This chapter presents a summary of the issues discussed during the one day workshop on “Support Vector Machines (SVM) Theory and Applications” organized as part of the Advanced Course on Artificial Intelligence (ACAI ’99) in Chania, Greece [19]. The goal of the chapter is twofold: to present an overview of the background theory and current understanding of SVM, and to discuss the papers presented as well as the issues that arose during the workshop.

Support vector machines: Theory and applications

Specific emitter identification (SEI) enables the discrimination of individual radio emitters with the external features carried by the received waveforms. This identification technique has been widely adopted in military and civil applications. However, many previous methods based on hand-crafted features are subject to the present expertise. To remedy these shortcomings, this paper presents a novel SEI algorithm using deep learning architecture. First, we perform Hilbert-Huang transform on the received signal and convert the resulting Hilbert spectrum into a grayscale image. As a signal representation, the Hilbert spectrum image has high information integrity and can provide abundant information about the nonlinear and non-stationary characteristics of signals for identifying emitters. Thereafter, we construct a deep residual network for learning the visual differences reflected in the Hilbert spectrum images. By using the residual architectures, we effectively address the degradation problem, which improves efficiency and generalization. From our analysis, the proposed approach combines high information integrity with low complexity, which outperforms previous studies in the literature. The simulation results validate that the Hilbert spectrum image is a successful signal representation, and also demonstrate that the fingerprints extracted from raw images using deep learning are more effective and robust than the expert ones. Furthermore, our method has the capability of adapting to signals collected under various conditions.

Specific Emitter Identification Based on Deep Residual Networks

Specific emitter identification is a technique that distinguishes different emitters using radio fingerprints. Feature extraction and classifier selection are critical factors affecting SEI performance. In this paper, we propose an SEI method using the Bispectrum-Radon transform (BRT) and a hybrid deep model. We propose BRT to characterize the unintentional modulation of pulses due to the superiority of bispectrum distributions in characterizing nonlinear features of signals. We then apply a hybrid deep model based on denoising autoencoders and a deep belief network to perform further deep feature extraction and discriminative identification. We design an automatic dependent surveillance-broadcast signal acquisition system to capture signals and to build dataset for validating our proposed SEI method. Theoretical analysis and experimental results show that the BRT feature outperformed traditional features in characterizing UMOP, and our proposed SEI method outperformed other feature and classifier combination methods.

/pdf/specific-emitter-identification-via-bispectrum-radon-1ayv7dqiwu.pdf

Specific Emitter Identification via Bispectrum-Radon Transform and Hybrid Deep Model

Aimed at the deficiency of traditional feature extraction techniques in radar emitter recognition, a novel deep feature extraction and recognition architecture is proposed. To fit into the model, the time-domain emitters are transformed into unique time-frequency images correspondingly. Since auto-encoders restrict the input data to be vector-form and convolutional model is hard to optimize, denoising auto-coders are stacked in a convolutional manner in the pre-training stage and the proposed framework is trained by greedy layer-wise algorithm. The optimized network parameters are employed to initialize convolutional neural networks. By layers of mapping and pooling, deep time-frequency features are extracted, which are fed into the collaborative representation-based classifier to implement classification task. Experimental results on simulated data validate the feasibility of the proposed architecture. Furthermore, compared with conventional shallow algorithms, the proposed one can obtain higher recognition accuracy and more robust performance. Taking advantage of collaborative representation, the proposed algorithm is more applicable to the small-sample-size case.

