光谱学与光谱分析, 2020, 40 (9): 2981, 网络出版: 2020-12-01   

基于紫外-可见光谱与深度学习CNN算法的水质COD预测模型研究

Study on the Predication Modeling of COD for Water Based on UV-VIS Spectroscopy and CNN Algorithm of Deep Learning
贾文珅 1,2,3,4张恒之 2马洁 2梁刚 1,3,4王纪华 1,3,4刘鑫 5,*
作者单位
1 北京农业质量标准与检测技术研究中心, 北京 100097
2 北京信息科技大学自动化学院, 北京 100192
3 农业部农产品质量安全风险评估实验室, 北京 100097
4 农业农村部华北都市农业重点实验室, 北京 100097
5 北京海关技术中心, 北京 100026
摘要
水是生命之源, 作为人类生活的必需品, 水质的优劣直接关系到人们的健康生活。 目前, 关于水质COD在线检测方法的研究主要集中在光谱数据预处理和光谱特征提取, 而针对光谱数据建模方法的研究较少。 卷积神经网络(CNN)作为目前深度学习领域应用最广泛的网络模型, 具有强大的特征提取和特征映射能力, 本文将CNN与紫外-可见光谱分析法相结合, 建立了基于CNN的水质COD紫外-可见光谱预测模型。 模型使用Savitzky-Golay平滑滤波方法去除光谱噪声, 光谱输入卷积层提取光谱数据特征、 池化层降维、 全连接层映射全局特征, 通过ReLU和Adam优化方法, 从而得到模型的预测值。 通过实验发现, CNN模型具有较强的水质COD预测能力, 具有较高的预测精度和回归拟合优度, 通过与BP, PCA-BP, PLSR和RF等模型比较后发现, CNN模型的预测样本的RMSEP和MAE最小, R2最大, 模型拟合效果最优。 在与训练样本的模型效果评价对比后发现, 模型具有较强的泛化能力。 针对吸收光谱的波峰偏移对预测结果所造成的预测结果不准确的问题, 作者还提出了一种基于CNN的分类回归模型卷积神经网络增强模型(CNNs), 去噪后的光谱数据通过CNN分类模型按照吸收波峰的不同特征分为三类, 分别输入对应CNN回归模型进行预测。 实验结果表明, 分段式CNNs模型比整体式CNN模型的拟合效果更好, 预测精度更高, R2达到0.999 1, 测试样本的MAE和RMSEP分别为2.314 3和3.874 5, 比CNN分别下降了25.9%和21.33%, 效果显著。 通过对预测模型的性能测试, 计算得出检出限为0.28 mg·L-1, 测量范围为2.8~500 mg·L-1。 本文创新的将卷积神经网络与光谱分析方法相结合, 为光谱分析方法在水质COD吸收光谱建模的应用开拓了新思路。
Abstract
Water is vital for human life, and the quality of water is directly related to people’s quality of life. At present, research into chemical oxygen demand (COD) methods for determining water quality is mainly focused on spectral data preprocessing and spectral feature extraction, with few studies considering spectral data modeling methods. Convolutional neural networks (CNN) are known to have strong feature extraction and feature mapping abilities. Thus, in this study, a CNN is combined with UV-visible spectroscopy to establish a COD prediction model. The Savitzky-Golay smoothing filter is applied to remove noise interference, and the spectral data are then input to the CNN model. The features of the spectrum data are extracted through the convolution layer, the spatial dimensions are reduced in the pooling layer, and the global features are mapped in the fully connected layer. The model is trained using the ReLU activation function and the Adam optimizer. A series of experiments show that the CNN model has a strong ability to predict COD in water, with a high prediction accuracy and good fit to the regression curve. A comparison with other models indicates that the proposed CNN model gives the smallest RMSEP and MAE, the largest -R2, and the best fitting effect. It is found that the model has strong generalization ability through the evaluation effect of the training samples. To counter the inaccuracy of the predicted results caused by the peak shift of the absorption spectrum, a regression model based on a strengthened CNN (CNNs) is also developed. After denoising, the spectral data can be divided into three categories according to the different characteristics of absorption peaks, and the corresponding CNN regression model is input respectively for prediction. When the corresponding regression model is applied, the experimental results show that the sectional CNNs model outperforms our original CNN model in terms of fitting, prediction precision, determination coefficient, and error. Not only does R2 increase significantly, reaching 0.999 1, but also the MAE and RMSEP of the test samples also reduced to 2.314 3 and 3.874 5, respectively, which were reduced by 25.9% and 21.33% compared with out original CNN. Performance testing of the prediction model, indicates that the detection limit is 0.28 mg·L-1and the measurement range is 2.8~500 mg·L-1. This paper describes an innovative combination of a CNN with spectral analysis and reports our pioneering ideas on the application of spectral analysis in the field of water quality detection.

贾文珅, 张恒之, 马洁, 梁刚, 王纪华, 刘鑫. 基于紫外-可见光谱与深度学习CNN算法的水质COD预测模型研究[J]. 光谱学与光谱分析, 2020, 40(9): 2981. JIA Wen-shen, ZHANG Heng-zhi, MA Jie, LIANG Gang, WANG Ji-hua, LIU Xin. Study on the Predication Modeling of COD for Water Based on UV-VIS Spectroscopy and CNN Algorithm of Deep Learning[J]. Spectroscopy and Spectral Analysis, 2020, 40(9): 2981.

本文已被 1 篇论文引用
被引统计数据来源于中国光学期刊网
引用该论文: TXT   |   EndNote

相关论文

加载中...

关于本站 Cookie 的使用提示

中国光学期刊网使用基于 cookie 的技术来更好地为您提供各项服务,点击此处了解我们的隐私策略。 如您需继续使用本网站,请您授权我们使用本地 cookie 来保存部分信息。
全站搜索
您最值得信赖的光电行业旗舰网络服务平台!