Real-time Evaluation On Compaction Quality Of Pavement Based On Machine Learning

Compaction work is crucial for the safety and durability of a pavement structure.Traditional compaction quality control relies on spot tests measured at several spots along a roadway after compaction work,such as sand replacement method and drilled core method.By using these methods,it is difficult to ensure the uniformity of compaction quality,leading to over or under compaction.Besides,it could cause a waste of man-work and resources if the compaction quality does not meet the requirement and need to be re-compacted.In order to address problems in existing compaction quality control,this study firstly proposed an evaluation method of compaction quality based on modal identification of ’vibratory drum-compacted material system’,and realize the method using machine learning.This method may provide technical reference for the research of intelligent compaction(IC).Firstly,by using a classic mechanical model of ’vibratory drum-compacted material system’ the relation between the natural frequency and stiffness of the system was analyzed.Then vibrational signal of the system was collected during construction of subbase in a field test.Signal processing technique was implemented to find the inherent correlation between the natural frequency and stiffness of the system.Characteristics of vibration signal on the scale of time-frequency was analyzed.The main model parameter(natural frequency)of the vibratory system was obtained using modal identification method.The experimental results show that the natural frequency of the system increases relatively fast at first and then slowly with the increase of compaction times,which is similar to the growth law of degree of compaction.It is proved to be feasible and reliable using the change of the natural frequency of the vibrational system to evaluate the compaction quality.The moving average difference of natural frequency of the system is presented as an evaluation index of compaction quality,and the process for evaluation of compaction quality based on this index is given.Secondly,artificial neural network(ANN)and support vector machine(SvM)are used to realize evaluation of compaction quality control(CCC)on pavement in real time.A database for machine learning training is established after pre-processing vibration signal.Through the determined of structural parameters of BP neural network model,a BP neural network model for classification and recognition of compaction quality was established.The ability of the network model for classification was tested using the testing sample.Combined with the ROC curve and confusion matrix,the ability of neural network model for classification and recognition of compaction quality is evaluated.The results show that the established model has relatively reliable recognition performance,but due to the limited number of labeled samples,the overall recognition accuracy of the network model is 65.3%,the recognition accuracy of 94-98%and>98%degree of compaction(DOC)are 92.3%and 76.0%in respective,the identification results are not ideal.Then,a compaction quality evaluation model based on support vector machine(SVM)was established through design of multi-classification model,determination of learning model parameter and optimization of the optimal parameter pair.The testing results show that the overall recognition accuracy of the evaluation model is 87.6%,and the recognition accuracy of the label>95%DOC is 91.2%.The results indicate that the SVM modal owns better performance of classification and prediction of compaction quality in comparison with the ANN modal.This is mainly due to the ANN model need enough training data to obtain satisfactory recognition performance.However,the SVM modal has more outstanding performance in classification and recognition when the database is in a small amount.Finally,a method based on the SVM modal is applied to evaluate compaction quality of asphalt mat.Through a field vibration compaction test of asphalt pavement,the vibration signals in the process of vibration compaction on asphalt pavement were collected,and the training database,as well as the pavement compaction quality evaluation model based on support vector machine,were established.The analysis results show that the SVM learning model also has good performance in the classification and identification of compaction quality of asphalt mat,especially in the middle and late stage of compaction when the DOC is close to or meets the requirements.In addition,based on the SVM learning model,a real-time analysis system of asphalt pavement compaction quality with real-time processing function of multi-source data was developed,which realized the real-time continuous monitoring of asphalt compaction quality with multi-dimension,multi-index and all-direction,which can guarantee the stability and uniformity of compaction quality.