Adaptive deep convolutional neural networks with wide first-layer kernel bearing fault diagnosis algorithm based on adaptive batch normalization

Abstract

The adaptive deep convolutional neural networks (ADCNN) with wide first-layer kernel algorithm model for processing one-dimensional bearing vibration signals has low diagnostic accuracy for data noise and variable load. Therefore, the adaptive deep convolutional neural networks with wide first-layer kernel rolling bearing data diagnosis algorithm based on adaptive batch normalization (ADABN) adaptation is proposed. The algorithm first builds 64 × 1 the first wide convolution kernel layer, and then construct all 3 × 1, and finally build the full connection layer and SoftMax layer. At this time, the mean and variance of the target domain samples are used in each BN layer of the network to replace the mean and variance of the source domain samples used in the original BN layer, so as to achieve the purpose of data domain adaptation. The research on the adaptive deep convolutional neural networks with wide first-layer kernel fault diagnosis algorithm based on adaptive batch normalization is carried out on the ship propeller bearing data set provided by CSIC. The results show that the fault diagnosis accuracy reaches more than 99% in the case of variable data domain, which is better than other algorithms, and improves the adaptive ability of adaptive deep convolutional neural networks with wide first-layer kernel algorithm in the data domain.