【作者】 杨磊；

【导师】 向峰；

【作者基本信息】 武汉科技大学 ， 机械工程， 2023， 硕士

【摘要】 人工智能技术的迅速发展为质量预测领域提供了新的解决思路,应用大数据分析等方法对生产数据的深度挖掘可有效地解决产品质量与工艺参数间复杂的非线性关系,提高产品质量预测的准确性。本文对复杂生产过程的生产特点、数据特点与现有数据驱动的质量预测方法进行分析,提出了一种数据驱动下模型融合的产品质量预测方法,以提升复杂生产过程中产品质量的预测准确性,主要研究内容如下。(1)分析了复杂生产过程的生产特点、数据特点以及现有基于数据驱动方式下质量预测研究的不足之处,提出了一种融合对生产过程以两种角度所建模型的模型融合质量预测方法。(2)从生产全过程整体建模预测最终成品质量的角度出发,提出了一种基于特征增强的整体式产品质量预测模型。通过所提基于互信息的特征选择方法,有效提高了模型输入特征的强相关性、低冗余性,增强了模型的预测准确性。在强相关、低冗余的特征选择结果基础上建立了基于改进PSO-LSTM的生产整体式产品质量预测模型。(3)从生产过程分阶段建模预测在制品质量的角度出发,提出了一种基于误差修正机制的分段式产品质量预测模型。对生产各阶段分别建立基于改进RF的在制品质量预测模型,结合所提基于历史生产数据的误差修正机制串联各个预测模型,有效解决了分阶段预测过程误差累积问题。(4)从单个模型在复杂生产过程中预测能力较为薄弱的角度出发,采用集成学习的思想,提出了基于Stacking的模型融合产品质量预测方法。集成融合上述两种对生产过程所建的质量预测模型,以五折交叉验证与多项式特征处理方法对元学习器训练过程做出改进,得到预测性能更加准确的质量预测融合模型。(5)通过对实际案例的仿真分析,验证所提数据驱动下模型融合的质量预测方法的有效性与准确性。选用了武汉卷烟厂烟丝生产过程中对烘丝机入口烟丝含水率的预测研究作为案例分析,通过对比模型融合前后的预测性能,验证了所提模型融合方法相较于单模型的预测准确性有所提高。通过对比其他烘丝机入口烟丝含水率预测研究方法,验证了所提方法在复杂生产过程中具有较高准确性与可行性。通过对案例的研究分析表明,本文所提的数据驱动下模型融合质量预测方法具有较高的准确性与有效性,拓展了复杂生产过程产品质量预测领域的研究内容,具有较强的理论意义与实际应用意义,为质量预测领域提出了一个新的思路。更多还原

【Abstract】 The rapid development of artificial intelligence technology has provided new solutions for the field of quality prediction.Through deep mining of production data using methods such as big data analysis,the complex nonlinear relationship between product quality and process parameters can be effectively addressed,improving the accuracy of product quality prediction.This paper analyzes the production characteristics,data characteristics,and existing data-driven quality prediction methods of complex production processes,and proposes a model fusion-based data-driven product quality prediction method to enhance the accuracy of product quality prediction in complex production processes.The main research contents are as follows.(1)Analyzing the complex production characteristics,data characteristics,and shortcomings of existing data-driven quality prediction research,a model fusion-based quality prediction method is proposed by integrating models built from two different perspectives of the production process.(2)A feature-enhanced whole product quality prediction model is proposed based on modeling the entire production process to predict the final product quality.The proposed feature selection method based on mutual information effectively enhances the strong correlation and low redundancy of the model’s input features,thereby improving the prediction accuracy.A production whole product quality prediction model based on improved PSO-LSTM is established based on the strong correlation and low redundancy feature selection results.(3)A segmented product quality prediction model based on error correction mechanism is proposed from the perspective of modeling the production process in stages to predict the quality of work in progress.Work-in-progress quality prediction models based on improved RF are established for each stage of production.The error correction mechanism based on historical production data is proposed to effectively solve the problem of error accumulation in the segmented prediction process.(4)From the perspective that the predictive ability of a single model is weak in complex production processes,a model fusion-based product quality prediction method based on stacking is proposed using the idea of ensemble learning.The two quality prediction models built for the production process are integrated,and the training process of the meta-learner is improved using five-fold cross-validation and polynomial feature processing methods to obtain a more accurate quality prediction fusion model.(5)Through simulation analysis of practical cases,the effectiveness and accuracy of the proposed data-driven model fusion-based quality prediction method are verified.The prediction of the moisture content of tobacco entering the dryer in the production process of Wuhan Tobacco Factory is selected as the case study.By comparing the prediction performance before and after model fusion,the proposed model fusion method is found to be more accurate than a single model.By comparing with other research methods for predicting the moisture content of tobacco entering the dryer,the proposed method is found to be highly accurate and feasible in complex production processes.The research and analysis of the case study demonstrate that the proposed data-driven model fusion-based quality prediction method has high accuracy and effectiveness,expands the research content of product quality prediction in complex production processes,and has strong theoretical and practical significance,providing a new way for the field of quality prediction.更多还原