A CNC precision 3-axis automated PCB soldering system with vision processing and AI solder recognition for in-app QA/QC. Built as a capstone project; awarded 1st place 🏆
📋 Table of Contents
This project implements an autonomous PCB soldering robot that combines precision CNC motion control, computer vision, and deep learning to automate and verify solder joint quality. The system features a 3-axis CoreXY positioning system, custom electronics, and a MobileNetV2-based AI classifier that can detect good, bad, and missing solder joints in real-time.
1st Place Capstone Project • McMaster University
| 🏭 Mechanical | 🔌 Electronics | 💾 Firmware |
|---|---|---|
| Belt-driven Core XY stage with leadscrew Z-axis and a stepper-gun solder dispenser for precise wire feed and tip placement | Custom Altium PCB integrating TMC2209 motor drivers, endstops/limit sensors, and a 5 V buck for regulated logic rails | C++/PlatformIO with UART telemetry for current limiting, microstepping and diagnostics; motion planner with dynamic error handling and soft-limits |
| 🎥 Vision | 🤖 Machine Learning | 🖥️ Application |
| Raspberry Pi + HQ camera under controlled LED ring lighting; manual bounding-box labeling of solder joints and dataset augmentation for robustness | Morphological pre-processing + MobileNetV2 (TFLite) for real-time Good / Bad / Missing joint classification | Tkinter GUI with live video/image inputs, overlays, per-joint confidences, and CSV logging for on-the-fly assessment and future auto re-solder |
Fig 01. System modules: mechanics, electronics, firmware, vision, ML, and operator app.
The system integrates six major subsystems: mechanical positioning (CoreXY motion + Z-axis), custom electronics for motion control, firmware for kinematics and safety, computer vision pipeline, machine learning classification, and the operator GUI. Each subsystem is designed for modularity and independent development.
Rigid V-slot frame; 12 mm Y-rods for stiffness, 8 mm X-rods to reduce moving mass, GT2 belts on XY; Z by leadscrew. Toolhead integrates a temperature-controlled iron with a geared solder-wire feeder. A kitted tray fixtures boards repeatably.
Fig 02 — Toolhead CAD. Iron + geared feeder align wire at the pad; compact, low-mass carriage.
Fig 03 — CoreXY CAD. Symmetric belt routing with idlers; X-rail on 8 mm rods.
 Fig 05 — Electronics enclosure. PSU + control PCB; thermal/EMI shielding for stable operation. |
 Fig 06 — Adjustable PCB tray. Hard-stop datums for repeatable imaging and soldering. |
Custom controller: Raspberry Pi Pico + TMC2209 drivers on a single board; 24 Volt motion rail with 5 Volt buck for logic. UART access to driver current, microstepping, and diagnostics enables safe current limits and smooth feeds.
Key Electronics Features:
- 🎛️ TMC2209 silent stepper motor drivers with UART configuration
- 🔋 Dual-rail power supply (24V motion, 5V logic)
- 🛡️ Integrated endstops and limit sensors
- 📊 Real-time telemetry and diagnostics
- CoreXY kinematics and soft/hard limit handling
- Variable motor speed (vs. variable steps) for stable ramps
- UART utilities for telemetry & tuning (foundation for future auto re-solder)
- Dynamic error handling with safety interlocks
axis_motion_960.mp4
vision_setup_960.mp4
Lighting-controlled captures (Pi HQ + microscope lens + LED ring) → histogram equalization → BGR→HSV → Otsu threshold on Hue to isolate solder → V-channel gating to drop dim pixels → median filter → morphological cleanup → per-joint crops at 224×224.
 Fig 09. Equalize → HSV → Otsu(H) → V-gating → median → morph → 224×224 crops. |
 Fig 10. Overlay of joint boxes with Good/Bad/Missing labels and confidences. |
- Manual labeling in LabelImg with rectangular boxes around each solder joint; class set: Good, Bad (e.g., bridges/voids/insufficient wetting), Missing (pad present, solder absent).
