Phenol red (PR) is generally used as an acid-base indicator and a printing and dyeing colorant. When its content exceeds a certain concentration in water, it will cause great damage to the human body. Therefore, it is very important to detect the content of PR in water. The advantage of surface enhanced Raman scattering (SERS) is detecting samples quickly, non-destructive and high sensitivity without sample pre-treatment. SERS has attracted great attention in all fields of detection and analysis. In this paper, the method of attaching silver nanoparticles to metallic single-walled carbon nanotubes form carbon nanotubes/silver nanoparticles (CNTs/AgNPs) structure and then combining it with silica sheet is proposed. SERS substrate with silica/carbon nanotubes/silver nanoparticles (SiO₂/CNTs/AgNPs) composite structure has extremely high reinforcement effect. In the quantitative analysis of the detected substance, mathematical fitting or machine learning is used to find the relationship between the intensity of Raman signal and the concentration of the detected substance. The BP neural network optimized by genetic algorithm (GA-BP) is designed in this study. The weights of GA-BP to enhance the robustness of BP neural network, the method of adaptive learning rate and the number of hidden nodes is set to solve the problem that GA-BP is easy to fall into local optimum, thus establishing a quantitative analysis model of PR solution concentration. The model can detect different concentrations of PR solutions with high accuracy quickly, simply and sensitively. Finally, compared with other published quantitative models, GA-BP correlation coefficient R² determined by the training results of the model is 0.99996, and the root mean square error of the prediction is RMSEP = 0.002510, which is 0.0005 higher than the mathematical fitting method, it shows better performance. A reliable idea for the preparation of SERS substrate and online detection of PR concentration in water proposed in this study.

酚红（PR）一般用作酸碱指示剂和印染着色剂。当其在水中的含量超过一定浓度时，会对人体造成很大的伤害。因此，检测水中PR的含量非常重要。表面增强拉曼散射（SERS）的优点是无需样品预处理即可快速、无损、高灵敏度地检测样品。SERS在检测和分析的各个领域都引起了极大的关注。在本文中，提出了将银纳米粒子附着到金属单壁碳纳米管上形成碳纳米管/银纳米粒子（CNTs/AgNPs）结构，然后将其与二氧化硅片结合的方法。具有二氧化硅/碳纳米管/银纳米粒子（SiO ₂/CNTs/AgNPs)复合结构具有极高的增强效果。在对被检测物质的定量分析中，通过数学拟合或机器学习来寻找拉曼信号强度与被检测物质浓度之间的关系。本研究设计了遗传算法优化的BP神经网络(GA-BP)。GA-BP的权值增强BP神经网络的鲁棒性，设置自适应学习率和隐藏节点数的方法，解决GA-BP容易陷入局部最优的问题，从而建立定量分析PR溶液浓度模型。该模型能够快速、简单、灵敏地高精度检测不同浓度的PR溶液。最后，与其他已发表的量化模型相比，²由模型的训练结果确定为0.99996，预测的均方根误差为RMSEP = 0.002510，比数学拟合方法高0.0005，表现出更好的性能。本研究提出了 SERS 衬底制备和水中 PR 浓度在线检测的可靠思路。