Data set preprocessing methods for the artificial intelligence-based diagnostic module

摘要

The paper presents the application of statistical (econometrics-originated) methods to process learning and testing data sets used by the artificial intelligence (AI) methods in the diagnostics of analog systems. Before the training and evaluation of the intelligent module is performed, the measurement data are analysed to minimize the number of attributes (symptoms) required to distinguish between different states of the System Under Test (SUT). This way the knowledge extracted from the set is simplified, increasing the operation speed and minimizing the threat of overlearning. Also, elimination of unnecessary symptoms from the set allows for decreasing the set of test points where measurements are taken (which is economically desirable). Preprocessing operations include elimination of constant or quasi-stationary symptoms and finding their minimal set, allowing for the efficient fault detection or parameter identification. The paper focuses on the Hellwig and Multiple Correlation Coefficient methods adjusted to the technical diagnostics applications. They are implemented to optimize data sets obtained from simulation of the fifth order lowpass filter. Their usefulness is tested using the artificial neural network (ANN) and Rough Sets (RS) classifiers responsible for detection, and identification of parametric faults.