Research Of Defect Depth Inversion Based On Machine Learning

The main purpose of the nondestructive testing(NDT)technology is to detect the damage degree of the detection object by using various principle techniques without destroying the original function of the detection object.Magnetic flux leakage(MFL)detection technology is an important branch of nondestructive testing technology(NDT),which is mainly used to detect the object of ferromagnetic materials.Oil and natural gas pipeline mainly composed of ferromagnetic materials,so the main detection method for oil and gas pipeline is magnetic flux leakage detection method.In the case of magnetic flux leakage detection for oil and gas pipelines,the amount of data in the test results tends to be large because the length of a pipe is usually long.Therefore,the determination of the location of the corrosion defects and the exact size of the corrosion defects is important.The main contents of this paper are as follows:Firstly,aiming at the problem of determining the position of the corrosion defect in the pipeline magnetic flux leakage detection data,a calibration method and a filtering method for the leakage detection data of the pipeline are proposed.In this paper,the anomaly region detection algorithm of differential size window is designed,and the defect feature is extracted and the sample library is generated by using the test field data.Secondly,a feature reduction dimension based on Pearson correlation coefficient and feature reduction algorithm based on principal component analysis and factor analysis are proposed for the balance of defect depth inversion and calculation of computational resources.A sample library is designed.Three machine learning algorithms,decision tree algorithm,support vector machine algorithm,and a deep learning network of AlexNet,and achieved good experimental results.Thirdly,this paper proposes a size information fusion algorithm based on feature selection and dimensionality reduction for defect depth inversion accuracy.The experiment proves the effect of the algorithm.