Research on Abnormal Detection of Rolling Bearing Based on Convolutional Neural Network

This paper proposes an anomaly detection method for rolling bearings based on a convolutional neural network (CNN) and support vector data description (SVDD). The main research contents are as follows. Firstly, to address the problems of pixel point waste and image resolution determination in the image preprocessing process of the current CNN-based bearing anomaly detection method, a transformation method of signal conversion into two-dimensional grayscale images is proposed to improve the image data generation efficiency. Secondly, to solve the problems that CNN-based anomaly detection requires the manual definition of labels and the lack of anomaly data in the actual detection process, a CNN-SVDD method that combines LeNet-5 and SVDD is adopted. By only defining normal sample labels, bearing anomaly detection can be performed with very little or even no anomaly data. The average AUC value reaches 96.50% after the validation of the public dataset, which has a high testing performance. Finally, the practicality of the CNN-SVDD algorithm model is further verified by experiments. Relying on the generator set of the test platform, a comparative test of motor bearing vibration with missing bearing slip oil and sufficient slip oil was designed. After several tests of the actual data in the algorithm model, the average AUC of the test index reached 98.82%. The whole testing system training and testing process is fast, highly accurate, and has strong practical application value.