Robust Multiple-Fault Diagnosis of PMSM Drives Under Variant Operations and Noisy Conditions

Abstract

With the rapid development of industrial applications using permanent magnet synchronous motors (PMSMs) and the Internet of Things, the demand for using robust fault diagnosis methods working in noisy conditions has increased significantly. The current data-driven methods depend mainly on deep learning (DL) models due to the effectiveness of automated feature extraction. However, these models have shallow depths compared with benchmark convolution neural networks, limiting their accuracy in final predictions, and they are established based on the hypothesis that the measured data are noiseless. Despite this, electric machinery is subjected to various noise sources that interfere with measurements during operation. This article proposes a new scheme combining a transfer-learned pretrained residual neural network (ResNet) and supervised machine learning (S-ML) to enhance the performance of DL models in noisy industrial environments. The effectiveness of the proposed scheme is validated using an in-house setup of a PMSM drive with demagnetization and intern-short circuit faults at variant operating conditions. The results show that the proposed method significantly reduced the computational burden by tenfold on average while improving the average accuracy to 96.84% across all the datasets compared with other DL and S-ML methods, with high robustness in noisy working conditions.

Keywords

• Deep transfer learning (TL)
• data-driven fault diagnosis
• hybrid learning
• permanent magnet synchronous motors (PMSMs)
• pretrained deep models
• wavelet analysis