Winding Short-Circuit Diagnosis of Permanent Magnet Synchronous Machines Using Artificial Intelligence
CEME Collaborator Professor Julia Zhang and Michael Nye – The Ohio State University
The goal of this project is to implement machine learning techniques to predict the location of inter-turn shorts within a permanent magnet synchronous machine (PMSM). We decided to frame the problem using a neural network. First we use time-series data of simulated PMSMs as inputs to the developed neural network in order to classify which windings in the machine are experiencing the short-circuit fault. This includes information such as three-phase currents, dq-axis currents, measured torque output, and other data as necessary. While there are many types of neural networks, some are better suited for time-series classification problems, including recurrent neural networks (RNNs), temporal convolution neural networks (TCNNs), and long short-term memory networks (LSTMs).
The data are collected from simulations using MATLAB Simulink. We design a magnetic circuit of a surface PMSM. Figure 1 shows an 8-pole, 24-slot surface PMSM with concentrated windings used in this study. Each stator tooth has 12 turns under a healthy condition. The magnetic circuit is then coupled with a control system that controls the three-phase currents of the SPMSM’s windings. We can then selectively short-out certain turn(s) of any of the windings to model the effect of a fault.
Figure 2 shows the magnetic circuit of the SPMSM. Rag,Rtg,RT and Ry are the reluctances of the air gap, tooth gap, tooth, and yoke, respectively; all are dependent on the machine’s physical parameters. Using the Simulink model, simulations are run at various combinations of dq current commands, at different angles and magnitudes. These simulations are for both a healthy machine and a machine with various inter-turn short locations between windings. Data from the simulations are collected to use for the neural network later on.