Intelligent frequency spectrum calculation method and device based on multi-neural network fusion
Through the intelligent spectrum calculation method of multi-neural network fusion, the problem of insufficient accuracy and timeliness of traditional electromagnetic spectrum measurement methods is solved, efficient and accurate electromagnetic spectrum measurement is achieved, and spectrum perception ability and combat effectiveness are improved.
Patent Information
- Application Number
- CN202410983044.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-22
- Publication Date
- 2024-11-12
AI Technical Summary
Traditional electromagnetic spectrum measurement methods cannot meet the accuracy and timeliness requirements of electromagnetic spectrum calculation, and there are problems such as measurement error, decreased aging, high calculation complexity and insufficient anti-interference ability.
Using intelligent spectrum calculation method based on multi-nerve network fusion, we can create a training time series by establishing a channel model and electromagnetic signal characteristics, performing cyclic spectrum analysis and feature extraction, and using multi-nerve fusion network model for learning and real-time analysis, establishing Cognitive knowledge tree to improve the accuracy and efficiency of spectrum measurement.
It realizes electromagnetic spectrum measurement with low computational complexity, high measurement accuracy, strong timeliness and strong anti-interference ability, and improves spectrum perception and combat effectiveness.
Smart Images

Figure CN118944784A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of wireless communication technology, and in particular to an intelligent spectrum calculation method and device based on multi-neural network fusion. Background Art
[0002] Electromagnetic spectrum resources exist in the entire time and space of the entire modern electronic countermeasure environment, are distributed in all elements of war, and are reflected in the entire process of combat. The increasing congestion of electromagnetic frequency bands and the relative scarcity of electromagnetic resources have gradually deepened the contradiction between supply and demand of electromagnetic spectrum resources in the electronic countermeasure environment. How to make full use of limited electromagnetic spectrum resources and enhance the competitiveness of my country's electronic information equipment in the entire frequency band has extremely far-reaching research value.
[0003] Due to the increasing complexity of the electromagnetic environment of the modern electronic battlefield and the gradual increase in the number of objects using the electromagnetic spectrum, the requirements for electromagnetic spectrum detection efficiency have further increased. Rapid and accurate spectrum measurement of unknown signals in complex electromagnetic environments will enhance the spectrum perception capability of the transceiver system in complex electromagnetic environments, thereby further improving the combat effectiveness of electronic equipment.
[0004] In the study of traditional electromagnetic spectrum measurement methods, the energy detection method requires certain manual operations, which will bring certain measurement errors and reduce the measurement time efficiency; the matched filter detection method requires certain prerequisites to be obtained in advance and a more complicated demodulation process; the cyclostationary feature detection method has certain noise resistance, but increases the detection time and calculation complexity. Since the traditional electromagnetic spectrum measurement method cannot meet the accuracy and time efficiency requirements of electromagnetic spectrum calculation, there is an urgent need to study more efficient electromagnetic spectrum calculation methods. Summary of the invention
[0005] The purpose of the present invention is to provide an electromagnetic spectrum measurement method and device based on multiple neural network fusion learning with low computational complexity, high measurement accuracy, strong timeliness, strong operability, strong anti-interference ability and high intelligence.
[0006] The technical solution to achieve the purpose of the present invention is: an intelligent spectrum calculation method based on multi-neural network fusion, comprising the following steps:
[0007] Step 1: Establish a training framework based on the channel model and electromagnetic signal characteristics in an actual electromagnetic countermeasure environment, and generate a training time series for learning under the framework;
[0008] Step 2: Send the generated training time series to the learning preprocessing calculation module to obtain the cyclic spectrum of the training time series and the digital features corresponding to the cyclic spectrum;
[0009] Step 3: Send the cyclic spectrum of the training time series and the digital features corresponding to the cyclic spectrum into the multi-neural fusion network model for learning, classify and store the output results, and establish a cognitive knowledge tree;
[0010] Step 4: sending the unknown signal sequence collected in real time to a real-time analysis preprocessing module to obtain a cyclic spectrum of the unknown signal sequence;
[0011] Step 5: Input the cyclic spectrum of the unknown signal sequence into the multi-neural fusion network model for analysis to obtain the cyclic spectrum features corresponding to the unknown signal sequence and the corresponding real-time analysis results;
[0012] Step 6: If no valid result is obtained, it is recorded as an invalid unknown signal sequence, and the invalid unknown signal sequence is used as a training time series. The training time series is sent to the learning preprocessing calculation module, and the invalid unknown signal sequence is used to re-learn the existing multi-neural fusion network model, and the cyclic spectrum and digital features of this type of invalid unknown signal sequence are recorded, the output result is saved, the scope of the cognitive knowledge tree is expanded, and return to step 4.
