Unsupervised Learning for Machinery Adaptive Fault Detection Using Wide-Deep Convolutional Autoencoder with...

Area of Science:

• Machine Learning
• Artificial Intelligence
• Industrial Engineering

Background:

• Unsupervised deep learning for bearing fault diagnosis in complex industrial settings is challenging.
• Traditional fault detection methods require costly and labor-intensive labeled data.
• Imbalanced datasets pose significant robustness challenges for fault detection models.

Purpose of the Study:

• To propose a novel unsupervised deep learning approach for bearing fault diagnosis.
• To enhance fault detection accuracy and robustness in complex industrial environments.
• To address the limitations of traditional methods requiring labeled data.

Main Methods:

• Developed a Wide Kernel Convolutional Autoencoder (WDCAE) integrated with a Large Kernel Attention (LKA) mechanism.
• Incorporated an adaptive threshold module utilizing a Multi-Layer Perceptron (MLP) for dynamic threshold adjustment.
• Validated the model on the CWRU dataset and a customized ball screw dataset.

Main Results:

• Achieved an average diagnostic accuracy of 90.29% on the CWRU dataset and 72.89% on the customized ball screw dataset.
• Demonstrated remarkable robustness under imbalanced data conditions.
• Outperformed traditional and state-of-the-art methods, reducing training time by 10-26% and improving accuracy by 5-10%.

Conclusions:

• The WDCAE-LKA model offers a robust and effective unsupervised solution for intelligent fault diagnosis.
• The proposed method significantly improves diagnostic accuracy and model robustness in industrial applications.
• This approach alleviates the need for extensive labeled data, making fault diagnosis more practical.