Analysis of Explainable Convolutional Neural Network for Weak Radar Signal Detection

Abstract

In contemporary electronic warfare, the importance of accurate signal detection continues to grow. Recently, detection techniques using convolutional neural network (CNN) have been applied to effectively detect signals. In this paper, we analyze the CNN-based signal detection model using an explainable artificial intelligence (XAI) technique. By employing the XAI technique, we can determine the specific regions within the network's input data that exert a significant impact on prediction through the heatmap. Simulation analysis shows that high weights of heatmap are distributed to areas where signals exist in all layers. In particular, in the layers close to the input, the heatmap significantly reflects the features of the data. In the layers close to the output, the heatmap resolution decreases due to sampling. In addition, analysis results showed that the noise area is flattened due to the activation function.