Understanding IFRA for Detecting Synchronous Machine Winding Short Circuit Faults Based on Image Classification and Smooth Grad-CAM++

Abstract

Fault diagnosis of the synchronous machine is essential for the safe operation of the modern power system. Recently, impulse frequency response analysis (IFRA) has been used to detect synchronous machine winding short circuit (SC) faults since the failure will significantly alter the IFRA signature of machines. However, there is no standard and reliable code for interpreting the IFRA signatures. It often refers to the frequency response analysis of power transformer standards, the mathematical-statistical indicator of relative factor is directly used to perform the judgment by the standards’ threshold, which might not be appropriate. Besides, the internal mechanism of frequency response analysis diagnosis is unclear. Therefore, this study proposes a technique based on image classification and smooth grade and gradient-based class activation maps (Smooth Grad-CAM++) to understand and interpret the IFRA method. It trains and analyzes the visualization results of the fault data set of a 5 kW synchronous machine’s winding. The experimental results show that the average accuracy of the image classification model based on Resnet18 reaches 99.63%. We performed an IFRA difference analysis according to the visualization results and gave some suggestions about IFRA. The training and detection process can be accelerated based on these suggestions. To illustrate the generalization of the suggestions, another 8 kW synchronous machine is used for the case study, and the experimental results show that these suggestions are still effective. The main contributions of this study are understanding the internal mechanism of IFRA diagnosis from another perspective and providing some conclusions different from previous research results on the IFRA method for the synchronous machine.