#NUERAL NETWORK RADAR FULL#

Each unit in the feature graph is obtained by 2*2 down-sampling and non-linear mapping operations, and the down-sampling sliding step is 2.Ĭ3 layer is a convolution layer with 12 feature graphs, which is similar to the structure of C1 layer. S2 layer is a pooling layer with 6 feature graph. Each unit in the feature graph is obtained by the convolution kernel operation, and the sliding step is 1. For the classification and recognition of radar signal, the selection of 32*2-size image can reduce the complexity of model training on the premise that the main features are preserved.Ĭ1 layer is a convolution layer with 6 feature graphs.

The network model consists of six layers, INPUT is a uniform 32*32-size input image. Next we will introduce three modules in detail.

#NUERAL NETWORK RADAR FULL#

CNN network uses convolution layer and pooling layer for feature autonomous learning, and full connection layer uses multi-layer perceptron as classifier, making full use of the advantages of high efficiency of feature extraction and fast classification of CNN. Reasonable image processing method can effectively remove the influence of noise and redundant information, and enhance the expression of feature information, which makes the neural network classifier extract image features better. In order to adapt to the input of neural network and reduce the interference of noise to time-frequency information, the time-frequency image should be preprocessed. In the time-frequency processing part, all signals are transformed into 2-D time-frequency images by Short-time Fourier transform (STFT). *This work is supported by Natural Science Foundation for Young Scientists of Jiangsu Province, China, Grant BK20160148 Natural Science Foundation for Young Scientists of Jiangsu Province, China, Grant BK20160147. Section 6 is the conclusion.Īutomatic identification system for reconnaissance signal based on CNN Simulation analysis will be introduced in section 5. Detailed algorithm is in sections 3 and 4.

In this paper, the second section is the system overview. Finally, CNN-LeNet-5 is designed to realize the autonomous learning and recognition of features. Then, the principal component information of the image is extracted by image processing method. Firstly, the radar signal is transformed into time-frequency image to highlight the modulation type information. The proposed algorithm for automatic recognition of radar signals consists of three steps. The purpose of this paper is to improve the success rate of signal recognition under the low signal-to-noise ratio (SNR) or complex electromagnetic environment conditions. CNN has a high tolerance for translation, scaling, rotation and distortion of information features, which is very helpful for modulation recognition of signals. The rapid development of deep CNN opens up a new way for image feature autonomous learning and recognition of radar signals. Under such complex conditions, if artificial image feature selection is adopted, uncontrollable factors will be added to the recognition of unknown signals, which will lead to the failure of signal recognition. For example, in signal transform domain, time-frequency image contains modulation information of signal, such as color, shape, texture, etc., but also contains a lot of redundant information. However, with the complexity of the electromagnetic environment and the diversity of signal features, in order to achieve higher recognition accuracy, the artificial feature extraction method needs to increase the type of feature extraction. Traditional signal recognition mainly uses artificial feature extraction method, which can achieve high recognition rate for reconnaissance signals with different waveform characteristics.

Autonomous learning and recognition of unknown signal features in electromagnetic spectrum is an important research direction in the field of cognitive electronic warfare.

Nueral network radar