A study of multi-dimensional defect size estimation of metal materials using LIBS spectral characterization

Abstract

Metal defect detection has always been one of the key links in the metal manufacturing industry. However, the existing metal defect detection methods have the problems of single size detection and low detection accuracy. To overcome these challenges, we proposed a novel method for detecting the size of metal defects using spectral quantification. By analyzing the response spectra of defect depth and width in a continuous gradient distribution, we systematically explored the variation in spectral physical characteristics for different defect size combinations and established a defect spectral ratio model. In addition, according to the high-dimensional characteristics of the spectrum, a feature layer weighted fusion scheme was proposed to improve the accuracy of metal defect discrimination. We conducted 1000 average accuracy tests on the training set, validation set, and test set. The results indicated that the training set achieved an average accuracy of 99.92 %, while the validation set and test set achieved average accuracies of 95.83 % and 95.5 2%, respectively. Finally, we provided a quantitative estimation of the size of metal defects. The relative error range for defect size estimation in the test sample ranges from 0.3314 % to 5.6371 %. The maximum values for average error and RMSE were 0.089 and 0.0995, respectively. The comprehensive results indicate that this method possesses high stability and accuracy, enabling effective identification of metal defects and estimation of size errors. Furthermore, this method introduces a novel idea and framework to the field of metal defect detection, showcasing remarkable scalability and positive impact.