基于迁移学习的FDR土壤水分传感器自动标定模型研究

Automatic Calibration Model of FDR Soil Moisture Based on Transfer Learning

针对频域反射技术(FDR)传感器人工标定数据拟合误差大的问题,引入其他地区数据作为辅助数据,建立了基于迁移学习的自动标定模型。该模型将FDR目标使用地点采集的数据作为源域数据,结合辅助数据与少量源域数据,使用TrAdaBoost算法即可得到准确的FDR传感器标定模型。将面向分类问题的TrAdaBoost算法改进为适用于本文面向回归的TrAdaBoost算法,将TrAdaBoost算法的基学习器由AdaBoost改为XGBoost,改进了更新权重误差率的计算方法。首先使用XGBoost对辅助数据进行训练,得到初始标定模型;然后在目标地点采集少量数据,使用改进后的TrAdaBoost算法对初始标定模型进行校准,即可得到准确的FDR标定模型。将10个不同地区站点数据作为辅助数据,训练得到初始标定模型,将沈阳地区6个站点分别作为目标使用地点,取80%数据作为源域数据,进行模型校正,其余20%数据用于测试。测试结果的平均准确率为99.1%,说明基于迁移学习的自动标定模型是有效和准确的。

Aiming at the problem of large fitting error of manual calibration data for FDR sensors,the data from other regions were introduced as auxiliary data,and an automatic calibration model based on migration learning was established.In this model,historical data from other regions were introduced as auxiliary data.Data collected from FDR targets were used as source data.Combined with auxiliary data and a small amount of source data,an accurate FDR sensor calibration model can be obtained by using TrAdaBoost algorithm.TrAdaBoost algorithm for classification problem was improved to TrAdaBoost algorithm for regression.The basic learner of TrAdaBoost algorithm was changed from AdaBoost to XGBoost,which improved the calculation method of error rate when updating weight.Firstly,XGBoost was used to train the auxiliary data to get the initial calibration model,and then a small amount of data was collected from the target location of FDR,and the improved TrAdaBoost algorithm was used to calibrate the initial calibration model,so that the accurate FDR calibration model can be obtained.The data of 10 different regional sites were trained as auxiliary data to obtain the initial calibration model.For the six sites in Shenyang,the target sites were used respectively.Totally 80% of the data were used as the source domain data for model correction,and the remaining 20% were used for testing.The results showed that the average preparation rate using the calibration method was 99.1%,which indicated that the automatic calibration model using migration learning was effective and accurate.