摘要
新一轮工业革命带来的数字技术推动质量管理与数学建模仿真等工具结合,利用数据驱动方式对产品质量进行预测以保障最终产品质量已成为制造业转型升级的重要趋势。其中构建准确性高且稳定性好的质量模型是实施质量预测的前提与关键,高斯过程模型作为典型的质量模型之一,由于其在实践应用中表现出的优越性能得以广泛应用。然而,由于核函数类型选择的不确定性,可能导致似然函数中的超参数估计无法近似逼近真实值,进而难以获得可靠的质量预测结果。为此,针对高斯过程模型中的核函数选择问题,提出了基于选择性集成核学习算法的模型构建策略。首先,采用Bootstrap方法对训练样本进行重复抽取,利用各训练样本分别获得不同核函数下的超参数近似值,构建多维高斯过程模型。其次,利用质量工具Pareto图分析不同核情形下高斯过程模型的预测性能,从而确定集成核的元素。然后,将集成参数融入似然函数中,构建改进的似然函数,进而确定集成核高斯过程模型的超参数。最后,结合仿真测试和工业实例验证了所提方法的有效性。分析结果表明,基于选择性集成核高斯过程模型不仅为核选择的问题提供了可行的优化路径,而且也提升了质量预测的准确度和精确度,从而为后续的质量优化或贝叶斯优化提供模型基础。
Due to the new round of industrial revolution,digital technology has further promoted the combination of quality management tools with mathematical modeling and simulation techniques.Quality prediction with data-driven methods has become an important trend in the modern manufacturing industry.Among them,building a quality model with high accuracy and good robustness plays a vital role in implementing quality prediction.As one of the popular quality models,the Gaussian process model is widely used due to its superior performance in practical applications.However,due to the uncertainty in selecting the type of kernel functions,the hyper-parameter estimation in the likelihood function may be not obtained precisely,which leads to unreliable quality prediction results.To deal with this issue,a model construction strategy based on a selective ensemble kernel learning algorithm is proposed in the framework of Gaussian process models.First,the Bootstrap method is used to repeatedly extract the training samples,and each training sample is used to obtain the approximate values of hyper-parameters under different kernel functions.Then,a multi-dimensional Gaussian process model can be constructed based on the above information.Second,the predictive performance of the Gaussian process model under different kernel scenarios is analyzed,and then the kernel functions can be determined by the quality tool Pareto plot.Meanwhile,the ensemble parameters are incorporated into the construction of an improved likelihood function.The hyper-parameters of the ensemble kernel Gaussian process model can be obtained by maximizing the improved likelihood function.Finally,the effectiveness of the proposed method is verified by simulation tests and case studies.The comparison results show that the Gaussian process model based on the selective ensemble kernel not only provides a feasible method for the determination of kernel functions,but also improves the accuracy and precision of the quality prediction.Then the proposed model can be used for the following quality optimization or Bayesian optimization.
作者
欧阳林寒
陶宝平
马妍
OUYANG Lin-han;TAO Bao-ping;MA Yan(College of Economics and Management,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China;School of Economics and Management,Nanjing University of Science and Technology,Nanjing 210094,China)
出处
《中国管理科学》
CSSCI
CSCD
北大核心
2023年第3期69-80,共12页
Chinese Journal of Management Science
基金
国家自然科学基金重点项目(71931006)
国家自然科学面上项目(72072089,71872088)
江苏高校“青蓝”工程资助项目。
关键词
质量预测
高斯过程模型
集成核学习
参数优化
quality prediction
Gaussian process model
ensemble kernel learning
parameter optimization
