【摘要】 模型库是高端工业软件的核心组件.“无免费午餐”定理证明没有一种模型能在所有场景中表现良好,根据对象特性确定合适的模型至关重要.然而,常见的模型选择过程繁琐,需要逐一训练候选模型,并根据性能评估来确定最佳模型,这需要大量的标签信息和计算资源.本研究的重点是通过不依赖逐一实验和评估的方式,快速为无标签的目标数据选择合适的模型,提高工业软件模型库对各种工业对象快速支撑和规模应用的能力.本文首次将电子商务推荐系统的理念引入工业领域,提出了工业模型推荐系统的概念,指出并解决了构建工业模型推荐系统的主要挑战,即工业时序数据缺乏直观推荐特征.具体地,本文提出了一套推荐特征挖掘方法,其能够从工业时间序列数据中挖掘与设备无关但与监控性能相关的特征.其次,设计了基于宽度学习和大语言模型的双通路模型推荐方法,其能够建立特征和监控模型性能之间的映射关系,且能够综合多种评价维度给出推荐结果.此外,构建了一套工业模型推荐系统,包括数据集、特征和模型,为工业时间序列监测任务的模型推荐领域提供了一个基准.通过大量实验验证,展示了工业模型推荐的可行性以及本文方法的有效性.更多还原

【Abstract】 Model libraries are core components of high-level industrial software. The “no free lunch” theorem proves that no single model can perform well in all scenarios, and it is crucial to determine the appropriate model based on objects. However, the common model selection process is cumbersome and requires training candidate models one by one and determining the best model based on performance evaluation, which requires label information and computational resources. The focus of this research is to quickly select appropriate models for unlabeled target data by not relying on trial-by-trial and evaluation, improving the ability of industrial software model libraries to quickly support and scale applications for various industrial objects. In this paper, we introduce the concept of e-commerce recommendation systems into the industrial domain for the first time, propose the concept of an industrial model recommender, and address the main challenge of building an industrial model recommender, i.e., the lack of intuitive recommendation features for industrial time-series data. Specifically, this paper proposes a set of recommendation feature mining methods, which are capable of mining subject-independent but monitoring performance-related features from industrial time series data. Secondly, a dual-path model recommendation method based on broad learning and LLMs is designed, which is able to establish the mapping relationship between features and monitoring model performance, and is able to integrate multiple evaluation dimensions to give recommendation results. In addition, the proposed industrial model recommendation system,including datasets, features and models, provides a benchmark for the field of model recommendation for industrial time series monitoring tasks. The feasibility of industrial model recommendation and the effectiveness of our approach are demonstrated through extensive experimental validation.更多还原