Improving Anomaly Detection Fusion Method of Rotating Machinery Based on ANN and Isolation Forest

Abstract

Early anomaly detection of rotating machinery, which is used to avoid major machine failures, has been gradually concerned. In many industrial scenarios, failure samples are difficult or expensive to obtain, so that anomaly detection methods show great advantages. In order to solve the problems caused by imbalanced samples in model training in industrial applications, the paper proposes an anomaly detection fusion method based on artificial neural network (ANN) and isolation forest (iForest) algorithm. The original data set is extended by upsampling to keep the number of different kinds of samples consistent. Then, the statistical features of each sample and a special energy feature are obtained through the Empirical Mode Decomposition (EMD) method. A feature data set will be combined. In order to better represent the intrinsic characteristics of the signal, the proposed method trains an ANN network to extract hidden features and reduce the dimensions of the feature data set. Finally, the extracted hidden features are fed to the isolated forest algorithm for anomaly detection. The proposed method is tested on the CWRU bearing datasets. The results show that the proposed method has better performance than the traditional method.