[0013] An intelligent spectrum calculation device based on multi-neural network fusion, the device is used to implement the intelligent spectrum calculation method based on multi-neural network fusion, the device includes a cognitive learning system and a real-time analysis system, wherein:
[0014] The cognitive learning system is based on a training framework established based on the channel model and classical electromagnetic signal characteristics in an actual electromagnetic confrontation environment. Under the framework, a training time series for learning is generated, and a learning preprocessing calculation module is entered to obtain the cyclic spectrum of the sequence and the digital features corresponding to the cyclic spectrum as input parameters. Learning is performed in a multi-neural fusion network model, and the output results are classified and stored to establish a cognitive knowledge tree.
[0015] The real-time analysis system sends the unknown sequence collected in real time to the real-time analysis preprocessing module to obtain the cyclic spectrum of the unknown signal, inputs it into the multi-neural fusion network model for analysis, and obtains the corresponding cyclic spectrum characteristics and the corresponding real-time analysis results; if no valid results are obtained, it is recorded as an invalid unknown signal sequence, and the invalid unknown signal sequence is sent to the learning preprocessing calculation module as a training time series, and the existing multi-neural fusion network model is re-learned using the invalid unknown signal sequence, and the cyclic spectrum and digital characteristics of this type of invalid unknown signal sequence are recorded, the output results are saved, and the scope of the cognitive knowledge tree is expanded.
[0016] A mobile terminal includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the intelligent spectrum calculation method based on multi-neural network fusion is implemented.
[0017] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps in the intelligent spectrum calculation method based on multi-neural network fusion.
[0018] Compared with the prior art, the present invention has the following significant advantages:
[0019] (1) Using neural networks and deep learning technology, we can effectively analyze the characteristics, spectrum patterns, decision-making rules, etc. of various complex electromagnetic signals, and establish a cognitive knowledge base based on this;
[0020] (2) In the actual electromagnetic spectrum measurement process, the time waste caused by manual operation in the traditional process is omitted, and the electromagnetic spectrum recognition ability of the system and the intelligence level of spectrum measurement are improved;
[0021] (3) It overcomes the problems of low measurement efficiency and high limitations of matched filtering methods brought about by common energy measurement methods. While ensuring measurement accuracy and measurement stability, it improves measurement efficiency and anti-interference ability. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] Figure 1 It is a flow chart of an intelligent spectrum calculation method based on multi-neural network fusion of the present invention.
[0023] Figure 2 It is a processing flow diagram of the learning preprocessing calculation module in the present invention.
[0024] Figure 3 It is a processing flow diagram of the real-time analysis preprocessing module in the present invention.
[0025] Figure 4 It is a three-dimensional structural diagram of the BPSK cycle characteristics in a 0dB signal-to-noise environment in an embodiment of the present invention.
[0026] Figure 5 It is a structural diagram of a multi-neural fusion network model in an embodiment of the present invention.
[0027] Figure 6 It is a structural diagram of a convolutional neural network model CNN in an embodiment of the present invention.
[0028] Figure 7 It is a schematic diagram of the structure of a recurrent neural network RNN in an embodiment of the present invention.
[0029] Figure 8 It is a schematic diagram of the structure of the hidden layer cells of the LSTM network of the long short-term memory network model in an embodiment of the present invention.
[0030] Fig. 9It is a schematic diagram of the forward propagation process of the long short-term memory network model LSTM network in an embodiment of the present invention.
[0031] Fig.10 It is a structural diagram of a deep neural network model DNN in an embodiment of the present invention. DETAILED DESCRIPTION
[0032] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.
[0033] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.
[0034] Combination Figure 1 The present invention provides an intelligent spectrum calculation method based on multi-neural network fusion, comprising the following steps:
[0035] Step 1: Establish a training framework based on the channel model and electromagnetic signal characteristics in an actual electromagnetic countermeasure environment, and generate a training time series for learning under the framework;
[0036] Step 2: Send the generated training time series to the learning preprocessing calculation module to obtain the cyclic spectrum of the training time series and the digital features corresponding to the cyclic spectrum;
[0037] Step 3: Send the cyclic spectrum of the training time series and the digital features corresponding to the cyclic spectrum into the multi-neural fusion network model for learning, classify and store the output results, and establish a cognitive knowledge tree;
[0038] Step 4: sending the unknown signal sequence collected in real time to a real-time analysis preprocessing module to obtain a cyclic spectrum of the unknown signal sequence;
[0039] Step 5: Input the cyclic spectrum of the unknown signal sequence into the multi-neural fusion network model for analysis to obtain the cyclic spectrum features corresponding to the unknown signal sequence and the corresponding real-time analysis results;
[0040] Step 6: If no valid result is obtained, it is recorded as an invalid unknown signal sequence, and the invalid unknown signal sequence is used as a training time series. The training time series is sent to the learning preprocessing calculation module, and the invalid unknown signal sequence is used to re-learn the existing multi-neural fusion network model, and the cyclic spectrum and digital features of this type of invalid unknown signal sequence are recorded, the output result is saved, the scope of the cognitive knowledge tree is expanded, and return to step 4.
