Abstract: In the operation of business process information systems,anomalies arising from factors such as software failures and human errors are common,significantly impacting system performance and posing substantial risks to organizations.As a result,anomaly detection is a crucial aspect of business process management.However,some organizations,due to the low frequency of new business operations,face challenges such as insufficient accumulation of business data or the absence of certain potential anomalies in historical data.When these anomalies occur,they are often difficult to manage.Meanwhile,existing anomaly detection methods struggle to proactively address anomaly detection challenges in complex temporal dependencies and high-dimensional data.To overcome these limitations,a Generative AI Business Process Anomaly Detection method (GAI-BPAD) was proposed.The framework comprised three primary modules:sensing,decision-making and execution.Generative Adversarial Networks (GANs) was used to enhance the diversity of business process behavior samples,and a Bi-directional Gated Recurrent Unit with attention mechanism (Att-Bi-GRU) was utilized for anomaly detection.Experimental evaluations on nine real-world datasets demonstrated that GAI-BPAD achieved notable improvements in both accuracy and robustness compared to traditional methods,effectively identifying anomalies in business processes.

Key words: process mining, anomaly detection, AI Agent, artificial intelligence

摘要： 在业务流程信息系统运行过程中,由于软件故障、操作人员失误等因素引发的异常现象十分普遍,这些异常会显著影响服务系统运行状态,并给企业组织带来风险,因此异常识别是业务流程管理中的关键环节。然而,部分组织机构由于新业务开展频率较低,存在业务数据积累不足或者某些潜在异常尚未在历史数据中展现的问题,这些异常一旦发生往往难以应对。同时,现有的异常识别方法难以主动应对复杂时序依赖和高维数据中的异常检测问题。为了解决上述问题,本文提出了基于生成式AI Agent的业务流程异常主动识别框架(GAI-BPAD),该框架分为感知、决策和执行3个主要模块,通过生成对抗网络(GAN)增强业务流程行为样本的多样性,并结合基于注意力机制的双向GRU神经网络(Att-Bi-GRU)进行异常识别。在9个真实数据集上进行了评估,实验结果表明,该方法相较于传统的异常识别方法,在准确性和鲁棒性方面均表现出显著提升,能够有效识别业务流程中的异常行为。

关键词: 流程挖掘, 异常识别, AI Agent, 人工智能