Real-Time Physiological Fatigue Prediction for Human–Robot Collaborative Manufacturing Using Wearable Sensor Fusion and Hybrid Deep Learning: An In Silico Digital Twin Study

Musculoskeletal fatigue precedes much of the injury burden in manufacturing, yet most monitoring schemes register an injury only once it has occurred, which becomes critical where operators share a workspace with collaborative robots. This paper presents a wearable sensing platform and a learning pipeline that track operator fatigue continuously during human–robot collaborative assembly, evaluated in silico. Eight surface electromyography (sEMG) channels, six inertial measurement units, and four force-sensitive resistors feed a 127-dimensional descriptor computed over 30s windows. Eight classifiers were trained on 1536 simulated working hours from 24 anthropometrically diverse synthetic operators. Under leave-operators-out cross-validation, a 1D-CNN–LSTM hybrid reached 89.3% three-class accuracy and 87.1% balanced accuracy with 73.9 k parameters and 22ms inference on a Jetson Nano; a CNN–BiLSTM–attention model gained 0.3 percentage points for 1.6 times the parameters, a difference that was not statistically significant. Ablation attributed 7.0 points of balanced accuracy to sEMG and 5.0 points to three contextual variables requiring no sensor. Alerts preceded severe fatigue by roughly 12min, and alert-triggered task reallocation cut peak shoulder load by 43% while retaining 94% of baseline throughput. Every result characterizes a simulated environment: the study establishes internal consistency, latency feasibility, and design trade-offs, and prospective validation with human operators remains a prerequisite for deployment.