Anomaly Detection Model – Edge AI for Casting Defect Inspection

Overview

Model Details

• Architecture: Convolutional Autoencoder
• Framework: PyTorch
• Training Objective: Minimize reconstruction loss (MSE) for normal samples
• Optimization: ONNX and OpenVINO IR export for edge inference
• Task: Unsupervised anomaly detection
• Domain: Industrial visual inspection

Repository Structure

├── casting_autoencoder.pth    # Trained PyTorch model
├── casting_autoencoder.onnx   # ONNX export
├── model.bin                  # OpenVINO IR model (bin)
├── model.xml                  # OpenVINO IR model (xml)
├── model_card.yaml
├── requirements.txt           # Dependencies
├── inference.py               # inference code
└── README.md                  # Model card (this file)

Dataset

• Classes: Defective / Normal
• Modality: Grayscale industrial images
• Training Strategy: Only normal samples used for training the autoencoder.

Key Configuration Parameters

• Image Size: 304×304 pixels
• Batch Size: 32
• Learning Rate: 1e-3
• Epochs: 10
• Loss Function: MSE Loss
• Optimizer: Adam

Model Outputs

• casting_autoencoder.pth - PyTorch model weights
• casting_autoencoder.onnx - ONNX export for deployment
• Calibrated anomaly threshold based on defective samples

Anomaly Detection Process

1. Training Phase: Model learns to reconstruct normal casting images
2. Threshold Calibration: Uses defective samples to determine optimal threshold
3. Inference: Images with reconstruction error > threshold are flagged as defective

Performance

• Final Training Loss: 0.0005
• Suggested Threshold: 0.0004
• Model Type: Unsupervised anomaly detection
• Architecture: Convolutional Autoencoder

Applications

• Quality control in metal casting manufacturing
• Real-time defect detection on production lines
• Automated visual inspection systems
• Edge deployment in industrial environments

Model Features

• Unsupervised Learning: Trained only on normal samples
• Real-time Capable: Optimized for edge deployment
• ONNX Compatible: Ready for production deployment
• Automatic Thresholding: Self-calibrating anomaly detection
• Industrial Grade: Tested on real manufacturing data

Technical Details

• Stride-2 convolutions for downsampling
• Transposed convolutions for upsampling
• ReLU activation in hidden layers
• Sigmoid output activation for pixel reconstruction.

Export & Deployment

Intended Use

• Automated visual inspection for manufacturing/QA systems.
• Real-time edge deployment in industrial environments.

• Non-industrial datasets.
• Scenarios with significant domain drift (e.g., lighting changes or non-casting objects).

Limitations

• Accuracy depends on lighting and background consistency.
• Model trained primarily on grayscale casting images.
• Thresholds for anomaly detection must be tuned for specific deployment environments.

License

Author