An Ensemble of Filters and Classifiers for Fault Diagnosis

Filters, a class of feature selection methods, play a crucial role in mechanical fault diagnosis since they have the ability to improve diagnosis accuracy and reduce computational costs. Most existing fault diagnostic methods only take advantage of a single filter and a classifier, which may lead to a variable performance easily. To overcome this shortcoming, an ensemble of filters and classifiers approach is proposed for fault diagnosis. In the proposed method, six filters, based on distinct metrics, as well as a K-nearest neighbor (KNN) classifier are adopted. Each filter is paired with a KNN to form a base model. For each base model, the filter first selects a subset of features which are then utilized to train and to test KNN. The outputs of six base models are aggregated by a developed weighting technique. The performance of the proposed approach is examined by means of two bearing datasets. Experimental results consistently demonstrate that the proposed an ensemble of filters and classifiers approach significantly outperforms the methods that use a single filter and a classifier on recognizing different bearing fault categories.