Real-time quality assurance using AI at the edge—no cloud dependency, no delays.
Project Overview
A precision metal parts manufacturer required a fully automated defect detection system to eliminate human errors in quality checks across casting and forging lines. Their challenge was to catch surface defects early—without relying on manual inspection or internet-based cloud processing.
Problem Statement
- Manual quality checks were slow, inconsistent, and error-prone, especially across 3 shifts.
- Visual defects like cracks, underfill, blowholes, and shrinkage were sometimes missed or misclassified.
- The client wanted real-time, on-device AI inference without sending image data to the cloud (due to factory network limitations and latency).
- They needed a cost-effective, scalable solution that could be deployed across multiple inspection points.
Our Solution: Edge-AI Based Defect Detection System
Kelectron Technologies designed a fully integrated Edge-AI inspection unit with the following architecture:
| Layer | Component |
|---|---|
| Hardware | NVIDIA Jetson Nano + HD industrial camera + LED ring light |
| AI Model | CNN-based custom defect classifier (trained on 10k+ labeled defect images) |
| Software Stack | Python, OpenCV, TensorFlow Lite |
| Edge Deployment | On-device image capture, real-time inference < 200ms |
| Connectivity | Local dashboard + USB/RS485 alerts to PLC |
Implementation Highlights
- Custom Dataset Preparation: We helped the client label 10,000+ real defect images using a semi-automated annotation tool.
- Model Training: We trained a deep learning model (CNN) to detect over 7 defect classes including surface cracks, blowholes, and overfill.
- Model Optimization: Converted the model to TensorFlow Lite for fast edge inference on Jetson Nano.
- System Integration: Camera + Jetson Nano + custom enclosure for factory floor installation.
- Alerts & Logging: Defect images logged locally and flagged in real time to the operator via PLC buzzer/relay.
Results & Benefits
- 99.3% detection accuracy achieved in production line.
- Detection speed < 200ms per part, enabling inline inspection at conveyor speed.
- Reduced false accept rate by 90%, eliminating manual judgment errors.
- Offline-capable: Zero reliance on internet/cloud, enabling 24/7 uptime.
- Easy scaling: Each unit is standalone; client plans to deploy 12 more units.
Technologies Used
- Hardware: NVIDIA Jetson Nano, Logitech Industrial Camera, LED lighting
- Software: Python, TensorFlow, OpenCV, Flask dashboard
- AI: Convolutional Neural Networks, Image Augmentation
- Protocols: RS485, USB Serial, Local Web Interface