Open Set Intrusion Event Recognition Using Anchor Point Learning for Distributed Optical Fiber System

Distributed optical fiber sensing systems are widely used in the field of security monitoring for their electromagnetic interference resistance and high sensitivity. Due to the complexity of the environment, unknown signals often arise during the monitoring process to interfere with the recognition model. However, the vast majority of existing research focuses on closed-set identification, which ignores the fact of open-set environments, and thus fails to effectively meet the needs of security monitoring in real scenarios. In addition, existing open-set identification methods as well as the traditional Softmax method are unable to control the orientation of convergence of the samples in the output space during the training stage, which increases the risk of open-set space. To tackle this problem, this article proposed a novel deep learning model based on convolutional neural networks called anchor point learning (APL) for improving classification robustness and reducing the distributional overlap between known and unknown category samples. With APL, different categories will be divided into different regions of the output space, which can be directly determined by the anchor points. Besides, we add an adversarial regularization term to improve the intraclass compactness of the feature representation by forming an adversary with the APL. In our experiments, our model is applied to the task of recognizing pavement intrusion signals. We compare it with the traditional Softmax method and the existing state-of-the-art open-set identification methods respectively. The results confirm that the proposed method not only outperforms the traditional methods in terms of classification performance but also can effectively achieve the recognition of intrusion events under low false alarm rate conditions in the open set environment.