Assessing the Feasibility of Damage Prediction and Detection through Multi-Modal Machine Learning in Digital Freight Forwarding Operations
- Project Background: Master's thesis project for Cargoboard GmbH, in collaboration with Fraunhofer IEM.
- Goal: Develop a machine learning framework to detect and predict shipment damage in digital freight forwarding.
- Framework Design: This method combines tabular data and image analysis to efficiently and accurately identify damaged shipments.
- Efficiency Considerations: Focuses on reducing the high computational costs of image processing.
- Validation Approach: Tested models on an imbalanced dataset that combines tabular data with limited images of damaged shipments.
- Technologies Used:
- Tabular Analysis: Random Forest Classifier, showing effective damage prediction.
- Image Analysis: A Vision Transformer (ViT) is used for damage classification, and a YOLO detector is used for damage segmentation. Image diversity limits these methods.
- Challenges: Addressed data imbalance and a scarcity of image data while demonstrating the framework's potential to improve damage detection in logistics.