Raw vibration signal pattern recognition with automatic hyper-parameter-optimized convolutional neural network for bearing fault diagnosis

Abstract

Bearing fault diagnosis is of great significance for evaluating the reliability of machines because bearings are the critical components in rotating machinery and are prone to failure. Because of non-stationarity and the low signal-noise rate of raw vibration signals, traditional fault diagnosis methods often construct representative fault features via the technologies of feature engineering. These methods rely heavily on expertise and are inadequate in actual applications. Recently, methods based on convolutional neural networks have been studied extensively to relieve the demands of hand-crafted feature extraction and feature selection. However, the raw vibration signal is rarely taken as a direct input. This study combines a convolutional neural network with automatic hyper-parametric optimization and proposes two deep learning models for time-series pattern recognition to achieve “end-to-end” bearing fault diagnosis: a one-dimensional-convolutional neural network and a dilated convolutional neural network. The architecture of the two models are tweaked by automatic optimization rather than manual trial or grid search. Further, we try to figure out the inner operating mechanism of the proposed methods by visualizing the automatically learned features. The proposed methods are applied to diagnose roller bearing faults on a benchmark experiment and a prototype experiment. The results verify that our methods can achieve better performance than other intelligent methods via a Gaussian-noise test.