从零样本学习理论模型到工业应用——动机、演变与挑战

From zero-shot learning theoretical model to its industrial application:Motivation,evolution and challenges

随着工业大数据技术的发展,应用于工业对象的有监督方法得到广泛研究.真实数据往往遵循长尾类分布,导致传统有监督模型在实际应用过程中存在模型退化以及模型失效等问题.零样本学习(zero-shot learning,ZSL)技术的提出为这一问题提供了一种新的解决思路.零样本学习的目标是使用收集到的已见类别数据训练模型,使得训练好的模型对于收集不到数据的未见类别同样适用.零样本学习通过将故障的文本描述等辅助知识引入到模型中,一定程度上缓解了模型在实际工业场景中对训练数据收集的依赖,提高了模型的泛化性能.然而,目前领域内尚缺乏对零样本学习在工业领域应用的系统梳理与讨论,而工业零样本学习在辅助知识的收集和处理、研究方法、应用场景等方面与其他领域的零样本学习相比也具有独特性.鉴于零样本学习在工业领域潜在的巨大应用价值和未来的发展潜力,系统性梳理和展示了从零样本学习理论模型到工业应用的动机、演变与挑战.首先,回顾零样本学习设定与相关方法的发展脉络,分析零样本学习与其他任务设定之间的关联,并指出本文与前人综述的区别.接下来,回顾工业领域零样本学习的研究现状,介绍典型的工业零样本学习任务和辅助知识,分析工业零样本学习的特征和典型问题,梳理工业领域零样本任务中应用的现有方法.此外,梳理工业零样本任务的基准数据集和开源工作.最后,在现有研究的基础上总结工业零样本任务面临的问题与挑战,并对该领域的研究做出展望.

With the advancement of industrial big data technology,supervised methods applied to industrial objects have been widely studied.However,real data often follow the long-tailed distribution phenomenon,which poses problems such as model degradation and failure in practical applications for traditional supervised models.The proposal of zeroshot learning(ZSL)technology provides a new approach to solve this problem.The objective of ZSL is to train the model using collected seen category data so that it can also be applied to unseen categories whose data cannot be collected.By incorporating auxiliary knowledge such as fault text description into the model,ZSL reduces the dependence of the model on training data collection in practical industrial scenarios and enhances its generalization performance.However,there is still a lack of systematic review and discussion on the application of ZSL in the industrial field.Compared with ZSL in other fields,industrial ZSL is unique in terms of auxiliary knowledge collection and processing,research methods,and application scenarios.Given the potential great application value and future development potential of ZSL in the industrial field,this paper systematically summarizes and presents the motivation,evolution,and challenges of ZSL theoretical models for industrial applications.Firstly,this paper reviews the development of ZSL settings and related methods,analyzes their correlation with other task settings,and highlights differences between this paper’s review and previous ones.Next,this paper reviews the current state of zero-shot learning research in the industrial field,introduces typical industrial zero-shot learning tasks and auxiliary knowledge,analyzes the features and typical problems of industrial zeroshot learning,and summarizes the existing methods used in industrial zero-shot tasks.In addition,this paper also presents the benchmark datasets and open-source works for industrial zero-shot tasks.Finally,based on the existing research,this paper summarizes the problems and challenges faced by industrial zero-shot tasks,and provides some prospects for the research in this field.