立足于成功研制的紫外-可见光谱水质检测多参数测量系统, 针对紫外-可见光谱水质多参数原位实时检测在精度、 灵敏度、 稳定性等方面的实际需要, 开展了基于二维重组和动态窗格的水质检测紫外-可见光谱去噪算法的研究, 以此提高紫外-可见光谱水质检测的测量精度。 光谱法水质检测系统通常使用工业级低成本光谱仪, 其输出光谱包含明显的非平稳噪声。 传统去噪法难以在滤除噪声的同时保留谱线细节。 而且, 原位实时水质检测条件下, 被测水样可能快速变化, 传统去噪法中常用的多次采样求均值法将产生额外的测量误差。 引入的去噪算法通过对水样光谱进行等间隙连续采样, 将光谱数据张成由光谱轴和时间轴构成的二维矩阵, 经过二维小波变换后, 设置一个可变宽度的窗格在系数矩阵中水平滑动, 使用窗格内的小波系数计算得到动态去噪阈值, 并随窗格滑动构建去噪阈值向量, 由此进行光谱去噪。 其中, 窗格宽度由相邻区域的噪声方差变化率决定, 变化率较高的区域缩小窗格宽度, 反之则扩大宽度。 实验结果表明, 这种去噪算法不仅能有效去除光谱中的非平稳噪声, 而且能保留光谱的细节信息, 有助于提高仪器的测量精度。 与此同时, 由于该算法并未使用时域平均, 样本的快速变化对去噪性能的影响较小, 适合在线或原位水质检测的水样本环境。

Based on the successful development of multi-parameter water quality detection system of UV-visible spectroscopy and the actual needs of accuracy, sensitivity, stability and other aspects in the measurement, the research is carried out to create a denoising algorithm of UV-visible spectroscopy on water quality detection based on Two-Dimension(2-D) restructuring and dynamic pane. As spectrometry water quality detection systems typically use low-cost industrial grade spectrometer, the CCD photon efficiency and stability are lower than research grade spectrometer, which is built with back-illuminated CCD and internal cooling thermostats. The output spectrum contains significant non-stationary noise, especially in the UV section and IR section. With the traditional denoising method, it is difficult to filter out the noise and retain spectral details at the same time. What’s more, in the case of online or in-situ real-time water quality measurement, the multiple-sample averaging method that commonly used in traditional denoising method may incur additional measurement error due to rapidly changed water sample. The new denoising algorithm proposed by this paper uses the continuous sampling with isochronous gap to expand spectral data into a 2-D matrix composed of spectrum and time axes. After a 2-D wavelet transformation, a variable-width pane which is able to slide horizontally in the coefficient matrix is set. The width of the pane is determined by the change rate of noise variance: the more rapid the rate changes, the narrower the width is. A dynamic denoising threshold is calculated with the wavelet coefficients in the pane and a threshold vector is created with pane sliding. Finally, the spectrum can be denoised by the threshold vector with wavelet shrinkage method. The experimental results show that this denoising algorithm not only removes the spectral non-stationary noise effectively, but also retains the spectral details, which is helpful to improve the accuracy of the instrument. Meanwhile, since time-domain average is not used here, the impact to the denoising performance on fast-changing of water samples is small, which is suitable for the online or in-situ water quality detection environment.