Identification Research For Obtaining Spatiotemporal Information Of Vehicle Loads On Bridges Based On The Deep Learning Theory And Computer Vision Technology

As one of the typical inverse problems in structural vibration field,conducting effective distinguish to vehicle loads would play a significant role in the performance assessment of bridges.The current technology is mainly relied on the bridge weigh-in-motion(BWIM)system,which would be restricted in the applications of mid-and shorten-span bridges due to its high-cost devices and complex installation process.Therefore,a novel vehicle loads identification methodology was proposed in the manuscript based on the deep learning theory and computer vision technology,which was mainly conducted as follows:(1)A novel vehicle loads identification methodology was proposed based on the deep learning theory and structural characteristic responses extraction.Specifically,the time-frequency analysis was employed to extract the feature patterns of bridge acceleration responses induced by the driving vehicles,and the convolutional neural network(DCNN)model was utilized to dig the potential pattern details of extracted features to achieve the system outputs identification based on captured inputs.Therefore,the solution of inverse problem could be transferred by replacing the DCNN as middle structural system,where the inputs could be further obtained by only utilizing the collected output responses.(2)The numerical analysis,experimental and field test were conducted to further evaluate the effectiveness of proposed method,respectively.A satisfied identification accuracy was investigated according to the test results,and the performance of trained DCNN model could be further enhanced by utilizing the image processing technology to increase the amount of test samples.Meanwhile,it was also revealed that the features extracted from the natural object images dataset by DCNN could be effectively transferred into the classification field of digital signal images.Furthermore,a 97.21% identification accuracy was obtained in the ambient test,which has demonstrated that the vehicle loads information could be directly distinguished from structural responses by the proposed approach.(3)A vehicle loads identification system was established based on the non-contact computer vision technology.Specifically,with the proposed rough-grained classification standard of common vehicles,the mapping relationship was established between vehicle types and corresponding weights interval.And the Faster-RCNN model trained by the HNU-Vehicle Dataset was utilized to detect vehicle objects,loading condition and real-time position from monitoring videos.Furthermore,the motion trajectory of driving vehicles could also be detected by the compressive sensing technology.Experimental and field test were both conducted to evaluate the effectiveness of proposed method,respectively.(4)A novel vehicle spatiotemporal information identification technology was achieved based on the non-contacted photogrammetry technology.The vehicle loads under driving state was selected as the calibration object,and the corresponding vanish points were detected by the parallel coordinate-based cascade Hough transform technology,which has achieved the camera calibration process without the support of targets.The effectiveness of proposed method was verified by distinguishing the actual position of vehicle loads,showing that this technology could be effectively utilized to obtain vehicle loads position even under the complex test condition.