Abstract: In the process of product production,accurate and fast prediction of product quality helps enterprises to adjust the manufacturing process in time and reduce losses.Aiming at the problems of high dimensionality,complex correlation and difficult to accurately predict with traditional methods in the actual production process,a method based on Improved Multiple hidden Layer Extreme Learning Machine（ML-ELM）named LCGWO-DMKEA-BLSTM was proposed.The collected feature parameters of the production process were screened by Mutual Information（MI）method to form the initial feature set of the model input.The Gaussian kernel function and the inverse cosine kernel function were weighted and combined to construct a new hybrid kernel function,and an auto-encoder was introduced to improve the extreme learning machine.The Deep Multi-Kernel Extreme learning machine Auto-encoder（DMKEA）feature mining model was established,and the key feature information that best reflected the product quality from the high-dimensional complex process feature set was extracted to input into the Bidirectional Long Short Term Memory Network（BLSTM）for quality prediction.In DMKEA learning training,an improved Grey Wolf Optimizer based on Levy flight strategy and Circle chaos mapping（LCGWO）was used to optimize the penalty coefficients,kernel parameters and kernel function combination weights,and the feature mining ability of DMKEA was improved.Finally,the effectiveness of the proposed method was experimentally verified with process data from semiconductor thin-film transistor liquid crystal display production line.The research results helped enterprises to realize accurate product quality prediction and also provide reference for data empowerment of enterprise production.

Key words: quality prediction, mutual information method, improved multiple hidden layer extreme learning machine, hybrid kernel function, bidirectional long short term memory network, Circle chaotic mapping, Levy flight, improved grey wolf algorithm

摘要： 在产品生产过程中,准确快速地预测产品质量有助于企业及时调整制造工艺,降低损失。针对实际生产过程中,现场采集的工艺数据存在维度高、相关性复杂且用传统方法难以准确预测的问题,提出一种基于改进多隐层极限学习机（LCGWO-DMKEA-BLSTM）的方法。首先,通过互信息法（MI）对采集的生产工艺特征参数进行筛选,组成模型输入初始特征集。其次,将高斯核函数与反余弦核函数加权结合,构造出新的混合核函数,并引入自动编码器对极限学习机进行改进,建立深度多内核极限学习机自编码器（DMKEA）特征挖掘模型,从高维复杂工艺特征集中提取最能反映产品质量的关键特征信息,输入决策层双向长短时神经网络（BLSTM）中进行质量预测。在DMKEA学习训练中,采用基于Circle混沌映射和Levy飞行策略改进的灰狼算法（LCGWO）,优化惩罚系数、核参数以及核函数组合权重,提高DMKEA的特征挖掘能力。最后用半导体薄膜晶体管液晶显示器生产线的工艺数据实验验证了所提方法的有效性。研究成果有助于企业实现准确地产品质量预测,也为企业生产的数据赋能提供参考。

关键词: 质量预测, 互信息法, 改进多隐层极限学习机, 混合核函数, 双向长短时神经网络, Circle混沌映射, Levy飞行, 改进灰狼算法