Two-Stage Anomaly Detection in Super High-Rise Fire Sensor Networks Using Low-Rank Completion and RV-GAN

Abstract

With the increasing intelligence of fire protection systems in super high-rise buildings, fire monitoring schemes based on sensor networks have raised higher requirements in terms of perception accuracy and response efficiency. However, complex spatial environments and instability in the performance of sensing equipment often led to anomalies such as missing, drifting, and distorted data collection. To address these challenges, this study develops a two-stage anomaly detection framework based on low-rank completion and generative adversarial learning. In the first stage, the Correlation-Regularized Schatten-p Precompletion (CRSP-Pre) model leverages feature correlations and low-rank constraints to structurally repair abnormal or missing samples. In the second stage, a Residual-Variational Generative Adversarial Network (RV-GAN) is designed to capture high-order and composite anomalies through variational generation and residual discrimination mechanisms. Experimental results show that when the data missing rate reaches 50%, the mean absolute error (MAE) and structural similarity index (SSIM) are 0.131 and 0.671, respectively-representing improvements of approximately 15% and 12% over conventional completion and single-stage detection methods. When the number of nodes is 40, the false alarm rate of the proposed method is 5.4%, the model parameter size is 7.8M, and the inference time is 22.4 ms. The results show that the method has good robustness and engineering adaptability, effectively enhancing the processing capability and response reliability of fire monitoring systems for abnormal data. It provides a technical path for intelligent management of sensing data in complex building environments.