[0041] As a specific example, in step 1, a training framework is established based on the channel model and classical electromagnetic signal characteristics in an actual electromagnetic confrontation environment, and a training time series for learning is generated under the framework, as follows:
[0042] Step 1.1: The channel model in the actual electromagnetic countermeasure environment is:
[0043]
[0044] Wherein, s(n) is the target signal to be measured, x(n) represents the sampling value of the external signal obtained from the receiving system at time t=n, w(n) represents the noise signal from the outside world, and h(n) represents the channel gain during the propagation of the electromagnetic signal of the target signal to be measured; H0 represents the situation in which the system receives valid external target signals, and H1 represents the situation in which the system fails to receive valid external target signals;
[0045] Step 1.2: Based on the channel model in the actual electromagnetic countermeasure environment, the modulated signal that experiences flat fading received by the receiving system is modeled using the electromagnetic signal characteristics:
[0046]
[0047] In the formula, s k represents a complex sequence that obeys independent and identical distribution and has a mean of 0, specifically representing the transmitted symbols within k cycles; Δf c is the carrier frequency deviation, T is the symbol period, t0 is the delay, θ is the phase offset, p(t) is the pulse function, and n(t) is the function with a mean of 0 and a variance of σ 2 Additive Gaussian noise;
[0048] Step 1.3: Under this training framework, the original time series obtained according to the modulation type of different types of signals are sampled to obtain a training time series X of N samples. i =[x i1 x i2 … x iN ] 1×N , where i = 1, 2, ..., M. On this basis, the received signal is sampled sequentially M times, and the data matrix X used for training is obtained as follows:
[0049]
[0050] The label matrix Y corresponding to X is:
[0051]
[0052] The sampled training time series is divided into a training data set and a test data set in a ratio of 7:3.
[0053] As a specific example, in step 2, the generated training time series is sent to the learning preprocessing calculation module to obtain the cyclic spectrum of the training time series and the digital features corresponding to the cyclic spectrum, as follows:
[0054] For the mean function E(x(t)) and autocorrelation function R of a random signal x(t) with a period of T x (t, τ), satisfying:
[0055]
[0056] R x (t+T0,τ+T0)=R x (t, τ)
[0057] If the random signal x(t) satisfies the above conditions, then x(t) is considered to have cyclostationary characteristics, and R x (t, τ) also satisfies the cyclic stability characteristics. x (t, τ) is expanded by Fourier series to obtain the Fourier expansion coefficients for:
[0058]
[0059] Where α = n / T represents the cycle frequency, i.e., the multiple of the reference frequency, n∈Z; is the circular autocorrelation function;
[0060] According to the relevant cyclic Wiener filtering theory, the cyclic autocorrelation function Perform Fourier transform to obtain the spectral correlation function SCF, which contains the second-order statistical moment characteristics of x(t) as a significant feature of the signal x(t) and is defined as:
[0061]
[0062] When the cyclic frequency α = 0, the spectral correlation function degenerates into the power spectral density PSD:
[0063]
[0064] Since the signal sequence involved in the training is discrete and finite, the cyclic period method is used to perform the spectrum smoothing process. In this process, the spectrum correlation function is:
[0065]
[0066] Where g(n) is a unit weighting function, with a width of Δt = NT s ; T s is the sampling time, f1 and f2 are the filter center frequencies, X T (r,f| 1,2 |) is the time-varying Fourier transform of x(n)
[0067]
[0068] In the formula, when the duration is NT s When α(r) represents the window of data gradually shrinking at the corresponding time, α(r) is normalized accordingly, thereby obtaining the cyclic spectrum of the corresponding sequence that satisfies time smoothness and has no deviation, and the α-f characteristic parameters are statistically analyzed in detail.
[0069] As a specific example, the multi-neural fusion network model in step 3 is as follows:
[0070] The characteristic structures of the long short-term memory network model LSTM, the convolutional neural network model CNN and the deep neural network model DNN are integrated and applied to obtain a multi-neural fusion network model;
[0071] The fused spectrum features of the signal are input into the convolution layer of the first CNN network to reduce the frequency domain changes of the signal; the output results of the first layer enter the second LSTM neural network to overcome the time domain changes of the input information; and then sent to the third DNN layer to map the input features and output the results, while recording the results.
