Fault Diagnosis Method for Different Types of Rolling Bearings

Abstract

Rolling bearings are in normal operating state most of the time, and their types are various. The failure data of a certain type of rolling bearing with labels is scarce or even impossible to be obtained, resulting in poor diagnostic accuracy or failure to diagnose faults. Aiming at this problem, a deep feature transfer fault diagnosis method was proposed for different types of rolling bearings. Firstly, short-time Fourier transform was used to process the vibration signals of different types of rolling bearings, and a two-dimensional image data set was constructed, the data of one type was selected as the source domain, and other types data as the target domain; Secondly, the domain-shared improved AlexNet deep convolutional network was constructed, and the conditional adversarial mechanism was introduced, the optimization method of joint distribution of features and labels was improved into random linear combination, and deep features were extracted, and so the features and labels of source and target domains can be adaptive simultaneously and the purpose of transfer can be achieved; Finally, the Nesterov accelerated gradient descent optimization algorithm was used to accelerate the gradient convergence during the training process, and the fault diagnosis models of different types of rolling bearings were established. The experimental results show that the proposed method can achieve multi-state classification of normal, different damaged location and degrees of different types of rolling bearings, and higher accuracy can be obtained.