DCATCN: A temporal convolutional network and cross-attention based UAV sensor anomaly detection method

Abstract

To address the issues of severe high-dimensional redundancy and interference in UAV sensor data, insufficient ability to model temporal dependencies, and anomaly decisions relying on fixed empirical thresholds, this paper proposes a lightweight anomaly detection framework called DCATCN (Dual Cross-Attention Temporal Convolutional Network) based on a bidirectional temporal convolutional network (BiTCN) and a Cross-Attention mechanism. First, the method uses the Maximal Information Coefficient (MIC) to adaptively select a feature subset that is highly correlated with target anomalies, effectively reducing data redundancy; then it constructs a bidirectional temporal convolutional network to extract forward and backward features of the time series data in parallel, introducing a Cross-Attention mechanism to dynamically integrate bidirectional information and enhance the model’s representation of temporal dependencies; finally, it employs Extreme Value Theory to statistically model the prediction residuals and determine the anomaly decision threshold, achieving robust and reliable anomaly detection. Comprehensive experiments on the public ThorFlight93 dataset demonstrate that this method outperforms various mainstream models in both detection accuracy and computational efficiency, showcasing strong potential for engineering applications. Code release: https://github.com/ZCchou/DCATCN.git