Model-driven deep unrolling: Towards interpretable deep learning against noise attacks for intelligent fault...

Area of Science:

• Mechanical Engineering
• Artificial Intelligence

Background:

• Deep learning (DL) methods have advanced intelligent fault diagnosis (IFD), but their "black box" nature limits industrial application, particularly in aero-engine IFD.
• Vibration signal-based IFD is susceptible to noise, significantly reducing diagnostic accuracy.
• Interpreting learned features in DL models remains a significant challenge for practical implementation.

Purpose of the Study:

• To develop a model-driven deep unrolling method for interpretable and noise-robust aero-engine IFD.
• To address the limitations of current DL-based IFD methods, focusing on interpretability and noise resilience.
• To propose a novel approach that integrates model-based principles with deep learning for enhanced fault diagnosis.

Main Methods:

• Developed a model-driven deep unrolling method by unfolding a general sparse coding (GSC) optimization algorithm into a neural network (LGSC-Net).
• Introduced a layered GSC (LGSC) algorithm inspired by multi-layer sparse coding (ML-SC).
• Investigated the relationship between the proposed LGSC-Net and traditional convolutional neural networks (CNNs).

Main Results:

• The proposed LGSC-Net demonstrated effectiveness in aero-engine fault diagnosis experiments, including bevel and helical gear faults.
• The method showed robustness against three types of adversarial noise attacks, maintaining high accuracy.
• Interpretability was achieved through the model-driven nature of deep unrolling and its inductive reconstruction properties.

Conclusions:

• The model-driven deep unrolling approach offers a solution for interpretable and noise-robust IFD in critical applications like aero-engines.
• LGSC-Net provides a naturally interpretable alternative to traditional "black box" DL models for fault diagnosis.
• The proposed method enhances the reliability and applicability of IFD systems in challenging industrial environments.