摘要
实际工业过程往往具有多变量、非线性和动态性等特点,建模数据包含过多冗余信息和时序依赖特征,从而导致建模复杂度增加和模型性能下降。因此,提出一种基于非负绞杀的稀疏化有序神经元长短时记忆网络(ordered neurons long short-term memory,ONLSTM)用于工业软测量建模。将非负绞杀收缩系数嵌入ONLSTM输入层权重矩阵,对其进行收缩绞杀,剔除冗余输入节点的同时实现变量选择。将非负绞杀收缩系数与ONLSTM隐藏层权重矩阵相结合,根据不同隐藏神经元重要性设计权重分配规则,剔除网络隐藏层冗余节点及其对应的信息传递通路,进行网络结构稀疏优化。通过数值仿真验证了所提算法的有效性,并将其应用于某火电厂烟气脱硫过程排放净烟气SO2浓度预测。实验结果表明所提算法能有效实现变量选择,并在保证预测性能的前提下,使模型结构得到稀疏优化,展现出比较广阔的应用前景。
Industrial processes often exhibit characteristics such as multivariable interactions,nonlinear behaviors,and time-varying changes.Thus,the resulting modeling data contain excessive redundant information and complex time-dependent patterns,which increase modeling complexity and degrade the model performance.To address these challenges,an ordered neurons long short-term memory(ONLSTM)network integrated with non-negative garrote-based regularization is proposed herein for industrial soft sensor modeling.First,the shrinkage coefficients of the non-negative garrote are embedded into the weight matrix of the ONLSTM input layer to eliminate redundant input nodes and enable variable selection.Second,these coefficients are integrated into the weight matrix of the ONLSTM hidden layer to assign weights based on the importance of hidden neurons.Consequently,redundant nodes and their corresponding information pathways are pruned,achieving sparse optimization of the network structure.The proposed method is validated via numerical simulations and subsequently employed to predict the SO2 concentration in flue gas emissions from a desulfurization process in a thermal power plant.Experimental results demonstrate that the algorithm effectively selects variables and sparsely optimizes the model structure while maintaining high predictive performance,offering promising prospects for broader industrial applications.
作者
郭迎宸
隋璘
熊伟丽
GUO Yingchen;SUI Lin;XIONG Weili(School of Internet of Things Engineering,Jiangnan University,Wuxi 214122,China;Key Laboratory of Advanced Process Control for Industry(Ministry of Education),Jiangnan University,Wuxi 214122,China)
出处
《智能系统学报》
北大核心
2025年第6期1366-1378,共13页
CAAI Transactions on Intelligent Systems
基金
国家自然科学基金项目(61773182)
江南大学“轻工技术与工程”双一流学科与支撑学科协同发展支持计划(QGJC20230203)。
关键词
软测量
长短时记忆网络
有序神经元
非负绞杀
冗余信息
变量选择
稀疏优化
深度学习
soft sensor
long short-term memory networks
ordered neurons
non-negative strangulation
redundant information
variable selection
sparse optimization
deep learning
