The treatment of organic pollutants in surface water, drinking water, and wastewater is one of the urgent social problems to be solved in the development of human society. Three-dimensional excitation-emission matrix (3D-EEM) fluorescence spectroscopy technology has been widely used to detect fluorescence components in surface water, sewage, and other samples. There are a lot of interference noises and fluorescence overlap information in the original 3D-EEM data, so there is an urgent need for a fast and accurate method to extract and analyze the useful information in 3D-EEM spectra. At present, parallel factor analysis (PARAFAC) is commonly used to decompose the overlapping fluorescence signals in 3D-EEM, but the analysis process of this method is complex, and the data set is strict, which greatly limits the on-line monitoring and analysis of samples. In this study, according to the results of PARAFAC, we propose a convolutional fast classification and recognition network model, which can quickly obtain water sample types, mass concentration grades, and fluorescent component maps by using only two convolutional neural network (CNN) models. As a result, it provides effective technical means for rapid detection of scenes such as surface water, drinking water, wastewater monitoring, and so on.

In this study, a method of water sample classification and fluorescence component fitting based on MobileNetV2, VGG11 component fitting (CF-VGG11) CNN, and PARAFAC is proposed. The 3D-EEM data of four types of water samples including surface water (DB), treated industrial wastewater (FS), sewage treatment plant inlet and outlet water (WS), and rural drinking water (XCYY) are collected, and the multi-output classification model of different water samples and the prediction and fitting model of fluorescence component maps are established with the results of PARAFAC as labels. The prediction of types and components is completed in two steps. In the first step, the MobileNetV2 algorithm is used to predict and classify different water samples. The second step is to use the CF-VGG11 network to fit the fluorescence component map.

The data sets of all kinds of water samples are analyzed by PARAFAC, and four fluorescence components are shown (Fig. 6). Then, the PARAFAC results are uploaded to the OpenFluor database to obtain possible substances of various types of fluorescence components in water samples (Table 2). The similarity comparison scores of all components are more than 95%. Combined with the PARAFAC results as network labels, the MobileNetV2 classification network and CF-VGG11 component fitting network obtain a classification accuracy of 95.83% and a component fitting accuracy of 98.11%, respectively (Table 3). In order to show that the trained model has good classification and fitting performance, a part of untrained 3D-EEM data is selected for the test, and the results show that MobileNetV2 and CF-VGG11 can classify and fit the 3D-EEM of water samples very well (Fig. 7), and MobileNetV2 and CF-VGG11 network models have certain advantages compared with PARAFAC in terms of time cost, data requirement, and analysis process (Table 4).

In this study, a fast CNN classification and recognition algorithm based on fluorescence spectrum is proposed to predict the types and mass concentrations of different water samples, as well as the overlapping fluorescence components in 3D-EEM. This study relies on PARAFAC for preliminary data preparation and MobileNetV2 network for classification of water sample types and mass concentration grades, which can achieve water pollution traceability and exceedance warning, and the CF-VGG11 network is used to fit the fluorescence component map of water samples. The results show that the fast classification and identification network model based on the results of PARAFAC can quickly predict the types and mass concentration grades of water samples and fit their specific fluorescence components by inputting 3D-EEM data of a single water sample, and there is no need to repeat the complex PARAFAC. Therefore, this study provides certain theoretical support for detecting water pollution by three-dimensional fluorescence spectrometry and is of a certain practical significance.