[0072] As a specific example, the real-time analysis preprocessing module in step 4 is as follows:
[0073] The time series collected by the real-time system is used to extract the characteristic spectrum of the cyclic spectrum of the data to be tested according to the method in step 2, and the calculation process is the same as the calculation process of the cyclic spectrum in step 2.
[0074] The present invention also provides an intelligent spectrum calculation device based on multi-neural network fusion, which is used to implement the intelligent spectrum calculation method based on multi-neural network fusion. The device includes a cognitive learning system and a real-time analysis system, wherein:
[0075] The cognitive learning system is based on a training framework established based on the channel model and classical electromagnetic signal characteristics in an actual electromagnetic confrontation environment. Under the framework, a training time series for learning is generated, and a learning preprocessing calculation module is entered to obtain the cyclic spectrum of the sequence and the digital features corresponding to the cyclic spectrum as input parameters. Learning is performed in a multi-neural fusion network model, and the output results are classified and stored to establish a cognitive knowledge tree.
[0076] The real-time analysis system sends the unknown sequence collected in real time to the real-time analysis preprocessing module to obtain the cyclic spectrum of the unknown signal, inputs it into the multi-neural fusion network model for analysis, and obtains the corresponding cyclic spectrum characteristics and the corresponding real-time analysis results; if no valid results are obtained, it is recorded as an invalid unknown signal sequence, and the invalid unknown signal sequence is sent to the learning preprocessing calculation module as a training time series, and the existing multi-neural fusion network model is re-learned using the invalid unknown signal sequence, and the cyclic spectrum and digital characteristics of this type of invalid unknown signal sequence are recorded, the output results are saved, and the scope of the cognitive knowledge tree is expanded.
[0077] The present invention also provides a mobile terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the intelligent spectrum calculation method based on multi-neural network fusion when executing the program.
[0078] The present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that when the program is executed by a processor, the steps in the intelligent spectrum calculation method based on multi-neural network fusion are implemented.
[0079] In order to make the steps of the present invention more detailed and clear, the method of the present invention is further explained below in conjunction with the accompanying drawings and specific implementation cases.
[0080] Example
[0081] Combination Figure 1 This embodiment provides an intelligent spectrum calculation method based on multi-neural network fusion, and the specific steps are as follows:
[0082] Step 1: Establish a training framework based on the channel model and classical electromagnetic signal characteristics in the actual electromagnetic confrontation environment, and generate a training time series for learning under this framework, as follows:
[0083] Step 1.1: The channel model in the actual electromagnetic countermeasure environment is:
[0084]
[0085] Wherein, s(n) is the target signal to be measured, x(n) represents the sampling value of the external signal obtained from the receiving system at time t=n, w(n) represents the noise signal from the outside world, and h(n) represents the channel gain during the propagation of the electromagnetic signal of the target signal to be measured; H0 represents the situation in which the system receives valid external target signals, and H1 represents the situation in which the system fails to receive valid external target signals;
[0086] Step 1.2: Based on the channel model in the actual electromagnetic countermeasure environment, the modulated signal that experiences flat fading received by the receiving system is modeled using the classic electromagnetic signal characteristics:
[0087]
[0088] In the formula, s k represents a complex sequence that obeys independent and identical distribution and has a mean of 0, specifically representing the transmitted symbols within k cycles; Δf c is the carrier frequency deviation, T is the symbol period, t0 is the delay, θ is the phase offset, p(t) is the pulse function, and n(t) is the function with a mean of 0 and a variance of σ 2 Additive Gaussian noise;
[0089] Step 1.3: Under this training framework, the original time series obtained according to the modulation type of different types of signals are sampled to obtain a training time series X of N samples. i =[x i1 x i2 … x iN ] 1×N , where i = 1, 2, ..., M. On this basis, the received signal is sampled sequentially M times, and the data matrix X used for training is obtained as follows:
[0090]
[0091] The label matrix Y corresponding to X is:
[0092]
[0093] The sampled training time series is divided into training data set and test data set in a ratio of 7:3.
[0094] In practical applications, the types of electromagnetic signals used for training include but are not limited to: single-point signal, linear frequency modulation signal LFM, nonlinear frequency modulation signal NLFM, pseudo-random phase coded signal and agile frequency radar signal. The specific signal expressions of the five basic signals are:
[0095] (1) Single point signal:
[0096] x(t)=Aexp(j(2πf0t+Δθ))+n(t)
[0097] Where A is the signal amplitude, f0 is the center frequency, and Δθ represents the preset phase offset;
[0098] (2) Linear frequency modulation signal:
[0099]
[0100] Where A is the signal amplitude, f0 is the center frequency, K is the frequency modulation slope, and Δθ represents the preset phase offset;
[0101] (3) Nonlinear frequency modulation signal:
[0102]
[0103] Wherein, A is the signal amplitude, f0 is the center frequency, a, b, c represent the modulation parameters of nonlinear frequency modulation, and Δθ represents the preset phase offset;
[0104] (4) Pseudo-random phase coded signal:
[0105]
[0106] Among them, A is the signal amplitude, f0 is the center frequency, τ l represents the signal delay corresponding to the lth code element, Represents the absolute phase of the lth symbol, and the phase encoding is pseudo-random;
[0107] (5) Frequency Agile Radar Signal:
[0108] x(t)=Aexp(j(2πf n t+Δθ))+n(t)
[0109] Where A is the signal amplitude, f n The center frequency point of the signal corresponding to the n-th signal is the center, and Δθ represents the preset phase offset.
