Unsupervised Feature-Construction-Based Motor Fault Diagnosis.

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Bearing faults are a primary cause of motor damage, leading to significant economic losses.
• Early and precise fault identification is crucial for preventing catastrophic equipment failure and maintaining industrial process continuity.
• Motor vibration signals offer a viable data source for diagnosing bearing health conditions.

Purpose of the Study:

• To propose a novel bearing fault detection method utilizing motor vibration data.
• To enhance the performance of bearing fault detection, particularly for challenging fault types.
• To leverage advanced neural network architectures for improved diagnostic accuracy.

Main Methods:

• The study employs a two-stage neural network approach: an autoencoder for feature extraction and a feed-forward network for classification.
• An autoencoder neural network is utilized to construct a new, informative motor vibration feature.
• A feed-forward neural network performs the final fault detection based on the enhanced features.

Main Results:

• The proposed method significantly improves the performance of the feed-forward neural network in bearing fault detection.
• The constructed signal feature effectively enhances prediction performance, especially for difficult-to-detect fault types.
• Experimental results on the CWRU bearing datasets demonstrate superior performance compared to other machine learning algorithms.

Conclusions:

• The developed neural network-based method provides an effective solution for accurate and rapid bearing fault detection.
• This approach contributes to preventing motor damage and ensuring uninterrupted industrial operations.
• The study highlights the potential of combining autoencoders and feed-forward networks for advanced condition monitoring.