基于机器学习算法的农田挥发氨多传感器阵列检测技术研究

【目的】设计能够快速、低成本、便捷检测农田挥发氨装置。【方法】构建基于二氧化锡(SnO_(2))半导体气体传感器阵列检测系统,并在新鲜空气(氨气质量浓度为0 mg·m^(-3))和氨气质量浓度分别为75.9、151.8、303.6 mg·m^(-3)条件下,以及混合有乙醇的空气、纯乙醇气体(质量浓度为151.8 mg·m^(-3))、混合有氨气的空气和纯氨气气体(质量浓度为151.8 mg·m^(-3))样品下,通过主成分分析法(principal component analysis,PCA)、K-最近邻算法(K-nearest neighbors,KNN)和支持向量机算法(support vector machine,SVM)对多传感器阵列响应稳态阶段和暂态阶段的数据进行分类处理,分析该系统对不同质量浓度氨气和混合气体环境下氨气的区分效果。【结果】该装置能够明显区分不同质量浓度氨气,稳态阶段的主成分1值超过90%。KNN与SVM算法稳态阶段平均准确率超过97%,暂态阶段平均准确率68%,KNN与SVM平均分类准确率为68%。【结论】该多传感器阵列检测系统不需要等待传感器进入稳态阶段便可以读取数据,有助于农田环境中氨气快速和连续检测。

Evolutionary Experience-Driven Particle Swarm Optimization with Dynamic Searching

Particle swarm optimization(PSO)algorithms have been successfully used for various complex optimization problems.However,balancing the diversity and convergence is still a problem that requires continuous research.Therefore,an evolutionary experience-driven particle swarm optimization with dynamic searching(EEDSPSO)is proposed in this paper.For purpose of extracting the effective information during population evolution,an adaptive framework of evolutionary experience is presented.And based on this framework,an experience-based neighborhood topology adjustment(ENT)is used to control the size of the neighborhood range,thereby effectively keeping the diversity of population.Meanwhile,experience-based elite archive mechanism(EEA)adjusts the weights of elite particles in the late evolutionary stage,thus enhancing the convergence of the algorithm.In addition,a Gaussian crisscross learning strategy(GCL)adopts cross-learning method to further balance the diversity and convergence.Finally,extensive experiments use the CEC2013 and CEC2017.The experiment results show that EEDSPSO outperforms current excellent PSO variants.