Information
- Title: A Non-Local Adaptive Network for Cross-Domain Intelligent Fault Diagnosis Leveraging Multi-Source IOT Data
- Publisher: SPRINGER NATURE Link
- Journal: Journal of Mechanical Science and Technology
- Author:
Hanshu Shao, Yongwen Tan, Jingbo Li, Hengkai Gao, and Huiying Yin - Status: Under Review
Abstract
The recent advancements in domain adaptation have sparked a shift towards practical applications of intelligent fault diagnosis, particularly in scenarios where label information for a specific machine remains elusive. Nonetheless, those techniques still require massive unlabeled data in the target domain, which often proves to be rather demanding for industrial applications due to the time-varying working conditions and unpredictable fault patterns. As such, this paper introduces a novel framework named non-local domain adaptation network (NLDAN) to tackle cross-domain fault diagnosis tasks with limited data for real-world applications. The framework explores the idea of non-local deep convolutional operation associated with the attention mechanism, which can learn both local and non-local features based on the multi-range feature extraction process. Furthermore, the framework introduces the weighted domain adaptation (WDA) module to align the distributions of the features considering the prior probability distribution of different categories, offering enhanced reliability in distribution matching, particularly for tasks confronted with a scarcity of samples. Combining those strategies, the proposed non-local domain adaptation model provides an end-to-end solution for cross-domain tasks with distribution mismatches and data scarcity. To assess the effectiveness of our approach, we conduct validation experiments tailored to cross-domain rolling bearing fault diagnosis. Notably, our methodology surpasses the best-performing deep learning models, demonstrating robust performance across data-rich and data-constrained scenarios
