基于多目标PSO混合优化的虚拟样本生成

Virtual Sample Generation Method Based on Hybrid Optimization With Multi-objective PSO

受限于检测技术难度、高时间与经济成本等原因,难测参数的软测量模型建模样本存在数量少、分布稀疏与不平衡等问题,严重制约了数据驱动模型的泛化性能.针对以上问题,提出一种基于多目标粒子群优化(Multi-objective particle swarm optimization, MOPSO)混合优化的虚拟样本生成(Virtual sample generation, VSG)方法.首先,设计综合学习粒子群优化算法的种群表征机制,使其能够同时编码用于连续变量和离散变量;然后,定义具有多阶段多目标特性的综合学习粒子群优化算法适应度函数,使其能够在确保模型泛化性能的同时最小化虚拟样本数量;最后,提出面向虚拟样本生成的多目标混合优化任务以改进综合学习粒子群优化算法,使其能够适应虚拟样本优选过程的变维特性并提高收敛速度.同时,首次借鉴度量学习提出用于评价虚拟样本质量的综合评价指标和分布相似指标.利用基准数据集和真实工业数据集验证了所提方法的有效性和优越性.

Due to the difficulty of detection technology,and high time and economic cost,the modeling samples of soft-sensing model with difficult parameters have some problems,such as small numbers,sparse distribution,and imbalance,which seriously restrict the generalization performance of data-driven models.To solve the above problems,a virtual sample generation(VSG)method based on multi-objective particle swarm optimization(MOPSO)hybrid optimization is proposed.First,the population representation mechanism of the integrated learning particle swarm optimization algorithm is designed,so that it can simultaneously encode the continuous and the discrete variables.Then,the fitness function of the integrated learning particle swarm optimization algorithm with multi-stage and multi-objective characteristics is defined to minimize the number of virtual samples while ensuring the generalization performance of the model.Finally,a multi-objective hybrid optimization task is generated for virtual samples to improve the integrated learning particle swarm optimization algorithm,so that it can adapt to the variable dimension characteristics of the virtual sample optimization process and improve the convergence speed.At the same time,the comprehensive evaluation index and distribution similarity index are proposed for evaluating the quality of virtual samples by referring to metric learning for the first time.In this paper,two benchmark datasets and an actual industrial dataset are used to verify the effectiveness and superiority of the proposed method.