Fault Types and Sizes Recognition of Rolling Bearing Based on Multitask Semi-Supervised Learning

Deep-learning models have shown their performance in bearing fault diagnosis. Those studies often focus on the fault types, but ignore the defect sizes of bearings. Moreover, deep-learning models rely on a large amount of data. However, it is still a challenge to obtain a mass of data in practical industrial applications. Besides, most of them are supervised learning models that require a number of labels. To solve these problems, a novel semi-supervised learning model named multitask multilevel feature fusion convolution shrinkage autoencoder (MMCSAE) is proposed for fault type and size recognition of rolling bearings in this study. First, a multiscale convolution shrinkage (MS-ConvS) is developed to extract features from vibration signals, where a shrinkage module is used to filter out the redundant components. Second, a multilevel feature fusion method is proposed for feature enhancement. The adaptability of the model is improved by feature fusion under small sample conditions. Third, a multitask weighting method is proposed to adaptively determine the convergence rate of different tasks in the training process. The effectiveness of MMCSAE is validated on a rolling bearing test bed and Case Western Reserve University (CWRU) dataset. The experimental results demonstrate that MMCSAE has a better fault diagnosis performance than representative models such as 1-D residual convolutional autoencoder (1DRCAE) and self-calibrated convolutional network (SCNet) under varying operational conditions and limited samples.