Multiscale Pooled Convolutional Domain Adaptation Network for Intelligent Diagnosis of Rolling Bearing Under Variable Conditions

Deep learning-based fault diagnosis methods usually require samples to meet the conditions of independent and identical distribution. In actual industrial occasions, the data distribution of mechanical equipment under variable operating conditions is different, which results in the degradation of diagnostic performance. To overcome the above shortcomings, a fault diagnosis method based on multiscale pooled convolutional domain adaptation network is proposed. In the method, a novel parallel multiscale pooled module is designed to replace the traditional convolution module in 1-D-CNN. The four parallel branches in the structure contain the pooling layers with different scales, as well as different modes, which can extract diversified features with different properties. Afterward, considering the influence of decision boundary on target feature matching, two independent fault classifiers are constructed and trained to decrease the misclassification of the samples. Simultaneously, the local maximum mean discrepancy (LMMD) metric is combined with the domain adversarial network to reduce the difference between the marginal and conditional distributions of samples so as to achieve intraclass and interclass alignment between the source and target domains. Experimental results on the Jiangnan University (JNU) bearing dataset show that, compared with the 1-D-CNN method, the average diagnosis accuracy of the proposed method on the six transfer tasks is increased by 9.43%, which indicates the effectiveness of the proposed fault diagnosis method.

I. Introduction

As the most common core component in rotating machinery, rolling bearings will bring huge economic losses and even casualties in case of failure [1], [2], [3]; therefore, real-time monitoring and early fault identification of rolling bearings are undoubtedly necessary, which can ensure the smooth and normal operation of equipment and prevent the occurrence of major accidents [4], [5]. Most of the traditional fault diagnosis methods use signal processing techniques to extract data features so as to identify faults in machinery; however, these methods not only require a certain level of expert knowledge but also lack the generalization ability in the face of increasing data [6], [7], [8]. In order to solve these problems, intelligent fault diagnosis methods based on deep learning have emerged. Deep learning models can automatically extract data features from the collected data, replacing the manual method of extracting features, and can realize an end-to-end diagnosis. Many deep learning algorithms have been widely used in the field of mechanical fault diagnosis, such as deep belief network (DBN), convolutional neural network (CNN), and recurrent neural network (RNN) approaches [9], [10], [11].