Temporal learning in predictive health management using channel-spatial attention-based deep neural networks

This article proposes a novel model, the multi-scale attention (MSA) block, to improve predictive health management (PHM) capabilities. The proposed MSA block incorporates channel and spatial attention mechanisms to improve temporal learning in industrial procedures. We further optimize the model by using strategies to increase the network's receptive field, boosting its adaptability across diverse channels. To validate the effectiveness of the MSA block, we develop a deep neural network, the Channel-Spatial Attention-Based Temporal (CSAT) network, which comprises 18 layers rooted in the MSA block. Our experimentation uses the AI4I 2020 predictive maintenance and Microsoft Azure predictive maintenance datasets for training and validation. The results reveal that our approach surpasses other leading models in performance, thereby establishing its efficacy and superiority. The ablation study shows that the model can benefit from the proposed components. This article contributes to the advancement of PHM by showcasing the benefits of integrating neural networks and attention mechanisms in temporal learning.