An accurate and low-cost abnormal data detection framework for structural health monitoring using hybrid neural network

Continuous bridge monitoring generates long-duration, noisy, and multi-channel acceleration streams in which abnormal events are sparse yet safety-critical. This work targets reliable event classification under operational conditions while preserving the time-series nature of the data. A hybrid one-dimensional convolutional model with channel-wise recalibration and attention analyzes fixed-length windows from a cable-stayed bridge monitoring system, combined with peak picking, batching, and lightweight inference that suit day-scale records. Evaluation on a held-out test set shows balanced class performance: precision across classes ranges from 90.5% to 98.4%, and recall across classes ranges from 91.4% to 99.6%, yielding precision of 94.6% and recall of 94.9%. At the sensor level, classification accuracy ranges from 93.6% to 99.7%, indicating consistent behaviour across heterogeneous channels with different noise and placement conditions. The dataset is split into training, validation, and test subsets, and the test set remains untouched during model selection. The results suggest robust detection across classes and sensors, stable operation on continuous 24-h records, and practical readiness for field deployment in bridge monitoring.