Fault diagnosis for pump equipment based on transfer learning: the domain generalization method

Area of Science:

• Machine Learning
• Artificial Intelligence
• Signal Processing

Background:

• Conventional data-driven fault diagnosis requires identical training and test data distributions, limiting its application in engineering.
• Inconsistent data distributions between source and target domains pose a significant challenge for accurate fault diagnosis.

Purpose of the Study:

• To propose a novel fault diagnosis method, FOCAL-TFAM-ResNet-LSTM, capable of handling inconsistent data distributions.
• To enhance the accuracy and robustness of machine learning-based fault diagnosis in engineering applications.

Main Methods:

• Developed a mapping-based deep transfer learning method incorporating a feature-oriented correlation alignment layer (FOCAL) for nonlinear distributional alignment.
• Introduced an adaptive metric function with gradient dynamic weight optimization to balance domain adaptation and classification accuracy.
• Integrated a time-frequency attention module (TFAM) into a ResNet-LSTM network to mitigate gradient disappearance in long-term sequence training.

Main Results:

• The FOCAL-TFAM-ResNet-LSTM method achieved 96.83% classification accuracy and a loss of 0.1024 on a reciprocating pump dataset under adverse conditions (SNR of -5 dB).
• Demonstrated a significant improvement over conventional domain alignment techniques like MMD and MK-MMD, with an 8.42% gain in accuracy and a 0.3594 reduction in loss.
• Validated the algorithm's effectiveness using reciprocating pump and Case Western Reserve University datasets.

Conclusions:

• The proposed FOCAL-TFAM-ResNet-LSTM method effectively addresses the challenge of inconsistent data distributions in fault diagnosis.
• This novel approach offers superior performance compared to existing methods, particularly in complex and noisy engineering environments.
• The study highlights the potential of deep transfer learning and attention mechanisms for robust fault diagnosis systems.