Anomaly Detection and Correction Framework for Pavement Maintenance and Rehabilitation Data Based on LSTM

Pavement maintenance and rehabilitation (M&R) data anomalies critically undermine pavement management system (PMS) reliability, yet conventional artificial neural network (ANN)-based correction methods fail to capture inherent temporal dependencies in pavement deterioration and treatment sequences. This study develops a novel anomaly detection and correction framework using long short-term memory (LSTM) networks to address this gap. The framework first identifies abnormal M&R records through three logical scenarios: (1) unrecorded treatments despite pavement improvement; (2) documented treatments with deterioration; and (3) implausible pavement condition changes given specific treatments (determined via interquartile range bounds). For correction, an LSTM model processes sequential pavement data to infer context-aware M&R actions. A comparative ANN model with identical hyperparameters validates LSTM’s temporal advantage. Applied to 4,100 Shanxi freeway records (2017–2019), the detection module identified 31% anomalies, primarily omissions and physically implausible pavement condition changes. The LSTM correction model achieved 96.83% accuracy, significantly outperforming the ANN (91.87%), with superior recall for key treatments such as hot in-place recycling (100% versus 56.82%) and milling and replacing with one layer (100% versus 20%). Granular error analysis revealed ANN’s systematic misclassifications due to static data processing, while LSTM’s gating mechanisms leveraged historical pavement condition trajectories to minimize errors. Post-correction, anomalies dropped to 2.6%, demonstrating the proposed framework’s practical efficacy. Theoretically, this research establishes LSTM as a paradigm-shifting solution for temporal-aware PMS data cleaning; practically, it enables reliable M&R records for optimized resource allocation and performance prediction.