Novel deep cross-domain framework for fault diagnosis or rotary machinery in prognostics and health management

Abstract

Improving the reliability of engineered systems is a crucial problem in many applications in various engineering fields, such as aerospace, nuclear energy, and water declination industries. This requires efficient and effective system health monitoring methods, including processing and analyzing massive machinery data to detect anomalies and performing diagnosis and prognosis. In recent years, deep learning has been a fast-growing field and has shown promising results for Prognostics and Health Management (PHM) in interpreting condition monitoring signals such as vibration, acoustic emission, and pressure due to its capacity to mine complex representations from raw data. This doctoral research provides a systematic review of state-of-the-art deep learning-based PHM frameworks, an empirical analysis on bearing fault diagnosis benchmarks, and a novel multi-source domain adaptation framework. It emphasizes the most recent trends within the field and presents the benefits and potentials of state-of-the-art deep neural networks for system health management. Besides, the limitations and challenges of the existing technologies are discussed, which leads to opportunities for future research. The empirical study of the benchmarks highlights the evaluation results of the existing models on bearing fault diagnosis benchmark datasets in terms of various performance metrics such as accuracy and training time. The result of the study is very important for comparing or testing new models. A novel multi-source domain adaptation framework for fault diagnosis of rotary machinery is also proposed, which aligns the domains in both feature-level and task-level. The proposed framework transfers the knowledge from multiple labeled source domains into a single unlabeled target domain by reducing the feature distribution discrepancy between the target domain and each source domain. Besides, the model can be easily reduced to a single-source domain adaptation problem. Also, the model can be readily updated to unsupervised domain adaptation problems in other fields such as image classification and image segmentation. Further, the proposed model is modified with a novel conditional weighting mechanism that aligns the class-conditional probability of the domains and reduces the effect of irrelevant source domain which is a critical issue in multi-source domain adaptation algorithms. The experimental verification results show the superiority of the proposed framework over state-of-the-art multi-source domain-adaptation models.