Application Research Of Variational Bayesian Decomposition Algorithm In Infrared Nondestructive Testing

The eddy current pulsed thermography system has been widely used to detect and evaluate the defects of the various materials,such as steel blade,railway and carbon fiber reinforced polymer.Many researchers proposed different feature extraction algorithms(e.g.principal component analysis,pulsed phase thermography,thermographic signal reconstruction,independent component analysis,etc)to obtain the feature of the defect.The feature extraction algorithms can reduce the interference of background,noise information and improve the accuracy of defect detection.However,the feature extraction algorithms cannot eliminate the interference of background and noise information.The detection accuracy of the defects is low.In order to solve these problems,the main work of the paper are:1)Analysing the surface temperature variation of the specimen with defects.The independent characteristic can be obtained on the physical level.The linear combination model of the different patterns can be proposed.The characteristic of the defects is sparse on the physical level.The sparsity linear combination model of the different patterns can be proposed.The model matches with the variational Bayesian matrix decomposition model,therefore the variational Bayesian matrix decomposition model can be used to extract the feature of the defects.The algorithm adopts the Bayesian framework and uses the variational Bayesian method to automatically update the parameters.2)Due to the different defects of specimens with the different sparsity.The variational Bayesian matrix decomposition method cannot adapt to all the situations.In order to solve the proplem,the adaptive variational Bayesian matrix decomposition algorithm is proposed.The method imposes automatically sparseness control as well as sub-group so that the decomposition can be iteratively optimized.This overcomes the problem of under-and over-sparse factorization.The results by using the algorithm show that the model can improve the defect detection precision and reduce the interference of the background(such as: edge,coil,etc.)and noise information.The control parameter is learned and adapted as part of the matrix factorization by using variational Bayesian approach.This bypasses the need of manual selection.3)The computational complexity of the adaptive variational Bayesian matrix decomposition algorithm is related with the size of the image,therefore,the high dimension of the data is high computational complexity.In order to solve the problem,the automatic relevance determination of adaptive variational Bayesian matrix decomposition algorithm is proposed.The sub-group zone can be deleted in the iterative,which do not contain the information of the defect.The model can reduce the dimension of the image,therefore,it can reduce the computational complexity and the running time.4)In order to objectively evaluate the detectability of different feature extraction methods,the event based F-score is computed.Experimental tests on both artificially and nature defects,the F-score of the adaptive variational Bayesian method is higher than other feature extraction algorithms,but its computational complexity is higher than other features extraction algorithms.The automatic relevance determination of adaptive variational Bayesian matrix decomposition methods can balance the two problems,the detection accuracy of the algorithm is higher than other feature extraction algorithms,slightly lower than the adaptive variational Bayesian matrix decomposition algorithm.However,the computational complexity is far lower than the adaptive variational Bayesian matrix decomposition algorithm.