Anomaly Detection Method Of Business Process Based On Attention Mechanism

At present,most enterprises and organizations use process-aware information systems to support daily business process management.Process-aware information systems can provide modeling,deployment,execution,and delivery of business processes to ensure efficiency and reliability.With the increased complexity and flexibility of business processes,business processes frequently cause anomalies.Anomalies can cause interruptions to business processes or deviations in execution,resulting in inconsistent execution results.How to detect anomalies in business process execution is a challenge for enterprises and organizations.Many anomaly detection methods require high-quality business process models,which are difficult to maintain;Other methods cannot find the source of anomalies.This paper,starting from the event log generated by business process execution,using the attention mechanism of the deep learning model,proposes a business process anomaly detection method.The work is as follows:Many business process representation methods only encode the business process from a single perspective,resulting in incomplete representation information,or the need to maintain a reference process model,which is costly.This paper introduces an attention-based business process representation method that encodes business processes from multiple perspectives.This method directly encodes the event log without a reference process model.Mining multi-perspective dependencies of business process using attention mechanism and proposing three representation strategies.Compared with other methods,the representation method proposed in this paper can effectively improve the representation quality of the business process.At present,most business process anomaly detection methods can detect case anomalies or event anomalies,but cannot detect the attribute anomaly.This paper introduces a business process anomaly detection method based on anomaly score.First,get the data set from the control flow and data flow of the event log.Then,use the representation method proposed in this paper to build the next event prediction model of the business process based on the Transformer model.Next,calculate the anomaly score for the attribute by comparing the actual occurrence of the event with the results of the event prediction model.then,Use the anomaly proportional gradient method to obtain an adaptive threshold.Finally,if the anomaly score is greater than the threshold,this attribute is an anomaly attribute,and the source of the anomaly is located as this event attribute.The experimental results show that on the public real business process data set,this paper introduces an anomaly detection method that can accurately detect the event and attribute anomaly,and locate the attribute anomaly.A prototype system for business process anomaly detection is developed.This prototype system demonstrates the practicability of the anomaly detection method in this paper.The anomaly detection method proposed in this paper is meaningful for improving the stability,security,and consistency of enterprises and organizations.