Research On Strength Prediction And Component Morphological Characteristic Analysis Of Concrete Based On Deep Learning

Concrete strength is an important indicator of concrete performance,which is the key object in academic research and engineering application.The accurate prediction of concrete strength can reduce the cost of manpower,material and time in the process of concrete strength design,and reduce the hidden dangers of concrete during service.However,concrete as a complex artificial material.the characteristics that affect the strength of concrete include but not limited to the constituent materials and their properties,mixing methods,curing conditions,etc.At present,due to the quantitative limitation of characteristics,the prediction of concrete strength is mainly dependent on the amount of constituent materials,i.e.mix proportion.With the help of machine learning,concrete strength prediction based on mix proportion has achieved a lot,but it only considers the amount of constituent materials and establishes the numerical relationship between the amount of material and concrete strength.Therefore,it is of great significance for the concrete strength prediction system to interpret the characteristics,add new quantitative characteristics and optimize the quantitative process of characteristics.Combined with deep learning technology,this paper had done concrete strength prediction,fly ash microsphere feature quantization and concrete pore recognition optimization based on explainable feature quantization process for explainable feature,newly added quantization feature and optimized feature quantization,and had drawn the following conclusions:(1)Based on the existing literature mix proportion and the basic principles of concrete,nine explainable features were selected and the prediction methods with convolutional neural network were proposed.The results showed that the convolution neural network model combined with prediction accuracy of strength of three kinds of concrete(ordinary concrete,high-strength concrete below 100MPa,recycled concrete)was good,and the determination coefficient R2 could reach 0.9735,and the influence analysis of three main explainable features of water cement ratio,sand ratio and paste-aggregate ratio on concrete strength was consistent with the existing concrete theoretical results.(2)Fly ash mixed with water in mass ratio 1:100 and ultrasonic vibration for 3 minutes and were taken photos.We obtained 78 photos with uniform dispersion and clear target at last.The trained PANet based on 78 photos could accurately identify the fly ash microspheres and mAP could reach 49.5.The particle size distribution results of the same kind of fly ash were consistent,using PANet,scanning electron microscope with threshold method and laser particle size tester.The repeatability and reproducibility analysis results are%GageR&R of 22.73%,%P/T of 18.03%,which shows the effectiveness and stability of the method.And the content of high-quality fly ash microspheres tends to high,and the particle size tends to small.(3)Collecting the images of concrete surface pores based on zero degree annular light source could add a new edge feature bright halo for concrete pores without other process except polishing.Combined with the instance segmentation network PANet could improve the identification accuracy,speed and distinguish the non-connected pore from connected pore in concrete.And mAP could reach 53.7.The PANet model trained by the data set of concrete pore collected by zero degree annular light source could recognize the pore of cement-based composite materials,and the recognition accuracy is higher.mAP could reach 65.3 in the cement paste,and 60.9 in the mortar.