Challenge
FRISA, a global leader in seamless rolled rings and open die forgings, supplies high-performance components to aerospace turbine manufacturers. At its plant, over 150 complex rings were manually inspected daily, a process that was:
- Labor-intensive and costly
- Dependent on subjective inspector judgment
- Prone to human error and inconsistent classification
- A bottleneck for engineering teams and decision-making
To reduce costs and improve precision, FRISA sought an AI-driven solution to automate inspection and classification.
Solution
FRISA partnered with Ensitech to develop a custom deep learning system for quality control. The project focused on:
- Aligning and centering 3D point cloud scans of forged parts
- Defining machining height, optimal centering zones, and layout classification
- Building a two-stage pipeline:
- Stage 1: Alignment using classical optimization and data cleaning
- Stage 2: Layout using neural networks trained on labeled inspection data
- Leveraging Microsoft Azure for development and AWS SageMaker for deployment
- Applying transfer learning with architectures like ResNet-34, ResNet-50, Inception, and DenseNet to reduce false negatives
The solution was built collaboratively, with Ensitech embedding mathematicians and engineers to deeply understand FRISA's inspection challenges.
Results
- Achieved 95% precision in identifying parts with potential quality issues
- Eliminated false negatives and reduced false positives to just 5%
- Cut inspection time dramatically, freeing up engineering capacity
- Improved decision-making accuracy and consistency across the plant
- Enabled reuse of AI models and insights in other FRISA development initiatives