.png){kind=icon}
.png){kind=icon}
Return
A sequence-to-sequence model for joint bridge response forecasting
.png){kind=icon}
OA
AI
DOI：10.1016/j.ymssp.2023.110690.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Knowledge of the structural response of bridges is extremely important for highway asset management and bridge structural health monitoring. Instrumenting every bridge in a road network and maintaining the monitoring instrumentation over decades of service can be financially infeasible. Mechanical intuition suggests a significant relationship exists between responses of two sets of bridges reasonably similar in design exposed to an identical load. This study explores the use of data-driven models to forecast the response of one bridge to a given truck load using the response of another bridge to the same loading profile. By deploying a modern monitoring system in multiple bridges in the same highway corridor integrated in a cyber-physical systems (CPS) framework, and utilizing advanced computer vision algorithms, the authors have gathered a unique dataset consisting of pairs of bridge responses to the same truck load from live traffic moving across a 32.2 km (20 miles) stretch of the I-275 highway in southeast Michigan. Signal processing techniques have been employed to isolate the response of the bridges to a single truck load in a time series of recorded responses. Then, a deep-learning-based time series forecasting framework using the encoder-decoder architecture with gated recurrent unit (GRU) and long short-term memory (LSTM) cells has been used for bridge response forecasting. Baseline models based on linear time series models are also developed to which the deep-learning forecasting models can be compared. After training the models, it is observed that deep-learning-based models can accurately forecast the response of one bridge using the response of another and reduce the forecasting root-mean-squared error (RMSE) by at least 20% relative to baseline linear models. The forecasting capabilities of the encoder-decoder architecture proposed herein outperform traditional approaches to response forecasting. Trained versions of the encoderdecoder forecasting model can be used to provide reliable estimates of bridge response using a single instrumented bridge in a corridor, thereby enhancing the value of data from instrumented bridges for asset management of bridge networks.
Keywords:
Time Series Forecasting
Bridge Monitoring
Encoder-Decoder
Recurrent Neural Networks
Deep Learning
Sequence-to-Sequence
Journal
IF:
8.9
Papers:
1.3W
Citations:
6.6W