Chemical enterprises generate significant amounts of organic pollutants during production, often due to inadequate or aging pollution treatment facilities. These shortcomings pose a risk of organic contamination to nearby groundwater. Among these pollutants, polycyclic aromatic hydrocarbons (PAHs) are persistent organic compounds known to be carcinogenic, teratogenic, and mutagenic, causing significant harm to ecosystems and human health. Groundwater contamination with PAHs is often difficult to detect due to complex sample characteristics and the limitations of traditional detection methods, such as cumbersome operational procedures and secondary pollution caused by chemical reagents. Therefore, there is an urgent need for a rapid, reliable method for detecting PAHs in groundwater. Three-dimensional fluorescence spectroscopy offers a rapid and sensitive approach to water analysis. However, the overlapping fluorescence characteristics of PAHs make it challenging for conventional linear models to achieve accurate qualitative and quantitative analysis of multi-component samples. In addition, real-world groundwater conditions often introduce fluorescence interferences, further complicating the analysis. In this context, an effective solution for detecting multi-component PAHs based on three-dimensional fluorescence spectra is essential.

In this paper, we utilize the fluorescence characteristics of organic pollutants to measure the three-dimensional excitation?emission matrix (3DEEM) of PAH solutions using a fluorescence spectrophotometer. A convolutional neural network (CNN)-based approach is proposed to enable rapid qualitative and quantitative analysis of PAHs in groundwater. Two distinct CNN models are developed for the qualitative identification and quantitative measurement of PAHs. The fluorescence characteristics of eight representative PAHs are analyzed using correlation statistical methods applied to deionized water. To enhance the spectral dataset, two 3DEEM data augmentation techniques, superposition and interpolation, are applied. The expanded dataset is then used to train the CNN models. The method is validated in laboratory conditions with PAH solutions and further tested on groundwater samples collected from areas surrounding chemical enterprises.

For the PAH solutions prepared with deionized water under laboratory conditions, the fluorescence characteristics of eight PAHs are preliminarily analyzed. Two CNN models are then trained using the proposed method. The model achieves a qualitative analysis accuracy of 99.8% and an average relative error of 10.71% for quantitative analysis. The models are further retrained to account for the fluorescence background of actual groundwater samples, and the method is applied to detect PAH mass contamination in real-world groundwater samples. In cases where the fluorescence background of the groundwater is unknown, the method still provides reliable qualitative analysis results. The relative error in quantitative analysis, despite interference from water background disturbance and varying levels of pollution, is maintained within an acceptable range. For instance, in groundwater sample 1, which has a low fluorescence background, the model achieves 100% qualitative accuracy, and the quantitative analysis errors range from 0 and 40 μg/L. In groundwater sample 2, which has a higher fluorescence with a broader range of overlapping PAH characteristics, the model achieves 95.83% accuracy for qualitative analysis. After removing PAHs that are significantly affected by the groundwater background spectrum, the average relative error in quantitative analysis drops to 27.53%, demonstrating the model’s effective prediction capabilities and high analysis efficiency.

In this paper, we propose a novel method for the detection of PAHs in groundwater using fluorescence spectroscopy and CNN models. By leveraging an extensive spectral dataset for training, the proposed method achieves effective qualitative and quantitative analysis of PAHs in both deionized water and real-world groundwater samples. Detailed results from different groundwater locations demonstrate the method’s practical applicability. The experimental results show that this method offers a viable and efficient solution for the rapid on-site analysis of groundwater contamination by multi-component organic pollutants. Improvement measures based on model performance in real-world scenarios are also discussed.