基于WDNN的温室多特征数据融合方法研究

目前物联网监测产品在温室生产中大量应用产生海量数据,但现有用于温室数据融合算法对高维特征及混合特征(数据同时包含稀疏特征和连续特征)处理精度较低且泛化能力较弱,鲜有利用深度学习模型对温室数据进行顶层融合并提供准确的决策信息。本文提出了一种基于宽-深神经网络(Wide-deep neural network,WDNN)的两级温室环境数据融合算法。首先利用温室内多点多特征数据训练WDNN深度学习模型,输出形式为多点单特征,再将该输出数据按照少数服从多数原则进行融合,得到温室环境状态的整体评估结果。试验结果表明,该融合方法对预测集中混合特征的决策准确率达到98. 90%,融合特征类型的增加,可用于监测参数更多、环境更复杂的温室,将WDNN模型用于温室混合数据融合是可行有效的,在保证决策精度的同时丰富了可融合特征类别,进一步提升温室融合系统的智能化程度,对温室智能调控提供有效技术支撑。

Multi-feature Data Fusion Method of Greenhouse Based on WDNN

The IoT monitoring products are widely used in greenhouse production, which could generate massive data. The existing data fusion algorithms for greenhouses had low fusion accuracy and weak generalization capability for high-dimensional features and mixed features (combined with sparse features and continuous features). It was rare to use the deep learning model to top-level fusion of greenhouse data and provide accurate decision information. Aiming at the above problems, a two-level greenhouse environment data fusion algorithm was proposed based on wide-deep neural network (WDNN). Firstly, integrating multi point multi-features mixed data in the greenhouse and marking the data categories. Then the constructed training set and test set were input into the WDNN deep learning model for 2000 step iteration training. The model structure was set as 7-100-50-7, the output form was multi point single feature, which was the first-level fusion result as decision information of each area of the greenhouse, and then the output data was second level fusion according to the minority obeyed majority principle, and the overall evaluation decision of the greenhouse environmental state was obtained. For comparison purposes, the other three fusion models were trained as deep neutral network (DNN), BP neural network (BPNN) and random forest (RF). The experimental results showed that the loss value of the initial segment of the WDNN network was higher than that of DNN network, but the loss function curve had a faster rate of decline and the model parameters were better. The accuracy of the model after training was 4.32 percentage points higher than that of DNN, but the training time was increased by 21.36%;the accuracy of BPNN model was 82% and its parameter optimization was the slowest, parameter optimization required more iteration steps;RF model training speed was the fastest, but its model fusion accuracy was 3.39 percentage points lower than that of WDNN. The fusion accuracy was insufficient;above comparison results proved that it was feasible and excellent to use the WDNN model to fuse the mixed data in the greenhouse. Inputting the mixed situation information contained the sensor anomaly and meteorological data under various conditions into the fusion system, then the context decision rate reached 98.90%. The realization of the WDNN fusion system could be used to monitor greenhouses with more parameters and more complex environments, and enrich the fusion feature categories while ensuring the accuracy of decision making. It could further improve the intelligence degree of the greenhouse fusion system.