Communication Signal Recognition Method and System Based on Adaptive Feedforward Neural Network

Abstract

The present invention relates to a communication signal recognition method and system based on an adaptive feedforward neural network, belonging to the technical field of signal recognition, and solves the problems of inaccurate recognition of existing small-sample signals and difficult adjustment of model parameters. It includes randomly generating the activation states of each hidden-layer neuron of binary type, the input weights of each hidden layer of real number type, and the hidden-layer bias based on a single-hidden-layer feedforward neural network, and forming multiple initial network individuals in a hybrid coding form as the parental population; using the root mean square error and network complexity as the objective function, and based on the preprocessed signal sample set, an improved NSGA-III algorithm is used to obtain the Pareto optimal population; according to the root mean square error, multiple network individuals are obtained from the Pareto optimal population as base classifiers, and the preprocessed communication signals collected in real time are input into the base classifiers for ensemble learning. After weighted summation of the output results, the category corresponding to the maximum value is taken as the communication signal recognition result. Accurate signal recognition is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of signal recognition, and in particular, to a communication signal recognition method and system based on an adaptive feedforward neural network. Background Art

[0002] Data classification and recognition is an important task in machine learning, and usually needs to process datasets in fields such as biological information, multimedia, and statistics. In the field of signal recognition, by extracting and analyzing signal features, the target category of the transmitted signal and specific target individuals are gradually recognized.

[0003] In existing machine learning, neural network-related algorithms are mostly used. However, it is difficult to simultaneously find the optimal values of connection weights and the number of hidden layer neurons, which makes it difficult to build a basic model when using a neural network to solve practical problems; when training a feedforward neural network using a gradient-based method, it may fall into a local minimum, converge slowly, and depend on the initial solution, resulting in a large error when using this neural network model to solve practical problems.

[0004] Extreme learning machine (ELM) is a machine learning method for training single-hidden layer feedforward neural networks (SLFNs), which is applied to many fields such as data classification and biomedical engineering. ELM randomly assigns input weights and hidden layer biases, then uses the Moore-Penrose (MP) generalized inverse to calculate the output weights, and uses a parameter learning method without tuning parameters to improve the training speed. However, the randomly assigned network parameters may lead to unstable network performance.

[0005] At the same time, when the target categories of signal samples are unbalanced, or the sample data volume is small, it is easy to make the evaluation of the classifier accuracy meaningless, and the correct category of the signal emission target cannot be recognized. In single-objective optimization during recognition, only one optimal solution is determined in one run, and multiple solutions need to be re-solved to obtain. Summary of the Invention

[0006] In view of the above analysis, embodiments of the present invention aim to provide a communication signal recognition method and system based on an adaptive feedforward neural network to solve the problems of inaccurate recognition of small-sample signals and difficult adjustment of model parameters in the prior art.

[0007] On the one hand, embodiments of the present invention provide a communication signal recognition method based on an adaptive feedforward neural network, including the following steps:

[0008] Based on a single-hidden layer feedforward neural network, randomly generate the activation states of each hidden layer neuron of binary type, the input weights of each hidden layer of real number type, and the hidden layer biases, and form multiple initial network individuals through a mixed coding form;

[0009] Taking the initial multiple network individuals as the parent population, using the root mean square error and network complexity as the objective functions, based on the preprocessed signal sample set, an improved NSGA-III algorithm is used for multi-objective optimization to obtain the Pareto optimal population;

[0010] According to the root mean square error, multiple network individuals are obtained from the Pareto optimal population as the base classifiers. After preprocessing the real-time collected communication signals, they are input into the base classifiers for ensemble learning. After weighted summing the output results, the category corresponding to the maximum value is taken as the communication signal recognition result.

[0011] Based on the further improvement of the above method, the maximum number K of hidden layer neurons in a single-hidden layer feedforward neural network is 2N + 1, where N represents the number of input neurons.

