Data Fused Motor Fault Identification Based on Adversarial Auto-Encoder

In this paper, a fault identification algorithm for motors is proposed. The Adversarial Auto-Encoder is adopted as the main structure of deep neural network, extracting the latent vectors of input signals. The latent vectors obey the designed prior distribution instead of random distribution in the traditional Auto-Encoder. The structure of neural network is improved to be non-fully-connected for data fusion, which improves the stability and reliability of fault identification. To solve the diverge training problem and accelerate the training processing, a cyclic training method with variable learning rate is proposed for the non-fully-connected network. Finally, the algorithm is testified by open source data of bearing fault motor whose current and vibration signal are fused for fault identification. The proposed algorithm has the ability to determine the fault type with accuracy reaching 85%, 10% much higher than the traditional machine learning models.