Pavement Disease Detection Method Based On Time-Frequency Features And Capsule Network

Pavement disease detection can help support the scientific maintenance decisions of roads to ensure traffic safety.Due to the long time dimension and nonlinear and nonsmooth characteristics of acceleration sensor data,the classification of pavement disease using traditional deep learning methods faces great challenges.In the thesis,a road condition detection method is proposed based on acceleration sensor data by extracting the time-frequency features and using capsule networks.The main research work is as follows.(1)C-TFS,a temporal data classification model based on time-frequency features and capsule network,is proposed.Considering that the acceleration sensor data is nonlinear and non-smooth,a time-frequency feature extraction method based on HilbertHuang Transform(HHT)is explored,which can extract features from the original signal.Then,capsule network is introduced to the time series classification.C-TFS is compared with the benchmark method on UCR time-series public datasets,and the experiments show it has the highest average accuracy and F1 score.(2)The model C-TFS is optimized using attention mechanism,and an extended model AC-TFS is proposed.Owing to the large number of parameters of C-TFS model and the high complexity of the dynamic routing algorithm of the capsule network,C-TFS often consumes a lot of resources.To address this problem,the thesis introduces attention mechanisms into the capsule network.This can effectively reduce model parameters and computing resource consumption while improving the model performance.(3)The AC-TFS-based road condition detection method is proposed.In order to accurately classify and identify different types of pavement diseases,the AC-TFS model is applied to the road condition detection.Experiments show that the classification accuracy of the AC-TFS model reaches 98.58%,which is an improvement of 7.09% and9.22%,respectively,compared with the SVM and KNN methods commonly used in existing pavement anomaly detection.Moreover,compared with the C-TFS model,its classification accuracy can also improve by 1% and its resource consumption is reduced.