Researchers from the Institute of Artificial Intelligence and Digital Sciences at the Faculty of Computer Science of HSE University have created a method that makes it possible to automatically detect faults in three-phase induction motors. The accuracy of determining the presence of a failure is 99%, while the classification accuracy for the type of fault is 86%. The development is patented until 2044, and the results were published in the journal Engineering Applications of Artificial Intelligence.
Three-phase induction motors are widely used in industry: they drive pumps, compressors, conveyors, and ventilation systems at metallurgical plants, water supply facilities, and automotive manufacturing sites. The failure of such equipment can lead to production line shutdowns and significant economic losses.
The traditional diagnostic approach is based on analyzing the electrical current signal consumed by the motor. Specialists manually study the frequency spectrum, identifying characteristic signs of faults. The process requires high qualifications, complex equipment setup, and significant time costs. An alternative option, the use of machine learning algorithms, faces a lack of sufficient data on real failures needed to train neural networks.
The neural network receives artificial but realistic examples of failures and learns to recognize them. At the same time, our method relies on the physical laws governing the motor's operation and does not require complex computer models or experiments with real equipment faults.
The developed approach was named Signature-Guided Data Augmentation (SGDA). The effectiveness of the method was confirmed during tests on two motors. The method can be applied to motors with various technical characteristics. To configure the system, it is sufficient to record the operating signal of a specific unit in normal mode. Early fault detection makes it possible to reduce repair costs, shorten downtime, and improve production safety.
Read more on the topic:
- Drones to Get Super-Fast AI: New Type of Neural Networks Created
- Neural Network That Will "Resurrect" Ancient Icons Is Being Created at RTU MIREA
- Ministry of Defense Patents Ground Drone with Neural Network Capable of Operating Without Operator Under Electronic Warfare