Fiche de révision : Automated PCB Fault Diagnosis

📋 Course Outline

1. PCB Structure and Manufacturing Defects
2. Research Objectives and Evaluation Scope
3. Computer Vision and Roboflow Workflow
4. Models and Literature Evidence
5. Experimental Methodology and Tools
6. Performance Metrics and Statistical Tests
7. Detection Workflow and System Specifications
8. Results, Limitations, and Future Development

📖 1. PCB Structure and Manufacturing Defects

🔑 Key Concepts & Definitions

• Printed circuit board : A glass-reinforced plastic board with etched copper tracks that connect electronic components into complex circuits.

📝 Essential Points

★ Must-know

• The study identifies missing components, misalignment, wrong polarity, short circuits, open circuits, solder bridges, component damage, scratches, and damaged traces as possible PCB defects.

Further detail

• PCBs are used in virtually all commercially produced electronic devices because they are rugged, inexpensive, and highly reliable.

🔄 Process — Bare PCB fabrication includes copper foil coverage, photoresist film formation, etching, and film stripping.

💡 Memory Hook

Dense boards create harder inspections

📖 2. Research Objectives and Evaluation Scope

📝 Essential Points

★ Must-know

🔄 Process — The research develops and evaluates a computer-vision fault-diagnosis system using Roboflow for image annotation, dataset management, preprocessing, and model training.

• The system is evaluated using accuracy, precision, recall, F1-score, mean Average Precision, and inference time.

⚡ The study compares automated computer-vision inspection with conventional manual inspection in terms of defect-detection accuracy, inspection time, and consistency.

Further detail

📌 The study excludes PCB repair, real-time industrial deployment, hardware integration, three-dimensional analysis, thermal imaging, X-ray inspection, electrical testing, and economic or maintenance-cost evaluation.

📖 3. Computer Vision and Roboflow Workflow

📝 Essential Points

★ Must-know

🔄 Process — The Roboflow workflow uploads PCB images, annotates defects with bounding boxes, preprocesses and augments images, versions datasets, splits data into training, validation, and testing sets, and exports datasets for model training.

🔄 Process — The complete data-processing cycle consists of PCB image collection, Roboflow annotation, preprocessing and augmentation, computer-vision model training, automated defect detection and classification, quantitative evaluation, statistical analysis, and interpretation of results.

Further detail

• The conceptual framework treats the PCB image dataset, Roboflow annotation tool, defect categories, training parameters, and CNN or object-detection algorithm as inputs; processing produces detected defects, classifications, performance metrics, and the automated diagnosis system.

📌 The study assumes that images are correctly labeled, sufficiently clear, representative of actual defects, and independent between training and testing.

💡 Memory Hook

Collect, annotate, augment, train

📖 4. Models and Literature Evidence

🔑 Key Concepts & Definitions

• Convolutional neural network : A deep-learning architecture used in visual inspection to learn complex visual defect features automatically.

📝 Essential Points

★ Must-know

⚡ YOLO, Faster R-CNN, and RetinaNet provide different trade-offs between object-detection inference speed and accuracy.

• The identified research gaps include limited Roboflow integration, insufficient practical deployment evaluation, dependence on complex frameworks, limited simultaneous multi-class classification, incomplete metric reporting, and limited dataset generalization.

Further detail

• A cited study reported that YOLOv4 achieved 92.3% mAP on a custom PCB dataset containing 2,500 annotated images across eight defect categories.

• The literature reports CNN accuracy exceeding 95% on standardized PCB defect datasets.

💡 Memory Hook

Speed versus accuracy

📖 5. Experimental Methodology and Tools

📝 Essential Points

★ Must-know

🔄 Process — The methodology collects defective and non-defective PCB images, preprocesses and annotates them in Roboflow, splits them into training, validation, and testing datasets, trains an object-detection model, and evaluates its results quantitatively.

Further detail

• The study reports interviewing thirty respondents, including IT experts and laboratory personnel, and using interview materials and evaluation questionnaires as data-gathering instruments.

