Application of A Taylor Expansion Criterion-based Pruning Convolutional Network for Bearing Intelligent Diagnosis

摘要

Deep learning methods provide an effective tool for processing massive data obtained in data-driven machine health monitoring because of their powerful feature self-learning ability from raw data. As one of popular deep networks, the convolutional neural networks (CNNs) show successful applications in various fields. However, a full CNN with a large amount of parameters suffers from the problem of long latency and high memory cost and a much smaller network can achieve similar generalization ability. In this paper, the network pruning based on the Taylor expansion criterion is introduced to process bearing vibration data and then conduct intelligent fault diagnosis. Experimental data collected from different bearings are used to access the efficiency and accuracies of the original and pruned networks. The results demonstrate that the network pruning on large pretrained CNN can keep a balance between the classification accuracy and compact structure and can be further extended to applications of online condition monitoring and intelligent fault diagnosis of bearings.