Unsupervised anomaly detection for gearboxes based on the deep convolutional support generative adversarial network.

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Gearboxes are critical industrial components prone to failures due to anomalies.
• Precise anomaly localization is essential for complex industrial automation systems.
• Existing condition monitoring methods may lack accuracy in detecting subtle gearbox faults.

Purpose of the Study:

• To propose a novel, unsupervised deep convolutional support generative adversarial network (DCSGAN) for gearbox condition monitoring.
• To enhance the precision and reliability of anomaly detection in industrial gearboxes.
• To provide an effective method for early identification of gearbox abnormalities.

Main Methods:

• Utilized high-dimensional gearbox data to train a generator within the DCSGAN framework.
• Calculated reconstruction errors from generated samples to train a one-class support vector machine (OCSVM).
• Applied normalized reconstruction errors to the trained OCSVM for anomaly detection during testing.

Main Results:

• The proposed DCSGAN method demonstrated superior performance in anomaly detection compared to other models.
• Experimental validation on a real gearbox dataset confirmed the effectiveness of the DCSGAN approach.
• The method successfully identified anomalies, indicating its potential for real-world applications.

Conclusions:

• The unsupervised DCSGAN combined with OCSVM offers a robust solution for gearbox condition monitoring.
• This approach significantly improves anomaly detection accuracy in industrial settings.
• The public release of the DCSGAN code facilitates further research and development in gearbox diagnostics.