euro-pravda.org.ua

A neural network has been trained to quickly and accurately adjust the operating modes of electric motors.

Electric motors, particularly direct current (DC) motors, are utilized in the operation of elevators, presses, metal-cutting machines, and various other electromechanical systems. These motors convert incoming electrical energy into mechanical rotation. Special electronic controllers manage critical parameters such as current, speed, and position, ensuring they remain stable and within desired limits without deviations. However, there are instances where these controllers fail to deliver the necessary precision and speed, resulting in delayed responses to changing conditions, jerks, jolts, and unstable operation of the mechanism. Consequently, neural networks are increasingly being integrated into such systems. Researchers at Perm Polytechnic University have developed an effective training approach for these networks, enabling fine-tuning of the controller and minimizing the occurrence of errors.
Нейросеть обучили быстро и точно переключать режимы работы электродвигателя.

The article is published in the journal "Electrical Engineering". The research was conducted as part of the strategic academic leadership program "Priority 2030".

Elevators, presses, metal-cutting machines, and other manufacturing mechanisms are powered by direct current motors. These processes are managed by an electronic controller. If improperly configured, it may not respond quickly and accurately to changes in operating conditions, leading to delays, jolts, and increased wear on mechanical parts and bearings.

All of this can result in inaccurate processing of parts on metalworking machines, uneven material feeding on conveyors, uncomfortable movements, and even pose risks to passengers in elevators. For precise control of such objects, regulators based on neural networks can be employed.

Neural networks can be trained in various ways: with a "teacher" and without. In the first case, they are taught to simply "mimic" an already configured regulator. The downside is that they won't be able to perform better than their prototype. In the second case, the network independently analyzes the input data fed into it and looks for patterns. However, this presents another challenge: selecting the right data sample for training can be difficult.

Scientists at Perm Polytechnic have developed an original approach where the neural network is trained not on the actual industrial object, but on its "digital twin." For this purpose, specialists from the enterprise work together with scientists to create special simulation models that describe the movement and interaction processes of various parts of the mechanism, including the motor and its elements, using mathematical equations. This allows for the tuning of the controller without disrupting the conditions of the real production process.

Training always occurs through the method of "trial" and (importantly) "error": in the early stages, AI does not know which action will be correct, so it simply tests random parameters and attempts to apply them to the system. If they turn out to be incorrect, it will adjust them and try again. However, in real production conditions, allowing it to experiment in this way and disrupt the workflow is not feasible, as it could lead to emergency situations. Therefore, using a model is a way to train the regulator more delicately on a large amount of diverse data.

“In our approach, we apply a ‘loss function’ that evaluates the discrepancy between the model's predictions and the actual values collected from the existing object. This provides a way to more finely ‘explain’ to the neural network what we want it to achieve. Importantly, unlike traditional neural regulator operations, in our scheme, this function is not embedded in the AI; it acts as an ‘external observer’ that compares the network's forecast with the truth and reports how well the network performed,” comments Dmitry Dadenkov, Associate Professor of the Department of "Microprocessor Automation Means" at PNIPU, PhD in Technical Sciences.

Polytechnic engineers tested the training process using a speed regulation system in a motor. They created stringent conditions: the neural network had to regulate the speed of rotation, first, when it changed suddenly and unpredictably, and second, under varying loads, which is the resistance the motor must overcome to rotate.

“Such conditions can arise in machines, conveyors, or emergency situations when it is necessary to quickly switch speed modes or even stop operation altogether. This requires the device to respond quickly and accurately. Tests showed that the regulator trained according to our scheme functions correctly: when the load changes, the motor's speed hardly drops, and when needing to set a different speed, there is only a slight overshoot – about one percent. To check the neural regulator's performance in real conditions, ‘noise’ – random incorrect data – was superimposed on the measured state of the object. The regulator trained on the model without noise successfully managed speed control even on the noisy object,” explains Igor Schmidt, Associate Professor of the Department of "Microprocessor Automation Means" at PNIPU, PhD in Technical Sciences.

The application of such regulators is not limited to direct current motors; they are meaningful wherever classic regulators struggle: if the control object is a complex nonlinear, interconnected system, or if there are additional quality control criteria.

The approach developed by scientists at Perm Polytechnic offers virtually unlimited opportunities for fine-tuning the neural regulator. Additionally, by obtaining information about factors that may lead to errors, the neural network can preemptively prevent their occurrence. This allows for effective management of processes in electric drive systems for elevators, conveyors, metal-cutting machines, rolling mills, and lifting and transport machines.