Diagnostic intelligent des défauts liés aux arbres tournants

Abstract

The thesis addresses the problem of the classification of mechanical shaft defects in rotating machines. In this context, four operational classes are needed, namely: normal condition (where the machine has no faults), unbalanced (where the load of the machine has its weight unevenly distributed), and misalignment (when rotors) and the machine is dislocated from its natural concentric position) vertical and horizontal. We have explained in this study to combine the use of different signal processing techniques and diagnostic methods by pattern recognition to analyze the severity of faults related to rotating shafts. The proposed approach has been added to a database of faults related to rotating shafts (Machines Fault Database MaFaulDa). A data set consisting of 880 records is used for training and testing the classification system for the four paralyzed classes. Each recording contains 8 channels: a top lap signal, three acceleration signals recorded in the three directions on the bearing on the coupled side and three on the outer bearing, finally an acoustic signal. The form vectors used are composed of scalar indicators: average, kurtosis and RMS and frequency indicators: peaks corresponding to the rotation frequency and its first two harmonics. In total, the vectors formed encompassing 48 indicators calculated on each of the 8 channels. The classifier based on artificial neural networks called Percepton Multi-Layer MLP is used. The good classification rate obtained is 100% for the detection and 99.3% for the identification of defects, which gave better results than those shown inreported in the literature