[0110] Using the above five signals, the frequency measurement range is selected to be 6-18GHz, the step frequency is selected to be 10MHz step, the noise type is Gaussian white noise, the signal-to-noise ratio changes in 1dB steps from -40dB to 40dB, the power parameter is 1w, the modulation pulse width is 1us, the pulse repetition time changes in 1us to 5us at intervals of 1us, and the remaining parameters are adjusted specifically for different signal types. In this framework, using the above parameters and signal models, a single training signal sequence, two mixed signal sequences, and a combination of three mixed signal sequences are generated as the final training signal.
[0111] Step 2: Send the generated training time series to the learning preprocessing calculation module to obtain the cycle spectrum of the training time series and the digital features corresponding to the cycle spectrum, such as Figure 2 As shown, the details are as follows:
[0112] For the above discrete sequence, the cyclic period method is used to perform spectrum smoothing on the above training sequence and obtain the spectrum correlation function:
[0113]
[0114] Where g(n) represents a unit weighted function, the width of which is Δt=NT s , T s represents the sampling period, f1 and f2 represent the center frequency of the filter function, and X T (r,f| 1,2 |) represents the time-varying Fourier transform of x(n)
[0115]
[0116] In the formula, when the duration is NT s When α(r) represents the window of data gradually shrinking at the corresponding time, α(r) is normalized accordingly, thereby obtaining the cyclic spectrum of the corresponding sequence that satisfies time smoothness and has no deviation, and the α-f characteristic parameters are statistically analyzed in detail. Figure 4 This is the three-dimensional structure diagram of the bpsk cycle characteristics in a 0dB noise environment.
[0117] Step 3: Send the cyclic spectrum of the training time series and the digital features corresponding to the cyclic spectrum into the multi-neural fusion network model for learning, classify and store the output results, and establish a cognitive knowledge tree;
[0118] The characteristic structures of the long short-term memory network model LSTM, the convolutional neural network model CNN and the deep neural network model DNN are integrated and applied to obtain a multi-neural fusion network model, such as Figure 5 As shown;
[0119] After obtaining the corresponding cyclic spectrum and corresponding digital features of the training sequence, the statistical results are sent to the convolution layer of the first layer of the CNN network as a one-dimensional feature parameter for preliminary processing. The structure of the convolutional neural network model CNN is as follows: Figure 6 As shown in the figure, the convolution operation is the process of extracting image features by traversing the convolution kernel through the image. Increasing the number of convolution kernels means increasing the number of image features extracted by the convolutional neural network, which can significantly improve the network learning efficiency and recognition accuracy. In convolutional neural networks, more discrete convolution operations are used, and the convolution layer of the convolutional neural network is used to improve the frequency domain stability of the training parameters. The design parameters of the convolution layer of the CNN network are shown in Table 1.
[0120] Table 1 Composition and main parameters of the neural network model
[0121]
[0122] After the processing is completed, the result is sent to the second LSTM layer for long-term memory learning. The corresponding LSTM network structure is as follows Figure 7 shown.
[0123] For the classic recurrent neural network RNN, the network output h(t) at time t is related to the current input x(t) and the previous output h(t-1), so the recurrent neural network has a good storage characteristic of retaining information from all previous moments when performing model calculations. However, this structure has certain problems due to its own characteristics: the long span reduces the time correlation of data information, which further causes the influence of previous information to gradually decrease compared to future outputs, gradually causing the gradual loss of historical effective information, and ultimately causing the accuracy of network model judgment to decrease rapidly. In order to retain useful information as much as possible and remove redundant invalid information, the LSTM network is changed in structure, and three information control units are further added to the hidden layer, namely the input gate i(t), the forget gate f(t) and the output gate o(t), as shown in Figure 8 shown.
