Implementation of MLP-Based Classifier of Current Sensor Faults in Vector-Controlled Induction Motor Drive
Krystian Teler, Maciej Skowron, Teresa Orłowska-Kowalska
Abstract
This article deals with the neural networks (NN)-based classification of current sensor (CS) faults in an induction motor drive system designed according to the idea of fault-tolerant control (FTC). The CS-FTC system ensures continued system functionality, even after a failure has occurred, not only through its ability to detect and isolate faults but also through its ability to compensate for them. The article focuses on the development of the technical implementation of a CS faults classifier based on NN in considering the problem. The proposed method uses the principle structure of a multilayer perceptron (MLP) based on single samples of the stator phase current signal. The tests demonstrate techniques that allow the development of NN input vectors while eliminating the constraints resulting from the machine's operating conditions (load change, rotational speed change). In addition, the possibility of optimizing the input vector of the network in terms of the analyzed damages is demonstrated by using harmonic analysis. The article shows the possibility of full use of the mathematical model in the practical implementation of diagnostic systems implemented on a real object. MLP aims to determine the CS fault category (no fault, gain change, offset, saturation, open circuit) and fault localization (phase A or B). The experimental verification on a laboratory stand with an induction motor of 1.1 kW confirmed the efficiency of the proposed method in 97.2% of the cases.