A fault diagnosis network based on domain adversarial learning and distribution matching for rotating machine vibration signal with noise and across-load conditions
In the practice of deep learning-based rotating machinery fault diagnosis method, how to improve the accuracy of cross-working condition diagnosis under noise background is an urgent problem to be solved, which limits the application of deep learning method in engineering practice. Although domain adaptation methods have been widely researched to solve the problem, they often face the problem of domain shift. In this paper, two domain adaptation methods: domain adversarial learning and distribution matching methods, are used to train the deep network, to make the model realize the fault diagnosis across load conditions and signal-to-noise ratio. Moreover, in order to suppress the domain shift phenomenon which often occurred in domain adaptation tasks. A new network architecture based on the wide convolution kernel, multi-scale attention module, and self-adaptive soft threshold function is proposed so that the network is more suitable for feature extraction in cross-working condition fault diagnosis, and can avoid the influence of noise in vibration signals. Compared with other methods, the proposed method has good performance in faults diagnosis across load conditions and under noise. Feature visualization proved that the proposed network can effectively extract the conjoint fault features across different load conditions from the signal; hence, the fault diagnosis across load conditions and under the noise background can be realized and the domain shift can be suppressed effectively.
展开▼