Rolling bearing fault identification with acoustic emission signal based on variable-pooling multiscale convolutional neural networks

This paper propose a new fault identification method based on variable pooling multiscale CNN (VPMCNN), which solves the bearing industrial problem of huge variable features and inherent multiscale characteristics in acoustic emission (AE) signals. First, the pooling projection components (PPCs) of the signals are obtained through the variable pooling layer. The PPCs consider the curse of invariant feature weight in traditional CNN pooling layer, and select the more weighted features to enhance the classifying quality. Second, an improved multiscale fusion feature module is introduced to further extract the hidden features, which is called fused components (FCs). The FCs aims to automatically extract multiple scale features using different filter sizes from raw acoustic signals. Then the GAP (Global Average Pooling) layer is performed to realize classification. Finally, the fault identification using the proposed algorithm is performed by testing the bearing AE signals with single operating condition and variable operating conditions, and the results show the effectiveness of the proposed method, compared with existing AE based bearing fault identification methods.