Researches On The Weakly-Supervised Deep Learning Based Intelligent Signal Analysis And Processing Technologies

In recent years,the intelligent signal analysis and processing technology has received increasing attention,and it plays an important role in the fields of information reconnaissance,electronic warfare and precision strike.Deep learning technology has extremely high application value for signal analysis and processing tasks because of its excellent ability to automatically learn distinguishable features for tasks,such as signal detection and modulation recognition.However,since the actual electromagnetic environment is complex and changeable,we are facing the problems like insufficient data,inaccurate labels and inexact labels,which are in contradiction with the demand of deep learning on large-scale,accurate and complete supervision data.The weakly-supervised deep learning technology utilizes incomplete supervision data to train the deep learning model,and its demand on data is consistent with the data characteristics of the actual complex electromagnetic environment.This technology is expected to solve the problem of artificial intelligence application in the actual electromagnetic environment.Therefore,this dissertation researches intelligent signal analysis and processing technologies based on the weakly-supervised deep learning.Aiming at the problems of insufficient data,inaccurate labels and inexact labels in the modulation recognition and the wideband signal detection tasks,this dissertation proposes intelligent signal analysis algorithms based on the weakly-supervised deep learning.In orded to solve the application problem of intelligent model in the game environment,this dissertation explores an adversarial signal generation algorithm.In view of the problem that intelligent signal analysis and processing models cannot be deployed on devices with limited computing resources,this dissertation proposes a convolutional neural network lightweight algorithm based on progressive mimic learning.Through the above researches,it is expected to solve some problems in the application of the intelligent signal analysis technology.The main work contents are as follows:1.In order to solve the problem of insufficient data in the modulation recognition task,this dissertation proposes a few-shot multimode modulation recognition algorithm based on a self-supervised and deep denoise network.Firstly,this dissertation analyzes the influence of complex channel on the modulation signal and proposes nine modulation signal argument methods via modifying influence parameters in the frequency-domain,time-domain,and amplitude domain.Then,this dissertation constructs a deep denoise autoencoder based on self-supervised learning and utilizes channel effect modules of the GNU radio to generate self-supervised denoise data pair for the training of the autoencoder.The autoencoder is employed to remove the noise of the modulation signal,in order to minimize the influence of channel effects on the extraction of modulation domain features and reduce the demand of model training on the amount of supervision data.Finally,a few-shot modulation recognition model based on multimode inputs is constructed to fuse the denoised signal and the modulation signal for the improvement of the ability to recognize modulation types.Experimental results on RML2016.10 a and RML2016.10 b data sets show that compared with semi-supervised models,such as SSRCNN(Semi-supervised Signal Recognition Convolutional Neural Network)and VAT(Virtual Adversarial Training),the proposed algorithm increases the overall accuracy by 1.09%~28.24%.2.Aiming at the problem of inaccurate labels in the modulation recognition task,this dissertation proposes the modulation recognition algorithm robust to noise labels based on the determinant information entropy loss.Firstly,a determinant information entropy loss is designed.As the loss measures the mutual information between labels and model outputs,it can ignore noise labels that have less mutual information.Therefore,the loss is robust to noise labels.Then,the latent class-conditional noise estimation method is introduced into the determinant information entropy loss.The method can approximately estimate the noise transfer matrix through Gibbs sampling,and correct noise labels for the training of the modulation recognition model.Experimental results on RML2016.10 a and RML2016.10 b data sets show that: when noise labels account for 80% of all labels,compared with noise label learning algorithms,FL(Forward Loss),GCE(Generalized Cross Entropy),and LCCN(Latent Class-Conditional Noise),the overall accuracy of the proposed algorithm is increases by 3.50%~54.46%,respectively.3.Aiming at the problem of inexact labels in the wideband signal detection task,this dissertation proposes a weakly-supervised intelligent wideband signal detection algorithm based on label propagation.Firstly,in order to solve the problem of inaccurate proposals due to the lack of color and texture information in the time-frequency map of the wideband signal with low signal-to-noise ratio,this dissertation uses the significance of signal objects activated by the class label to calculate the coordinate of the narrowband signal object’s center and roughly predict the position of the object.Secondly,the label propagation algorithm based on pixel semantic affinity is employed to estimate the signal object’s boundary,where the object center is iteratively propagated within the object boundary for the pixel-level category prediction.Finally,the minimum circumscribed rectangle method is utilized to generate the minimum circumscribed rectangle of the pixel-level category prediction for the object bounding box.Experimental results on the SWSD signal data set show that compared with the weakly-supervised object detection algorithm,WSDDN(Weakly Supervised Deep Detection Network),the proposed algorithm increases the of 24.32%,the of 12.94%,and the of 7.15%,respectively.Compared with the object detection algorithm,YOLO(You Only Look Once)v4,the proposed algorithm increases the of 4.14%.4.Aiming at the application problem of the intelligent model in the game environment,this dissertation explores an adversarial signal generation algorithm based on vision and information constraints for the model attack and defense.First of all,in order to solve the problem that adversarial signals generated by existing algorithms have large visual deformation and large changes in transmission information,the constraint that only allows local disturbances in the neighborhood is proposed to generate adversarial signals.Secondly,the differential evolution algorithm based on the population evolution mechanism is introduced into search for the globally optimal adversarial signal.Experimental results on RML2016.10 a and RML2016.10 b data sets show that: compared with adversarial samples generation algorithms,FGSM(Fast Gradient Sign Method),BIM(Basic Iterative Method),JSMA(Jacobian-based Saliency Map Approach)and UAP(Universal Adversarial Perturbations),the proposed algorithm increases the attack rate of 2.50%~49.92%.5.Aiming at the problem that intelligent signal analysis and processing models cannot be deployed on devices with limited computing resources,this dissertation proposes a convolutional neural network lightweight algorithm based on progressive mimic learning.The algorithm regards knowledge learned by a teacher model and the model’s optimized path as the supervision information,and utilizes the mapping of the optimized path to constrain the optimization process of the lightweight neural network.Specifically,multiple nodes on the optimized path are regarded as optimization landmarks,and a designed landmark loss progressively uses the mapping of different landmarks to periodically train the lightweight neural network for the knowledge from the teacher model.Experimental results on RML2016.10 a and RML2016.10 b data sets show that compared with knowledge distillation algorithms,KD(Knowledge Distillation),PAA(Paying more Attention to Attention)and RL(Rocket Launching),the proposed algorithm increases the overall accuracy of 1.29%~4.25%.