Efficient Sensor Fusion For Non-Destructive Data Of Concrete Based On Machine Learning

Corrosion in reinforced concrete is an important impairment mechanism in civil engineering.This damage can affect the integrity of the concrete structure and cause changes in its stressing,resulting in a high risk to the building.Various non-destructive testing(NDT)methods exist for the detection of active corrosion.The performance of individual processes is methodically and physically limited and is often not reliable enough.Therefore,by inspecting the same specimen with various NDT methods a more comprehensive view of the part’s condition can be obtained.However,automatic data evaluation with multi-sensor data has poor detection performance caused by the complexity of the data and low computational efficiency and accuracy of traditional algorithms.Aiming at the problem of poor detection performance caused by the complexity of multisensor data in automatic data evaluation,this work designed the function models of decision tree(DT),Boosting and Bootstrapping based on machine learning(ML)to improve the NDT of reinforced concrete.The principal effects that are used for data fusion are shown theoretically.The algorithms learn optimal linear decision boundaries from multivariate labeled training data,to distinguish corrosion and non-corroding areas.And the results are compared with the relatively simple logistic regression algorithm.The experiments are implemented with Non-destructive Testing(NDT)data obtained from the multi-sensor detection of the corroded reinforced concrete.Their effectiveness is tested in case studies carried out on large-scale concrete specimens with built-in chloride-induced rebar corrosion.The data set consists of half-cell potential mapping(HP),Wenner-resistivity(WR),microwave moisture(MW)and ground penetrating radar(GPR)measurements.The unique possibility to monitor the deterioration,and targeted corrosion initiation,allows data labeling.This work demonstrated that the simple and robust logistic regression algorithm is superior to the DT algorithm with complex decision boundaries trained by small heterogeneous data sets.However,Boosting and Bootstrapping can be used to improve detection and the corrosion analysis of reinforced concrete,between which Bootstrapping had a superior performance with its property of balancing data.The results exhibit an improved sensitivity of the data fusion with Boosting and Bootstrapping compared to the simple linear classifier of LR.