ZPW-2000A track circuit anomaly detection based on dual-channel decision fusion

Abstract This paper proposes an anomaly detection method for the ZPW-2000 A track circuit that analyzes massive monitoring data to reflect the circuit’s operational status in real time. The study utilizes two independent parallel channels for anomaly detection: Extended Isolated Forest (EIF) and Variable Auto-Encoder (VAE), each employing a distinct detection mechanism. In order to enhance the training effect, the VAE model’s training dataset is initially refined through the EIF model for data purification. Anomaly detection outputs from the dual channels are mapped to a two-dimensional space, and using Kernel Density Estimation for effective decision fusion. The dual-channel decision fusion model exhibits exceptional performance, achieving a detection accuracy of 98.78%, a precision rate of 97.62%, a recall of 96.48%, and an F1 score of 97.05%, significantly surpassing the performance of the standalone EIF or VAE methods. This technique substantially enhances the accuracy and robustness of anomaly detection. It introduces a fresh perspective and approach to anomaly detection in track circuit.