面向工业实体抽取的联邦学习优化算法

Federated learning optimization algorithm for industrial entity extraction

人工智能技术在工业、医疗和金融等领域得到了广泛应用,并取得了巨大的成功。工业实体抽取任务是实现工业领域数字化转型的关键一环,然而其实现往往需要大量的数据支持,而这些数据往往分布在各个机构或组织之间。各行各业都产生了海量的有价值的数据,但是在实际的应用场景中,安全隐私、法律法规和行业竞争等多种因素往往导致各方的数据不能共享,从而形成所谓的“数据孤岛”。针对这一问题,联邦学习提供了一种解决方案,可以有效解决数据孤岛问题,但联邦学习目前仍然面临一些问题和挑战,其中最典型的问题就是数据异构问题。针对各行各业存在的数据孤岛问题以及联邦学习本身的数据异构问题,本文以工业领域实体抽取任务为对象研究联邦学习的异构问题,从本地优化的角度提出了一种基于本地修正的联邦学习算法FedAmend,改善该联邦学习框架在面对工业领域数据非独立同分布时的表现,并在某汽车集团的工业设备故障数据上验证了FedAmend的可行性。

Artificial Intelligence(AI)technologies have been widely used with great success in the industrial,healthcare and financial sectors.The industrial entity extraction task is a key part of achieving digital transformation in the industrial sector,yet its realization often requires the support of a large amount of data,which is often distributed among various institutions or organizations.All industries generate huge amounts of valuable data,but in actual application scenarios,security,privacy,laws and regulations,and industry competition often lead to the formation of so-called"data silos"where data cannot be shared.To address this problem,Federated Learning provides a solution that can effectively solve the problem of data silos,but Federated Learning is still facing a number of problems and challenges,the most typical of which is the problem of data heterogeneity.Aiming at the data silo problem existing in various industries and the data heterogeneity problem of federated learning itself,this paper studies the heterogeneity problem of federated learning with the entity extraction task in the industrial domain as the object,proposes a federated learning algorithm FedAmend based on local correction from the perspective of local optimization,improves the performance of this federated learning framework in the face of the non-independent and homogeneous distribution of the data in the industrial domain and verifies the performance of this federated learning algorithm on industrial equipment failure data of an automobile group.The feasibility of FedAmend is verified on the industrial equipment failure data of an automobile group.