赞
踩
点击蓝字关注我们
文章发表于JIM期刊,出版于2020年1月.
链接:https://doi.org/10.1007/s10845-019-01530-8
摘要
现有的质量预测方法可分为两类,物理方法和数据驱动方法。由于产品质量与工艺参数之间存在不确定性和非线性关系,因此物理方法难以以足够的准确性和适应性来预测产品质量。而随着物联网的快速发展,历史制造和装配数据爆炸式增长,设备计算能力的提高和人工智能的发展提高了数据挖掘的能力,基于数据驱动方法的研究已成为产品质量预测领域的主流。但是,现有的数据驱动的质量预测方法存在许多限制。一方面,制造和装配过程的多工序特征导致了不同工序之间在时间方向上的强耦合;另一方面,由于产品定制技术和设备的灵活使用,不同产品甚至同一产品的工序可能会有所不同。
为了把握制造和装配过程中各个过程之间的时间交互作用,本文提出了一种称为QTD的端到端统一产品质量预测框架。它由三个模块组成:质量嵌入模型库(quality embeddingmodel pool, QEMP)、时间交互模型(temporal-interactive model)和解码模型(decoding model)。此外,为了处理并行过程在时间方向上的信息传递和集成问题,设计了一种新型的双向串并行LSTM(Bi-SP-LSTM)作为时间交互模型的实例化模型。Bi-SP-LSTM是双向长期短期记忆的扩展,同时设计了无监督任务和对抗性焦点损失的损失函数,以使框架能够评估归因于数据固有不确定性的分类任务中的异方差不确定性。最后,利用基于Kaggle竞赛的公共数据集的子集进行了实验,证明了所提出框架的有效性。经过验证,与其他方法相比,该框架更加准确和可靠。
文章导读
QTD框架示意图如下图所示,首先,根据产品通过站点的顺序从QEMP中选择QEM;其次,利用QEM 来获取嵌入质量的向量;然后根据站点的时间步长和每个时间步长上的并行站点位置,将质量嵌入矢量作为A-Bi-SP-LSTM中LSTM单元的输入,从而获得潜在状态向量;最后,通过解码模型的输入以获得预测结果。
QTD框架
算法伪代码
原文信息
Abstract
In order to capture temporal interactions among processes in manufacturing and assembly processes, an end-to-end unified product quality prediction framework called QTD is proposed in this paper. It consists of three modules: quality embedding model pool, temporal-interactive model, and decoding model. Besides, to handle the information transfer and integration problems in the time direction of parallel processes, a novel bidirectional serial–parallel LSTM (Bi-SP-LSTM) is devised as an instantiated model of temporal-interactive model. Bi-SP-LSTM is an extension of bidirectional long short-term memory. Moreover, an unsupervised task and a loss function named adversarial focal loss have been designed to give the framework the ability to assess heteroscedastic uncertainty in classification task due to intrinsic uncertainty in data. Furthermore, experiments are devised based on a subset of a public dataset from Kaggle competition to demonstrate the validity of the proposed framework. Compared with other latest methods, the proposed framework is verified to be more accurate and robust. Taking Matthews correlation coefficient as an example, the adversarial Bi-SP-LSTM-based QTD framework is superior to the best existing methods with 95% confidence interval in most cases, and its mean MCC is 4.88% higher than the best existing method. The results suggest that the proposed framework has a broad application prospect for quality prediction in manufacturing and assembly processes.
Keywords
Quality prediction
Manufacturing and assembly processes
Bidirectional long short-term memory
Temporalinteractions
Parallel processes
Cite this article as:
Liu, Z., Zhang, D., Jia, W. et al. An adversarial bidirectional serial–parallel LSTM-based QTD framework for product quality prediction. J Intell Manuf (2020). https://doi.org/10.1007/s10845-019-01530-8
/本期编辑/
徐思迪
沈阳工业大学 硕士在读
研究方向:质量控制、可靠性研究
指导老师:姜兴宇教授
扫码关注我们
JIM期刊公众号
聚焦智能制造;传播学术观点;促进合作交流
左边点击“阅读原文”,跳转至原文链接
右边给我一朵小花花
Copyright © 2003-2013 www.wpsshop.cn 版权所有,并保留所有权利。