Research On Flaw Recognition Technology Of Laser Ultrasonic Signal

Laser ultrasonic detection technology with high sensitivity,non-contact,rich pattern andhigh resolution,become the important guarantee of quality inspection and safety production inthe industry.In the laser ultrasonic surface defect detection experiments, the ultrasonic signalwe get is often high-dimensional and small samples.Application of traditional signalprocessing and pattern recognition methods for such high-dimensional and small samples datais often inefficient or even ineffective.Therefore,the study of feature extraction andclassification algorithms suitable for high-dimensional and small sample data is a majorproblem of laser ultrasonic flaw identified.In this paper, laser ultrasonic flaw detection experiments system has been build for defectdetection experiments.Our aim is to find effective feature extraction methods,highgeneralization ability and accuracy classification algorithm,utilizing the reflected andtransmitted wave signal with surface defect information obtained experimentally.The maincontents of this article include:(1)The traditional signal processing methods is limited by the signal stability,workingconditions and other factors, tend to have certain limitations and instability when applied tolaser ultrasonic signals.Meanwhile, the measurement and assessment for the characteristics ofdefect signal is still in the exploratory stage.Therefore, we propose a supervised Kohonennetwork.This network with high nonlinear mapping ability and intelligence capabilities canachieve adaptive learning and classification for ultrasonic signals.Multiple cross experimentalresults show that the supervised Kohonen network can effectively achieve characteristiclearning and classification for the defect signal,and has good generalization ability.(2)Because the high-dimensionality of laser ultrasonic flaw signals directly affects thetime complexity and space complexity of the algorithm, which may lead to invalidalgorithm.Accordingly,we adopt KPCA, a nonlinear principal component extraction algorithm, to reduce the dimension of the defect signal data,and then apply the SVM classificationalgorithm.Multiple cross experimental results show that, the algorithm based on KPCA andSVM can identify defects quickly and accurately.(3)KPCA method based on variance for the principal component extracted.With respectto KPCA,we introduced an principal component extraction method(KECA) based oninformation entropy.This kind of feature extraction and dimension reduction method isapplied to the laser ultrasonic flaw signal processing.Multiple cross experiments show thatthis method based KECA and SVM has high accuracy rate,low complexity,and stronggeneralization ability.Comparing to KPCA,KECA method can achieve higher contributionrate,when you select the same number of principal components.This showed that therepresentative of principal component for KECAis more pronounced.