An Adaptive Self-Configuration Scheme for Severity Invariant Machine Fault Diagnosis

Vibration signals, used for abnormality detection in machine health monitoring (MHM), exhibit significant variation with varying fault severity. This signal variation causes overlap among the features characterizing different types of faults, which results in severe performance degradation of the fault diagnostic model. In this paper, a wavelet based adaptive training set and feature selection (WATF) self-configuration scheme is presented, which selects the optimum wavelet decomposition level, and employs adaptive selection of the training set and features. Optimal wavelet decomposition level selection is such that the maximum fault signature-signal energy bands are achieved. The severity variant features, which could cause detrimental class overlap for MHM, are avoided using adaptive selection of the training set and features based on the location of a test data in feature space. WATF uses Support Vector Machines (SVM) to build the fault diagnostic model, and its performance and robustness has been tested with data having different severity levels. Comparative studies of WATF with eight existing fault diagnosis schemes show that, for publicly available datasets, WATF achieves higher fault detection accuracy, even when training and testing datasets belong to different severity levels.