Interactive Drift Visualization in Sensor Data Streams for Explainable Process Outcome Prediction

In real-world process scenarios such as manufacturing and logistics, the process outcome is frequently predicted by IoT sensor data streams and their drifts, e.g., the quality of a product can be affected by the temperature during the production process. In particular, drifts can explain variations in the process outcome, and hence, their early detection can support the definition of mitigation actions, e.g., canceling the production of probably low quality products. As in most cases, humans have to define such actions, it is crucial to support them in making decisions in the context of outcome prediction. Hence, this paper aims to support the interactive visualization of drifts in specific points of sensor data streams to show the development of critical sensor measurement points over different traces. Furthermore, these critical points can be identified based on drifts between distinct groups of traces. Being able to visualize drifts and identify critical points enables (early) outcome prediction, ranging from the analysis of drifts at single points or at specific timestamps in a trace to the investigation of average time series of traces representing different outcomes, e.g., OK/NOK. Three different methods to derive and visualize drift points are presented and evaluated based on a prototypical implementation and a survey with users.