Fall accidents in industrial and construction environments require an immediate reaction, to provide first aid. Shortening the time between the fall and the relevant personnel being notified can significantly improve the safety and health of workers. Therefore, in this work, an IoT system for real-time fall detection is proposed, using the ASSIST-IoT reference architecture. Empowered with a machine learning model, the system can detect fall accidents and swiftly notify the occupational health and safety manager. To train the model, a novel multimodal fall detection dataset was collected from ten human participants and an anthropomorphic dummy, covering multiple types of fall, including falls from a height. The dataset includes absolute location and acceleration measurements from several IoT devices. Furthermore, a lightweight long short-term memory model is proposed for fall detection, capable of operating in an IoT environment with limited network bandwidth and hardware resources. The accuracy and F1-score of the model on the collected dataset were shown to exceed 0.95 and 0.9, respectively. The collected multimodal dataset was published under an open license, to facilitate future research on fall detection methods in occupational health and safety.

Continuous, real-time monitoring of occupational health and safety in high-risk workplaces such as construction sites can substantially improve the safety of workers. However, introducing such systems in practice is associated with a number of challenges, such as scaling up the solution while keeping its cost low. In this context, this work investigates the use of an off-the-shelf, low-cost smartwatch to detect health issues based on heart rate monitoring in a privacy-preserving manner. To improve the smartwatch’s low measurement quality, a novel, frugal machine learning method is proposed that corrects measurement errors, along with a new dataset for this task. This method’s integration with the smartwatch and the remaining parts of the health and safety monitoring system (built on the ASSIST-IoT reference architecture) are presented. This method was evaluated in a laboratory environment in terms of its accuracy, computational requirements, and frugality. With an experimentally established mean absolute error of 8.19 BPM, only 880 bytes of required memory, and a negligible impact on the performance of the device, this method meets all relevant requirements and is expected to be field-tested in the coming months. To support reproducibility and to encourage alternative approaches, the dataset, the trained model, and its implementation on the smartwatch were published under free licenses.

Currently, deploying machine learning workloads in the Cloud–Edge–IoT continuum is challenging due to the wide variety of available hardware platforms, stringent performance requirements, and the heterogeneity of the workloads themselves. To alleviate this, a novel, flexible approach for machine learning inference is introduced, which is suitable for deployment in diverse environments—including edge devices. The proposed solution has a modular design and is compatible with a wide range of user-defined machine learning pipelines. To improve energy efficiency and scalability, a high-performance communication protocol for inference is propounded, along with a scale-out mechanism based on a load balancer. The inference service plugs into the ASSIST-IoT reference architecture, thus taking advantage of its other components. The solution was evaluated in two scenarios closely emulating real-life use cases, with demanding workloads and requirements constituting several different deployment scenarios. The results from the evaluation show that the proposed software meets the high throughput and low latency of inference requirements of the use cases while effectively adapting to the available hardware. The code and documentation, in addition to the data used in the evaluation, were open-sourced to foster adoption of the solution.

Federated learning (FL) involves joint model training by various devices while preserving the privacy of their data. However, it presents a challenge of dealing with heterogeneous data located on participating devices. This issue can further be complicated by the appearance of malicious clients, aiming to sabotage the training process by poisoning local data. In this context, a problem of differentiating between poisoned and non-IID data appears. To address it, a technique utilizing data-free synthetic data generation is proposed, using a reverse concept of adversarial attack. Adversarial inputs allow for improving the training process by measuring clients' coherence, favoring trustworthy participants. Experimental results, obtained from the image classification tasks for MNIST, EMNIST, and CIFAR-10 datasets are reported and analyzed.