Determination of factors affecting induction motor equivalent circuit parameters by machine learning methods

Induction motor (IM) has been used in production areas for years. However, today's developing technology and the increasing demands of production areas increase the importance of studies on these motors day by day. There are many studies on early detection and control of faults of IMs. However, the important thing is to prioritize the uninterrupted use of these motors, whose costs are increasing day by day, in production areas. In order to meet the demands, IMs must be used efficiently for a long time without malfunctioning, and for this, mathematical modeling must be done precisely and accurately. The accuracy of the mathematical modeling depends primarily on the accurate determination of the Equivalent Circuit Parameter (ECP) of the IMs under all variable conditions. When the studies in the literature are examined, it is seen that they mostly focus on a single variable operating condition or a limited number of motor powers. It is also a fact that the variation in the runtime parameters of IMs is inadequate when analyzed from a single factor. However, the variability and diversity of operating conditions (motor power, load, rotor speed, etc.) make mathematical modeling of IMs a very complex problem. Therefore, it is necessary to identify the most influential operating conditions. In this way, the complexity of modeling can be eliminated and a simple and accurate analysis can be performed. The main direction of this work is to perform a simple and accurate full analysis by determining the influence ratios of various variable input parameters on ECPs. Thus, bridging various fields related to IM, such as model predictive control (MPC), which is the main area of power electronics research, fault diagnosis and data accuracy in post-failure repair. First, a comprehensive analysis of ECP prediction and fault detection methods is presented, with numerous references to contemporary literature. The ECP method of an IM is described in full detail, covering both the theoretical foundations, computer simulations and experimental application aspects. Challenging and time-consuming experiments were carried out by building the system to obtain an effective motor data stack including different powers, different rotor speeds and different load ratios and are documented in detail in Chapter 3. After the presentation of detailed parameter plots and rates of change in the experiments, the next focus is on the effectiveness of the input variable parameters on the output parameters as the second main aspect of this study. Without the need for all the data of the demanding and time-consuming experiments performed here, the Taguchi analysis method was restricted by coding and experiment restriction. Thus, it is shown that with fewer experiments, approximations to the results obtained in all experiments are obtained by avoiding the difficulties such as cost, time and laboratory environment. The focus is then on the detection of the variable input parameters of an IM and its effectiveness on the output parameters, narrowing the range and reducing the complexity of the mathematical modeling. It was found that power, speed and load have different degrees of influence on each engine parameter. A mathematical modeling is presented by Regression Analysis method. Mathematical modeling is presented by increasing the accuracy ratio (R2) of the mathematical equations created by including winding temperature and slippage in the input parameters. Thus, the problem of difficulty, inaccuracy and complexity in mathematical modeling due to the diversity in input parameters, which is also mentioned in the literature, has been overcome. The last focus, the large amount of parameter data obtained with the experimental setup, led the study to Artificial Neural Network (ANN) as the third main direction for full analysis. A thorough evaluation was carried out during the algorithm selection process for the predictive modeling of the data analyses. In this context, among various machine learning algorithms, it was decided to test the potential of tree-based models in forecasting success. The research and literature review concluded that tree models provide consistent and reliable results, especially in forecasting applications. In neural network applications such as K-nearest neighbor and Random Forest algorithms, the difficulties in the learning process and the lack of generalization ability of the model have come to the forefront in our dataset. While the application of tree-based models alleviated these problems to a large extent, it presented new challenges in terms of computational cost and time, especially as the size of the dataset increased. Although these challenges existed, the improvement in accuracy rates achieved justified this cost increase. As the diversity and volume of our dataset increased, the accuracy of the model steadily improved. With the adjustments and improvements implemented, the efficiency of the model increased significantly. Ensuring an even distribution of data led to a significant increase in the prediction accuracy of the model, while at the same time minimizing imbalances in data representation. Among the tree-based models, the Backpropagation Neural Network (BPNN) algorithm showed high training success rates for the entire data set. The rates of other algorithms were also high and close to each other. This shows that the BPNN algorithm is the most suitable algorithm for engine parameter analysis. When these results are evaluated, it is seen that the mathematical equation to be created for the prediction of IMECPs should be based on the inputs determined as a result of the activity analysis.