Radar signal recognition exploiting information geometry and support vector machine

Abstract

Aiming at the recognition of low-probability-of-intercept (LPI) radar signals, a support vector machine (SVM)-based algorithm is proposed, where the information geometry theory is utilised to optimise the kernel function of the SVM. Since signals with different modulations have different characteristics in the time-frequency domain, the timefrequency transformation result of the LPI radar signal is considered as an image, referred to as the time-frequency image, and computer vision techniques are utilized to perform recognition. Specifically, the time-frequency images of different LPI radar signals are obtained via the Choi-Williams distribution (CWD) transform, and the AlexNet network, one improved convolutional neural network (CNN), is used to extract time-frequency features. Then, an SVM is adopted to recognise LPI radar signals due to its superiority in addressing the dimension disaster and non-linear inseparability issue. The extracted time-frequency features are fed into the SVM for classification and recognition. Note that the classification performance of SVM depends on the kernel function. Therefore, in the proposed algorithm, information geometry theory is exploited to improve the Gaussian kernel function, and the maximum margin between different categories of samples is further enlarged. As a consequence, the recognition accuracy for LPI radar signals with similar time-frequency images is effectively improved. In addition, the proposed algorithm has better robustness to small samples than other deep learning-based algorithms, since the SVM method minimises the structural risk instead of the empirical risk. Simulation results verify the effectiveness of the proposed algorithm.

Hand gesture recognition based on micro‐Doppler radar using graph neural network

“…All models in the Table 3 are trained in the same environment. The traditional machine learning algorithms constructed from the the scikit‐learn library, KNN [13], SVM [14], random forest [15], decision tree [16], Gaussian naive bayes [17] and AdaBoost [18], show inferior accuracies compared to our multi‐scale GNN model. For CNN [19] and MLP [20] models, we only replaced SAGEConv in our proposed model (all other hyperparameters are the same) and used the same training hyperparameters.…”

Structural inversion of radar emitter based on stacked convolutional autoencoder and deep neural network

“…We also employ the mapping relationship between the RF structural features and the internal structure of the emitter to realise the inversion of the emitter structure. In the forward modelling process of the radar emitter, although the radar signals of two radar emitters types are different in the time, frequency, and spatial domains, the RF structural features extracted from radar signals generated by the same radar emitter with different modulation modes and the modulation parameters remain the same. Although the radar signals obtained in the process of forward modelling are different in the time, frequency, and spatial domains, the RF structural characteristics of the same radar emitter are only related to the structure itself and are unrelated to the changes in the time, frequency, and spatial domains of the radar signal. Generally, inversion and forward modelling are employed together [25–27]. After completing the forward modelling process, the SCAE network is exploited to extract RF structural features, and a deep neural network (DNN) is used to realise the structural inversion of the radar emitter.…”

“…Generally, inversion and forward modelling are employed together [25–27]. After completing the forward modelling process, the SCAE network is exploited to extract RF structural features, and a deep neural network (DNN) is used to realise the structural inversion of the radar emitter.…”

“…Although the radar signals obtained in the process of forward modelling are different in the time, frequency, and spatial domains, the RF structural characteristics of the same radar emitter are only related to the structure itself and are unrelated to the changes in the time, frequency, and spatial domains of the radar signal. � Generally, inversion and forward modelling are employed together [25][26][27]. After completing the forward modelling process, the SCAE network is exploited to extract RF structural features, and a deep neural network (DNN) is used to realise the structural inversion of the radar emitter.…”

Time-frequency preprocessing method of radiation source signal based on cyclic statistics

“…Among them, the method of extracting time-frequency domain image features to obtain the modulation type and core parameters of radiation source signals has been widely used in radiation source devices such as detection, guidance and communication [2] .However, with the increasing sampling rate of radiation source signal and the increase of computing time, it is difficult for time-frequency analysis to meet the requirements of high real-time signal processing. References [3][4][5][6] use the time-frequency image characteristics of the signal as input to the artificial neural network to identify the signal features, which takes a long time to compute in the time-frequency transformation process. In order to solve the above problems, this paper proposes a time-frequency feature preprocessing method of radiation source signals based on low-order cyclic statistics and CWD time-frequency analysis.…”