基于神经网络模型的海水硝酸盐测量方法研究

doi:10.3964/j.issn.1000-0593(2020)10-3211-06

Abstract
References
Related Citation (15)

Download: PDF (2818 KB)
Export: BibTeX | EndNote (RIS)

Abstract Nitrate concentration is an important indicator for the marine ecosystem. Compared with laboratory chemical methods such as Cadmium-Reduction method, in-situ nitrate optical sensor is much faster and reagent-free in a long time and continuous monitoring. Partial Least Squares (PLS) method is often used in ultraviolet absorption spectrum modeling, which is difficult to optimize and has low generalization ability. The neural network can compel any no-linear function by any precision, which has high generalization ability in the modeling. A neural network model is established in the in-situ nitrate sensor to measure the nitrate concentration in seawater in which the nitrate concentration range is 30～750 μg·L^-1. Double-hidden layer neural network model is determined to adopt by contrasting performance of single-hidden layer and double-hidden layer to measure nitrate concentration，the input layer is absorption spectrum from 200 to 275 nm, the output layer is nitrate concentration, and sigmoid function is used as the activation function. Gradient descent method is used to update weighting parameters for the neural network of each layer, after 55 000 times iteration, network training is conducted based on the learning rate of 0.26. After validation for the blind test of the model through 8-group randomized validation data, the nitrate concentration using double-hidden layer neural network model is higher in linear correlation to its actual concentration (R²=0.997) in which the Root Mean Squared Error is 10.864, average absolute error is 8.442 μg·L^-1, average the relative error is 2.8%. Compared with single-hidden layer neural network model, the double-hidden layer neural network model has higher accuracy in which the average relative error is reduced by 4.92%, the Root Mean Squared Error of PLS is 4.58% using the same spectral data, while the mean relative error is 11.470. The result shows that the neural network model is much better than the Partial Least Squares model under certain conditions. It verifies the superiority of the neural network model applied to the nitrate concentration measurement by ultraviolet absorption spectrometry. The application test was carried out on the “Environmental Monitoring 01” monitoring vessel of the Ministry of Natural Resources，the measurement results are basically identical with the laboratory method in 11 stations, which is further proved from the reliability and practicality.

Key words：Ultraviolet; Absorption spectroscopy; Seawater; Nitrate; Neural network; Spectral analysis

Received: 2019-09-06 Accepted: 2020-01-10

ZTFLH:

X853

Corresponding Authors: FENG Wei-wei E-mail: wwfeng@yic.ac.cn

Cite this article:

URL:

https://www.gpxygpfx.com/EN/10.3964/j.issn.1000-0593(2020)10-3211-06 OR https://www.gpxygpfx.com/EN/Y2020/V40/I10/3211