Dynamic Graph Neural Networks with Coordinate Attention for Human Activity Recognition in Manufacturing Industry

In today's rapidly advancing manufacturing industrial landscape, the application of wearable sensor-based human activity recognition (HAR) technology has become essential for enhancing productivity, ensuring safety, and optimizing human-machine collaboration. However, extracting robust feature from wearable sensing data that is multi-positional presents significant challenges, due to the implicit and disordered correlations among data from different locations. Considering graph neural networks (GNNs) could represent data in structured graph format and uncover complex relationships with message passing between graph nodes, a novel human activity recognition strategy based on dynamic graph neural networks and coordinate attention mechanism is proposed, which uses a dynamic graph mechanism to capture the associations between different time slots and extracts spatial-temporal features from multi-dimensional data, utilizing the coordinate attention mechanism. Experiments are conducted on the public OPPORTUNITY HAR dataset and a self-constructed worker activity dataset in manufacturing industry (WAMID), the experimental results demonstrate that the proposed dynamic graph neural network with coordinate attention (CA-DGNN) achieves a more competitive classification accuracy than other existing HAR deep learning-based models.