Quantitative analysis method of steel composition based on LIBS and CNN

Laser-Induced Breakdown Spectroscopy (LIBS) combined with traditional mathematical methods for quantitative analysis of material composition often relies on specific prior knowledge, leading to issues such as poor analysis accuracy and weak generalization performance. In order to mitigate the impact of prior knowledge on the analysis method and simultaneously enhance analysis accuracy, this study proposes a 1d CNN model based on the ResNet-18. The model is specifically designed for the quantitative analysis of four elements (Cu, Mo, P, S) in steel. Subsequently, to better serve edge computing scenarios, the research focuses on lightweight processing of the algorithm model, making it easily transferable to low computational power devices. A total of 40 steel samples were collected for constructing the dataset, and optimizations were made regarding the loss function utilized during model training. Experimental results indicate that the proposed method achieves RMSE values of 0.0096 for the four elements. Even after reducing the model parameters by almost 90%, the RMSE for the four elements remains at 0.0131, demonstrating no significant performance degradation. This validates the effectiveness of the proposed method.