Monitoring the condition of road infrastructure is crucial for maintaining its structural integrity and ensuring safe transportation.This study proposes a deep learning framework based on Temporal Convolutional Networ...Monitoring the condition of road infrastructure is crucial for maintaining its structural integrity and ensuring safe transportation.This study proposes a deep learning framework based on Temporal Convolutional Networks(TCN)integrated with Adaptive Parametric Rectified Linear Unit(APReLU)to predict future road subbase strain trends.Our model leverages time-series strain data collected from embedded triaxial sensors within a national highway,spanning August 2021 to June 2022,to forecast strain dynamics critical for proactive maintenance planning.The TCN-APReLU architecture combines dilated causal convolutions to capture long-termdependencies and APReLU activation functions to adaptively model nonlinear strain patterns,addressing limitations of traditional ReLU in handling bidirectional strain signals(compressive and tensile).Comparative experiments demonstrate TCN-APReLU’s superior performance.These improvements highlight its enhanced accuracy in predicting strain accumulation under cyclic traffic loads,enabling maintenance teams to prioritize interventions 5-7 days before critical thresholds(e.g.,>100με)are exceeded.This work provides a robust data-driven solution for urban road health monitoring,emphasizing scalability through parallelizable convolutions and adaptability to sensor noise.Future extensions will integrate multi-modal data to further generalize predictions across diverse infrastructure scenarios.展开更多
基金Supported by open project fund of National Engineering Research Center of Digital Construction and Evaluation Technology of Urban Rail Transit(2024023).
文摘Monitoring the condition of road infrastructure is crucial for maintaining its structural integrity and ensuring safe transportation.This study proposes a deep learning framework based on Temporal Convolutional Networks(TCN)integrated with Adaptive Parametric Rectified Linear Unit(APReLU)to predict future road subbase strain trends.Our model leverages time-series strain data collected from embedded triaxial sensors within a national highway,spanning August 2021 to June 2022,to forecast strain dynamics critical for proactive maintenance planning.The TCN-APReLU architecture combines dilated causal convolutions to capture long-termdependencies and APReLU activation functions to adaptively model nonlinear strain patterns,addressing limitations of traditional ReLU in handling bidirectional strain signals(compressive and tensile).Comparative experiments demonstrate TCN-APReLU’s superior performance.These improvements highlight its enhanced accuracy in predicting strain accumulation under cyclic traffic loads,enabling maintenance teams to prioritize interventions 5-7 days before critical thresholds(e.g.,>100με)are exceeded.This work provides a robust data-driven solution for urban road health monitoring,emphasizing scalability through parallelizable convolutions and adaptability to sensor noise.Future extensions will integrate multi-modal data to further generalize predictions across diverse infrastructure scenarios.
