A deep learning-empowered data anomaly detection method based on a symmetric dot pattern for bridge monitoring data

Abstract

With the widespread applications of bridge structural health monitoring systems (SHMS), massive amounts of abnormal data caused by sensor faults have significantly reduced the reliability of structural health diagnosis. Traditional sensor anomaly diagnosis methods suffer from issues such as strong time-frequency coupling and a scarcity of real fault samples. To address the issue, this study proposes a deep learning-empowered data anomaly detection method based on a symmetric dot pattern (SDP), validated using vibration acceleration data collected from the SHMS of a bridge. First, the SDP algorithm converts one-dimensional signals into two-dimensional polar coordinate snowflake images, enhancing the visual distinguishability of anomaly features through parameter optimization. It is found that insufficient sensor resolution can cause time-series data to exhibit staircase-like anomalies. Next, a sensor failure mathematical model is used to generate a labeled simulated dataset with seven distinct patterns, addressing the scarcity of real anomaly samples. A classification model is then constructed using DenseNet, which performs well on the simulated data, but exhibits insufficient generalization capability when directly applied to real data (average F1-score of 0.82). Finally, a two-stage fine-tuning strategy from transfer learning is introduced, using a small amount of practical labeled data to correct the model, significantly improving the accuracy of practical testing to approximately 96%. Compared to models that directly utilize time-series image data for anomaly detection, this method achieves an F1-score improvement of 0.14, particularly for the staircase-like anomalies. In comparison with short-time Fourier transform and continuous wavelet transform, SDP also demonstrates a distinct advantage in terms of the average F1-score across all classes. The integration of SDP with deep learning and transfer learning methods provides an efficient and universal solution for the preliminary screening of sensor anomalies in the field of SHM, which holds practical engineering value for improving the quality and reliability of SHMS data.