Neural Network Modeling Based on Multimodal IIoT Sensing Data: Psychological Stress Assessment and Industrial Human-Machine Collaboration Early Warning System for University Students

Abstract

With the rapid development of industrial Internet of Things (IIoT), its educational applications extend from equipment monitoring to mental health management. Addressing the limitations of traditional methods (e.g., subjective self-assessment scales) in real-time psychological stress evaluation, this paper proposes a neural network integrating multimodal IIoT data—physiological signals (EEG, HRV), behavioral data (expression, gesture), and interaction logs (text, clickstream)—to build a dynamic fusion and lightweight warning system. The model employs a cross-modal attention mechanism to adaptively allocate data weights (e.g., prioritizing EEG signals by 58% in exam scenarios) and a tensor fusion network (TFN) for feature extraction. An edge-cloud collaborative framework based on federated learning enhances generalization while ensuring privacy through AES-256 encryption, local feature preprocessing, and differential privacy protections during model updates. Experiments on a campus dataset show 89.2% stress classification accuracy (10.7% higher than unimodal approaches), sub-105 ms alert latency, and a 4.7% false alarm rate. Personalized interventions (e.g., counseling) improved stress alleviation by 61.7% for moderate-to-severe cases. This study advances IIoT's role in intelligent mental health management and adaptive human-computer interaction systems.