Unification of Multi-class Fault Classification from Diverse Domain Features of Gear Using SVM Algorithms

摘要

In the present work, the health monitoring of a gear box has been attempted by the support vector machine (SVM) learning technique with the help of time, frequency and time-frequency (wavelet) vibration data. Four fault conditions in the gear were considered including the no defect case. The multi-fault classification capability of the SVM has been suitably demonstrated and is based on the selection of SVM parameters. Different optimization methods (i.e., the grid-search method (GSM), the genetic algorithm (GA) and the artificial bee colony algorithm (ABCA)) have been performed for optimizing SVM parameters. Time domain vibration signals were measured from the gearbox casing operated in a suitable speed range and was transformed in frequency domain. The continuous wavelet transform (CWT) and the wavelet packet transform (WPT) are extracted from time domain signals. A set of statistical features are extracted from signals in three domains. The prediction of fault classification has been attempted at the same angular speed as the measured data as well as innovatively at the intermediate and extrapolated angular speed conditions, since it is not feasible to have measurement of data at all speed of interest. The classification ability has been noted and compared among all domains, and in general these show excellent overall prediction performances especially with time-frequency domain data. Classifications so obtained are unified for a more reliable prediction by a voting technique. A fault for a particular data point is assigned according to the majority in voting of the fault defined by different methods and by using different domain data. The unified prediction is compared first with the best accuracies obtained using the time, frequency and time-frequency domain data and subsequently with the average of all accuracies obtained using three domains. It is observed that this technique is better than the average perditions obtained from different domains and classification methods, and many times they are competitive with the best fault perdition results.