紫外-可见吸收光谱法测量水质化学需氧量(COD), 本质是对大量水质光谱数据建模, 以此模型为基础引入待测的水质光谱数据进行预测的过程。 而实测的邻苯二甲酸氢钾COD标准溶液在200～300 nm存在两个特征吸收峰, 标准溶液在不同浓度下的峰值也不同, 利用此特性对该波段进行特征波长的选择, 用其表征光谱信息, 降低数据冗余度的同时提高了预测精度。 针对实测水质光谱信号容易受到仪器本身和外界干扰, 光谱数据存在大量非平稳噪声, 且特征吸收峰及其临近信号频率较高, 常规去噪算法直接舍弃高频信号以及无法准确判断信噪分量界限, 导致有效信号缺失这一实际问题。 提出了一种基于完全自适应噪声集合经验模态分解(CEEMDAN)和双树复小波变换(DT-CWT)的联合去噪算法。 该联合算法利用CEEMDAN将信号分解为本征模态函数(IMF), 并通过归一化自相关函数和互相关系数进行线性相关性分析, 得到各阶IMF分量之间的自相关性以及IMF分量与原始信号的互相关系数, 以确定高频含噪分量与低频信号分量的界限; 进而应用DT-CWT阈值去噪算法对含噪高频IMF分量进行处理, 将DT-CWT处理之后的IMF高频分量与CEEMDAN分解得到的IMF低频分量进行信号重构, 获得最终去噪后的水质光谱信号。 实验结果表明: 基于CEEMDAN联合双树复小波变换的去噪算法适用于紫外-可见光谱水质检测的数据处理。 对于化学需氧量COD标液为100 mg·L-1的邻苯二甲酸氢钾溶液, 将实测的紫外-可见光谱数据应用该算法去噪后的SNR=24.201 5 dB, RMSE=0.024 0, NCC=0.999 4, PSNR=37.573 6, 不仅去噪效果显著优于CEEMDAN和双树复小波阈值算法, 还有效地保留了原始COD标液的吸收特征峰, 遏制了平移敏感性现象, 提高了重构信号的平滑度, 改善了重构信号质量。 为紫外-可见光谱法检测水质COD提供了一种新的数据预处理方法。

The essence of measuring water quality COD by UV-vis absorption spectrometry is to model a large number of spectral data, and then introduce the measured spectral data to predict the process. However, there are two characteristic absorption peaks in the measured COD standard solution of potassium hydrogen phthalate at 200～300 nm, and the peak and peak values of the standard solution are also different at different concentrations. This feature is used to select the characteristic wavelength of this band and use it to characterize the spectral information, which reduces the data redundancy and improves the prediction accuracy. Because the measured water quality spectral signal is easily disturbed by the internal and external interference, resulting in a large number of non-stationary noise in the spectral data, and the characteristic absorption peak and its adjacent signal frequency is high, conventional denoising algorithms directly abandon high-frequency signals and can not accurately judge the limits of signal-to-noise components, resulting in the lack of effective signals. A joint denoising algorithm based on fully adaptive noise set empirical mode decomposition CEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) and dual-tree complex wavelet transform DT-CWT (The Dual-Tree Complex Wavelet Transform) is proposed. The algorithm uses CEEMDAN to decompose the signal into intrinsic mode function IMF (Intrinsic Mode Function). It makes linear correlation analysis through normalized autocorrelation function and cross-correlation number to determine the boundary between high-frequency noise components and low-frequency signal components. Then the DT-CWT threshold denoising algorithm is used to process the noisy high-frequency IMF component, and the IMF high-frequency component after DT-CWT processing is reconstructed from the IMF low-frequency component demarcated by CEEMDAN, and the final denoised signal is obtained. The experimental results show that the denoising algorithm based on CEEMDAN combined with dual-tree complex wavelet transform is suitable for data processing of UV-Vis spectrum water quality detection. For potassium hydrogen phthalate solution whose chemical oxygen demand (COD) standard solution is 100 mg·L-1, the denoising effect of SNR=24.201 5 dB, RMSE=0.024 0, NCC=0.999 4 and PSNR=37.573 6 denoised by the combined algorithm is significantly better than that of CEEMDAN and double-tree complex wavelet threshold algorithm. Moreover, it effectively retains the characteristic absorption peak of the original COD standard solution, suppresses the translation sensitivity and improves the smoothness of the reconstructed signal. The quality of the reconstructed signal is improved. It provides a new data pre-processing method for detecting water quality COD by UV-Vis spectrum.