[0124] i(t)=σ(W ix (t)x(t))+W ih (t)h(t-1)+b i (t))
[0125] f(t)=σ(f ix (t)x(t))+W fh (t)h(t-1)+b f (t))
[0126] o(t)=σ(f ox (t)x(t))+W oh (t)h(t-1)+b o (t))
[0127] Where W(t) is the weight coefficient matrix (for example, the weight coefficient matrix of i(t) is W ix (t), and others are similar), b(t) represents the bias matrix (for example, the weight coefficient matrix of i(t) is b i (t), others are similar), σ is the sigmoid function, tanh is the hyperbolic tangent function, and in the process of state updating, the state information function C(t), the hidden layer output function h(t) and the neural network output function y(t) can be further expressed as:
[0128] C(t)=f(t)C(t-1)+i(t)tanh(W cx (t)x(t)+W ch (t)h(t-1)+b c (t))
[0129] h(t)=o(t)tanh(C(t))
[0130] y(t)=W yh (t)h(t)+b y (t)
[0131] The forward propagation process of the long short-term memory network model is as follows Fig. 9 shown.
[0132] After the training data is preprocessed accordingly, it is used as the input of the LSTM model. After mapping in the hidden layer, higher-dimensional training data features are further extracted. At the same time, the high-dimensional feature mapping obtained by the network model is output through the fully connected layer and the Softmax layer, where P0 represents the probability that the current input sample belongs to H0, and P1 represents the probability that the current input sample belongs to H1.
[0133] For n pairs of training data {(x (1) ,y (1) ), (x (2) ,y (2) ),…,(x (n) ,y (n) )} is calculated. The forward propagation process starts from the input layer of the network model. After the data passes through the hidden layer, the fully connected layer and the softmax layer, the output of the neural network is finally mapped into the high-dimensional features of the training sequence of the data. The final mapping relationship is:
[0134]
[0135] And when there is a certain error between the corresponding label of the expected result and the calculated result of the forward propagation output, further model adjustment is performed by calculating the cross entropy error loss function. The actual error loss function can be finally expressed as:
[0136]
[0137] After that, the calculated error loss function is back-propagated again, and the relevant parameter values of the LSTM network are optimized layer by layer according to the function. In the actual application of the specific model, the specific steps of this part are:
[0138] (1) Complete the preliminary construction of the model and initialize the structural parameters in the network model, while presetting the number of iterations and other optimization parameters.
[0139] (2) Input the training matrix and verification data into the training network respectively, complete the forward propagation, and calculate the Loss function of this propagation process.
[0140] (3) During the back-propagation process, the gradient is continuously calculated to update the parameters according to the gradient descent method until the iteration is completed, thereby obtaining relatively stable high-order data features.
[0141] Some of the corresponding model parameters used in this example are shown in Table 1.
[0142] After obtaining high-order data features from LSTM, they are sent to the DNN neural network for a deeper spectrum analysis process.
[0143] In the classic deep network model, the layers of DNN are connected through full connection, that is, any neuron in the i-th layer is connected to any neuron in the i+1-th layer. In this sense, each individual neuron can be considered as an independent perceptron, that is, it satisfies:
[0144]
[0145] y=σ(z)
[0146] Among them, XW i They represent the input of the i-th neuron and its connection weight, b represents the bias of the neuron, σ(·) represents the activation function, and y represents the output of the neuron.
[0147] In the specific propagation process, when the number of layers of the deep neural network is l-1 and the number of neurons is m, the weight matrix corresponding to the neuron numbered j in the l-1 layer to the neuron numbered k in the l layer is The bias matrix of neuron j in layer l is set to , then the corresponding j-th neuron output can be obtained The expression is:
[0148]
[0149] At this time, the coefficient matrix composed of the linear coefficients of the lth layer is represented as W l (the matrix size is n×m), and the vector composed of the bias of the lth layer is b l Represents (size is n×1), and represents the output vector a of the l-1th layer as a l-1 , then the output of the l network can be obtained as:
[0150] a l =σ(z l )=σ(Wa l-1 +b l)
[0151] The algorithm for forward propagation of deep neural networks is summarized as follows:
[0152] Initialize a l =x
[0153] When the number of layers l is within the range 2-L, we can calculate
[0154] a l =σ(z l )=σ(Wa l-1 +b l )
[0155] That is, the final output layer outputs a l Where the number of network layers is L, each hidden layer and output layer output matrix is W, the bias matrix is b, and the input matrix is x.
[0156] In the process of back propagation, in order to find the appropriate linear coefficient matrix W and bias matrix b, the loss of the output function can be reduced as much as possible. The loss function is selected to measure the output loss of the training sample, and the gradient descent method is used to solve the minimum value of the loss function under this condition. The corresponding linear coefficient matrix W and bias matrix b are the results that need to be obtained. The specific algorithm steps are shown in Table 2:
[0157] Table 2
[0158]
[0159] The DNN neural network in this embodiment adopts a two-layer DNN neural network, such as Fig.10 As shown in the figure, the first layer uses Relu as the activation function, and the second layer uses the Softmax function as the final loss function.
