Abstract: The sensory quality indexes of baked cigarette products are important indicators for evaluating the quality of tobacco manufacturing,and accurate prediction of sensory quality evaluation indexes during the production process can help to improve the cigarette manufacturing process in a timely manner and improve the stability of quality.Aiming at the characteristics of cigarette sensory quality evaluation data,such as feature sparsity,redundancy,multi-noise and missing data,this paper proposes a method based on Multimodal Feature Fusion Deep Bayesian Regularized Artificial Neural Network for predicting sensory quality evaluation indexes.Firstly,the encoder-decoder structure is used to build on the input data to extract deep features;secondly,mutual information loss based on information entropy is introduced into the encoder of the product chemical composition data and the process parameter data,which is trained to obtain a shared feature representation of the input samples;and lastly,the Deep Bayesian Regularized Artificial Neural Network is introduced into the decision-making layer as a predictor.In this predictor,the regularization coefficients as well as the display expressions of the prediction error weight coefficients are adaptively updated based on the Bayesian formulation;in the learning training,the Broyden-Fletcher-Goldfarb-Shanno (BFGS) algorithm is used for the objective function optimization,while using the estimation of the inverse of the Hessian matrix in the BFGS algorithm to adaptively update the regularization coefficients and the prediction error weighting coefficients,to improve the generalization ability of this neural network and reduce the risk of overfitting.The effectiveness of the proposed method is verified by taking the main brand cigarette products of a domestic tobacco company as the test object,and the method is helpful for the intelligent prediction of the sensory evaluation indexes of cigarette products and the data empowerment of the production process,so as to enhance the precise manufacturing capability of the production process.

Key words: sensory quality assessment prediction, Bayesian regularization, multisource feature fusion, stacked autoencoder, adaptive regularization coefficients

摘要： 烤烟型卷烟产品的各项感官质量指标是评价烟草制造品质的重要依据,在生产过程中对各项感官质量评价指标进行准确地预测有助于对卷烟制造工艺过程进行及时改进,提高质量稳定性。针对卷烟感官质量评价数据具有特征稀疏、冗余性、多噪声和数据缺失的特点,提出一种基于多源特征融合的深度贝叶斯正则化神经网络（Multisource Feature Fusion Deep Bayesian Regularized Artificial Neural Network,MFF-DBRANN）的方法对感官质量评价指标进行预测。首先,利用编码器-解码器结构,对输入数据建立堆叠自动编码器提取深层特征;其次,在产品化学成分数据与工艺参数数据的编码器中引入基于信息熵的互信息损失,训练得到输入样本的共享特征表示;最后,在决策层引入深度贝叶斯正则化神经网络作为预测器。在该预测器中,基于贝叶斯公式自适应更新正则化系数以及预测误差权重系数的显示表达式;在学习训练中,使用Broyden–Fletcher–Goldfarb–Shanno（BFGS）算法进行目标函数的优化,同时利用BFGS算法中对Hessian矩阵逆的估计,自适应更新正则化系数以及预测误差权重系数,提高该神经网络的泛化能力,减少过拟合风险。以国内某烟草公司主要卷烟品牌产品为试验对象验证了所提方法的有效性,该方法有助于卷烟产品感官评价指标的智能化预测和生产过程数据赋能,从而提升生产过程的精准制造能力。

关键词: 感官质量评价预测, 贝叶斯正则化, 多源特征融合, 堆叠自编码器, 自适应正则化系数