Model-driven optimisation of monitoring system configurations for batch production

The increasing need to monitor asset health and the deployment of IoT devices have driven the adoption of non-desctructive testing methods in the industry sector. In fact, they constitute a key to production efficiency. However, engineers still struggle to meet requirements sufficiently due to the complexity and cross-dependency of system parameters. In addition, the design and configuration of industrial monitoring systems remains dependent on recurring issues: data collection, algorithm selection, model configuration and objective function modelling. In this paper, we shine a light on impact factors of machine vision and signal processing in industrial monitoring, from sensor configuration to model development. Since system design requires a deep understanding of the physical characteristics, we apply graph-based design languages to improve the decision and configuration process. Our model and architecture design method are adapted for processing image and signal data in highly sen sitive installations to increase transparency, shorten time-to-production and enable defect monitoring in environments with varying conditions. We explore the potential of model selection, pipeline generation and data quality assessment and discuss their impact on representative manufacturing processes