The paper describes an approach to detect anomalies in the timestamps of events that happen during process execution. The approach is based on historical observations of usual behavior and uses a stochastic model to find outliers in a non-parametric way. It helps process analysts to identify and filter errors from process logs, and also to distinguish between measurement errors (where only one timestamp is the outlier) and real delays (where succeeding activities are also shifted).
The analysis of business processes is often challenging not only because of intricate dependencies between process activities but also because of various sources of faults within the activities. The automated detection of potential business process anomalies could immensely help business analysts and other process participants detect and understand the causes of process errors.
This work focuses on temporal anomalies, i.e., anomalies concerning the runtime of activities within a process. To detect such anomalies, we propose a Bayesian model that can be directly inferred form the Petri net representation of a business process. Probabilistic inference on the above model allows the detection of non-obvious and interdependent temporal anomalies.