- Augmentation: rotations, flips, Gaussian noise, slight blur, exposure jitter, and random spatter/dropout (~×7 expansion) to generalize across pad geometry and lighting.
- Splits: balanced train/val/test with class weights to mitigate skew; crops normalized before inference.
- MobileNetV2 (α = 0.75) in TFLite for low-latency inference on Pi; trained on augmented crops.
- Loss: categorical cross-entropy with early stopping; metrics tracked per-class to ensure Missing stays separable from Bad.
Fig 11. Loss/accuracy vs. epochs; early stopping at convergence.
Fig 12. Clear separation for Missing vs Bad.
Tkinter GUI accepts camera / image / video inputs, renders detections with class/confidence, shows FPS, and writes CSV logs. Interface is structured to publish UART messages to the Pico for closed-loop re-solder in the next revision.
Fig 13. Batch assessment: per-joint labels & confidences across multiple boards; all decisions logged to CSV for QA and future closed-loop re-solder.
For comprehensive technical details, architecture decisions, design rationale, and implementation specifications, refer to the complete Capstone Final Report:
📄 S05 - Capstone Final Report.pdf
This 70+ page report covers:
- 📐 Complete mechanical design and CAD models
- 🔌 Detailed electronics schematics and PCB layout
- 💻 Firmware architecture and motion control algorithms
- 🧠 ML model development, training methodology, and evaluation
- 📊 Experimental results, testing procedures, and performance metrics
- 🔮 Future work and extension opportunities
- Motion Control: CoreXY belt-driven positioning system
- Actuators: NEMA 17 stepper motors with TMC2209 drivers
- Microcontroller: STM32H503 (ARM Cortex-M33)
- Vision System: Raspberry Pi 4 + HQ Camera Module
- Motion System: GT2 belts, linear rods (8mm/12mm), leadscrew Z-axis
- Firmware: C++17 with PlatformIO
- Communications: UART telemetry for motor control
- GUI: Tkinter-based desktop application
- Computer Vision: OpenCV for image preprocessing
- Model: MobileNetV2 (α = 0.75) with TensorFlow Lite
- Training: Two-phase fine-tuning with data augmentation
- Framework: TensorFlow/Keras
- Preprocessing: HSV color space, Otsu thresholding, morphological operations
- Design Tool: Altium Designer for custom PCB
- Power Management: 24V motion rail, 5V buck converter
- Motor Drivers: TMC2209 with UART configuration
- Sensing: Endstops, limit switches, temperature monitoring
- Labeling Tool: LabelImg for bounding box annotation
- Augmentation: Rotations, flips, noise, blur, exposure variations (~7x expansion)
- Dataset: Balanced train/val/test splits with class weighting
| Metric | Performance |
|---|---|
| 📏 Imaging Precision | Repeatable at ~10″ working distance with low-glare LED fill |
| 🤖 Classification Accuracy | Robust segmentation + accurate 3-class decisions |
| 🔗 System Integration | Full motion + solder-feed PoC with UART telemetry |
| ⚡ Real-time Processing | Live video feed with overlay and CSV logging |
| 🛡️ Safety Features | Soft/hard limits, current monitoring, error handling |
Additional Highlights:
- ✅ Complete end-to-end pipeline from image capture to quality assessment
- ✅ Modular architecture enabling easy upgrades and modifications
- ✅ Production-ready GUI with exportable results for QA/QC workflows
- ✅ Comprehensive telemetry system for diagnostics and tuning
- 🔄 Unify capture + inference on Picamera2 end-to-end
- 🔗 Close the loop: UART-triggered re-solder routines
- ⚡ Accelerate on Jetson/Coral; expand dataset with more boards/edge-cases
- 📊 Real-time analytics dashboard for production monitoring
- 🌐 Web-based interface for remote monitoring
Arji Thaiyib, Arjun Bhatia, Ahmad Ali, Abdullah Hafeez, Mayar Aljayoush
Supervisors: Dr. S. Shirani, Dr. C. Chen
This project is licensed under the MIT License - see the LICENSE file for details.
Made with ❤️ by the PCB Solder Robot Team