Deep joint convolutional neural network with double-level attention mechanism for multi-sensor bearing performance degradation assessment

The deep learning methods with data fusion are promising to deal with the performance degradation assessment (PDA) of rotating machinery with multi-sensor data reliably. However, there are still two challenges: (1) each sensor that is mounted at a different position makes a different contribution to the task, (2) there is much conflicting information between the signature owing to strong background noise. To address these two challenges, a deep joint convolutional neural network (DJ-CNN) including the feature extractor and the predictor is proposed for intelligent PDA tasks. Within this framework, multi-sensor data are input to the feature extractor network in parallel. Then, the predictor, whose attention module refines and recalibrates the feature maps in sensor-wise attention and signal-wise attention, is trained with input being multi-sensor data again. Finally, the trained DJ-CNN, which not only could naturally extract deep features from raw multi-sensor but also enhances the more important parts of feature maps in a double-level attention structure, is constructed for performance degradation assessment. The effectiveness and superiority of the proposed DJ-CNN are demonstrated on a run-to-failure bearing experiment.