Dynamic Identification Methods Of Overloaded Vehicles On Long-span Bridges Based On Temporal Convolutional Network And Multi-task Learning

With the rapid development of bridge construction,the importance of bridges in the national economy continues to rise,and the health of bridges is getting more and more attention.The impact of overloaded vehicles is significant.In severe cases,overloaded vehicles might lead to the collapse of the bridge.Therefore,the identification of overloaded vehicles has become a research focus.However,the existing static identification methods will interrupt the traffic flow,while the existing dynamic identification methods are mainly for small-span bridges.There are few studies on the dynamic identification methods of overloaded vehicles on long-span bridges.This thesis proposes a deep-learning overloaded vehicle identification algorithm based on temporal convolutional network,named DOVI.In this algorithm,a convolution layer is used to capture the spatial correlation between sensors while the temporal convolutional network is employed to make full use of the time correlation between displacement response data at multiple adjacent sampling instants.The DOVI algorithm is capable of learning certain information from the displacement response data obtained by structural health monitoring systems to determine if any overloaded vehicle exists on an identification region on the bridge.Compared with traditional dynamic identification methods,the DOVI algorithm does not need the influence line,speed information and axle information,while the occurrence of multiple vehicles is under consideration.The effectiveness of the algorithm is verified on a simply supported beam model and a long-span cable-stayed bridge model.Results indicate that compared with baseline algorithms,the proposed DOVI algorithm is more capable of capturing the spatial and temporal correlation in the time-series data of displacement response and can obtain a better identification accuracy.In order to consider the information that if any overloaded vehicle exists on other identification regions on the bridge,this thesis improves the DOVI algorithm and proposes an overloaded vehicle identification algorithm based on multi-task learning,named M-DOVI.The M-DOVI algorithm utilizes the multi-task learning method to simultaneously determine if there are overloaded vehicles on multiple identification regions on the bridge.In the M-DOVI algorithm,features with more generalization ability can be learned and therefore the generalization ability of the algorithm can be improved.Results demonstrate that compared with baseline algorithms and DOVI,the proposed M-DOVI algorithm is able to achieve a better identification accuracy.Concerning the similarity between the attention mechanism and the bridge influence line mechanism,this thesis proposes two algorithms on the basis of the DOVI and M-DOVI algorithms,named A-DOVI and AM-DOVI respectively.In the A-DOVI and AM-DOVI algorithms,an attention module is added to the DOVI and M-DOVI algorithms.The attention module is used to learn importance scores of displacement response data at different sampling instants in the original time-series data.Importance scores are utilized to scale the original timeseries data and amplify the impact of important information.Meanwhile,in the proposed approach,prior knowledge,where the bridge influence line is employed,is added to the attention module,so that deep learning algorithms is able to utilize bridge influence line knowledge to assist in decision-making.Results show that with the attention module and the prior knowledge,the identification accuracy of algorithms can be improved.Also,comparing the influence line with the curve obtained by visualizing the trained parameters in the attention module,the similarity between the attention mechanism and the bridge influence line mechanism can be observed.