Research On Machine Learning Based Strength Regression Method For Concrete Testing

Rebound method and ultrasonic-rebound combined method have been widely used in the concrete strength testing of practical engineering in China for more than 50 years due to their simplicity and flexibility.However,the problem of low accuracy of rebound method is one of the hot issues that research workers have been concerned about.Based on the analysis of related research,this paper conducts a correlation analysis of the influence of concrete mix proportion and related parameters of rebound method and ultrasonic-rebound combined method on the compressive strength of concrete.Combining with the characteristics of machine learning algorithm,six machine learning regression models were established based on the existing concrete specimen data using MATLAB software,and the fitting of strength related data is carried out.The accuracy of the regression model is discussed,and the application of small sample engineering example data was also demonstrated to improve the accuracy and reliability of predicting concrete compressive strength in practical engineering.The main research contents of this paper are as follows:(1)The impact factors of testing the compressive strength of concrete using rebound method and ultrasonic-rebound combined method were analyzed,and the correlation between various parameters of concrete mix proportion and compressive strength was analyzed using three correlation coefficient methods: Pearson,Spearman,and Kendall.The analysis showed that water and admixture had a greater influence on the compressive strength of concrete,and mechanism sand had a smaller influence on the compressive strength of concrete;(2)Through analyzing the correlation between rebound method parameters and compressive strength of concrete,it was found that the rebound value had the greatest influence on the compressive strength of concrete,followed by age,and then carbonation depth.The correlation analysis of ultrasonic-rebound combined method parameters and compressive strength of concrete showed that the rebound value had the greatest influence on the compressive strength of concrete,followed by ultrasound velocity,and then age;(3)Based on MATLAB software,six regression analysis algorithms commonly used in machine learning were used to fit the relevant data,and the MLP,random forest and Gaussian kernel regression models were fitted better,and the errors were smaller than the error range specified by conventional linear regression methods and specifications;(4)Using six trained models,combined with small sample engineering example data,the strength is predicted and verified.The results are compared with the regional strength curve and the national strength curve given in the specification.The results show that the errors of the six models are much smaller than the error range specified by the national strength curve and the specification in the actual engineering data,and they also have corresponding accuracy in small sample data;(5)In the process of data regression comparison and validation,combined with the calculation results of the relative mean error and relative standard deviation of the fitting and prediction results,it was analyzed that the MLP regression model had better prediction effect than other models,followed by the random forest regression model.