基于模型驱动的田间数据压缩采集方法研究

Model-driven in-situ data compressive gathering

基于模型驱动的数据采集方法可有效降低数据传输能耗。阐述了差分自回归移动平均模型(ARIMA)、支持向量回归模型(SVR)、线性模型(DBP)和时钟驱动循环神经网络(CW-RNN)等数据预测模型的运行机制。结合空气温度、土壤湿度、果实膨大和风速等数据,探索预测模型的训练参数和误差阈值设置方法。综合考虑数据采集误差、业务数据传输率、模型更新及预测代价等指标,运用熵权逼近最优排序法(TOPSIS)评价模型适用性。结果表明:最佳训练参数与模型机制和数据对象有关,基于前期采样值自动获取误差阈值可行。模型的适用性与数据对象、节点运算资源和网络带宽有关。常量模型Constant的适用性最高,DBP模型次之。ARIMA模型可应用于带宽受限、节点运算资源较为充沛的应用场景,SVR模型可应用于高带宽、节点运算资源受限的应用场景。

The energy consumption of data transmission is effectively reduced when model-driven in-situ data compressive gathering method is employed.The mechanism was elaborated on about several data-prediction models,such as autoregressive integrated moving average model(ARIMA),support vector regression(SVR),derivative-based prediction(DBP)and clock work recurrent neural networks(CW-RNN).Moreover,their setting strategies of training parameters and error thresholds were explored based on air temperature,soil moisture,fruit growth and wind speed.Based on the gathered data error,data transmission ratio,model update and prediction overhead,the applicability of data-prediction models was evaluated using entropy weight technique for order preference by similarity to and ideal solution(TOPSIS).It was demonstrated that the optimal training parameters were relevant to model mechanism and data objects,and it was feasible to automatically acquire error thresholds based on previously gathered data.The model applicability was dependent on data objects,computation resources of nodes and network bandwidth.More specifically,Constant model had the greatest applicability,followed by DBP.ARIMA model was suitable to scenarios with limited bandwidth,but abundant computation resources.On the contrary,SVR model could be applied to scenarios with high bandwidth,but limited computation resources.