Evaluation of Time-Frequency Representations for Deep Learning-Based Rotating Machinery Fault Diagnosis

Abstract

This study evaluates and compares five time-frequency representation (TFR) methods for fault diagnosis in rotating machinery, aiming to ensure operational reliability and reduce unexpected downtime. The methods—short-time Fourier transform (STFT), continuous wavelet transform (CWT), modified S-transform (MS-transform), smoothed pseudo Wigner-Ville distribution (SPWVD), and Hilbert-Huang transform (HHT)—are investigated. Vibration signals from benchmark bearing and gearbox datasets are converted into two-dimensional TFR data and classified using a convolutional neural network (CNN). Results show that MS-transform achieves the highest accuracy (up to 99.87%) under ideal conditions. STFT and CWT demonstrate better robustness in noisy environments, maintaining over 99% accuracy at 15 dB signal-to-noise ratio (SNR). SPWVD is computationally intensive with moderate performance, while HHT performs poorly under noise. Renyi entropy, energy conservation, and training time are also used to assess TFR quality. These findings support selecting appropriate TFR methods for industrial fault diagnosis.