[0160] Step 4: Send the unknown sequence collected in real time to the real-time analysis preprocessing module to obtain the cyclic spectrum of the unknown signal;
[0161] Furthermore, the real-time analysis preprocessing module is as follows Figure 3 As shown, the details are as follows:
[0162] The time series collected by the real-time system is used to extract the characteristic spectrum of the cyclic spectrum of the data to be tested according to the method in step 2, and the calculation process is the same as the calculation process of the cyclic spectrum in step 2.
[0163] Step 5: Input the cyclic spectrum of the unknown signal into the multi-neural fusion network model for analysis to obtain the cyclic spectrum characteristics corresponding to the unknown signal and the corresponding real-time analysis results;
[0164] Step 6: If no valid result is obtained, it is recorded as an invalid unknown signal sequence, and the invalid unknown signal sequence is used as a training time series. The training time series is sent to the learning preprocessing calculation module, and the invalid unknown signal sequence is used to re-learn the existing multi-neural fusion network model, and the cyclic spectrum and digital features of this type of invalid unknown signal sequence are recorded, the output result is saved, the scope of the cognitive knowledge tree is expanded, and return to step 4.
[0165] In summary, the intelligent spectrum calculation method proposed in the present invention in the spectrum analysis calculation has built a reasonable neural network model by combining the structural characteristics of various neural networks, established a complete training signal system, and provided an effective signal feature extraction method. On the basis of the existing methods, it further overcomes the measurement errors caused by human factors, improves the measurement method, and effectively improves the spectrum calculation efficiency.
[0166] The above examples are only some of the better examples of the present invention and should not be limited thereto. It should be pointed out that for ordinary practitioners in this industry, based on the principles of the present invention, the improvements and enhancements made to the invention should be regarded as within the scope of protection of the invention.
Claims
1. An intelligent spectrum calculation method based on multi-neural network fusion, characterized in that: The following steps are involved: Step 1: Establish a training framework based on the channel model and electromagnetic signal characteristics in an actual electromagnetic countermeasure environment, and generate a training time series for learning under the framework; Step 2: Send the generated training time series to the learning preprocessing calculation module to obtain the cyclic spectrum of the training time series and the digital features corresponding to the cyclic spectrum; Step 3: Send the cyclic spectrum of the training time series and the digital features corresponding to the cyclic spectrum into the multi-neural fusion network model for learning, classify and store the output results, and establish a cognitive knowledge tree; Step 4: sending the unknown signal sequence collected in real time to a real-time analysis preprocessing module to obtain a cyclic spectrum of the unknown signal sequence; Step 5: Input the cyclic spectrum of the unknown signal sequence into the multi-neural fusion network model for analysis to obtain the cyclic spectrum features corresponding to the unknown signal sequence and the corresponding real-time analysis results; Step 6: If no valid result is obtained, it is recorded as an invalid unknown signal sequence, and the invalid unknown signal sequence is used as a training time series. The training time series is sent to the learning preprocessing calculation module, and the invalid unknown signal sequence is used to re-learn the existing multi-neural fusion network model, and the cyclic spectrum and digital features of this type of invalid unknown signal sequence are recorded, the output result is saved, the scope of the cognitive knowledge tree is expanded, and return to step 4.
2. The intelligent spectrum calculation method based on multi-neural network fusion according to claim 1 is characterized in that: In step 1, a training framework is established based on the channel model and classical electromagnetic signal characteristics in the actual electromagnetic confrontation environment, and a training time series for learning is generated under the framework, as follows: Step 1.1: The channel model in the actual electromagnetic countermeasure environment is: Wherein, s(n) is the target signal to be measured, x(n) represents the sampling value of the external signal obtained from the receiving system at time t=n, w(n) represents the noise signal from the outside world, and h(n) represents the channel gain during the propagation of the electromagnetic signal of the target signal to be measured; H0 represents the situation in which the system receives valid external target signals, and H1 represents the situation in which the system fails to receive valid external target signals; Step 1.2: Based on the channel model in the actual electromagnetic countermeasure environment, the modulated signal that experiences flat fading received by the receiving system is modeled using the electromagnetic signal characteristics: In the formula, s k represents a complex sequence that obeys independent and identical distribution and has a mean of 0, specifically representing the transmitted symbols within k cycles; Δf c is the carrier frequency deviation, T is the symbol period, t0 is the delay, θ is the phase offset, p(t) is the pulse function, and n(t) is the function with a mean of 0 and a variance of σ 2 Additive Gaussian noise; Step 1.3: Under this training framework, the original time series obtained according to the modulation type of different types of signals are sampled to obtain a training time series X of N samples. i =[x i1 x i2 … x iN ] 1×N , where i = 1, 2, ..., M. On this basis, the received signal is sampled sequentially M times, and the data matrix X used for training is obtained as follows: The label matrix Y corresponding to X is: The sampled training time series is divided into a training data set and a test data set in a ratio of 7:
3.
