Research On Stress Measurement Methods Of Ferromagnetic Materials Based On Barkhausen Noise

Magnetic Barkhausen noise(MBN)detection is an electromagnetic non-destructive testing method for ferromagnetic materials.Ferromagnetic materials are subject to stress in different directions during manufacture and service,resulting in performance degradation or fracture.Barkhausen noise detection can be used to detect stress without destructing the materials.It is important for overhaul of rails and special equipment,which has practical engineering application value.Based on the Barkhausen noise signals under different stress conditions,feature generation,feature selection and model establishment are studied in this thesis.The main contents are as follows:(1)The mechanism of Barkhausen noise is studied,and the influence of material microstructure on Barkhausen noise is analyzed.The material stress detection system based on Barkhausen noise is introduced,including excitation coil,detection coil,signal conditioning circuit,stress loading platform,etc.The experimental data of Barkhausen signal used in this thesis is introduced.(2)Based on the feature generation of Barkhausen noise under different stress conditions,a data set of BN features and stress is established to find the important features in the process of feature selection.For BN signals,the extracted features are root mean square(RMS),mean value,energy,pulse count,standard deviation,interquartile range and entropy.For BN profile,the extracted features are full width at half maximum(FWHM),skewness,kurtosis,Gaussian function fitting parameters,peak value,profile area and reconstructed remanence.For BN time-frequency domain,the extracted features are marginal spectral peak 1,marginal spectral peak 2 and marginal spectral energy.(3)The feature selection algorithms are used to eliminate redundant and uncorrelated MBN features.The algorithms used in the thesis are stepwise regression,ReliefF algorithm and neighbor component analysis.By analyzing and comparing the results of three feature selection algorithms,it is found that the features of entropy envelope area,interquartile range and Gaussian function fitting parameters are sensitive to material stress.(4)The relationship between BN and material stress is not clear,this thesis compares three modeling methods: partial least squares regression(PLSR)model,principal component regression(PCR)model and convolutional neural network(CNN)model.The model of BN features and material stress is established by using features selected before.The results show that the established model can be effectively used to predict material stress,and the prediction accuracy is better than traditional modeling.