GNMLC-Net: a graph neural inference and multi-layer context-aware fusion network for high-precision detection of avionics thermistor solder joints

Abstract

In the aviation fields, the integrity of the thermistor solder joints is critical to system reliability. The defect detection of avionics thermistor solder joints stills faces significant challenges from the complex background noise, extreme scale variations between defect types, and low-contrast characteristics of tiny defects. To solve these issues, a Graph Neural Inference and Multi-layer Context-Aware Fusion Network (GNMLC-Net) is proposed. A Dynamic Hypergraph Block is proposed to capture long-range non-local dependencies by constructing a semantically-aware interaction space, while a Unified Cross-Scale Feature Integration and Remapping network is developed to optimize cross-level information integration through an Aggregate-Refine-Redistribute strategy. Furthermore, a Cascaded Refinement and Context-Gated Semantic Injection mechanism is designed to enhance sensitivity to small targets by refining key feature information and improving the quality of feature representations. This design effectively mitigates the semantic-spatial mismatch issue in deep networks while enhancing multi-scale feature representation capabilities. On a custom-built avionics thermistor solder joint dataset acquired via active infrared thermography, GNMLC-Net achieves a state-of-the-art mAP@0.5 of 93.4%. Notably, for the highly challenging tiny pinhole defect, its accuracy reaches 86.1%, showing an improvement of 12.8% over the baseline model. To validate its generalization, GNMLC-Net achieves a 97.8% mAP@0.5 on the PKU-Market-PCB dataset, demonstrating superior performance over all current comparison methods.