Deep clustering domain adaptation for fault diagnosis of rolling bearings in mining belt conveyors

• Published in: Measurement : journal of the International Measurement Confederation Vol. 248; p. 116878
• Main Authors: Li, Xin, Kou, Ziming, Han, Cong, Huang, Shuai
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 15.05.2025
• Subjects: Deep learning; Domain adaptation; Intelligent fault diagnosis; Roller bearing; Vibration signal; Deep learning; Roller bearing; Domain adaptation; Intelligent fault diagnosis; Vibration signal
• ISSN: 0263-2241
• DOI: 10.1016/j.measurement.2025.116878

•An unsupervised deep domain adaptation method is proposed for rolling bearing diagnosis in mining conveyor belts.•A novel approach combines deep clustering and generative adversarial networks for improved classification.•Experiments in real mining settings validate the method’s effectiveness for conveyor belt diagnostics. Existing domain adaptation methods effectively address domain shift challenges in publicly available datasets but face substantial hurdles in mining belt conveyors. Challenges include complex mechanical interference, irregular coal flow impacts, and environmental noise, impeding fault signal extraction and diagnostic model accuracy. To combat domain shift and low diagnostic accuracy due to noise, we introduce a novel roller bearing fault diagnosis approach, Deep Clustering and Domain Adversarial Learning (DCDA). DCDA autonomously learns target domain features, enhancing pattern recognition. A pseudo-label correction scheme, uniting generative adversarial and deep clustering models, bolsters clustering quality and shared information learning. Experimental results in real mining settings demonstrate DCDA’s superior diagnostic accuracy and robustness, boasting a 18.16% improvement over current domain adaptation methods. DCDA excels at diagnosing roller bearing faults in noisy mining environments and circumvents labeled data scarcity in underground coal mines.