A Novel Generative Adversarial Networks via Music Theory Knowledge for Early Fault Intelligent Diagnosis of Motor Bearings
Peien Luo, Zhonggang Yin, Dongsheng Yuan, Fengtao Gao, Jing Liu
Abstract
Weak signal features of early bearing faults are interfered by environmental noise, which seriously affects the accuracy of diagnosis results. Moreover, a large amount of data calculation and manual parameter adjustment during model training will affect the timeliness and intelligence of diagnosis results. Aiming at the above problems, an intelligent diagnosis method for motor bearing fault based on music theory knowledge novel generative adversarial networks (MTKGAN) is proposed for the first time. First, the game between generation and discrimination models is used to generate fault samples. The Earth-Mover distance is used to measure the distance between the real and generated distribution. The method generates and enhances weak signal features, and the interference of environmental noise on the signal is effectively solved to improve the accuracy of fault diagnosis. Second, inspired by music theory knowledge, the fault feature affine invariance migration method based on adaptive chord transformation strategy is proposed. The problems of Big Data training and manual parameter adjustment are effectively solved to improve the timeliness of fault diagnosis. Finally, the advantages of MTKGAN in early fault diagnosis of motor bearings are verified by comparing the public dataset and motor bearing fault experiment platform with the existing advanced methods.