Working Mode Recognition of Non-Specific Radar Based on ResNet-SVM Learning Framework.

Area of Science:

• Radar Systems Engineering
• Machine Learning Applications
• Signal Processing

Background:

• Radar mode recognition is crucial for interpreting multi-functional radar behavior.
• Existing methods often require large neural networks and struggle with training-test data mismatches.
• Signal defects pose challenges for accurate radar mode identification.

Purpose of the Study:

• To develop an effective radar mode recognition framework for non-specific radars.
• To address the limitations of purely data-driven approaches and improve robustness against data mismatches and signal defects.
• To enhance the accuracy and reliability of radar behavior interpretation.

Main Methods:

• A multi-source joint recognition framework (MSJR) combining residual neural network (ResNet) and support vector machine (SVM).
• Embedding prior radar knowledge into the machine learning model for targeted feature learning.
• A two-stage cascade training method to leverage ResNet's data representation and SVM's classification capabilities.
• Combining manual intervention with automatic feature extraction.

Main Results:

• The proposed MSJR model achieved an average recognition rate improvement of 33.7% compared to purely data-driven models.
• Recognition rate increased by 12% compared to other state-of-the-art models (AlexNet, VGGNet, LeNet, ResNet, ConvNet).
• MSJR maintained over 90% recognition rate with 0-35% leaky pulses in the test set, demonstrating robustness.

Conclusions:

• The MSJR framework effectively improves radar mode recognition accuracy and robustness by integrating prior knowledge.
• The model demonstrates superior performance and resilience, particularly under conditions of signal defects and data variability.
• This approach offers a significant advancement in interpreting unknown radar signals with similar characteristics.