HSE scientists train neural network to 'hear' faults in electric motors

Researchers at the AI and Digital Science Institute of the HSE Faculty of Computer Science have developed a new method—the Signature-Guided Data Augmentation (SGDA) framework—that achieves 99% accuracy in motor fault detection and 86% accuracy in fault classification. The application of this approach can reduce industrial equipment repair costs, minimise downtime, and improve production safety. The study results have been published in Engineering Applications of Artificial Intelligence.

Three-phase asynchronous induction motors are the backbone of modern industry. They power pumps, compressors, conveyors, and fans—for example, in metallurgical plants, urban water supply systems, and automotive assembly lines. Therefore, even a minor fault can halt production and lead to significant losses.

Today, engineers detect motor faults by analysing signals from the electrical current consumed by the motor. They examine the frequency spectrum and manually identify characteristic fault signatures. However, this approach requires complex setup and substantial expertise: specialists must spend considerable time processing the signal, selecting the relevant frequencies, and checking various motor parameters. This makes the process labour-intensive and relatively slow.

Another approach is to use machine learning algorithms. However, training these models requires data on how faulty motors behave. In real industrial settings, such data is scarce, so the algorithms simply do not have enough samples to learn from.

A team of researchers from the HSE Faculty of Computer Science—Artem Ryzhikov, Saraa Ali, Alexandr Khizhik, Stepan Svirin, and Denis Derkach—has proposed a solution: they trained an algorithm to introduce artificial faults into the signal of a functioning motor. To do this, specific frequencies—those that typically appear during real faults—are added to the signal.

This allows the neural network to learn to recognise defects automatically. As a result, the lengthy manual search for fault-related frequencies can be replaced with fast automated diagnostics that achieve near-perfect accuracy.

'The neural network is fed artificial yet realistic fault samples and learns to recognise them. Since our method is grounded in the physical laws governing motor operation, it does not require complex computational models or experiments involving real equipment failures,' explains one of the study authors, Denis Derkach, Head of the Laboratory of Methods for Big Data Analysis at the HSE FCS AI and Digital Science Institute.

The solution developed by HSE scientists, called Signature-Guided Data Augmentation (SGDA), was tested on datasets collected from two motors. In the task of determining whether a motor was in good condition, the accuracy reached 99%. In a more complex task—distinguishing between different types of faults—the accuracy was 86%.

'We train the system on data from normal motor operation and then obtain a fully functional fault-diagnosis tool. This approach is especially useful for enterprises that lack archives of failure data or experience in dealing with equipment malfunctions,' says Saraa Ali, postgraduate student at HSE FSC and one of the study authors. 

An advantage of this method is that it can be applied to motors with completely different specifications. It is sufficient to record the normal operating data of a particular motor, and the system will then be able to detect any anomalies.

The solution can detect motor faults in advance, even before equipment failure occurs. This will reduce repair costs, minimise downtime, and improve production safety. In the future, the scientists plan to validate the method on additional motors and test it under real-world industrial conditions.

The study was supported by a grant for AI Research Centres from the Russian Ministry of Economic Development. The method is patented until 2044.