Numerical simulation driving generative adversarial networks in association with the artificial intelligence diagnostic principle to detect mechanical faults

Abstract

Predictive maintenance (PM) is a potential outburst point in intelligent manufacturing. As the key issue of PM in recent years, artificial intelligence (AI) diagnosis has become a research focus in engineering research and developed a series of diagnostic methods with great application prospect. Complete faulty samples are essential to activate AI models. However, in engineering applications, faults are considerably more than the fault samples obtained in advance from real-world running mechanical systems. Therefore, a numerical simulation driving generative adversarial networks (GANs) is proposed to obtain relatively complete fault samples for building a bridge between AI models and real-world running mechanical systems and to present the new-type principle for AI diagnosis. First, the finite element method is employed to simulate missing fault samples of mechanical systems, which are combined with the fault samples obtained in advance to construct relatively complete fault samples. Then, a large number of new samples similar to the relatively complete fault samples are generated using GANs to improve the quality of fault samples. The new samples are further combined with the relatively complete fault samples to obtain complete fault samples (synthetic fault samples). The complete fault samples are employed as training samples of AI models, and the test samples of unknown faults are finally classified. The fault diagnostic performances are compared with the commonly used AI diagnostic models by using the public datasets of faulty bearings, the faulty gear test rig, and the faulty rotor test rig as examples. The effectiveness of the proposed AI diagnostic principle is verified and it brings the new hope to extend AI diagnostic models to detect faults in real-world running mechanical systems.