基于Attention-BLSTM的复杂产品制造质量预测方法

Attention-BLSTM based quality prediction for complex products

复杂产品制造过程中工艺数据的高维特性以及工艺数据间的复杂关联特性,使得工艺数据中的深层次关键工艺特征难以挖掘,限制了产品质量的准确预测。鉴于此,提出一种基于注意力机制(Attention)与双向长短期记忆网络(BLSTM)的复杂产品质量预测方法。首先,设计数据预处理环节进行工艺数据清洗以及互信息特征筛选。然后,运用BLSTM网络模拟产品制造过程误差的复杂传递特性,挖掘上下游工艺参数的关联关系,输出BLSTM所有时刻提取的关联化工艺特征;同时,设计了Self-Attention网络,自学习各时刻关联化工艺特征对最终产品质量贡献的差异,对不同时刻工艺特征分配不同注意力权值,以强化关键特征。通过以上两阶段特征处理方式,实现深层次关键工艺特征的挖掘。最后,以关键特征作为输入层,通过反向传播神经网络(BPNN)实现复杂产品质量的准确预测。实验表明,相较于BPNN、长短期记忆神经网络(LSTM)、BLSTM以及XGBoost、基于粒子群优化的支持向量(PSO-SVR)、随机森林-贝叶期优化(BO-RF)等主流质量预测方法,所提方法有效提高了预测精度。

Due to the high-dimensional characteristics of process data and complex correlation between process data in the manufacturing process of complex products,it is difficult to explore the deep-level key process features in the process data,thus limiting the improvement of the accuracy of complex product quality prediction.Thus,a complex product quality prediction method based on attention mechanism and Bidirectional Long Short-Term Memory(BLSTM)network was proposed.A data pre-processing link for process data cleaning and mutual information feature screening was designed.The BLSTM network was used to simulate the complex transmission characteristics of errors in the product manufacturing process to mine the correlation between upstream and downstream process parameters,and to output the correlated process features extracted by BLSTM at all moments.The self-attention mechanism was designed to self-learn the differences in the contribution of correlated process features to final product quality at each moment,and assign different attention weights to process features at different moments to reinforce key features.Through the above two-stage feature processing method,the mining of deep-level key process features was achieved.Finally,the key features were used as the input layer to achieve accurate prediction of complex product quality through BPNN.Experiments showed that the proposed method had the best performance compared with the mainstream quality prediction methods such as BPNN,LSTM,BLSTM and so on.