Enhanced labor activity recognition via wearable physiological signals: a comparative machine learning study

Labor activity recognition using wearable physiological signals can support safety, productivity monitoring, and human-centered decision-making in industrial workplaces. This study presents an exploratory comparative benchmark of twenty machine learning and deep learning models for two labor activity-recognition tasks: binary productive/non-productive recognition and multiclass recognition of specific productive tasks. The raw data comprised 29,187 synchronized one-second physiological readings collected from four workers during three shifts in a stone prefabrication facility using Empatica E4 and Zephyr BioHarness 3 wearable devices. Blood volume pulse, electrodermal activity/galvanic skin response, heart rate, respiration rate, and skin temperature were transformed into 35 physiological statistical features per labeled analysis window; an additional activity-window Duration predictor yielded 36 predictors per modeling window. The final binary dataset contained 400 windows (200 productive and 200 non-productive), and the multiclass dataset contained 393 windows (131 oiling, 131 moving, and 131 cleaning). All models were evaluated using a stratified 80/20 train-test split, with standardization fitted only on the training data. The MLP Classifier achieved the highest binary performance (accuracy = 0.8625, support-weighted precision = 0.8646, support-weighted recall = 0.8625, and support-weighted F1-score = 0.8623). For multiclass recognition, CatBoost, Extra Trees, and Bidirectional LSTM each achieved accuracy = 0.9873, support-weighted precision = 0.9878, support-weighted recall = 0.9873, and support-weighted F1-score = 0.9873. These findings demonstrate strong within-dataset classification performance while emphasizing the need for future worker-independent and multi-site validation.