Scalable AI Inference for Manufacturing
Nordex Standardizes Visual Quality Inspections Worldwide Using Azure Machine Learning
Scalable AI Image Inspection in Nacelle Assembly With Azure Machine Learning
Nordex, one of the world’s leading manufacturers of wind turbines, previously conducted visual inspections of key components—such as nuts, cable harnesses and fittings—during nacelle assembly. The goal was to introduce a unified, scalable image inference standard that increases inspection coverage and documentation quality across all global sites. b.telligent developed a modern Azure ML–based cloud inference architecture.
Quick Facts About the Project
Germany, Manufacturing
Enterprise
3 Months
Architecture & Cost Assessment, MVP Operations
Microsoft Azure
1
From analysis to optimization: Image processing accelerated to around 1 second per image and validated via benchmark.
Initial Situation & Challenge
Previously, Nordex conducted visual inspections of complex components during nacelle assembly—for example, essential elements such as bolts, cable harnesses, and fastenings. The goal was to establish a unified, reliable standard for image inference to ensure consistent inspections across all production sites, increase testing coverage, guarantee completeness, and enable comprehensive documentation.
In parallel, management requested reliable estimates on development and operating costs as well as projected savings from improved defect detection—hard numbers that are naturally limited in early project phases, whereas soft factors such as standardization, reproducibility, and governance already offer clearly measurable value.
b.telligent was tasked with addressing both needs: evaluating an optimal inference architecture focused on latency, quality, scalability, and cost, and ultimately creating a sound decision-making framework—particularly regarding Graphics Processing Unit (GPU) versus Central Processing Unit (CPU).
Solution
At the start of the project, the existing AI setup was thoroughly analyzed and significantly accelerated through a structured optimization phase. A reproducible benchmarking method was developed to compare various model approaches and realistically assess performance for high-resolution production images. This made image processing noticeably more stable and confirmed a target inference time of roughly one second per image.
Building on this, the appropriate MVP architecture was defined and implemented. Azure Machine Learning was chosen to create a scalable, secure, and flexible cloud solution that ensures consistently low latency even as volumes grow—without requiring high upfront investments.
For future productive use, a long-term target architecture was designed based on Azure Kubernetes Service. This operating model provides elastic scaling, high availability, and clear options for cost optimization—essential prerequisites for international 24/7 operations.
Finally, a complete MLOps and governance framework was established. Telemetry, automated quality and drift monitoring, versioning, and defined approval processes ensure transparency and security. The result is a ready-to-implement solution with a clear roadmap for global rollout—including consistent inspection processes that reliably support even the most demanding micro-feature detections.
Voices From the Project
With ONNX optimization, ensemble models, and Managed Endpoints, we were able to reduce latency to around one second per image—reproducible, scalable, and audit-proof. Through rapid iterations in the MVP and a clear AKS target architecture, we jointly transformed the successful proof of concept into an efficient production setup.
Martin Graf
Project Lead, b.telligent
b.telligent Services at a Glance
AI setup analysis & optimization
Performance analyses, benchmarking methods, and stabilization of image processing.
Model comparison & GPU/CPU evaluation
Reproducible tests for different model and hardware approaches.
Architecture design for the MVP
Development of a scalable Azure Machine Learning architecture for low latency.
Implementation using Managed Endpoints
Rapid deployment and iteration in the MVP phase through Azure ML Online Endpoints.
Target architecture on AKS
Definition of a production-ready operating model with elastic scaling and cost optimization.
MLOps & governance framework
Telemetry, automated quality checks, versioning, and defined approval workflows.
Results & Successes
Highlights
Efficient Production: Reliable detection of even small features such as bolts and cable harnesses.
Fast Development Cycles: Lower initial effort and rapid iterations through cloud deployment.
Optimized Operating Costs: Automatic scale-down and flexible provisioning scenarios (regional/global).
Implementing the architecture recommended by b.telligent enables Nordex to perform reliable, fast, and scalable image inference in production. The cloud-based approach significantly reduces investment and operational costs compared to on-premises setups while providing the flexibility needed for future extensions. The integration of Azure Machine Learning ensures efficient development cycles, reduced complexity, and an optimized cost structure during productive operations.
Want the Inside Scoop?
Tune in to the podcast now!
The Tech Behind the Success
Microsoft
Innovation and integration, as well as interoperability, are key factors of the Microsoft product development. Learn more about our collaboration!
Management Summary for Download
Would you like to explore the key insights from our project in a concise and easy-to-read format – or share them with your team?
Simply download the Management Summary as a PDF and discover the main highlights of our success story at a glance. Enjoy reading!
Inspired?
Did our success stories spark your interest? If you're facing similar challenges in data, analytics and AI and look for expert support, let’s talk. A brief call can reveal how we can help you move forward.
