Quantitative Analysis of NaCl Aerosols Based on Convolutional Neural Network and Filament-Induced Fluorescence Spectroscopy

Objective Quantitatively analyzing aerosol composition is key to monitoring air pollution. In this study, filament-induced fluorescence spectroscopy (FIFS) technology is used to detect NaCl aerosols remotely based on intense femtosecond laser pulses that excite materials and induce their fingerprint fluorescence. However, the fluorescence intensity does not vary linearly with the different mass concentrations of NaCl aerosols because of the self-absorption effect. Thus, a one-dimensional convolutional neural network (1D-CNN) is proposed for predicting the mass concentration of NaCl aerosols with FIFS spectra because deep neural networks can well fit non-linear relationships. In the future, FIFS technology with deep neural networks might be key for the development of nextgeneration laser lidar for remote air pollution detection systems. Methods An FIFS system with ultra-short, ultra-intense laser pulses that generate high-intensity filaments is first built. The method based on the use of a filament focused by a combined lens in air can be employed for the multi-component, remote, and rapid quantitative analysis of atmospheric aerosols. Next, a spectrum collection system based on FIFS is designed (Fig. 1) to acquire aerosol fluorescence spectra. The system uses a Coherent commercial Ti : sapphire femtosecond laser amplifier (Legend Elite) with a pulse wavelength of 800 nm, a pulse energy of 6 mJ, and a frequency of 500 Hz. Using the telescope focusing system, the filamentation position of the femtosecond laser is fixed in the cloud chamber at 30 m. Moreover, a NaCl aerosol is chosen as the experimental sample to simulate aerosols in the air. Specifically, an aerosol generator is used to generate the NaCl aerosol in the cloud chamber, where the femtosecond laser excites it and induces thin filaments. Fluorescent signals are collected using a spectrometer. Finally, a 1D-CNN model (Fig. 5) is designed to collect the FIFS spectra and predict mass concentration of the NaCl aerosol. To construct distinguishable features of the spectra, the 1D-CNN is set up with two convolution and two pooling layers, and the constructed features are inserted into the full connection layer to obtain the predicted value. To prevent gradient explosion, ReLU is selected as the activation function of the 1D-CNN. Results and Discussions A 10-fold cross-validation comparison experiment was conducted with traditional quantitative models, back propagation neural network (BPNN), and 1D-CNN on the full and characteristic spectral data. Generalized prediction experiments were performed for each model to further verify the reliability of the proposed model. According to the results of the 10-fold cross-validation experiments on NaCl spectral data (Tables 1 and 2), the 1D-CNN outperforms the other models in all prediction performance measures on the full spectrum dataset. It can achieve a root mean square error (RMSE), mean absolute error (MAE), coefficient of determination (R-2), relative percentage deviation (RPD), and accuracy (ACC) of 0.110, 0.073, 0.997, 18.478, and 0.99, respectively. Its performance is further improved when executed on the characteristic spectrum dataset. This result indicates that the 1D-CNN can fit a non-linear relationship. When comparing the results on the full and characteristic spectra, most of the models perform better on the characteristic spectrum than on the full spectrum (Fig. 6). In the generalization experiments (Tab. 5), the 1D-CNN performs poorly only on the lowest concentration (0.33 mg/m(3)) but performs well when predicting higher mass concentrations. The RMSE, MAE, and ACC of the 1D-CNN on the characteristic spectrum dataset are 0.34, 0.31, and 0.87, respectively, in the generalization experiments. The 1D-CNN outperforms the other models when predicting mass concentrations that have not appeared in the training data (Tables 3 and 4). The results indicate that the 1D-CNN model can be generalized to NaCl aerosols with unknown mass concentrations. Conclusions An experimental system is built to collect the FIFS spectra of NaCl aerosols and predict its mass concentration based on the proposed 1D-CNN. Compared with baseline models, the convolution and pooling layers of the 1D-CNN can generate spectral characteristics to improve prediction accuracy. The results of 10-fold cross-validation experiments show that the 1D-CNN and BPNN models have unique advantages over CR, MLR, and PLSR. In addition, the 1D-CNN performs significantly better than the other models in the generalization experiments. This indicates that FIFS technology and 1D-CNNs are suitable for the quantitative analysis of FIFS spectra of NaCl aerosols. Hence, they can be the core technology of next-generation laser lidar for monitoring air pollution.