Vibrodiagnostics of gearboxes using NBV-based classifier: A pattern recognition approach

Abstract

Gearbox faults are one of the major factors causing breakdown of industrial machinery and gearbox diagnosing is one of the most important topics in machine condition monitoring. This paper presents a new pattern recognition approach to the condition monitoring of technical objects working under time varying load. The approach shows potential for the fault detection of the high-power planetary gearbox used in bucket wheel excavators.In the presented pattern recognition approach, relations between spectral components of the gearbox vibration signal were investigated in the full range of gearbox operating conditions. A novel Noise-Assisted Feature Subset Evaluation (NAFSE) method addressed for the extraction of diagnostic parameters was introduced. The NAFSE method integrates the feature subset evaluation with the NBV-based classifier and extracts the diagnostic parameter set useful for this classifier. The NBV-based classifier conducted the final recognition of the gearbox condition on the basis of the diagnostic parameters obtained from the NAFSE method.The NBV-based classifier is, in its essence, the condensed 1-NN classifier based on Nearest Boundary Vector algorithm. The elaborated algorithms for determining basic and supplemental boundary vectors together with the original editing procedure of the training set reduction create the original hybrid prototype selection method. The effectiveness of this method has been confirmed in the classification task of the benchmark dataset. In contrast to the traditional hard classifier that assigns only a single-value class label to an investigated pattern, the NBV-based classifier enables the semi-soft classification which offers the possibility of evaluating classification certainty. The offered possibility of evaluating classification certainty has a significant diagnostic meaning. In diagnostic practice it is often not enough merely to recognize the object's condition, but the information about the certainty of the classifier's decision is also necessary.The effectiveness of the proposed pattern recognition approach is illustrated by the fault detection of the high-power planetary gearbox used in a mining machine. It was demonstrated that in order to effectively diagnose machines operating under non-stationary conditions, separate diagnostic relationships at various operating conditions are required. For this reason the extraction of diagnostic parameters was executed for every range of operating conditions separately. This enabled to perform error-free recognition of the gearbox condition including the cases of no load or small load.