Research And Implementation Of Real-Time Anomaly Detection Of Business Process Control Flow Based On LSTM And Incremental Learning

Enterprises and institutions will involve many business processes in business management.The execution of the process produces a series of logs and records.Using these logs and records,enterprises and institutions can analyze,monitor,and predict the process,find weaknesses,anomalies,and risks in the business process,to adjust,improve,and enhance the business process accordingly,which is business process mining.With the advancement of technology and the constant challenges in business process mining,the traditional offline process mining for process history logs cannot meet the growing demand for real-time analysis,prediction,and alert due to severe information lag.Online process mining has become a new trend.However,in real-life situations,the occasional noise in the control flow usually harms the performance of process mining algorithms.Therefore,real-time anomaly detection is the key to the smooth application of online process mining technology.The existing automata-based realtime anomaly detection algorithm can only use the activity transition information in the event stream,and the threshold for anomaly determination is single and fixed,resulting in low anomaly detection accuracy.It is also not flexible enough in the changeable online environment.Based on the above background,this thesis studies and implements a real-time anomaly detection algorithm for business process control flow based on LSTM and incremental learning.The main work is as follows:(1)Since the existing algorithms cannot make full use of data features and the anomaly determination threshold is not flexible enough,this thesis proposes to use deep neural networks(LSTM)to extract event context information to achieve more accurate activity transition probability prediction,then proposes an algorithm based on knee point detection to implement the dynamic anomaly determination.Experiments on simulated and real datasets show that the proposed algorithm has better accuracy and robustness than other algorithms in the scene without concept drift.(2)To deal with the concept drift in the online environment,this thesis further proposes using concept drift detection and incremental learning technology to promptly update the weights of models to deal with the decline of anomaly detection accuracy caused by concept drift.The experimental results show that the introduction of concept drift detection and incremental learning technology significantly improves the accuracy and robustness of the proposed algorithm in concept drift scenarios.(3)Based on the proposed algorithm,this thesis designs and implements a prototype system for real-time anomaly detection of business process control flow,which can detect abnormal behaviors in the event flow in real-time and label the event anomaly levels with distinct colors,making it convenient for users to find anomalies and the associated problems quickly.It supplies a blueprint for the practical application of the proposed algorithm and shows its practicability.