Near infrared spectroscopy quantitative analysis model based on incremental neural network with partial least squares

This paper proposes an near infrared spectroscopy quantitative analysis model based on incremental neural network with partial least squares. The proposed model adopts the typical three-layer back-propagation neural network (BPNN), and the absorbance of different wavelengths and the component concentration are the inputs and the outputs, respectively. Partial least square (PLS) regression is performed on the history training samples firstly, and the obtained history loading matrices of the independent variables and the dependent variables are used for determining the initial weights of the input layer and the output layer, respectively. The number of the hidden layer nodes is set as the number of the principal components of the independent variables. After a set of new training samples is collected, PLS regression is performed on the combination dataset consisting of the new samples and the history loading matrices to calculate the new loading matrices. The history loading matrices and the new loading matrices are fused to obtain the new initial weights of the input layer and the output layer of the proposed model. Then the new samples are used for training the proposed mode to realize the incremental update. The proposed model is compared with PLS, BPNN, the BPNN based on PLS (PLS-BPNN) and the recursive PLS (RPLS) by using the spectra data of flue gas of natural gas combustion. For the concentration prediction of the carbon dioxide in the flue gas, the root mean square error of prediction (RMSEP) of the proposed model are reduced by 27.27%, 58.12%, 19.24% and 14.26% than those of PLS, BPNN, PLS-BPNN and RPLS, respectively. For the concentration prediction of the carbon monoxide in the flue gas, the RMSEP of the proposed model are reduced by 20.65%, 24.69%, 18.54% and 19.42% than those of PLS, BPNN, PLS-BPNN and RPLS, respectively. For the concentration prediction of the methane in the flue gas, the RMSEP of the proposed model are reduced by 27.56%, 37.76%, 8.63% and 3.20% than those of PLS, BPNN, PLS-BPNN and RPLS, respectively. Experiments results show that the proposed model could optimize the construction and the initial weights of BPNN by PLS and has higher prediction effectiveness. Moreover, based on the information of the built model, the proposed model uses the new samples for incremental update without accessing the history samples. Hence, the proposed model has better robustness and generalization.