Abstract
A common challenge in real-world classification scenarios is the "cold start" problem, where target domain data is sequentially appended, leading to insufficient training datasets—specifically, missing defect classes. Conventional deep learning and transfer learning classifiers often fail when individual classes are not represented at the outset. This paper proposes a classification approach based on a CNN encoder trained with a modified triplet loss function. By utilizing a labeled source domain dataset alongside a non-defective target domain dataset, the model learns to transfer relevant defect features across highly different domains. Benchmarks in technical and non-technical domains demonstrate that this architecture improves domain generalization and classification accuracy even with significant domain shifts.
The "Cold Start" Problem in Industrial AI
In many industrial applications, such as the launch of a new production line or mechanical component, manufacturers face a data imbalance: they have an abundance of "non-defective" data but zero examples of "defective" parts because failures have not yet occurred. This creates a significant barrier for standard supervised learning models.
Existing solutions like Domain Generalization (DG) often struggle when the visual gap between the source (e.g., flat steel) and target (e.g., machined spindles) is too large. Domain Adaptation (DA) typically requires at least some target defect samples. The proposed method addresses the scenario where we have a fully labeled source dataset and only non-defective target data.
Methodology: Contrastive Learning with Modified Triplet Loss
The core of the approach is a CNN encoder trained using contrastive learning. The goal is to map images into a feature space where similar classes are clustered together regardless of their domain of origin.
Modified Triplet Loss
Standard triplet loss uses an Anchor, a Positive (same class), and a Negative (different class). However, using standard loss in this cross-domain scenario can lead the model to focus on domain-specific background features (e.g., material texture) rather than defect features.
The authors introduced a modified loss function that forces the encoder to ignore domain-specific context and isolate universal "defect features" like cracks or pitting by:
- Minimizing the distance between non-defective samples across both domains.
- Maximizing the distance between the target's non-defective samples and the source's defective samples.
Experimental Results
The researchers evaluated the model across two distinct scenarios: transferring features from Severstal strip steel to Ball Screw Drive (BSD) components, and transferring disease features from Apple leaves to Bean leaves.
Key Findings
- High Accuracy: In the steel-to-spindle transfer, the model achieved 100% accuracy on non-defective parts and 95% accuracy on defective parts in the target domain, despite never seeing a defective spindle during training.
- Robustness: The model outperformed state-of-the-art R-CNN object detectors, which failed completely (0% detection) when moved to the new domain without target-specific defect training.
- Efficiency: The approach proved effective even with small datasets, as contrastive learning leverages a high number of triplet combinations to build robust feature representations.
Conclusion and Future Work
This framework allows for the deployment of defect detection systems before defects are even observed in a new process. While currently limited to binary classification (defect vs. no-defect), future work aims to extend this to multi-class defect categorization and integrate the logic into object detection frameworks for precise defect localization.
Summary updated