An Efficient CNN with Tunable Input-Size for Bearing Fault Diagnosis

Abstract

Deep learning can automatically learn the complex features of input data and is recognized as an effective method for bearing fault diagnosis. Convolution neuron network (CNN) has been successfully used in image classification, and images of vibration signal or time-frequency information from short-time Fourier transform (STFT), wavelet transform (WT), and empirical mode decomposition (EMD) can be fed into CNN to achieve promising results. However, the CNN structure is complex and not efficient enough for different datasets. Furthermore, it is less efficient to process the input data by WT and EMD than by STFT. In this work, the low bound for input size of 2D data is analyzed by considering the relationship between the characteristic vibration frequencies and the window size of STFT to guide the determination of the minimum input size. Then a general adaptive CNN structure for different datasets is designed. According to the experimental results for four datasets, the proposed method is universal and the parameter settings can be guided by the low bound of input size. Surprisingly, all classification accuracies for the four datasets can achieve 100% in ten times of independent run without redesigning the CNN structure.

Keywords

• Bearing fault diagnosis
• Deep learning
• CNN
• STFT