A health image for deep learning-based fault diagnosis of a permanent magnet synchronous motor under variable operating conditions: Instantaneous current residual map

To take full advantage of a convolutional neural network (CNN) for deep learning-based fault diagnosis, many studies have examined the transformation of sensory signals into a two-dimensional (2D) input image. An important question to consider is: how can fault-related signatures in motor stator current signals be incorpo-rated into the 2D input image to a CNN model for fault diagnosis of a permanent magnet synchronous motor (PMSM)? To answer the question, this study newly proposes a novel health image, namely instantaneous current residual map (ICRM). Inspired by the idea that the phase and amplitude modulations in motor stator current signals are related to faulty states of a PMSM, the overall procedure for constructing ICRM includes two key steps: (1) to calculate current residuals (CRs); and (2) to spread the scaled CR pairs into a 2D matrix. A type of faults can be figured out by analyzing a degree or shape of spreading of the CRs in ICRM. Moreover, ICRM is robust to variable operating conditions in practical settings because the scaled CRs that the effects of the operating conditions are reduced can highlight fault-induced irregularities. To demonstrate the effectiveness of ICRM, it was experimentally validated using a surface mounted PMSM, operated under variable-speed and different load torque conditions.

Automated summary

This study proposes a novel method for fault diagnosis using deep learning and convolutional neural networks. By transforming fault-related signatures in motor stator current signals into a two-dimensional input image, the proposed method achieves accurate fault diagnosis of permanent magnet synchronous motors. The effectiveness of the proposed method is experimentally validated using a surface mounted PMSM under various operating conditions.