Matras English – The scientists from the University of Lyon, Jules Verne University of Picardy in France, and the Polytechnic University of Valencia in Spain have offered a new method for monitoring and diagnosing the electric motors using their digital twins. The equipment virtual copies updated in real time using sensors and computational models could become an optimal tool for an early damage detection and repair prediction.
Today, the maintenance of electric motors and generators remains conservative: it is carried out either according to a strict schedule or after the machine has already broken down. Laboratory experiments allow individual defects to be investigated, but they are time-consuming and costly, and, most importantly, they fail to reflect a full picture of equipment operation under real-world conditions. For example, in induction machines, bearing damage, interturn short circuits, or rotor imbalance often go unnoticed until an accident occurs.
According to the research done by the French and Spanish scientists, digital twins could be a solution to this problem, enabling predictive maintenance of their real-world prototypes thanks to modern sensors and machine learning algorithms. These twins constantly compare the behavior of the real machine with a reference virtual model and record the slightest deviations before they develop into serious breakdowns. For example, a digital twin can simulate operation of a motor under different loads, temperatures, and wear conditions to predict when and under what circumstances a failure may occur.
To create such twins, the researchers used different approaches: physical models based on the laws of electromagnetism and mechanics; circuit models with lumped parameters; machine learning methods trained on large arrays of historical and current data. However, the most effective approach was found to be a hybrid approach combining the accuracy of physical calculations with the speed of neural network algorithms.
The experiments have shown that digital twins are capable of monitoring current, vibration, temperature, and magnetic fields in real time, identifying the signs of wear or damage. At the same time, it was not necessary to shut down the equipment or use complex monitoring systems — it was enough to connect the conventional sensors to a cloud platform. In a number of models, it was possible to predict a defect even before it became noticeable through the standard signals. In particular, the use of physically informed neural networks accelerated modeling by 40 times.
The researchers are confident that these technologies will make it possible, in the foreseeable future, to move away from reactive maintenance and get down to intelligent management of the entire life cycle of electrical machines, from commissioning to decommissioning. This will enable the industry to reduce downtime and costs, and the energy sector to improve reliability and safety.
