Bearing Fault Diagnosis Using Weighted K-Nearest Neighbor

The rolling element bearings are essential components in rotating machines. Their condition can directly affect the machine operations. In this paper a new methodology for bearing fault diagnosis is proposed using weighted K nearest neighbor Classifier (WKNN). In WKNN, a squared inverse feature weighting technique is used to improve the performance of K-NN classifier. The proposed method uses simple time domain features to classify bearing condition using vibration signal. Multiple WKNN based classifiers are tested for optimizing the classification accuracy and computation complexity. Principle component analysis (PCA) is used to reduce the dimensionality of data set. Performance of WKNN for different K values and different distance metrics are compared with PCA and without PCA. Three bearing conditions, namely healthy, inner race fault and outer race fault were classified. The experimental results indicate that this method enables the fault detection in bearings with high accuracy. An accuracy of 100% was achieved using Mahalanobis distance metric.