A novel deep-learning model for detecting small-scale anomaly temperature zones in RDTS based on attention mechanism and K-Means clustering

With the increasing integration of RDTS technology into disaster monitoring systems, such as pipeline leak detection and fire surveillance, promptly and precisely identifying small-scale anomaly temperature zones characterized by short lengths and low temperature variations within RDTS data is crucial for effective early warning systems. Current anomaly detection algorithms for RDTS, including PCA and CNN-based approaches, are typically designed to identify large-scale anomaly temperature zones, which exceed spatial resolution and exhibit temperature values significantly above room temperature. To address this gap, we have introduced an innovative RDTS anomaly detection model that incorporates global and local feature extraction modules, a multi- head cross-attention fusion module, a self-attention module, and an AR module. Additionally, we developed a label generation method based on K-Means clustering that adaptively generates labels using anomaly scores. We collected four distinct types of RDTS data with varying temperature zone distributions and conducted performance evaluation experiments on our model. On test dataset, our proposed model achieved a peak F1 score of 0.772, which improved to 0.832 after employing the K-Means clustering-based label generation method. These findings demonstrate that our model possesses superior capability in detecting small-scale abnormal temperature zones in RDTS data. Moreover, the proposed K-Means clustering approach for data label generation significantly enhances the model's detection performance. The refined model consistently performs anomaly detection tasks on RDTS data with temperature zone lengths equivalent to or greater than the sampling interval (40 cm) and holds potential for widespread application in RDTS-based disaster monitoring scenarios.