An intelligent monitoring framework for crack anomaly detection and early warning in concrete bridges using GAF-based classification and GRU prediction

Crack development in concrete is a crucial indicator for the safety assessment of bridge structures, as abnormal crack behavior often reflects underlying structural deterioration. Under normal service conditions, a strong negative correlation typically exists between crack width and ambient temperature. This study presents an intelligent framework for concrete bridge crack monitoring by establishing a temperature–crack correlation model. Temperature-induced crack features are first extracted using wavelet packet decomposition and transformed into Gramian Angular Field images. These are input into a convolutional neural network for automated classification, which demonstrates high accuracy (98.67%) in identifying anomalous channels deviating from group-level behavior. This step facilitates the detection of sensor faults and localized structural anomalies, thereby ensuring reliable inputs for further modeling. Subsequently, a gated recurrent unit-based model is developed for crack width prediction, achieving R2 scores of 0.9637 and 0.9310 on the training and test sets, respectively. To enhance anomaly detection, a residual-based dynamic threshold warning method is introduced, maintaining a low outlier rate (0.56%–1.86%) on typical datasets, and robust performance (below 6%) even with significant noise. Field monitoring results confirm the robustness and practical applicability of the proposed method. Overall, this framework offers an effective approach for early warning of potential safety hazards in concrete bridges, contributing to improved operational safety and extended service life.