Research On Detection Method Of Aggregate Gradation Based On Machine Learning And Multi-Feature Data

Traditional methods for detecting aggregate gradation suffer from low detection efficiency,high error rates,and unstable results.To improve the automation level of aggregate gradation detection and ensure the quality of road construction,this research focuses on multi-feature data analysis of aggregates.The study constructs a machine learning-based model for predicting aggregate grading and quality,and conducts a detailed investigation of aggregate gradation detection methods.Firstly,the three-dimensional scanning system is used to obtain the original three-dimensional data of the aggregate,and to extract the aggregate feature information.A multi-feature data dataset of aggregates is constructed comprising 51 features.To address the anomaly data in the dataset,an anomaly detection method combing the isolation forest algorithm and the box plot method is proposed.The mean imputation algorithm is used to repair the anomaly data,which improves the quality of the data.To retain the highly correlated features relevant to the prediction target,a feature selection method is proposed that uses both the correlation coefficient and permutation importance for analyzing the correlation between features and grading,as well as features and quality.Based on this analysis,datasets for predicting aggregate grading and quality are constructed.Secondly,aggregate grading and quality prediction models are constructed using the TabNet,Light GBM,XGBoost,and Random Forest algorithms.Grid search algorithms are utilized to optimize the models’hyperparameters.The experimental results indicate that Light GBM exhibits the highest prediction accuracy for grading(Accuracy=0.9176),while TabNet has the highest prediction accuracy for quality(R~2=0.9742).Therefore,Light GBM and TabNet are chosen as the grading and quality prediction models for aggregate gradation detection experiments,respectively.Finally,to further improve the accuracy of aggregate gradation detection,an arithmetic optimization algorithm(AOA)is used to optimize the Light GBM and TabNet models,resulting in the AOA-Light GBM grading prediction model and the AOA-TabNet quality prediction model.Experimental results show that the accuracy of the AOA-Light GBM grading prediction model reaches 0.9346,which is 0.017 higher than that of the Light GBM model optimized by grid search algorithms.Furthermore,the AOA-TabNet quality prediction model achieved an R~2of 0.9867,which is 0.0125 higher than that of the TabNet model optimized by grid search algorithms.Based on these models,the quality of each aggregate grading can be computed,and then the sieving efficiency at each grading can be obtained.Four groups of aggregate gradation detection experiments are conducted using random sampling from the prediction dataset.The experimental results show that the predicted gradation curves and actual gradation curves of the four experimental groups are highly consistent.The maximum absolute error of the six grading levels is 2.8%,and the average absolute error is 0.66%.The detection accuracy satisfies the practical demand for aggregate gradation detection.This thesis researches on the multi-feature extraction of three-dimensional data for aggregate materials and proposes AOA-Light GBM aggregate grading prediction model and AOA-TabNet aggregate quality prediction model based on arithmetic optimization algorithm.Detection of aggregate gradation is achieved.Compared with the original models,the improved AOA-Light GBM model and AOA-TabNet model have improved the accuracy of gradation prediction and quality prediction by 0.017 and 0.0125,respectively,with higher detection efficiency,lower detection error and more stable detection results.The research results can provide theoretical basis and technical support for road construction enterprises to timely and accurately grasp aggregate gradation information and ensure construction quality.