Study Of Bridge Weigh-in-motion Method Based On The Encoder-Decoder Architecture

Bridges play a significant role in the transportation infrastructure.The safety of bridges ensures the orderly operation of the whole transportation system.The efficiency of public transport operations is of great importance to the development of the national economy.In recent years,the growing trend of overloading has resulted in the fact that a large number of vehicles’ actual loads are far beyond the capacity of bridges,leading to lots of bridges suffering from severe disease problems.Monitoring the vehicle load on bridges enables the administration to detect and control overloaded vehicles in time,thus to protect the safety of bridge structures and maintain the order of the transportation system.Bridge weigh-in-motion(BWIM)is a traffic load monitoring technology which estimates the weight of vehicles based on the bridge responses.BWIM systems can effectively and timely detect overloaded vehicles on bridges and issue safety warnings on them.Besides,the obtained statistics of traffic information are also of great sign ificance to the safety assessment of the related bridge structures.Compared with the conventional pavement weigh-in-motion systems,the BWIM systems have some advantages such as convenient installation and maintenance without the need of road damage,high identification accuracy,and long service life.Conventionally,extra axle detectors are required for BWIM systems to facilitate the identification of vehicles’ axle weights.However,problems arise from additional sensors such as strict synchronization,fragile stabili ty,and the limited scope of application,which might plague the implementation of BWIM systems in practice.While some studies have shown that using deep learning methods can estimate axle weights,speed,and some other traffic information of passing vehicles without the need of axle detectors,difficulties in acquiring the training dat a of axle weights impede the development of BWIM systems based on deep learning methods.To overcome these limitations,a novel BWIM method(a deep learning network Bwim Net)based on encoder-decoder architecture is proposed in the thesis,which predicts properties of vehicles(i.e.,velocity,wheelbase,axle weight,and count)merely based on the bridge dynamic strain responses without the requirement of axle weights for the training process.Finite element models of typical bridges and vehicles were established.Numerical simulations based on bridge-vehicle coupled vibration theory were performed to validate the applicability and reliability of Bwim Net,and the axle weight identification performances of the Moses algorithm and Bwim Net were compared.Additionally,the performance of Bwim Net under different road conditions,vehicle lateral positions,sizes of the training dataset,and degrees of the measurement errors in the training data was studied.The results show that Bwim Net can predict the properties of vehicles(i.e.,velocity,wheelbase,axle weight,and count)with good accuracy.Even under the circumstances that large errors exist in the training data,a robust identification performance of Bwim Net can be achieved.Besides,the effect of transfer learning on Bwim Net was also investigated.The results show that the pre-trained models trained by data from other bridges or synthetic data can be effectively reused by Bwim Net,and using transfer learning can further reduce Bwim Net’s requirement for training data and remarkably improve the identification performance and applicability of Bwim Net.In conclusion,a neural network structure(Bwim Net)based on the Encoder-Decoder architecture for bridge weigh-in-motion is proposed in the thesis.The proposed Bwim Net method can effectively detect the properties of vehicles(i.e.,vehicle weight,velocity,and wheelbase)with good accuracy and robustness,showing strong applicability and great potential in the BWIM field.The research contributes to the identification of vehicle loads on bridges and the safety maintenance of bridges.