Dual-input anomaly detection method based on deep reinforcement learning

Abstract

Aiming at the problem of low accuracy of unsupervised learning anomaly detection algorithm, a dual-input anomaly detection method based on deep reinforcement learning was proposed. The proposed model mainly consists of a feature extractor and anomaly detector. Based on the deep reinforcement learning framework, the feature extractor uses a dual-input deep neural network to form the current value network and the target value network, which are used to extract the low-dimensional feature vectors. Based on the 3σ principle, the reward function of reinforcement learning is designed to reward and punish the output results of the model during training. The model was trained only with the normal data, and the extracted feature vector of the normal class was used as the input of the anomaly detector to complete the learning of the detector. During the test, the input anomaly detection was realized based on the dual-input convolutional neural network, and the anomaly detector was completed by learning. To illustrate the generality and generalization performance of the proposed method, four sets of image data and two sets of rolling bearing fault data in different fields were verified respectively. At the same time, the proposed method is applied to the fault detection of a real aero-engine rolling bearing.The results show that the proposed model has high anomaly detection accuracy, which is superior to the current optimal method.