An Edge–Cloud Collaborative Framework for Anomaly Detection and Hierarchical Storage in Concrete Bridge Sensor Networks

Distributed monitoring systems generate massive high-frequency data, posing challenges for real-time processing and storage. Existing edge–cloud methods lack a hierarchical mechanism from noise filtering to feature-space anomaly detection and may struggle to identify subthreshold deviations. This paper proposes an edge–cloud collaborative hierarchical data processing and anomaly detection method, validated using field-measured data from two in-service concrete bridges. At the edge, static-threshold screening, persistence checking, and signal-activity analysis are integrated with Median Absolute Deviation (MAD)–Cumulative Sum (CUSUM) preprocessing. In the cloud, a Robust Random Cut Forest (RRCF)-based anomaly-scoring procedure incorporating time-domain, frequency-domain, and differential features generates Collusive Displacement (CoDisp) scores and couples anomaly-event output with differentiated storage. For Bridge 1, MAD–CUSUM preprocessing produced a mean estimated signal-to-noise ratio (SNR) change of +0.503 dB, a mean signal correlation of 0.972, and a mean normalized root-mean-square error (NRMSE) of 0.033. With the sensitivity-selected setting of α=2.0, the RRCF procedure identified 56 and 39 sensor-level anomaly-event sensor-data segments in the Bridge 1 and Bridge 2 datasets, respectively. Controlled fault injection, frequency-shift sensitivity, storage-window sensitivity, baseline comparison, and ablation analyses further characterized robustness and trade-offs; across 36 frequency-shift events, RRCF achieved a micro F1-score of 0.7768 and a window area under the receiver operating characteristic curve (ROC-AUC) of 0.7989. Logical record counts were reduced by 86.765% and 90.046%, respectively, while all flagged samples were retained.