Anomaly Detection Modeling Based on Self-Adaptive Threshold Voting Integrating DBN-LRs

Abstract

Due to the imbalance of positive and negative samples, high data dimension and huge data volume, anomaly detection for real multi-source operation and maintenance data is challenging. Thus, an anomaly detection model is proposed based on integrated learning of Deep Belief Nets. It solves the imbalance problem of positive and negative samples in data set, and makes use of the good feature extraction function of Deep Belief Net to effectively reduce the dimension of the multi-source KPIs data. Combined with Logistic Regression and Restricted Boltzmann Machine, the anomaly detection model is constructed. During the process of integrating multiple weak classifiers, a self-adaptive threshold voting algorithm is put forwards, it integrates multiple weak classifiers and improves the generalization of the model. The approach presented in this paper is evaluated based on a real operational data set. The accuracy of the anomaly detection model is more than 99.23%, and the recall of the model is 99.38%.