Hierarchical ensemble deep learning for data-driven lead time prediction
Hierarchical ensemble deep learning for data-driven lead time prediction
复制标题
DOI：
10.1007/s00170-023-12123-4
复制
发表时间：
2023-08
期刊：
影响因子：
--
通讯作者：
A. Aslan;G. Vasantha;Hanane El-Raoui;J. Quigley;Jack Hanson;J. Corney;A. Sherlock
中科院分区：
文献类型：
--
作者：
A. Aslan;G. Vasantha;Hanane El-Raoui;J. Quigley;Jack Hanson;J. Corney;A. Sherlock
在按订单生产的生产环境中，基于订单到达时刻获取的实时生产状态，研究了基于数据驱动的产品订单提前期预测。特别是，我们考虑了一个复杂的制造系统，其中有大量关于生产状态的测量(例如，传感器数据)。为了应对这一复杂的预测挑战，我们提出了一种由三个深度神经网络组成的新型集成分层深度学习算法。其中一个网络充当多面手，而另外两个网络则充当不同产品的专家。分层集成方法以前已被成功地用于解决各种多类分类问题。在本文中，我们将这种方法扩展到包括提前期预测的回归任务。我们在两个不同的案例研究中演示了我们的算法的适用性。第一个案例研究使用了现有最大的制造业数据集之一--博世生产线数据集。第二个案例研究使用了受博世生产线启发的多产品、按订单生产的多产品生产系统的基于可靠性的模型生成的合成数据集。在这两个案例研究中，我们证明了我们的算法提供了高精度的预测，并且显著优于选定的基准，包括单个深度神经网络。此外，我们发现在合成数据集中的预测精度显著更高，这表明在工业制造过程中存在着在人工模型中不易再现的复杂性(即微妙的相互作用
This paper focuses on data-driven prediction of lead times for product orders based on the real-time production state captured at the arrival instants of orders in make-to-order production environments. In particular, we consider a sophisticated manufacturing system where a large number of measurements about the production state are available (e.g. sensor data). In response to this complex prediction challenge, we present a novel ensemble hierarchical deep learning algorithm comprised of three deep neural networks. One of these networks acts as a generalist, while the other two function as specialists for different products. Hierarchical ensemble methods have previously been successfully utilised in addressing various multi-class classification problems. In this paper, we extend this approach to encompass the regression task of lead time prediction. We demonstrate the suitability of our algorithm in two separate case studies. The first case study uses one of the largest manufacturing datasets available, the Bosch production line dataset. The second case study uses synthetic datasets generated from a reliability-based model of a multi-product, make-to-order production system, inspired by the Bosch production line. In both case studies, we demonstrate that our algorithm provides high-accuracy predictions and significantly outperforms selected benchmarks including the single deep neural network. Moreover, we find that prediction accuracy is significantly higher in the synthetic dataset, which suggests that there is complexity (i.e. subtle interactions) in industrial manufacturing processes that are not easily reproduced in artificial models