一种基于Bagging-SVM的智能传感器集成学习方法

摘要

集成多个传感器的智能片上系统( SoC)在物联网得到了广泛的应用.在融合多个传感器数据的分类算法方面,传统的支持向量机( SVM)单分类器不能直接对传感器数据流进行小样本增量学习.针对上述问题,提出一种基于Bagging-SVM的集成增量算法,该算法通过在增量数据中采用Bootstrap方式抽取训练集,构造能够反映新信息变化的集成分类器,然后将新老分类器集成,实现集成增量学习.实验结果表明:该算法相比SVM单分类器能够有效降低分类误差,提高分类准确率,且具有较好的泛化能力,可以满足当下智能传感器系统基于小样本数据流的在线学习需求.%Intelligent system on chip( SoC)which integrates multiple sensors has been widely applied in Internet of Things. However,considering data fusion of multiple sensors,traditional SVM single-classifier can't directly support small sample incremental learning for sensor data stream. Aiming at above problem,put forward a kind of ensemble incremental algorithm based on Bagging-SVM,the algorithm supports incremental learning by combining original ensemble classifiers with the new ones which can reflect new information change of incremental data sets, while the new classifiers trained by the data sets which is extracted from incremental data sets by Bootstrap means. Experimental results show that the algorithm compared with the single-classifier can effectively decrease classification error,improve classification accuracy and has good generalization ability,which can smoothly meet requirements of intelligent sensor system for online learning based on small sample data flow.