Identifying anomalous monitoring data caused by unstable sensor performance is crucial for accurately assessing the operational condition of bridges. In practical monitoring, such anomalies often exhibit various complex patterns, such as slow-varying trends and missing data. However, traditional analysis methods based on unimodal data features struggle to simultaneously consider both the transient dynamics and the global evolutionary features of time-series data, which leads to insufficient identification capability for slow-varying anomalies such as drift and trend. To address this, a framework for diagnosing anomalous bridge data based on multimodal data feature fusion is proposed, which achieves fine-grained identification of complex anomaly patterns by fusing Markov Transition Field (MTF) image features with one-dimensional (1D) time-series features. This fusion dynamically combines features from two parallel branches: one branch extracts global state transition patterns from the MTF images, while the other captures key transient dynamics from the 1D time-series data. Experimental results show that the method achieves an overall mean Average Precision (mAP) of 99.83% on the main girder strain monitoring data from a highway cable-stayed bridge (across seven data classes), representing a significant improvement compared to models using only unimodal data features, with the image-only model achieving 94.63% and the time-series-only model achieving 91.34%. Notably, the F1-Scores for minority slow-varying anomalies (trend, drift) are improved by over 15%. Furthermore, the model demonstrates strong generalization, achieving 97.97% accuracy on a large-scale dataset collected from sensor locations that were used during training.
Chen et al. (Thu,) studied this question.