1 Our contribution: Edge-to-cloud, worker feedback, and rapid adaptation

Scalable evaluation platform and AI models. We have built a cloud architecture in which general models and application-specific calibrations interact. The system forms the basis for integrating new use cases on a plug-and-play basis – with clear interfaces and top-level architecture. The result: new applications can be fine-tuned in a matter of days or weeks.

Feedback system with worker view. Everything relevant is displayed directly at the operator's workstation – including a specific error description in the timeline, zooming for detailed inspection, and process selection for product changes. Administrative users adjust thresholds and hyperparameters themselves and show/hide classes/error types as needed. This reduces friction, increases the detection rate, and keeps the system flexible

‍

‍

History analysis for quality and process. In addition to live feedback, there is a collective view with filter functions for all recordings. Quality and process engineers can see frequencies, time sequences, and sample images—the basis for root cause analysis and sustainable process improvements.

‍

‍

2    Robust in practice: hard interfaces, stable pipelines

In joint testing at the laboratory measuring station, we hardened interfaces (including cleanly versioned REST APIs, robust authentication/authorization, retry logic) and stabilized the edge-to-cloud pipelines with buffering, QoS, and structured logging. In addition, timing/triggers between the camera, robot, and glue gun were synchronized. The result: stable data transmission and reproducible image quality – regardless of handling.

‍

‍

The stationary camera at the workplace proved to be practical: consistent image quality, efficient evaluation, lower data volumes. The interaction between the glue gun, test bench, and evaluation system met the requirements for accuracy, stability, and response speed.
‍

‍
3 Model and performance optimizations

Based on earlier practical tests, we addressed misclassifications in fluctuating lighting and complex backgrounds and reduced slight overfitting. At the same time, inferences were accelerated (including pipeline optimizations), memory consumption was reduced, and the web application was hardened in terms of loading times and fault tolerance.

‍

4 What users get today

· Worker interface with live feedback, configurable parameters, color coding, and clear error communication.

· Admin functions for model thresholds/hyperparameters and UI configuration without code changes.

· Documentation & manual: UI guides, roles/rights, error scenarios & troubleshooting; plus API specification for integration.

‍

5 Validated – and ready for rollout

Final tests at the laboratory measuring station confirm: high detection accuracy, stable communication, and industrial practicality. With defined ideal parameters (threshold values, tolerances, camera setups), the system can be quickly put into operation at new locations.

Bottom line: With Fluidworks, ANTICIPATE delivers a production-ready, scalable AI evaluation system including worker guidance and analytics. The system is designed so that new adhesive use cases can be quickly adapted – and the architecture provides a robust basis for transferring procedures to other processes involving fluid materials.

FKZ KK5556901, project duration 01.08.2023–31.10.2025. Partners: Reutlingen University, axiss GmbH.

‍