Explainable Graph Wavelet Denoising Network for Intelligent Fault Diagnosis.

Area of Science:

• Machine Learning
• Signal Processing
• Fault Diagnosis

Background:

• Deep learning (DL) methods enhance fault diagnosis but struggle with noisy signals, signal interactions, and lack of interpretability.
• Existing DL models often fail to extract discriminative features from noisy data and overlook inter-signal relationships.
• The 'black-box' nature of DL hinders the industrial adoption of intelligent fault diagnosis systems.

Purpose of the Study:

• To propose an explainable graph wavelet denoising network (GWDN) for robust intelligent fault diagnosis under noisy conditions.
• To address the limitations of existing DL methods in feature extraction, signal interaction analysis, and interpretability.
• To develop a method that can achieve state-of-the-art performance while maintaining explainability for industrial use.

Main Methods:

• Signals are transformed into graph-structured data to capture interactions between them.
• A novel graph wavelet denoising convolution (GWDConv) is introduced, based on the discrete graph wavelet frame.
• GWDN utilizes GWDConv for multiscale feature extraction on graph-structured data and performs signal denoising.

Main Results:

• The proposed GWDN achieves state-of-the-art performance in intelligent fault diagnosis compared to existing methods.
• Experimental results validate the network's efficacy in handling noisy working conditions.
• Analysis using the square envelope spectrum confirms that GWDConv retains fault-related components and effectively denoises signals, demonstrating GWDN's explainability.

Conclusions:

• The GWDN offers a significant advancement in intelligent fault diagnosis, particularly for noisy environments.
• The method successfully integrates signal interaction analysis and denoising capabilities with enhanced interpretability.
• GWDN provides a promising, explainable solution for industrial fault diagnosis applications.