Applying deep incremental learning-based posture recognition model for injury prevention in construction

This research investigated the feasibility of applying a Deep Incremental Learning model to achieve a high posture recognition performance with Wearable Sensors (WS). The authors use the recognition of Musculoskeletal Disorder (MSDs) related postures among construction workers as the testbed. This research proposed the Convolutional Long Short-Term Memory (CLN) model under Incremental Learning (IL), where a trained model adapts to new subject' postures to maintain high recognition performance. The model was evaluated on datasets from nine construction workers. Results show: i) CLN model with shallow convolutional layers achieved high recognition accuracy (Macro F1 score) under personalized (0.87) and generalized (0.84) modelling; ii) Generalized CLN model under “Many-to-One” IL strategy can adapt to a new subject and balance the forgetting of learnt subjects; iii) incremental CLN model gave close detection of posture proportion and holding time to ground-truth, which facilitates reliable MSDs risk assessment and further prevention through monitoring injury-related postures.