A Cooperative Convolutional Neural Network Framework for Multisensor Fault Diagnosis of Rotating Machinery

Multisensor data fusion techniques and advanced convolutional neural network (CNN) have contributed significantly to the development of intelligent fault diagnosis. However, few studies consider the information interactions between different sensor data, which limits the performance of diagnosis frameworks. This article introduces the novel convolution concept and the cross attention mechanism, proposing a cross attention fusion CNN (CAFCNN) diagnostic framework to improve the multisensor collaborative diagnostic technique. Specifically, a global correlation matrix is first developed to encode signals as images, highlighting the correlations between different points in the time-series data. Then, an attention mechanism called global spatial (GS) attention is proposed for extracting positional and spatial information in images. Finally, the developed interactive fusion module (IFM) utilizes cross attention to achieve information interaction of features from different sensors. The created gear dataset and the publicly available bearing dataset validate the effectiveness and generalization of the proposed methods. Moreover, the information interaction capability of CAFCNN is explained by visualizing the features.