Efficient Data Construction Method for Electromagnetic Fault Diagnosis of Radar Timing Control Boards

This work targets four core bottlenecks restricting electromagnetic susceptibility (EMS) fault diagnosis of digital radar boards, including cold-start modeling caused by scarce fault samples, low-efficiency scanning prone to omitting narrow sensitive frequencies, insufficient feature representation of electromagnetic signals, and severe overfitting induced by imbalanced data. An integrated EMS dataset construction framework is accordingly proposed for radar timing control boards based on closed-loop electromagnetic interference (EMI)injection and multidimensional key signal point (KSP) monitoring. The framework adopts an adaptive coarse-to-fine scanning scheme to rapidly lock sensitive bands and accurately resolve fault boundaries. It further incorporates the hybrid synthetic minority oversampling technique and random undersampling (SMOTE-RU) strategy and time–frequency feature filtering to balance sample distribution and remove redundant information, yielding standardized high-quality EMS feature datasets. We validate dataset quality on a suite of classic machine learning and deep learning models. Test results show that all baseline networks attain classification accuracy above 85.1%, while the convolutional block attention module (CBAM) attention-augmented model delivers the highest accuracy of 95.0%. Cross-board tests on dissimilar radar boards show that fine-tuning with 10% target samples drastically raises cross-domain accuracy, and moderate interference yields the best generalization above 91.2%. Incurring minimal retraining costs, this lightweight scheme achieves strong cross-hardware generalization to build EMS databases for radar timing boards and implement multiclass fault diagnosis, effectively solving the cold-start problem for new hardware with insufficient historical fault data.