A Convolutional Neural Network for Electrical Fault Recognition in Active Magnetic Bearing Systems.

Area of Science:

• Mechatronics
• Machine Learning
• Bearing Technology

Background:

• Active magnetic bearings (AMBs) are complex mechatronic systems with integrated mechanical, electrical, and software components.
• The intricate nature of AMBs necessitates robust fault detection mechanisms to ensure operational reliability.
• Classical rolling bearings lack the complexity and advanced control systems inherent in AMBs.

Purpose of the Study:

• To develop an efficient fault detection method for active magnetic bearings using machine learning.
• To create a fault dictionary based on system signal-derived image signatures.
• To identify and classify common soft electrical faults in AMB position sensors and actuators.

Main Methods:

• Construction of a fault dictionary using image representations of fault signatures derived from system signals.
• Training a convolutional neural network (CNN) to recognize these fault signature images.
• Real-time implementation for computationally convenient fault identification.

Main Results:

• A novel fault detection system was developed, achieving 93% accuracy on a test dataset.
• The method successfully identified seventeen distinct fault classes, including soft electrical faults.
• The system demonstrated real-time capability in determining the failed component and fault type.

Conclusions:

• The proposed machine learning-based fault detection method offers a computationally efficient and accurate solution for AMBs.
• This approach enhances system reliability by enabling timely and appropriate countermeasures.
• The study validates the effectiveness of using image-based fault signatures and CNNs for diagnosing faults in complex mechatronic systems like AMBs.