ResNet-AE for Radar Signal Anomaly Detection.

Area of Science:

• * Signal Processing
• * Machine Learning
• * Radar Systems

Background:

• * Traditional Autoencoder (AE) models exhibit low accuracy for radar signal anomaly detection.
• * Generative Adversarial Network (GAN) models present complex network structures, posing challenges for practical implementation.
• * Anomaly detection in radar signals is crucial for identifying potential threat targets.

Purpose of the Study:

• * To propose a novel anomaly detection method for radar signals using a ResNet-Autoencoder (ResNet-AE).
• * To enhance the accuracy and efficiency of radar anomaly detection compared to existing models.
• * To develop a robust and stable model with good performance across various signal-to-noise ratios (SNRs).

Main Methods:

• * Feature extraction using Convolutional Neural Networks (CNN) to learn data distribution.
• * Time dependence discovery using Long Short-Term Memory (LSTM) networks.
• * Gradient loss mitigation and deep network efficiency improvement using Residual Networks (ResNet).
• * Signal subsequence extraction based on pulse rising and falling edges.
• * Model training on normal radar signals and error calculation using mean square error (MSE).
• * Anomaly determination using an adaptive threshold.

Main Results:

• * The proposed ResNet-AE method achieved a recognition accuracy exceeding 85%.
• * Accuracy was improved by over 4% compared to AE, CNN-AE, LSTM-AE, LSTM-GAN, and LSTM-based VAE-GAN models.
• * Significant improvements were observed in Precision, Recall, F1-score, and Area Under the Curve (AUC).
• * The model demonstrated good performance across different signal-to-noise ratios (SNRs).

Conclusions:

• * The ResNet-AE model offers a superior anomaly detection solution for radar signals.
• * The method presents a simple structure, strong stability, and broad applicability.
• * This approach effectively addresses limitations of traditional AE and GAN models in radar anomaly detection.