Data-driven Methods with Stronger Domain Generalizability for Machine Fault Diagnosis

Description

Project Background:Efficient fault diagnosis is paramount to uphold the reliability and safety of machines in modern industry. Meanwhile, the increasing complexity of machines and application scenarios poses new challenges to conducting fault diagnosis via traditional physical or parametric models. The recent industrial digitalization unfolded the unprecedented opportunity of studying data-driven methods for more intelligent machine fault diagnoses. Numerous studies have been carried out to explore effective methods from signal processing, classical machine learning, and recent deep learning for modeling the relationship between data and machine faults. However, the non-explicit and dataset-specific nature of models developed via classical data-driven schemes have been frequently criticized about its performance robustness and generalizability. Recently, there have been promising endeavors in exploring innovative machine learning schemes for more adaptive and robust data-driven machine fault diagnosis models. Interesting ideas include designing novel transfer learning schemes to enable model domain adaptation, conducting data augmentation with generative adversarial-learning-based domain synthesis, developing domain-invariant representation learning schemes for domain alignment, etc. Despite these exciting developments, following critical issues still hinder the readiness of data-driven models for realizing the domain generalization in machine fault diagnosis applications.• Awareness of domain generalization success conditions.A systematic understanding on conditions that data-driven models can successfully realize machine fault diagnosis domain generalization, which is extremely important to the application success, is untapped.• Lack of interpretability.Previous studies focused on upgrading the domain generalization capability of models. To gain more trust from practitioners, it is highly valuable to study more interpretable modeling schemes for domain generalization in machine fault diagnosis.• Field data quality issue.Field data might be corrupted at both input and label by noises. Existing domain generalization studies have addressed such an issue with assuming that data are independent and identically distributed (IID). Yet, data are not necessarily IID in reality and data-driven models with domain generalization for such data corruption need new research efforts. Project Aim and Description:This project aims to tackle three critical issues and develop data-driven methods with domain generalization of a higher practicality for machine fault diagnosis. The plan includes three research tasks. First, we will investigate key factors influencing the domain generalization success from the theoretical aspect. In addition, we will advance the interpretability of data-driven modeling for machine fault diagnosis domain generalization via incorporating the causal learning concept. Finally, we will develop more powerful data-driven methods for realizing noise robust domain generalization in machine fault diagnosis.