Soft Classification Scheme with Pre-cluster Based Regression for Identification of Same-based Alloys using Laser-induced Breakdown Spectroscopy (LIBS)

Abstract

A novel soft classification scheme with pre-clustering based regression is proposed for identifying same-base alloys using laser-induced breakdown spectroscopy (LIBS). LIBS is a technique for analyzing the spectral signals of plasma generated by irradiating a target sample with a high-energy pulsed laser. With LIBS, real-time analysis is possible for all states without pre-preparation, and several elements can be detected simultaneously. The enhanced applicability of LIBS enables the quick identification of raw samples through a quantitative and qualitative analysis of the spectral data acquired by LIBS.

For alloys of different base metals, an investigation of the locations of the peaks in the spectrum allows us to distinguish them easily, since different elements have inherent characteristic wavelengths. On the other hand, alloys of a given base metal share major elements, such that additional quantitative analysis may be needed to accurately classify them. However, LIBS measurements may be impaired by complicated self-absorption and matrix effects, and the relationship between the peak intensity and the concentration is generally not straightforward. In the present study, spectral lines of other elements, particularly those which are less dominant but strongly correlated to the target, are also considered as part of the analysis to mitigate the effects of large variations in the peaks of major elements. By considering all the peaks under a complementary relationship, the accuracy with which the concentration of metal base elements can be estimated from the LIBS spectrum can be improved. To incorporate the relationship between multiple elemental concentrations, this study proposes a pre-cluster based regression and most probable classification scheme to enhance the classification performance of signal processing, whereby clusters are divided in advance based on the relative relationships of multiple elements in same-base alloys.

The proposed scheme is divided into three stages: 1) pre-cluster decision using prior information; 2) stepwise regression analysis; 3) most-probable classification based on soft information. In the first step, information on metal alloys as obtained from the NIST and Brammer databases is used to determine the distributions of metals as prior information. Since the metals are not uniformly distributed, multiple clusters are divided and determined in advance suing this prior information. In the next step, full-scale regression, which involves training all the data sets and testing them jointly, is applied first, after which the probability of belonging to each cluster is calculated from the regression result. The final regression result is obtained through the weighted sum of the probability and pre-clustering regression results which is used to train each cluster dataset. These are tested separately. Last, the most-similar type is identified through most probable classification using estimated regression results and soft information on the mean and standard deviation.

To numerically evaluate the performance improvement using the proposed method, the partial least square(PLS), and random forest(RF) regression methods are employed, and the root mean square error(RMSE) measure is calculated based on these machine learning methods. As a result, a comparison of the RMSE of simple regression analysis with the RMSE of the proposed regression analysis shows that the performance of the proposed scheme improves. It also shows that the accuracy of the most probable classification using soft information is improved. Likewise, the proposed scheme not only estimates the elemental content more accurately by in-depth regression analysis for unknown samples, but also provides information about which metal is closest to the unknown samples.

In summary, this thesis proposes a new statistical analysis method that improves accuracy by highlighting the peak information of other components related to the main component, which can compensate for the physical distortion of the signal.