毕业生知识库

制造业是我国国民经济和社会发展的重要支柱产业。制造业中的复杂工业过程是由物料、设备、工艺和环境等多个要素高度融合而成的复杂系统。然而，复杂工业过程面临着资源利用率低、能耗物耗高、产品质量差、成本高、环境污染严重等问题。因此，以高效化和绿色化生产为目标，实现复杂工业过程各环节的决策优化具有重要意义。目前，复杂工业过程的决策优化存在评估成本高、过程非平稳、多目标协同等问题，难以通过人工来决策优化。本文主要研究了基于深度高斯过程的多目标贝叶斯优化方法，同时探讨了面向复杂工业过程的贝叶斯优化方法，并将方法分别应用于铜冶炼过程和天然气净化过程的决策优化中。主要研究工作和贡献如下：

（1）针对复杂工业过程决策优化中存在的非平稳和多目标协同问题，提出了基于深度高斯过程的多目标贝叶斯优化方法DGMBO。该方法构建了深度高斯过程代理模型，学习输入与多个目标的复杂非平稳映射关系，并使用期望超体积改进采集函数来充分平衡探索与利用，快速找到最优解。在多个标准测试函数上的实验结果表明了所提方法的有效性和高效性。

（2）针对贝叶斯优化难以与实际工业过程结合的问题，提出了面向复杂工业过程的贝叶斯优化方法CI-DGMBO，旨在利用大量历史数据和少量仿真数据驱动贝叶斯优化。本文设计了仿真系统交互方法来评估反馈优化结果，提高决策优化的实时性和可靠性，并提出了基于Transformer自编码器的工业数据去噪与采样方法，实现端到端地去除数据噪声和缺失值，并提取历史数据中的关键点。实验结果表明了该方法的有效性。

（3）针对铜冶炼过程中存在效率低和污染严重的问题，进行铜冶炼过程决策优化设计与应用验证。本文根据闪速炼铜的反应机理和数学模型，基于分段建模思想搭建了闪速炼铜仿真系统，并实现了优化算法与仿真系统的实时交互。在闪速炼铜仿真系统上的实验结果表明CI-DGMBO方法能够提高决策优化效率，并在产量和环保方面提升了生产指标。

（4）针对天然气脱硫过程中存在的低效益和高能耗问题，进行天然气脱硫过程决策优化设计与应用验证。本文设计并构建了天然气脱硫过程的实时数据库和关系数据库，以存储与管理历史数据和工艺数据。针对优化过程中存在的已知约束和未知约束，对CI-DGMBO进行相应改进，扩展为CI-DGMCBO方法，分别使用搜索空间限制和可行性指示函数，使其在多个限制条件下寻找多目标最优解。在天然气脱硫仿真系统和实际系统上分别进行了验证，结果表明CI-DGMCBO方法能够实现快速优化，为提高企业经济效益和实现实时决策优化提供了可靠有效的方案。

Manufacturing is an essential industry for our country's economic and social development. Complex industrial process in the manufacturing industry is a complex system that is highly integrated with multiple elements such as materials, equipment, processes and environment. However, complex industrial processes face challenges such as low resource utilization, high energy and material consumption, poor product quality, high cost, and serious environmental pollution. Therefore, with the goal of efficient and green production, it is of great significance to realize the decision optimization of each link of the complex industrial process. At present, the decision optimization of complex industrial processes encounters issues like high evaluation cost, non-stationary process, multi-objective coordination, etc., making manual decision optimization difficult. This thesis mainly investigates the multi-objective Bayesian optimization method based on deep Gaussian processes, discusses the Bayesian optimization method for complex industrial processes, and applies the method to the decision optimization of copper smelting process and natural gas purification process, respectively. The main research work and contributions are as follows:

(1) Aiming at the non-stationary and multi-objective coordination problems in the decision optimization of complex industrial processes, the multi-objective Bayesian optimization method based on deep Gaussian process (DGMBO) is proposed. This method constructs surrogate model based on deep Gaussian processes, learning the complex non-stationary mapping relationship between the input and multiple objectives, and uses the expected hypervolume improvement as acquisition function to fully balance the exploration and utilization, quickly finding the optimal solution. Experimental results on several standard test functions demonstrate the effectiveness and efficiency of the proposed method.

(2) Aiming at the challenge of integrating Bayesian optimization with actual industrial processes, the Bayesian optimization method for complex industrial processes (CI-DGMBO) is proposed, which aims to use a large amount of historical data and a small amount of simulation data to drive Bayesian optimization. In this thesis, a simulation system interaction method is designed to evaluate the feedback optimization results, improving the real-time performance and reliability of decision optimization. An industrial data denoising and sampling method based on Transformer autoencoder is proposed, which achieves end-to-end removal of noise and missing data, and extracts key points in historical data. Experimental results show the effectiveness of the method.

(3) Aiming at the problems of low efficiency and serious pollution in the copper smelting process, the decision optimization design and application verification of the copper smelting process are conducted. According to the reaction mechanism and mathematical model of flash copper smelting, this thesis constructs a flash copper smelting simulation system based on the concept of segmented modeling and realizes the real-time interaction between the optimization algorithm and the simulation system. Experimental results on the flash copper smelting simulation system indicate that the CI-DGMBO method can improve the efficiency of decision optimization, and improve production indicators in terms of output and environmental protection.

(4) Aiming at the problems of low benefit and high energy consumption in the natural gas desulfurization process, the decision optimization design and application verification of natural gas desulfurization process are carried out. This thesis designs and constructs a real-time database and relational database for natural gas desulfurization process to store and manage historical data and process data. Aiming at the known and unknown constraints in the optimization process, CI-DGMBO is improved accordingly and extended to CI-DGMCBO method, which uses the search space limit and the feasibility indicator function, respectively, so that it can find multi-objective optimal solutions under multiple constraints. Verification is carried out on the natural gas desulfurization simulation system and the actual system, respectively. The results show that the CI-DGMCBO method can achieve rapid optimization and provide a reliable and effective solution for improving the economic benefits and realizing real-time decision optimization.