A Neural Network-Based Approach to Determining the Mechanical Design Dimensions of Asynchronous Machines

Abstract

It is crucial to figure out the parameters of asynchronous machines, which have an essential function in industry, to guarantee secure operation, control, and analysis. In order to address the time-consuming calculations associated with conventional approaches, the shortcomings in the manufacturer’s documentation, and the interruptions caused by experimental studies, the methods presented primarily were concentrated on determining electrical parameters. However, research concerning the estimation of mechanical parameters was restricted to a minor quantity of parameters and utilized a sample size that is insufficient to establish broad conclusions. Hence, in this research, it is aimed at developing a machine learning-based, high-accuracy, and fast prediction system that surpasses this restricted range. This approach was specifically developed to estimate 17 mechanical dimensions by evaluating three prediction algorithms—Multilayer Perceptron (MLP), Recurrent Neural Network (RNN), and Long Short-Term Memory (LSTM)—to choose the most effective one within a specified power range and enable parameter configuration in less than a minute for practical implementation. The RNN demonstrated the best performance by capturing dependencies effectively and achieving the highest accuracy, while MLP provided rapid results with a simpler structure but limited capacity for modeling complex relationships. LSTM, despite its theoretical advantages, fell short due to high computational demands and inconsistent test performance. A correlation coefficient of 0.99, mean absolute error values below 0.0025, and root mean square error values below 0.0045 were attained throughout the study, thereby signifying a statistically significant relationship between the variables. This research offers a remarkable framework for enhancing the design and operation of machines by improving a parameter determination approach.