Abstract: Due to the influence of various factors,business processes will inevitably have abnormal situations.Accurate detection of anomalies in execution can ensure the reasonable operation of the process.At present,advanced anomaly detection technology uses deep learning methods to input control flow and data flow information in the process through coding to achieve anomaly detection of activities,time and other attributes.However,it is difficult to characterize the behavioral relationship between process activities,and the interaction between control flow and data flow is not specific.For this reason,a Multi-level business process anomaly detection method based on Graph Transformer and AutoEncoder (MLGTAE) was proposed.The method first used Petri net to characterize the entity behavior in the process,and combined the data attributes such as time and resources in the process to construct multi-level trace.Then,Graph Transformer was used as the encoder to extract the behavior information and data information in the graph.Finally,the attention mechanism was used to realize the deep interaction between behavior and data,and the decoder reconstruction was used to detect anomalies.The proposed method was verified by several real data sets,and the results showed that the proposed method was superior to comparison methods in anomaly detection at both activity level and data attribute level.

Key words: Petri nets, Graph Transformer, behavioral footprint, behavioral relationship, business process anomaly detection

摘要： 由于各种因素的影响,业务流程不可避免地会出现异常情况,精准地检测出执行中的异常,能够保障流程的合理运行。目前先进的异常检测技术利用深度学习方法,通过编码的方式来输入流程中的控制流和数据流信息以实现对活动、时间等属性的异常检测,但是该方式难以表征流程活动间的行为关系且控制流和数据流交互不具体,针对该问题,本文提出一种基于Graph Transformer和自编码器的多层次业务流程异常检测方法(MLGTAE),该方法首先利用Petri网表征流程中的实体行为,并结合流程中的时间、资源等数据属性实现多层次迹图的构造,然后通过Graph Transformer作为编码器提取图中的行为信息和数据信息,最后利用注意力机制实现行为和数据的深度交互,并利用解码器重建检测异常。该方法通过多个真实数据集进行验证,结果表明提出的方法在活动层次和数据属性层次上的异常检测效果都优于对比方法。

关键词: Petri网, Graph Transformer, 行为足迹, 行为关系, 业务流程异常检测