基于Kalman-PNN协同融合的悬移质含沙量测量

刘明堂 陈健 刘书晓 刘佳琪 吴思琪 江恩慧 刘雪梅

摘 要：针对悬移质含沙量在线测量易受环境因素干扰的问题，通过分析卡尔曼滤波（Kalman filter）特性与概率神经网络（PNN）数据融合特点，提出基于卡尔曼滤波和概率神经网络（Kalman-PNN）的协同融合模型。首先应用卡尔曼滤波器对含沙量传感器输出值进行无偏估计，减少含沙量传感器的噪声干扰;然后将含沙量信息和水温、深度、流速等环境信息进行多源数据融合处理，进一步消除环境因素对含沙量测量的影响;最后经过Kalman-PNN协同融合处理，得到更加精确的含沙量实测值。为了验证Kalman-PNN协同融合模型的数据处理效果，在相同试验条件下进行了一元线性回归（ULR）模型、多元线性回归（MLR）模型、PNN模型与BP神经网络模型的含沙量数据处理。通过误差比较分析发现，基于Kalman-PNN协同融合模型的试验数据平均绝对误差仅为11.72 kg/m3，而一元线性回归模型、多元线性回归模型、PNN模型与BP神经网络模型的分别为103.12、56.02、12.47、49.81 kg/m3。试验结果表明，基于Kalman-PNN的协同融合模型对含沙量测量精度的提升具有积极作用。

关键词：悬移质含沙量;在线测量;卡尔曼滤波;概率神经网络;协同融合

中图分类号：TV149;TP274+.4 文献标志码：A

doi：10.3969/j.issn.1000-1379.2021.05.003

Abstract： Aiming at the issues of susceptibility to environmental factors in the online measurement of suspended sediment concentration， by analyzing the characteristics of Kalman filter and Probabilistic Neural Network （PNN） data fusion， a new method based on the collaborative fusion model of Kalman and PNN （Kalman-PNN） was firstly applied Kalman filter to unbiased estimation of the output of the sand content sensor to reduce the noise interference of the sand content sensor; multi-source data fusion processing of environmental temperature information such as water temperature， depth and flow velocity further eliminated the influence of environmental factors on sand content measurement. Finally， the Kalman-PNN fusion processing was performed to obtain more accurate measured sand content values. In order to illustrate the data processing effect of the Kalman-PNN fusion model， under the same experimental conditions， it conducted data processing of sediment concentration by using univariate linear regression （ULR） model， multiple linear regression （MLR） model， PNN model and BP neural network model. Through error comparison analysis， the average absolute error of experimental data based on the Kalman-PNN fusion model was only 11.72 kg/m3， while the univariate linear regression model， multiple linear regression model， PNN model and BP neural network model were 103.12 kg/m3， 56.02 kg/m3， 12.47 kg/m3 and 49.81 kg/m3 respectively. The experimental results show that the Kalman-PNN-based collaborative fusion model has a positive effect on improving the accuracy of sediment concentration measurement.

Key words： suspended sediment content; online measurement; Kalman filter; probabilistic neural network; collaborative fusion model

泥沙問题在黄河研究及其各类规划中占有不可或缺的重要地位[1-2]。长期以来，人们一直尝试利用各种方法对悬移质含沙量进行测量。比重瓶法属于人工取样比重法，其测量结果较为准确，但操作步骤烦琐、数据结果获取周期长，难以满足含沙量在线测量的时效性要求，通常只作为一段时间内的数据参考[3]。……