Real Time Neural Classifiers for Sensor Faults in Three Phase Induction

Induction motors can be modeled in different ways for correct operation and control, one of these is the $\alpha $ - $\beta $ representation, this model has six state variables that can be monitored: rotor position, rotor speed, $\alpha $ flux, $\beta $ flux, $\alpha $ current and $\beta $ current. Usually, only three of these variables can be measured directly with sensors. These sensors are subject to long periods of work and stress, so a failure in these sensors cannot be ruled out. Sensor failure can cause problems to control the motor, instability or motor performance degradation. That is why fault tolerant controllers are proposed to maintain the stability of the induction motor despite sensor failure, assuming that the error is classified correctly and in a short period of time. This paper is concerned with the detection and classification of sensor faults: rotor position, $\alpha $ and $\beta $ currents, in real time, considered faults can occur by sensor disconnection, sensor degradation, sensor failure, or connection damage, among other hardware or software phenomena. Different neural networks are proposed and compared for real-time classification, these are: Multilayer perceptron, convolutional neural network, the unidirectional Long short-term memory (LSTM) and bidirectional LSTM. The results show that the CNN neural network presents the best performance compared to the other methods, but the LSTM has a shorter classification time with high accuracy to classify the true class. The CNN used corresponds to a simple configuration of a convolution layer with 20 filters of $2\times 1$ , followed by a pooling layer and two dense layers. The results show that CNN has a classification accuracy above 99% and an average classification time per sample of 4.6236e-08 s. For its part, the LSTM shows a classification accuracy of approximately 99% and an average classification time per sample of 3.1298e-09 s, MLP shows a classification accuracy of 97.96% with a classification time of 5.5 e-10 s, while BiLSTM shows an accuracy above of 98% and a classification time of 4.47e-4 s.