• The proposed software environment includes Python 3.10 or later, Roboflow, OpenCV, NumPy, Pandas, Ultralytics YOLOv8 or YOLOv11, Matplotlib, Scikit-learn, PyTorch, and Jupyter Notebook or Visual Studio Code.

• The recommended hardware includes at least an Intel Core i5 or Ryzen 5 processor, 16 GB RAM, a 512 GB SSD, an NVIDIA RTX 3060 or higher GPU, a 12 MP or higher camera, adjustable LED lighting, and a Full HD display.

📖 6. Performance Metrics and Statistical Tests

📝 Essential Points

★ Must-know

🧮 Formula — Accuracy is calculated as (TP + TN) / (TP + TN + FP + FN), where TP, TN, FP, and FN are true positives, true negatives, false positives, and false negatives.

🧮 Formula — Precision is calculated as TP / (TP + FP) and indicates how many predicted defects are actually defects.

🧮 Formula — Recall is calculated as TP / (TP + FN) and indicates how many real defects the system detects.

🧮 Formula — The F1-score is calculated as 2 × (Precision × Recall) / (Precision + Recall) and balances precision and recall, especially when defect classes are imbalanced.

📌 Independent tests compare two models, one-way ANOVA compares three or more models, Pearson correlation examines augmentation and accuracy, and Tukey HSD is used after significant ANOVA results.

• The reported model achieved 96.8% accuracy and 97.3% [email protected].

Further detail

🧮 Formula — Inference time is calculated as total processing time divided by the number of images and is measured in seconds per image.

💡 Memory Hook

Accuracy, precision, recall, F1

📖 7. Detection Workflow and System Specifications

📝 Essential Points

★ Must-know

🔄 Process — The inspection flow captures a PCB image, resizes, enhances, normalizes, and crops it, sends it through the Roboflow pipeline, loads the trained model, performs detection, classifies any fault, displays results, and stores them for analysis.

📌 When no defect is detected, the system displays “No Defect”; when a defect is detected, it displays bounding boxes, class labels, and confidence scores.

• Roboflow simplifies image annotation, dataset management, augmentation, model training, and deployment within the proposed inspection workflow.

Further detail

• The system can classify missing holes, short circuits, mouse bites, open circuits, spurs, and spurious copper among its listed defect categories.

💡 Memory Hook

Capture, detect, classify, store

📖 8. Results, Limitations, and Future Development

📝 Essential Points

★ Must-know

• The developed system detected and classified multiple PCB defects while processing images more rapidly and consistently than traditional human visual inspection.

📌 Detection performance depends on high-quality annotated data, sufficient dataset diversity, image quality, lighting, camera position, PCB design complexity, and the representation of rare defects during training.

🔄 Process — Future improvement should expand datasets, compare detection architectures, enhance preprocessing and augmentation, integrate industrial inspection equipment, retrain models continuously, and evaluate multi-class and severity-based classification.

Further detail

• Real-time visualization of bounding boxes and class labels helps inspectors locate defects and verify classifications.

• Potential future integrations include industrial cameras, conveyor systems, robotic inspection platforms, Internet of Things devices, cloud computing, manufacturing execution systems, and predictive analytics.

💡 Memory Hook

Expand, retrain, deploy

📊 Synthesis Tables

Performance Metrics

⚠️ Common Pitfalls & Confusions

1. A PCB is the physical board and not the computer-vision system used to inspect it.
2. mAP is an object-detection metric, whereas accuracy, precision, recall, and F1-score summarize classification performance.
3. Annotation identifies defect locations and classes; augmentation changes training examples but does not replace labeling.
4. CNNs learn features from images, whereas traditional image processing relies more on manually designed features.
5. Accuracy includes both correct defect detections and correct non-defect classifications.
6. Open circuits and short circuits are opposite connectivity failures, not interchangeable defect labels.
7. The scope is comparative evaluation, not replacement of manufacturing hardware.