多源迁移学习的贝叶斯网络预测方法与应用研究

Bayesian network is currently one of the most effective approaches for indeterminacy knowledge representation and reasoning. In the forecasting process, some situations, such as repetition of similar tasks with different data distribution and lacking of training samples in target domain, appear inevitably, which bring challenges for the Bayesian network modeling. Transfer learning provides a new way to solve such problems through transferring the knowledge learned from source domain samples into the target domain. Followed by the research route of “theory, method and application”, this project is composed of three parts: Firstly, sharing feature selection strategy is researched to address the theoretical issue (domain adaptation) of transfer learning by some techniques including space reconstruction, probability measurement and information evaluation. This is a bridge (guarantee) of effective transfer. Secondly, a Bayesian network based on multi-source transfer learning is constructed, including Bayesian network learning in the source domain, domain similarity evaluation and multi-source transfer strategy research. This can realize the structure and parameters of Bayesian network transferring from the source domain to the target domain. Finally, the research achievements are applied for forecasting the trend of water pollution emergencies caused by rainstorm in typhoon season---case study in Dongjiang-Shenzhen water supply project. This project can not only solve the problems of Bayesian network learning in the case of incomplete data, but also provide reference basis for emergency decision-making, having significant theoretical and realistic meaning.

贝叶斯网络是目前不确定知识表示和推理最有效的模型之一，但在预测过程中不可避免会遭遇重复面临相似任务（数据分布不同）、目标训练样本不完备等情形，影响贝叶斯网络的建模效率和使用率。而迁移学习能够将从源域数据样本中获得的域知识迁移到目标域中，为解决此类问题提供了新途径。本项目遵循“理论→方法→应用”的思路，首先研究迁移领域共享特征选择策略，以共享特征为媒介，通过空间重构、概率度量、信息评估等途径研究领域适应性问题，为源域与目标域有效迁移搭建桥梁；然后构造一种贝叶斯网络的多源迁移学习框架，通过源域贝叶斯网络学习、领域相似性评估、多源迁移策略研究，实现贝叶斯网络结构和参数从源域到目标域的迁移；最后将此成果应用于突发事件演化预测，以东深供水工程为例，预测台风季暴雨引起的突发性水质污染趋势，既解决了数据不完备情况下的贝叶斯网络学习问题，也为突发事件的应急决策提供参考依据，具有显著的理论意义和现实意义。

为了提升贝叶斯网络的预测建模能力，本项目利用迁移学习从源域获得目标域所需信息特征，构建一种多源迁移学习贝叶斯网络，解决传统贝叶斯网络应对不同任务的适用性问题、提高模型使用率；同时解决先验知识缺失的网络参数学习问题，为贝叶斯网络学习提供一种新方法。本项目遵循“理论研究→方法研究→应用研究”的思路，首先研究了迁移领域共享特征选择策略，用最大均值差异理论对源域与目标域进行共性特征评估，为领域适应性问题的解决提供了理论支持和实现路径；然后研究了一种贝叶斯网络的多源迁移学习框架，通过改造最大均值差异（引入收缩因子），实现了源域贝叶斯网络的参数迁移，从而构建了目标域的新贝叶斯网络；最后将此成果应用于不同场景，包括流域水质预警、企业电力预测、装备故障预示等。本项目研究成果可为其他模型建模效率的提高、数据不完备的预测建模提供参考，也可为突发事件应急决策管理（样本量不足导致的训练“差”问题）、高质量制造生产管理（多变工况的标签样本选择“准”问题）、重大装备健康管理（苛刻环境下的数据采集“难”问题）等提供了数据决策支持。项目执行期间，共发表（含录用）论文25篇，其中SCI 18篇（中科院分区二区以上15篇）、EI 2篇、国家自然科学基金委员会管理科学A类重要期刊2篇，CSCD 3篇，均标注国家基金资助。授权中国发明专利1项、新申请中国发明专利3项、获得软件著作权1项。获得省级科技奖励二等奖1项。部分技术得到了企事业单位的应用。