Task-Oriented Comparison and Fusion of Spectra and Acoustic Signals in LIBS for Alloy Analysis: From Identification to Quantitative Measurement

Abstract

Laser-induced breakdown spectroscopy (LIBS) has been widely applied in alloy elemental analysis. However, reliance solely on spectra makes it difficult to fully capture all the information related to the laser-material interaction and plasma evolution processes, thereby limiting the analytical capability of this technique. To overcome this limitation, this study introduces acoustic signals generated during laser ablation as a complementary information source and proposes a fusion framework of spectra and acoustic signals for alloy characterization. The respective and combined capabilities of acoustic signals and spectra were systematically evaluated in both qualitative classification and quantitative regression tasks for micro-alloyed steel samples. A series of classification models (DT, KNN, and SVC) and regression models (SVR, PLSR) were constructed. The experimental results show that the fused data consistently outperform individual signals in both classification and regression tasks, exhibiting higher performance, robustness, and generalizability. Furthermore, for individual signals, acoustic signals and spectra each demonstrate strengths in different tasks. Specifically, acoustic signals, owing to their higher information density and lower dimensionality, perform better than spectra in relatively simple classification tasks. Among them, the SVC classification model built on acoustic signals achieved an accuracy of 1.0000 on both the training and test sets, comparable to the performance based on fused data. In contrast, for more complex quantitative analysis tasks, spectra show superior predictive capability due to their direct correlation with elemental content. For example, the PLSR regression model for Cr based on spectra performed better, with RMSE, STD, and RSD on the test set of 0.0207, 0.0197, and 15.0750% respectively, compared to 0.0337, 0.0329, and 27.3185% for the model based on acoustic signals. These findings fully confirm that each type of signal emphasizes distinct aspects of information representation, highlighting their complementary roles in capturing the physical and chemical properties of samples. Therefore, in practical applications, selecting an appropriate data type requires a comprehensive consideration of the scenario, task objectives, and performance requirements to balance performance, efficiency, and resource constraints.