Development of Anomaly Detection with Variable Contexts on Refrigeration Data

Anomaly detection in time series data is a well-known research topic and of great interest to the industry. However, industrial systems are complex due to the nature of real world plants and their interrelationships in the overall system. Industrial plants often cannot be considered in isolation because their environment affects them. Thus, industrial plants have variable contexts in which they are embedded. The combination of a variable context with the information of the plant poses a challenge in modeling AI approaches for anomaly detection.In the field of refrigeration, machine learning approaches are still in their early stages. In live systems, simple threshold values are often still used. There are only a few published approaches to detect error situations using machine learning on such data.In this work, the challenges to such a model architecture are analyzed and a dataset of real refrigeration plants is built. Then, models for detection and classification are explored and their applicability to this type of data is evaluated. Furthermore, models are combined to add context to improve the results compared to models without context information.