基于高斯过程建模的物联网数据不确定性度量与预测

物联网已经成为农业大数据最重要的数据源之一,自动观测数据的质量控制对农业生产分析以及基础科研数据应用非常重要。针对农业物联网观测的一类非平稳时间序列数据中的数据缺失、野值剔除、感知故障预警和长时间预测等问题,采用光滑弱假设高斯先验,构建了基于高斯过程的自回归模型表征的动态系统,并通过样本集学习,形成能考虑噪声干扰的传感变化规律建模,并可提供预测误差带用于预测数据的不确定性度量。针对原始数据的缺失和野值问题,采用基于高斯过程的短期预测,可补齐缺失数据,利用其不确定性度量可甄别数据野值,进行野值剔除与替换,并在此基础上判断感知故障;给出了基于输入数据不确定性传播的多步迭代预测方法,使长期预测仍可以跟踪农业数据的动态轨迹,并可为其预测值提供不确定性度量;将温室采集的真实传感数据用于分析试验,验证了高斯过程用于服务器端的农业时间序列数据采集质量控制的可行性。

Uncertainty Measurement and Prediction of IOT Data Based on Gaussian Process Modeling

The internet of things has become one of the most important data sources of agricultural big data, therefore automatic quality control of observational data is very important to agricultural production analysis and basic scientific data application. To solve the data missing, outliers excluding, perceived sensing failure and long-term prediction problems of the nonstationary time series data observed in agricultural systems, smooth Gaussian prior of weak assumptions on typical agricultural data was utilized; the dynamic system was built which was characterized by state space equations based on Gaussian process model; through the train set learning, the sensed variation models considered noise distribution were formed, and prediction error bar was provided with uncertainty measurement for the prediction data. For the problems of missing data and outliers excluding of raw data, short-term forecasts based on Gaussian process were adopted to fill with missing data, and its uncertainty measurement was used to detect outliers. Therefore, the outliers were removed and replaced with prediction value, and further sensing failure could be determined on the basis of accumulated outliers in certain time slice. The multi-step iterative method based on the uncertainty spread of input data was given for long-term prediction to track the dynamic trajectory of agricultural sensing data, and an uncertainty measurement could be provided for its predictive value. The data analysis of real sensed collection greenhouse microclimate verifies the feasibility of quality control of agricultural time series data based on Gaussian process in server-side.