Bridge Anomaly Diagnosis Based on a Feature-Augmented BiLSTM Autoencoder and Hotelling T2 Test

For bridges in long-term operation, only limited abnormal samples with clear labels are available, which makes it difficult to apply conventional supervised methods that require sufficient labeled data. To address this problem, this study proposes a bridge anomaly diagnosis method combining a BiLSTM autoencoder with the Hotelling T2 test. Multi-point vehicle-induced deflection signals are divided into samples using fixed time windows, and time-domain, frequency-domain, and energy features are extracted. Two BiLSTM autoencoder schemes, feature preprocessing and feature augmentation, are then developed. In the feature augmentation scheme, the original synchronous vehicle-induced deflection sequences are retained, while the statistical features are introduced as supplementary information. Healthy samples are used to learn the healthy reference pattern, and multidimensional reconstruction errors are further evaluated using the Hotelling T2 statistic. A case study is conducted using measured data from the G346 Grand Canal Bridge, with two types of controlled response-deviation samples constructed from normal monitoring data. The feature augmentation scheme achieves a maximum response-deviation-to-normal reconstruction error ratio of 4.968 and provides clearer reconstruction-error separation than the feature-only preprocessing scheme. At the 95% confidence level, recall values are 95.8% and 91.0%, precision values are 93.9% and 93.6%, and accuracy values are 94.8% and 92.4% for Type 1 and Type 2 samples. The results demonstrate the potential of the proposed framework for detecting controlled response-level deviations in multi-point bridge monitoring data, while further validation using longer-term monitoring data and naturally occurring abnormal responses is still required.