Human-Centered Artificial Intelligence to Improve Workers' Safety and Well-being

Industry 5.0 is a new era of industrial transformation, characterized by the integration of digital technologies and artificial intelligence (AI) into production processes. Compared to the previous revolution, workers' health and well-being have become crucial in ensuring sustainable production and safe workplaces where machines collaborate with operators, reducing stressful workloads and optimizing production processes. Today, advanced technologies allow continuous data collection regarding workers' physical and mental state that AI can elaborate for predictive analysis and preventive assistance. This thesis deals with human-centered AI in the field of Occupational Safety and Health (OSH), focusing on improving workers' well-being and safety. Regarding well-being, we introduced a new method to self-assess emotions intuitively, based on a personalized emoji set. We used this method to develop a mobile application that allowed people to tag physiological signals recorded by wearables at random moments of the day. Various AI models were investigated to distinguish emotions based on the physiological signals and corresponding labels obtained using our new self-assessment method. In another experiment, we stimulated emotions through audio-visual content while recording a 14-channel EEG signal, labeling emotions using the proposed approach based on emojis. This dataset was used to develop an AI-based system to detect happiness from features extracted from the EEG signal graphs, thereby measuring the time a worker feels positive emotions throughout the day. We also proposed a speech analysis system that uses deep learning to detect quarrels and prevent escalating arguments that could lead to stress. Concerning workers' safety, we focused on workplace ergonomics and introduced a system to track the leg position without using cameras or wearables, respecting privacy. The system uses a Light Detection and Ranging (LiDAR) sensor to determine desk workers' leg positions. It recognizes the correct posture and 14 incorrect positions, classifying possible other positions as anomalies. A recommendation module alerts workers when they are sitting in incorrect positions for too long, and a dashboard provides information on the most frequent incorrect positions, helping promote customized and effective training paths. The thesis ends by describing an advanced version of this system conforming to the human-centric view of Industry 5.0. This system estimates manufacturing workers' postures in assembly/disassembly lines, combining LiDAR and inertial data recorded by a smartwatch. We modeled various body posture combinations in compliance with the ISO 11226 standard. These postures can be analyzed through a dashboard highlighting which workers are at risk of musculoskeletal disorders (MSDs), which body parts are more stressed, and which posture habits contribute to that risk. Safety engineers can then set up targeted and customized interventions to improve postural habits and prevent the onset of MSDs and pain in the medium-long term.