K-Means-CNN-Based Classification Study of Mixed Alloy Samples of Complex Grades
MA Yao-an1, HUANG Yu-ting1, ZHANG Jian-hao1, QU Dong-ming1, HU Bei-bei2, LIU Bi-ye2, YANG Guang1, SUN Hui-hui1*
1. College of Instrumentation and Electrical Engineering, Jilin University, Changchun 130026, China
2. Beijing Oriental Institute of Metrology and Testing, Beijing 100000, China
Abstract：Laser-induced breakdown spectroscopy (LIBS) is a highly efficient elemental analysis method with simple sample preparation, non-contact measurement, strong field adaptability, and fast analysis speed by focusing an ultra short pulse of laser light on the surface of the sample to form a plasma, and then analyzing the emission spectrum of the plasma to determine the material composition and content of the sample. Using LIBS technology for elemental analysis, component classification, and identification is the key direction of the research. At present, LIBS technology is mainly used in rock and mineral detection, environmental monitoring, chemical identification, and related fields, while less research is conducted on the classification of mixed alloys with multiple components and complex grades. The commonly used high-performance, accurate classification algorithms usually require high computational resources and are difficult to mount on portable and miniaturized LIBS systems. A mixed sample of various grades of AL, FE, and CU alloys was excited by an MPL-T-1064 laser with a modulated optical path through a front mirror set to collect data. Data were preprocessed using Principal Component Analysis (PCA) and then input into the K-means clustering algorithm (K-means), a Convolutional Neural Network (CNN) model for classification. The K-means is unable to classify complex alloy grades finely, but has an accuracy of 99.97% in the work of large class differentiation.CNN can classify complex alloy grades finely with an accuracy of 99.15%, but it has a relatively high demand on computational resources. Aiming at the above problems, a fusion algorithm is designed to use the K-means algorithm to process the mixed alloy spectral data. It coarsely classifies samples of the same kind but different grades. Then, the data after the first-stage classificationis input into the CNN model to carry out fine classification. The accuracy of classification in the mixed alloy spectra of ten kinds of samples of grades of AL, FE, CU reaches 99.35%, and the accuracy in the 5-fold cross-validation reaches 99.52%, which verifies that the algorithm has better generalization ability while classifying accurately. The classification accuracy of the fusion algorithm is 39.65% higher than that of the K-means algorithm, and the running speed is 21.94% faster than that of the CNN algorithm. It provides an efficient, fast, and accurate method for the classification of mixed alloys with multiple compositions and complex grades. It provides a new idea for developing a more lightweight and portable LIBS system.
马耀安，黄裕婷，张健豪，曲东明，扈蓓蓓，刘碧野，杨 光，孙慧慧. 基于K-means-CNN的复杂牌号混合合金样品分类研究[J]. 光谱学与光谱分析, 2025, 45(07): 1946-1952.
MA Yao-an, HUANG Yu-ting, ZHANG Jian-hao, QU Dong-ming, HU Bei-bei, LIU Bi-ye, YANG Guang, SUN Hui-hui. K-Means-CNN-Based Classification Study of Mixed Alloy Samples of Complex Grades. SPECTROSCOPY AND SPECTRAL ANALYSIS, 2025, 45(07): 1946-1952.
