Time-Shift Denoising Combined With DWT-Enhanced Condition Domain Adaptation for Motor Bearing Fault Diagnosis via Current Signals

• Published in: IEEE sensors journal Vol. 24; no. 21; pp. 35019 - 35035
• Main Authors: Wang, Xindan, Liu, Zhiliang, Dai, Menghang, Li, Wenru, Tang, Jinlong
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.11.2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Adaptation; Artificial neural networks; Conditional domain adaptation (CDA); cross-working condition; current fault diagnosis; Data collection; Datasets; Discrete Wavelet Transform; discrete wavelet transformation; Discrete wavelet transforms; Fault diagnosis; Feature extraction; Machine learning; Motor stators; Motors; Noise reduction; Resonant frequencies; Sensors; Stators; time-shift denoising (TSD); Vibration; Wavelet transforms
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2024.3455099

Traditional fault diagnosis often relies on vibration signals for experimentation. However, vibration diagnostic techniques demand high precision from sensors and stringent requirements for data collection points. Moreover, in environments with significant noise, vibration signals are subject to severe interference. Therefore, this article proposes a novel method for diagnosing bearing faults in motor stator currents across different operating conditions. It leverages time-shift denoising (TSD) and discrete wavelet transform (DWT)-enhanced conditional domain adaptation (CDA) techniques. Initially, the TSD method effectively eliminates the fundamental frequency and odd harmonic components from the motor stator currents, extracting residual signals with more significant valid components. By integrating the DWT-enhanced CDA technique, the feature learning space of the convolutional neural network (CNN) is expanded into the wavelet domain, which enhances frequency domain decomposition, retains low-frequency components, and filters out noise and irrelevant information. This approach aligns low-frequency features and strengthens the network's adaptability to different data distributions without requiring target domain labels. We conducted experiments on two public datasets and one real-world dataset, achieving an accuracy of up to 93.94%. The results demonstrate that the proposed method offers significant advantages in diagnosing bearing faults in motor stator currents under varying operating conditions.