A Self-Attention Enhanced Deep Learning Framework for Fatigue Prediction

Fatigue, a form of weariness and lack of energy, affects many people daily. Early detection of fatigue is often the first milestone to accomplish. Detection can then help to reduce injuries, improve athletic performance and allow smarter decision making. Wearable devices provide rich data that can be utilized to detect fatigue. Deep learning architectures have shown strong success in taking advantage of the raw data that wearable devices provide. However, many architectures can still be improved through the incorporation of hybridization design strategies. In addition, many studies that predict fatigue use complex environments and setups that are impractical for a production environment. This study aims to improve the architecture of a popular hybrid model through the introduction of a self-attention mechanism in it's design. In addition, this study will evaluate many single-sensor setups to classify fatigued and non-fatigued states using the DUO-GAIT dataset. Our new architecture, DeepConvLSTM with Self-Attention (DCLSA), achieves 93.5% accuracy using only the left wrist sensor, improving over the DeepConvLSTM architecture by 1.7% and increasing inference time by only 0.34 ms. This study highlights the importance of architectural design and aims to popularize the usage of single-sensor setups to enable practical deployment of fatigue prediction systems.