One-Dimensional Multi-Scale Domain Adaptive Network for Bearing-Fault Diagnosis under Varying Working Conditions

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Data-driven bearing-fault diagnosis methods are crucial but often fail under varying working conditions due to differing data distributions.
• Existing methods require identical data distributions and large labeled datasets, which are impractical in real-world scenarios.
• The challenge lies in developing diagnostic tools that maintain accuracy despite changes in machine operating environments.

Purpose of the Study:

• To propose a novel deep transfer learning model, the one-dimensional Multi-Scale Domain Adaptive Network (1D-MSDAN), for robust bearing-fault diagnosis.
• To enable accurate fault diagnosis under varying working conditions without requiring identical data distributions or extensive labeled data for each condition.
• To enhance the adaptability and generalizability of bearing-fault diagnostic systems in dynamic industrial settings.

Main Methods:

• The proposed 1D-MSDAN utilizes a deep transfer learning approach combining multi-scale and multi-level feature adaptation.
• Domain-invariant features are learned by minimizing distribution discrepancies using Multi-kernel Maximum Mean Discrepancy (MK-MMD).
• Classifier adaptation is achieved through entropy minimization to align source and target domains, further reducing domain discrepancy.

Main Results:

• The 1D-MSDAN demonstrated superior diagnostic accuracy across 12 transfer tasks on the CWRU bearing database compared to mainstream transfer learning models.
• The model effectively handles varying working conditions, a common challenge in practical machine diagnostics.
• Validation confirmed the model's strong transfer learning performance for multi-target domain adaptation and its applicability to real industrial scenarios.

Conclusions:

• The 1D-MSDAN offers a powerful solution for data-driven bearing-fault diagnosis under variable working conditions.
• The combined feature and classifier adaptation strategies effectively bridge domain gaps, enhancing diagnostic reliability.
• This approach significantly advances the practical implementation of intelligent fault diagnosis systems in industry.