Study On Extreme Learning Machine Based Communication Signal Transmitter Recognition Technology

Communication signal transmitter recognition includes not only characteristicfine feature extraction from the sampled sequence generated by di?erent transmitterswith the same expected performance parameters and modulation mode, which actsas a comprehensive representation of the transmitters , but a highly e?cient andreal-time classification strategy to implement the online transmitter recognition. Itis of great significance and role in future electronic warfare, wireless network securityand communication resource management.Unlike usual e?ciency- and reliance- oriented communication theory, communi-cation signal transmitter recognition focuses on the hidden fine feature set associatedwith the variances in actually working performance parameters and nonlinear prop-erty of the device and cares no details about modulated information. This papersystematically and thoroughly discusses the generating mechanism and extractionmethods of fine features from the sampled signal observations of di?erent steady-working transmitters for the first time, proposes an adaptive online sequential learn-ing extreme learning machine to train the Single hidden-Layer Feedforward Neuralnetwork (SLFN) to fulfill the classification and identification of transmitter basedon the extreme learning machine, which attains good recognition rate. The maincontributions are as follows:1. From the theoretical and engineering point of view, the discrete HT-based theoryand method for communication signal measured feature (carrier and modulationparameters) extraction are presented, which includes unwrapped instantaneousphase based linear fitting carrier estimation and absolute envelope based symbolrate estimation for digital communication signal, etc. Lots of simulations andexperiments show the proposed methods are e?ective and simple to implement.2. To extract such modulation-relative features as FM modulation index and dig-ital modulation parameters, two generalized, DSP implementable FFT-basediteratively interpolated and CZT-based high performance single tone parameterestimator are proposed. The FFT-based one adaptively adjusts the estima- tion accuracy by the constant iteration of dynamic symmetric frequency o?setestimation to attain not only more accurate estimation but lower SNR thresh-old, higher speed and much more CRLB-approximating performance over thefull frequency range compared to traditional Rife-Jane's and Quinn's one. TheCZT-based one enhances the frequency resolution by means of certain priori toobtain an accurate, stable, consistent, closely CRLB-approximating single-toneparameter estimation as well as more robustness to noise in the case of short-timestationarity and adjacent single-tone interference3. Based on nonlinear parasitic amplitude modulation miscellaneous output and sub-sampling, featured bispectra e?ectively distinguish the spontaneous intrinsic fre-quency from that caused by nonlinear QPC, retain all property of HOS butgreatly reduce the computation e?ort meanwhile. Surrounding line Integration-based featured bispectra selection incorporates advantages of both integrated bis-pectra and selected bispectra: strong noise robustness, no cross interference andsimple, e?cient computation, accomplishes the translation from 2-D featuredbispectra into 1-D vectored feature which shows great individual transmitterdiscriminant ability.4. Based on the time-frequency property of individual transmitters, EMD time-frequency spectrum entropy- based miscellaneous feature extraction is proposed,which characterizes the nonlinear, nonstationary miscellaneous output by adap-tive time-frequency distribution. Correlation degree- based decomposition fin-ished criterion speeds up the EMD process and the concept of entropy e?ectivelydepicts the degree of uniform distribution of the energy in EMD time-frequencyspectrum as a simple scalar of some individual transimitter discriminant ability.5. Based on the extreme learning machine, contribution of randomly generated hid-den node in SLFN to network learning accuracy is proposed by regressive fittingand orthogonalization to realize the adaptability of network model one-by-oneor chunk-by-chunk and adaptive online sequential learning of SLFN which notonly runs automatically without manual parameter setting and special require-ment on the arriving mode and order of training samples but greatly reduces therequirements on storage and computation ability of hardware. Lots of simula- tions and experiments show that the proposed algorithm is e?ective and simpleto attain a recognition rate of above 85% for the collected both FM and FSKindividual transimitter observations.Moreover, based on COM and Database techniques, a communication signaltransmitter recognition software system is established to implement our proposedalgorithms, which identifies individual transmitters by carrier, modulation parametersand miscellaneous feature variance according to conditions and criteria of these finefeatures.