Research On Broadband Communication Signal Detection Technology Based On Deep Learning

Broadband communication signal detection is the first step of communication signal reconnaissance and surveillance tasks,and it is the key fundamental technology of both electromagnetic spectrum automatic supervision in civil field and electromagnetic spectrum warfare(EMSW)in military field.However,with the development of wireless communication technology,the number of wireless communication signals has increased dramatically and their types have become more complex and diverse in the limited spectrum space,making the broadband communication signals detection facing a huge challenge.In recent years,artificial intelligence technology has developed rapidly with the assistance of high-performance computing hardware.Deep learning,as one of the key technologies in the field of modern artificial intelligence,has a powerful learning capability and has been noticed by many scholars.So,deep learning has been applied to many fields with excellent research results.The purpose of this dissertation is to propose an efficient broadband noise floor estimation method and broadband carrier signal detection method based on the ability of deep learning to adapt to efficient feature characterization,which can fully explore and extract useful information from broadband communication signals,so the methods can adapt to more complex electromagnetic environments,and verify the effectiveness of the methods through simulation data and real data.The study of broadband signal detection in this dissertation has effectively improved the accuracy and efficiency of signal detection.The proposed signal detection algorithm can accommodate a wide range of bandwidth sizes and frequency resolutions,while taking into account both constant and burst signals,making the detection process no longer limited by the signal bandwidth size.The main work of this dissertation includes:1.This dissertation investigates the problem of estimating the noise floor of wireless broadband communication signals.The problem of fluctuating signal power spectrum noise floor due to noise interference in the transmission of wireless communication signals is addressed.The dissertation proposes a wideband power spectrum noise floor estimation method based on deep convolutional self-encoder.Using the corrupted data reduction property of the denoising self-encoder and the efficient feature extraction capability of the deep convolutional neural network(CNN),a one-dimensional deep convolutional self-encoding network is constructed to estimate the noise floor distribution of the power spectrum directly by using the received power spectrum of the noise-interfered broadband communication signal as the input of the network.The effectiveness of the method is verified experimentally,and compared with the traditional nonlinear recursive smoothing filters(NLR)noise floor estimation method,the performance and computational efficiency of the deep convolutional self-coding network are significantly better than the NLR method in terms of quantized noise floor estimation error and intuitive noise floor estimation results.2.This dissertation investigates the problem of carrier signal detection on the broadband power spectrum.To address the problems of leakage and false detection of traditional broadband communication signal detection algorithms,we introduce the concept of power spectrum semantic segmentation based on image semantic segmentation,and design a one-dimensional fully convolutional network(FCN)based on deep learning to detect all subcarriers on the broadband by simply processing the power spectrum semantic segmentation.Numerous experiments have verified the effectiveness of this method.Compared with the traditional algorithm,the method does not require a large amount of a priori knowledge and human intervention,and the one-dimensional FCN network model can adapt to the undulating power spectrum of broadband signals by simulating the generation of broadband communication signal samples and labeling them automatically.3.This dissertation investigates an end-to-end keypoint-based approach for broadband carrier signal detection.By considering the carrier signal detection problem as a target localization task in a one-dimensional broadband power spectrum image,a deep learning-based Spectrum Center Network(SCN)is constructed.The network introduces an attention mechanism in the backbone network for feature extraction,uses a residual network with better generalization performance to extract multi-scale broadband power spectrum features,then refines them by a Feature Pyramid Network(FPN),and finally uses a Regression Network(Reg Net)to directly obtain the broadband All subcarrier center frequencies and bandwidth sizes are obtained directly by Regression Network(Reg Net).The effectiveness of the method is verified by a large number of experiments.The SCN does not require post-processing to complete the detection of each subcarrier position,which effectively solves the problem of insufficient robustness of the carrier signal detection method caused by the overlapping of adjacent subcarrier signals,and the overall detection effect is better than traditional algorithms and other deep learning signal detection methods.4.This dissertation investigates the detection of multi-carrier modulated signals over broadband,and proposes a broadband carrier signal detection method based on multi-frame power spectrum feature fusion.On the basis of preserving the frequency resolution of the broadband signal power spectrum and using the power spectrum of multiple consecutive frames in a short period of time as the network input,the dissertation designs the Frame Fusion Spectrum Center Network(FFSCN)to extract the before-and-after relationship features of sub-carrier signals on multiple consecutive power spectra,which can detect the multi-carrier This dissertation is based on the group convolution,depth-adjustable,and the fusion of the sub-carrier signals.In this dissertation,a lightweight study of FFSCN network model is conducted based on group convolution,deep separable network and adaptive pooling layer.Through simulation experiments,FFSCN can accurately detect subcarrier signals on broadband,including multi-carrier and single-carrier modulated signals.Compared with other deep learning methods,the FFSCN model has better carrier signal detection performance and higher detection efficiency.This dissertation also proposes a joint multi-frame fusion method for wideband carrier signal detection,which uses FFSCN to perform secondary fusion of independent consecutive multi-frame wideband power spectrum signal detection results without slicing the time-frequency map into multiple small pieces for detection,and can efficiently achieve the frequency and temporal location of carrier signals.