45 / 2021-06-23 18:09:00

Generative Zero-shot Learning Compound Fault Diagnosis of Bearing

Because of the concurrency and coupling of various types of faults, and the exponential increase of the number of possible failure modes, compound fault diagnosis has always a difficult issue in fault diagnosis of bearings. The existing deep learning-based model requires a large number of labeled vibration samples of compound faults. In industrial scenarios, collecting and labelling enough compound fault samples is unpractical, while the single fault samples are easy to obtain. These issues motivated us to construct a novel Zero-shot Learning model trained on the single fault samples to identify unknown compound faults. According to the fault characteristics, we pioneer a semantic vector coding method for expressing single fault and compound fault. A deep convolutional neural network was designed to extract visual features in the time-frequency domain of fault vibration data. Then the semantic vectors and visual features of single faults were adversarial trained in GAN module, so that the trained generator can generate the compound fault visual feature using the compound fault semantics. Then the K-nearest neighbor method is used to measure the distance between the visual feature of the real compound fault and the generated compound fault feature to identify the unknown compound faults. To verify the proposed method, we conduct experiments on a self-built experimental platform. The results demonstrate that the accuracy of identifying compound fault reached 78.10% when the model was trained without any compound fault samples. This is the first attempt of compound fault diagnosis of bearing base on generative zero-shot learning.