Automatic Radar Waveform Recognition Based on Deep Convolutional Denoising Auto-encoders

Radar signal intra-pulse modulation recognition is an important technology in electronic warfare. A radar signal intra-pulse modulation recognition method based on convolutional denoising autoencoder (CDAE) and deep convolutional neural network (DCNN) is proposed in this paper. First, we use Cohen’s time-frequency distribution to convert radar signals into time-frequency images (TFIs). Then image preprocessing is applied to TFIs, including bilinear interpolation and amplitude normalization. Next, we design a CDAE to denoise and repair TFIs. Finally, we design a deep convolutional neural network based on Inception architecture to identify the processed TFIs. Simulation results demonstrate that CDAE effectively reduces the interference of noise on TFIs classification, and improves the classification performance at a low signal-to-noise ratio (SNR). The DCNN architecture we designed makes good use of computing resources and has a good classification effect. The approach has good noise immunity and generalization. It can classify twelve kinds of modulation signals and an overall probability of successful recognition is more than 95% when the SNR is −9 dB.

/pdf/radar-signal-intra-pulse-modulation-recognition-based-on-31eqbatewf.pdf

Radar Signal Intra-Pulse Modulation Recognition Based on Convolutional Denoising Autoencoder and Deep Convolutional Neural Network

To improve the recognition rate of radar emitters with complex signal system in an awful electromagnetic environment, a new recognition method based on short time Fourier transform (STFT) and convolutional neural networks (CNN) was proposed. In this method, STFT obtains the time-frequency distribution of radar emitter in-pulse modulated signals and CNN extracted the features of different radar signals with the processed data. Before the time-frequency distribution (TFD) arrays were input to the CNN, a base noise reduction was conducted after six-time zero-means scaling. In the end of the classification, an additional operation was made to distinguish NS and BPSK. Simulations were implemented to present the high recognition performance of the method.

Radar emitter recognition based on the short time fourier transform and convolutional neural networks

To cope with the complex electromagnetic environment and varied signal styles, a novel method based on the energy cumulant of short time Fourier transform and reinforced deep belief network is proposed to gain a higher correct recognition rate for radar emitter intra-pulse signals at a low signal-to-noise ratio The energy cumulant of short time Fourier transform is attained by calculating the accumulations of each frequency sample value with the different time samples Before this procedure, the time frequency distribution via short time Fourier transform is processed by base noise reduction The reinforced deep belief network is proposed to employ the input feature vectors for training to achieve the radar emitter recognition and classification Simulation results manifest that the proposed method is feasible and robust in radar emitter recognition even at a low SNR

Radar Emitter Recognition Based on the Energy Cumulant of Short Time Fourier Transform and Reinforced Deep Belief Network.

When signal samples are severely contaminated by interference noise, good emitter recognition can't be achieved in most cases simply by extracting distinctive features and improving the performance of a single classifier. Firstly, vectorized time-frequency features are extracted, and then representation coefficients are obtained in the frame of collaborative representation. Then, a decision-level fusion of multiple sensors is implemented under the maximum activity rule and recognition results are acquired by selecting the minimum residual. The simulation experiments validate the feasibility of the proposed algorithm and show that the recognition rate of fusion is higher than a single classifier, which indicates the good recognition performance.

An emitter fusion recognition algorithm based on multi-collaborative representations

Aimed at the performance decline of conventional radar emitter signals recognition techniques with the presence of noise and impulsive interference, a robust recognition algorithm based on Cauchy random projection is proposed. By virtue of sparse characteristics of impulse system radar and impulsive interference, the more robust l1 norm random projection is combined with the state-of-art sparse representation recognition. As a consequence, the recognition problem is converted into solving l1 norm minimization. The SaS interference model is built, and the recognition performance between l2 and l1 random mapping. Simulation results show the recognition validity of the proposed method in the range of impulse radar signals, and that it's more robust than sparse recognition method based on l2 norm random dimensionality-reduction, which is more applicable for the practical applications.

Zhiwen Zhou

Papers

Automatic Radar Waveform Recognition Based on Deep Convolutional Denoising Auto-encoders

Radar emitter recognition based on the short time fourier transform and convolutional neural networks

Radar Emitter Recognition Based on the Energy Cumulant of Short Time Fourier Transform and Reinforced Deep Belief Network.

An emitter fusion recognition algorithm based on multi-collaborative representations

A robust radar emitter signals recognition algorithm based on Cauchy random projection