Abstract: Aiming at the problems of low intelligent diagnosis efficiency of rolling bearings due to the excessive number of parameters of existing convolutional neural network models and the limited recognition accuracy due to the number of training samples, a rolling bearing fault recognition method based on multi-scale atrous convolution-convolutional neural network (MSAC-CNN) was proposed. In this method, a large atrous convolution kernel and a standard convolution kernel are used in the input layer of the model to extract the multi-scale sensitive features of one-dimensional vibration signals. Then, the sensitive features extracted in the input layer were further extracted by using the 1×1 and 3×1 small convolution kernel and the 2×1 maximum pooling operation. Finally, the fully connected layer in the traditional convolutional neural network is replaced by the global average pooling layer. The experimental verification results with bearing fault data from Western Reserve University and our laboratory show that the proposed method has good generalization performance and performs the task of fault identification with fewer training samples. Moreover, it can accurately identify weak bearing faults under certain noise interference.

Key words: multi scale atrous convolution-convolutional neural network, rolling bearing, fault identification, small samples, weak fault