Multiscale Conditional Adversarial Networks based domain-adaptive method for rotating machinery fault diagnosis under...

Area of Science:

• Artificial Intelligence
• Mechanical Engineering
• Machine Learning

Background:

• Deep learning is vital for rotating machinery health management and maintenance decision-making.
• Real-world applications face challenges like domain shifts and insufficient labeled data, hindering model effectiveness.
• Existing methods often assume consistent feature distributions and ample labeled samples, which are not always feasible.

Purpose of the Study:

• To propose a novel domain adaptive framework, deep Multiscale Conditional Adversarial Networks (MCAN), for machinery fault diagnosis.
• To address the limitations of domain shifts and the scarcity of labeled data in complex industrial environments.
• To enhance the transferability and stability of deep learning models in machinery health management.

Main Methods:

• Developed a shared feature generator using a multiscale module with an attention mechanism to capture rich, scale-aware features.
• Employed Bidirectional Long Short-Term Memory (BiLSTM) based domain classifiers to leverage spatiotemporal features for domain adaptation.
• Incorporated cross-covariance dependencies between feature representations and classifier predictions to improve discriminability.

Main Results:

• The MCAN model demonstrated superior performance in cross-domain transfer tasks compared to state-of-the-art methods.
• Evaluations on public datasets and experimental data confirmed the method's enhanced transferability and stability.
• The attention mechanism improved the model's dynamic adjustment and self-adaptation capabilities across different working conditions.

Conclusions:

• The proposed deep Multiscale Conditional Adversarial Networks (MCAN) framework effectively addresses domain shift challenges in machinery fault diagnosis.
• MCAN offers a significant advancement for deep learning applications in rotating machinery health management and maintenance.
• This approach holds potential to revolutionize future applications by enabling robust fault diagnosis with limited labeled data.