A Novel Multiscale Lightweight Fault Diagnosis Model Based on the Idea of Adversarial Learning

Big data fault diagnosis methods based on deep learning (DL) have been widely studied in recent years. However, the number of labeled bearing fault samples is limited in industrial practice, and these samples usually are contained with complex environmental noise. Therefore, it is necessary to develop a generalizable DL model with strong feature learning ability. To tackle the above challenges, this article proposes a multiscale lightweight fault diagnosis model based on the idea of adversarial learning. The multiscale feature extraction unit is applied to the vibration signal for learning complementary and abundant fault information at different time scales, increasing the width and reducing the depth of the network. Meanwhile, a novel easy-to-train module based on the idea of adversarial learning is utilized to strengthen the feature learning ability by competitive optimization. Besides, the depthwise separable convolution is introduced to reduce the size of the network and achieve the lightweight design. These measures strengthen the feature learning ability and generalization of proposed method, and further ensure its noise robustness in the case of limited samples. The effectiveness of the proposed method has been Verified on two bearing datasets, and experimental results show that the proposed method is robust to noise in the case of limited samples.

Automated summary

This article proposes a multiscale lightweight fault diagnosis model based on the idea of adversarial learning to address the limited labeled bearing fault samples and complex environmental noise in industrial practice. The model utilizes multiscale feature extraction unit, an easy-to-train module based on adversarial learning, and depthwise separable convolution to achieve a lightweight design, strengthening feature learning ability and generalization.