Non-destructive Hardness Detection For Ferromagnetic Material Based On Magnetic Barkhausen Signal

With the increasing of science and the gradual lack of resources,people are increasingly demanding the economy.Non-destructive testing,a technique that can detect the properties without destroying the material itself,has become a new direction of development.For the most widely used ferromagnetic materials in current society,the Barkhausen signal in the micromagnetic field is a typical signal that can be used to detect metal properties,so it is a hot topic in the current non-destructive testing.At present,people use the Barkhausen signal to detect metal hardness from time and frequency domains.The time domain method is a classical method.It uses the appearance characteristics of the Barkhausen signal in the time domain.However,because a variety of metal properties can change the shape of the Barkhausen signal in time domain,the prediction effect will be impaired when several metal properties of the test piece are different.The detection method in frequency domain is a new type of method proposed in recent years,in which one method finds that the peak heights of some peaks of the AR spectrum of Barkhausen only vary with one metal property.If these peaks are found,the relevant relationships will be used ro predicted the hardness qualitatively.This method solves the problem of inaccurate hardness prediction when multiple metal properties are not the same,but it requires manual extraction of features,and the characteristics are single.In addition,the number of wave peaks is different,which causes automatic prediction cannot be achieved,so it cannot be practically applied.This thesis redesigns a new method of automatic prediction based on the frequency rules.After improving the frequency domain features,we design two automatic hardness prediction methods.The research content of this article mainly includes the following parts:Firstly,the previous frequency domain prediction method and its used features are analyzed,and the prior knowledge is learned by finding the adventage of these features and methods.According to the defects and deficiencies of the methods and features,the overall improvement ideas are determined.Secondly,the peaks and valleys related to hardness on the first and second derivatives of AR spectra are designed as new feature resourse.The new features are the frequency distance between the peaks and valleys in first derivative,the frequency and the deep of the valleys on the second derivative.Compared with amplitude of the hardness-related peaks,the new features are automatic extracted,more diversity and have stronger anti-interference.The experimental results show that the use of new features is better than that of the old feature: the classification error rate of the old features is about 2%,while the classification error rate of the new features is less 0.67%.Thirdly,a discrete output hardness detection method based on signal similarity is designed.This method calculates the similarity of the signal represented unknown hardness and the signal represented known hardness,and outputs the most similarity known hardness with unknown hardness,thus the output is discrete.Therefore,the bag of word algorithm can be used to unify the dimensions of features.The algorithms of this method include BP neural network enhanced by ensemble learning and PCA feature dimensionality reduction.From the experimental results,the detection error rate of the traditional frequency domain method and the classical time domain method exceeds 20%,while the error rate of the new method approaches zero.These results demonstrate the effectiveness of this method.Fourthly,a new method for hardness detection of continuous output based on the relationship with new feature and hardness is designed.By learning the relationship between the new features and hardness,the value of the unknown hardness can be obtained.The hardness value of the method is continuous and no longer applies to the bag of words algorithm.Therefore,feature extraction algorithm based on observation points is designed.This algorithm uses the idea of the column sampling.It replaces the feature set with multiple sets of feature subsets of the same dimension.In order to reduce the impact of feature misalignment,a new feature extraction criterion,"the valley nearest to the observation point," was designed.The algorithm of this method is a random forest algorithm based on the regression tree.From the experimental results,two kinds of metals are tested in this thesis.The mean square error of the time-domain method are 229.8HV30 and 298.7 HV30,and the mean square error of the new method are 60.3 HV30 and 81.3 HV30,respectively.These results demonstrate the accuracy of the methodBased on the above research content,this thesis applies the new method of frequency domain prediction to the field of ferromagnetic material hardness prediction.