (M2M) and the Internet of Things (IoT) service layer specifications for globally interoperable M2M/IoT solutions. However, there exist a myriad of embedded sensor systems not complying with oneM2M, such as de facto IoT standards (e.g., AllJoyn and IoTivity). In this paper, we present our effort in global interworking of IoT systems across multiple M2M/IoT standards, in particular, on a global scale. By implementing specialized oneM2M entities for interworking, interworking proxy entities (IPEs), we demonstrate a promise of globally interoperable IoT systems interconnected with all kinds of embedded systems.

We present an empirical study of the long-term practicality of using human motion to generate operating power for body-mounted consumer electronics and health sensors. We have collected a large continuous acceleration data set from eight experimental subjects going about their normal daily routine for three days each. Each subject is instrumented with a data collection apparatus that simultaneously logs 3-axis, 80 Hz acceleration data from six body locations. We use this data set to optimize a first-principles physical model of the commonly used velocity damped resonant generator (VDRG) by selecting physical parameters such as resonant frequency and damping coefficient to maximize the harvested power. Our results show that with reasonable assumptions on size, mass, placement, and efficiency of VDRG harvesters, most body-mounted wireless sensors and even some consumer electronics devices can be powered continuously and indefinitely from everyday motion. We have optimized the power harvesters for each individual and for each body location. In addition, we present the potential of designing a damping- and frequency-tunable power harvester that could mitigate the power reduction of a generator generalized for "average” subjects. We present the full details on the collection of the acceleration data sets, the development of the VDRG model, and a numerical simulator, and discuss some of the future challenges that remain in this promising field of research.

We present an empirical study of the long-term practicality of using human motion to generate operating power for body-mounted consumer electronics and health sensors. We have collected a large continuous acceleration dataset from eight experimental subjects going about their normal daily routine for 3 days each. Each subject is instrumented with a data collection apparatus that simultaneously logs 3-axis, 80Hz acceleration data from six body locations. We use this dataset to optimize a first-principles physical model of the commonly used velocity damped resonant generator (VDRG) by selecting physical parameters such as resonant frequency and damping coefficient to maximize harvested power. Our results show that with reasonable assumptions on size, mass, placement, and efficiency of VDRG harvesters, most body-mounted wireless sensors and even some consumer electronics devices, may be powered continuously and indefinitely from everyday motion.

In this paper, we propose the UbiFloorII, a novel floor-based user identification system to recognize humans based on their stepping pattern, the arrays of the transitional footprints from heel-strike to toe-off. To obtain users' stepping pattern from their gait, we deployed photo interrupter sensors instead of switch sensors used in the UbiFloorI. We developed a software module to extract stepping pattern from users' gait. For user identification, we employed neural network trained with users' stepping samples. We achieved about 92% recognition accuracy using this floor-based approach. The UbiFloorII system may be used to automatically and transparently identify users in a home environment.