Quantitative Analysis of Spectral Data Based on Stochastic Configuration Networks

In quantitative analysis of spectral data, traditional linear models have fewer parameters and faster computation speed. However, when encountering nonlinear problems, their predictive accuracy tends to be lower. Nonlinear models provide higher computational accuracy in such situations but may suffer from drawbacks such as slow convergence speed and susceptibility to get stuck in local optima. Taking into account the advantages of these two algorithms, this paper introduces the single-hidden layer feedforward neural network named stochastic configuration networks (SCNs) into chemometrics analysis. Firstly, the model termination parameters, that is, the error tolerance and the allowed maximum number of hidden nodes are analyzed. Secondly, times of random configuration are discussed and analyzed, and then the appropriate number is determined by considering the efficiency and stability comprehensively. Finally, predictions made by the SCN are tested on two public datasets. The performance of the SCN is then compared with that of other techniques, including principal component regression (PCR), partial least squares (PLS), back propagation neural network (BPNN), and extreme learning machine (ELM). Experimental results show that the SCN has good stability, high prediction accuracy and efficiency, making it suitable for quantitative analysis of spectral data.