Performance Evaluation of the Time-Frequency Transformation Methods on Electrical Machinery Fault Detection

Abstract

Detecting faults in electrical machine systems is crucial for developing maintenance strategies. Modern technology enables personalized maintenance planning for system components by continuously or periodically monitoring systems with sensors. The first step in condition-based maintenance planning is predicting faults from sensor data. Monitoring vibration signals is one of the most preferred approaches for fault diagnosis in electrical machine systems. We have used a dataset containing vibration data recorded to detect intentionally created faults in an electrical machine system. The paper spots three popular methods to convert the time domain data into the frequency domain: power spectral density signal, spectrogram images, and scalogram images. Furthermore, we have analyzed the performance of the popular machine learning and deep learning methods with frequency-domain inputs. We have reported the results with accepted performance metrics such as accuracy, precision, recall, and F1 score. Our findings indicate that spectrogram images with the InceptionV3 model achieve maximum accuracy of over 98% accuracy among. The findings also highlight the necessity of carefully selecting model parameters based on the data type.