Inter-turn short circuit detection in PMS motor using artificial neural network

The PMS motor inter-turn short-circuit detection system using a neural network demonstrates the possibilities of using artificial intelligence in the control processes of AC electric drives at the inverter control level, right in the embedded system. This application is very different from typical applications for neural networks, which today often target the field of image processing. A significant difference lies in the different requirements on the response speed with a much smaller volume of processed data. In the case of well-processed input signals, relatively simple neural network structures, for which a short response time is required, are sufficient for the diagnostics of electric drives.
The prepared software demonstrates the use of the open-source platform TensorFlow for neural networks learning. Subsequent conversion of the neural network into source code for nVidia graphics cards (CUDA code). Finally, the source code is compiled on the target platform. The nVidia Jetson AGX Xavier platform was used to demonstrate inter-turn short circuit fault detection.
The system uses the inverter control algorithm, which is implemented in the AURIX TC397 microcontroller. The control system periodically sends data via the Ethernet interface using the UDP protocol. The data is processed on an external platform. The used solution offers the possibility of implementing a neural network in various platforms. A useful alternative for implementing a neural network may be a TPU or FPGA. In principle, it is also possible to implement neural networks directly in the free cores of the control microcontroller.
Prepared neural networks use advanced data preprocessing algorithms. Convolutional neural networks use detection based on actual motor currents in DQ coordinates. In this case, the detection takes place in average after one to two electrical revolutions. This network uses mixed input data. Some input data is passed directly to the perceptron layers that follow the convolution layers.
The multilayer perceptron network uses filtered voltage and current values for fault detection. The values of voltages and currents are filtered together with the extracted second harmonics, which are closely related to the inter-turn short circuit phenomena. Detection using this method is slightly slower compared to convolutional networks.