![case study](https://cdn.quantiphi.com/2024/03/4ecd7e1d-engine-defect-detection.png)
Engine Defect Detection
ManufacturingBusiness Impacts
Reduced manual effort
95%
Model accuracy
Streamlined inspection process
Customer Key Facts
- Location : North America
- Industry : Automotive
Problem Context
Quality assurance in manufacturing is a critical, yet traditionally manual, cumbersome and time-consuming process. The customer, one of the largest automotive manufacturers in the world, had a quality analysis process which involved a combination of manual and computer-based decisions across the plant.
Manual inspection is necessary due to a high error rate that leads to higher costs, lost productivity, and increased time to production. Therefore, they wanted to empower their Quality Inspectors by augmenting their plants’ capabilities with a machine learning-based model that detects potential defects on a V6 engine casing after the final assembly.
Challenges
- Multi-view point images
- High imbalance in defective versus non-defective classes
- Data augmentation through various techniques
![](https://cdn.quantiphi.com/2024/03/Q_Engine-Defect-Detection-Challenges.png)
Technologies Used
![Python](https://cdn.quantiphi.com/2024/02/5-2-python-logo-png-image-thumb-1.png)
Python
![TensorFlow](https://cdn.quantiphi.com/2024/02/TensorFlow.png)
TensorFlow
Google Cloud Platform
Implementing a Computer Vision Framework for Detecting Defects in Engines on the Assembly Line
Solution
Quantiphi developed a custom and highly accurate computer vision model that identifies the type of defect and localizes the region of the defect in the assembled engines. The model provides a real-time pass or fail test on defects on the V6 engine casing. An active learning pipeline was also built that allows the Quality Inspectors to label a sample set of data for incremental improvement of the model performance over time. The model retraining is scalable globally and enables plants to share information directly at a station level.
Result
- Improvement in model accuracy from 50 percent to 95 percent
- False Positive Rate (FPR) decreased by a factor of 5+ from an unreliable FPR of 50 percent to less than 10 percent
- Increased accuracy per class (defective/non-defective), now within 99 percent to 99.8 percent