Big data-driven pavement condition monitoring and anomaly detection using multi-sensor time series

Pavement condition monitoring is essential for road maintenance, but many practical inspection systems still rely on periodic surveys or manual visual checking. Such schemes are difficult to capture short-term pavement changes caused by traffic load, local damage, or environmental disturbance. Meanwhile, smartphone and vehicle-mounted sensors can provide continuous monitoring signals, but their data are often noisy, heterogeneous, and strongly time-dependent. To address this problem, this paper proposes PaveMTS, a multi-sensor time-series framework for pavement condition monitoring and anomaly detection. The framework first aligns and cleans heterogeneous sensor streams, and then learns temporal pavement responses together with cross-sensor correlation patterns. An anomaly score is further constructed by combining reconstruction and prediction errors, so that both gradual deterioration and sudden abnormal disturbances can be detected. Experiments on a public road monitoring dataset show that PaveMTS achieves higher Accuracy, F1-score, and AUC than several classical machine learning and deep time-series baselines. The results also indicate that the proposed method remains relatively stable under missing-data and noisy-data settings. These results suggest that multi-sensor temporal modeling can provide practical support for data-driven pavement maintenance.