Lightweight Adaptive Feature Aggregation Network for Cross-Domain Defect Detection in Data-Scarce L-DED Processes

Laser Directed Energy Deposition (L-DED) faces key challenges such as unstable printing processes, low forming quality, and poor interlayer consistency. An urgent need is to establish real-time monitoring and process control mechanisms to improve its industrial applicability. However, in practical applications, the construction of monitoring systems is constrained by two major bottlenecks. The scarcity of high-quality defect samples limits the generalization capability of deep learning models. Additionally, due to confidentiality concerns regarding printing paths and product defects, users generally refuse to share raw data, further hindering centralized model training and large-scale deployment. To address these challenges, this study proposes a lightweight neural network integrated into a federated transfer learning framework, termed SK-MNV4. The network incorporates a unified inverted bottleneck module, a lightweight multi-head attention mechanism, and a selective kernel fusion strategy to enable efficient multi-scale feature extraction and robust defect representation, significantly enhancing performance under data privacy and limited-sample conditions. The proposed federated transfer learning scheme treats independently collected melt pool image datasets from three representative alloys (IN738, TB6, Ti-10Mo) in L-DED experiments as heterogeneous clients for federated learning. It ensures local data processing and prevents raw melt pool image exposure, while enabling cross-material transfer learning with IN718 melt pool images as the target domain, thereby improving the model's generalization ability. We used T-SNE to compare and visualize each model. We compare our model with mainstream models (ConvNeXt, DenseNet, EfficientNet, and Swin Transformer) in terms of accuracy and inference efficiency. Comparative experiments show that SK-MNV4 outperforms these models and shows 97.3% accuracy and 0.001 seconds inference efficiency on average, demonstrating superior potential for applications in scenarios with strong real-time requirements, data sensitivity, and sparse defect distributions.