Akademik Veri Yönetim Sistemi
JOURNAL OF INTELLIGENT MANUFACTURING, 2025 (SCI-Expanded)
Recently, there has been a growing interest in utilizing deep learning-based models for equipment condition monitoring. However, many existing fault diagnosis techniques that employ deep learning require large amounts of labeled historical fault data, which is often challenging or impossible to obtain. Furthermore, access to long-term data during fault occurrences is typically limited in real-world industrial settings. To address these issues, the paper proposes a multi-signal data fusion-enhanced deep meta-learning method for intelligent fault diagnosis with limited data. Meta-learning, an emerging technique, shows promise in mitigating the challenges posed by data scarcity through the use of few-shot learning strategies. Coupling it with multi-signal fusion techniques improves the model's generalization capability, yielding higher testing accuracy and robustness. Thus, the proposed method not only addresses the problem of data scarcity but also provides better diagnostic performance across a variety of equipment conditions compared to single signal approaches. Experiments conducted on three industrial datasets demonstrate the efficacy of the proposed method across three use cases, encompassing various fault severities, types, and working conditions. Comparative analyses reveal that the proposed approach outperforms single signal techniques, with performance improvements observed as the number of signal types increases. Notably, the method achieves significant gains in diagnostic accuracy over existing approaches, including 5.83 and 6.07% for 3-way 1-shot and 3-way 10-shot tasks in the first use case; 11.69 and 10.26% for 5-way 1-shot and 5-way 10-shot tasks in the second use case; and 5.88, 6.07, 6.55, and 2.95% for 3-way 1-shot, 3-way 10-shot, 5-way 1-shot, and 5-way 10-shot tasks, respectively, in the third use case.