3. The intelligent spectrum calculation method based on multi-neural network fusion according to claim 2 is characterized in that: In step 2, the generated training time series is sent to the learning preprocessing calculation module to obtain the cyclic spectrum of the training time series and the digital features corresponding to the cyclic spectrum, as follows: For the mean function E(x(t)) and autocorrelation function R of a random signal x(t) with a period of T x (t,τ), satisfying: R x (t+T0,τ+T0)=R x (t,τ) If the random signal x(t) satisfies the above conditions, then x(t) is considered to have cyclostationary characteristics, and R x (t,τ) also satisfies the cyclic stability characteristics. x (t,τ) is expanded by Fourier series to obtain the Fourier expansion coefficients for: Where α = n / T represents the cycle frequency, i.e., the multiple of the reference frequency, n∈Z; is the circular autocorrelation function; According to the relevant cyclic Wiener filtering theory, the cyclic autocorrelation function Perform Fourier transform to obtain the spectral correlation function SCF, which contains the second-order statistical moment characteristics of x(t) as a significant feature of the signal x(t) and is defined as: When the cyclic frequency α = 0, the spectral correlation function degenerates into the power spectral density PSD: Since the signal sequence involved in the training is discrete and finite, the cyclic period method is used to perform the spectrum smoothing process. In this process, the spectrum correlation function is: Where g(n) is a unit weighting function with a width of Δt = NT s ; T s is the sampling time, f1 and f2 are the filter center frequencies, X T (r,f |1,2| ) is the time-varying Fourier transform of x(n) In the formula, when the duration is NT s When α(r) represents the window of data gradually shrinking at the corresponding time, α(r) is normalized accordingly, thereby obtaining the cyclic spectrum of the corresponding sequence that satisfies time smoothness and has no deviation, and the α-f characteristic parameters are statistically analyzed in detail.
4. The intelligent spectrum calculation method based on multi-neural network fusion according to claim 3 is characterized in that: The multi-neural fusion network model in step 3 is as follows: The characteristic structures of the long short-term memory network model LSTM, the convolutional neural network model CNN and the deep neural network model DNN are integrated and applied to obtain a multi-neural fusion network model; The fused spectrum features of the signal are input into the convolution layer of the first CNN network to reduce the frequency domain changes of the signal; the output results of the first layer enter the second LSTM neural network to overcome the time domain changes of the input information; and then sent to the third DNN layer to map the input features and output the results, while recording the results.
5. The intelligent spectrum calculation method based on multi-neural network fusion according to claim 4 is characterized in that: The real-time analysis preprocessing module in step 4 is as follows: The time series collected by the real-time system is used to extract the characteristic spectrum of the cyclic spectrum of the data to be tested according to the method in step 2, and the calculation process is the same as the calculation process of the cyclic spectrum in step 2.
6. An intelligent spectrum calculation device based on multi-neural network fusion, characterized in that: The device is used to implement the intelligent spectrum calculation method based on multi-neural network fusion as described in any one of claims 1 to 5, and the device includes a cognitive learning system and a real-time analysis system, wherein: The cognitive learning system is based on a training framework established based on the channel model and classical electromagnetic signal characteristics in an actual electromagnetic confrontation environment. Under the framework, a training time series for learning is generated, and a learning preprocessing calculation module is entered to obtain the cyclic spectrum of the sequence and the digital features corresponding to the cyclic spectrum as input parameters. Learning is performed in a multi-neural fusion network model, and the output results are classified and stored to establish a cognitive knowledge tree. The real-time analysis system sends the unknown sequence collected in real time to the real-time analysis preprocessing module to obtain the cyclic spectrum of the unknown signal, inputs it into the multi-neural fusion network model for analysis, and obtains the corresponding cyclic spectrum characteristics and the corresponding real-time analysis results; if no valid results are obtained, it is recorded as an invalid unknown signal sequence, and the invalid unknown signal sequence is sent to the learning preprocessing calculation module as a training time series, and the existing multi-neural fusion network model is re-learned using the invalid unknown signal sequence, and the cyclic spectrum and digital characteristics of this type of invalid unknown signal sequence are recorded, the output results are saved, and the scope of the cognitive knowledge tree is expanded.
7. A mobile terminal comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, an intelligent spectrum calculation method based on multi-neural network fusion as described in any one of claims 1 to 5 is implemented.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by the processor, the steps in the intelligent spectrum calculation method based on multi-neural network fusion as described in any one of claims 1 to 5 are implemented.
