A Fault Diagnosis Method for Rolling Bearing Based on 1D-ViT Model

Abstract

Rolling bearings are the key component of large rotating machinery. When such components fail completely, the equipment will be out of service, causing significant economic loss, and even leading to safety accidents, threatening the lives of workers. However, it is difficult to detect the early failure symptom of rolling bearings during the routine maintenance. Advanced fault diagnosis tools are limited in cost and technology. Therefore, researchers focus on how to implement accuracy and convenience of fault diagnosis method. This paper proposes a fault diagnosis method for rolling bearings based on 1D-vision transformer (1D-ViT) encoder structure, which applies vision transformer (ViT) model to the fault diagnosis area. The end-to-end fault diagnosis can be realized by directly inputting the original one-dimensional acquired data without additional time-frequency domain conversion. After the encoder ablation experiment, the model structure was optimized. With the rolling bearing data set of Southeast University and Case Western Reserve University (CWRU), the number of floating point operations (FLOPs) is as low as 0.169G, the parameter number is as low as 4.13M, and average accuracy is 99.9%, which also has better noise-resistance performance. Compared with common classification models, the 1D-ViT model has achieved great comprehensive advantages in fault diagnosis accuracy, time complexity, space complexity and noise-resistance performance, which verifies the effectiveness and convenience of the proposed fault diagnosis method.