Research On Characteristics And Model Uncertainty Of Barkhausen Noise For Fatigue State Prediction

Fatigue state prediction based on Magnetic Barkhausen Noise(MBN)detection is widely used to detect the microscopic properties of ferromagnetic materials,and the damage and fatigue state of materials are characterized by magnetic parameter characteristics.Parameter characteristics have observation uncertainty due to the influence of material characteristics,detection methods,environment,etc.In addition,there is uncertainty in the model of detecting and predicting material fatigue state,that is,the incompleteness and deviation of the model when describing the actual system mainly come from the model selection,parameter estimation,prior conditions and other aspects.Reducing the uncertainty of data and model can improve the reliability and accuracy of fatigue state prediction.And play an important role in the safe operation of industrial products and major machinery and equipment.Therefore,the magnetic parameter characteristics of MBN signal and fatigue state prediction are mainly studied.This paper mainly includes the following four aspects:1.In view of the observation uncertainty existing in MBN parameter features,new MBN signal features were extracted to improve the characterization ability of feature parameters to material properties.Feature selection was realized through mathematical statistical analysis,principal component analysis and feature correlation analysis.Variational encoder was used to reconstruct the feature space of magnetic parameters to reduce the uncertainty of observation data.2.Aiming at the problem of uncertainty in the general fatigue state prediction model,a Bayesian inference network model is proposed to apply a certain distribution function of trust space to the weight parameters of the model,so as to improve the prediction accuracy and reduce the uncertainty of the model parameters.Two different prior distribution forms,diagonal Gaussian and scale mixed Gaussian,are tried to discuss their influence on the prediction results.The model overfitting problem was discussed in the case of spatial discontinuity or incomplete data of experimental samples,and the NCP noise contrast prior model was proposed to solve the problem.3.Aiming at the problem that the large randomness of MBN shape class features is not suitable for time series prediction,GMF features based on Gaussian mixture model are proposed to fit the envelope,and component number optimization algorithm GCO is proposed to solve the problem that the Gaussian component number is difficult to determine.Based on GMF characteristics,the fatigue model of hidden Markov and KL divergence calculation methods is constructed to predict the evolution process of fatigue state.4.Aiming at the study of anisotropy of material plane,a test method of tensile stress was designed,and the curve of variation between tensile stress and two-dimensional magnetic parameters was calibrated;A calibration method based on two-dimensional plane biaxial is proposed,and the relationship between root mean square and stress distribution anisotropy is used to verify the calibration accuracy of the new method.R~2 is increased from 0.9664 to 0.9803 in 10-dimensional shape feature set 1,and from 0.9787to 0.9798 in 12-dimensional statistical feature set 2.Plane stress analysis was carried out based on a variety of magnetic parameters,and its relationship with the direction of easy magnetization axis and tensile stress was studied.Characteristic analysis and selection were carried out on the amplitude RMag and phase RPh of multi-frequency eddy current signal,the distortion coefficient K of tangential magnetic field intensity harmonic signal,coercive magnetic field Hco and other related characteristic parameters.