[0012] Based on the further improvement of the above method, the activation states of each hidden layer neuron of the binary type include: when the activation state of the hidden layer neuron is 1, it means that the hidden layer neuron is activated; when the activation state of the hidden layer neuron is 0, it means that the hidden layer neuron is not activated; the ranges of the input weights of each hidden layer and the hidden layer biases of the real number type are [-1, 1].

[0013] Based on the further improvement of the above method, the initial multiple network individuals are composed in a hybrid coding form as shown in the following formula:

[0014]

[0015] where, X i (G) represents the i-th network individual in the G-th generation, i = 1, 2,..., N pop , N pop is the population size, θ i,j (G) represents the activation state of the j-th hidden layer neuron of the i-th network individual, θ i,j (G) ∈ {0, 1}; represents the input weight of the j-th hidden layer neuron of the i-th network individual, b i,j (G) represents the bias of the j-th hidden layer neuron of the i-th network individual, b i,j (G) ∈ [-1, 1], j = 1, 2,..., K.

[0016] Based on the further improvement of the above method, using the root mean square error and network complexity as the objective functions means taking the minimum root mean square error and the minimum network complexity as the two conflicting objectives to be optimized, where the network complexity is represented by the average value of the activation states of each hidden layer neuron in the network.

[0017] Based on the further improvement of the above method, the improved NSGA-III algorithm is used for multi-objective optimization to obtain the Pareto optimal population, including:

[0018] S121: Perform mixed crossover and mixed mutation of binary and real numbers on the population P of the current iteration G to generate the offspring population Q G , where the number of network individuals in P G is the same as that in Q G ;

[0019] S122: Synthesize the new population R G = P G ∪ Q G , and input the preprocessed signal sample set into each network individual in the population R G . After calculating the corresponding output weights through the extreme learning machine, calculate the two objective function values of each network individual to obtain the objective function value vector;

[0020] S123: Perform non-dominated sorting on the population R G . Based on the reference point selection mechanism, according to the objective function value vector of each network individual, screen out network individuals from the population R G to obtain the population P G+1 ;

[0021] S124: Judge whether the loop has reached the maximum number of iterations. If not, return to S121, use the population P G+1 as the new P G to perform the loop. If it has reached, exit the loop, and the population P G+1 is the Pareto optimal population.

[0022] Based on the further improvement of the above method, the mixed crossover of binary and real numbers includes: binary crossover for K-dimensional binary vectors and real crossover for (N + 1)×K-dimensional real vectors;

[0023] The binary crossover is performed through the following formula:

[0024]

[0025] The real crossover is performed through the following formula:

[0026]

[0027] where α j ∈{0,1} is a randomly selected binary parameter; β j is a continuous parameter randomly selected in [0,1]; D is the total length of the mixed coding of network individuals, D = (N + 2)×K; x p,j and xq,j They are the j-th coding values in the p-th and q-th network individuals of the current iteration population P G respectively, and y p,j and y q,j are the j-th coding values in the p-th and q-th network individuals of the generated offspring population Q G respectively.

[0028] Based on the further improvement of the above method, the hybrid mutation of binary and real numbers includes: inverting the randomly assigned binary mutation, and adding the real number mutation of normal distribution random numbers;

[0029] The binary mutation is performed through the following formula:

[0030]

[0031] The real number mutation is performed through the following formula:

[0032]

[0033] where rand(0,1) is a random number in (0,1), μ is a preset control parameter, and N(0,σ)

[0034] represents a normal distribution random number with a mean of 0 and a variance of σ, and x t,j is the j-th coding value in the t-th network individual of the current iteration population P G respectively, and y t,j is the j-th coding value in the t-th network individual of the generated offspring population Q G respectively.

[0035] Based on the further improvement of the above method, according to the root mean square error, multiple network individuals are obtained from the Pareto optimal population as base classifiers. They are sorted in ascending order of the root mean square error, and according to the quantity threshold, the first few network individuals are selected as base classifiers, and the weights are set from large to small and used for the weighted summation of the output results.

[0036] On the other hand, an embodiment of the present invention provides a communication signal recognition system based on an adaptive feedforward neural network, including:

[0037] A signal preprocessing module, configured to preprocess historical communication signals to obtain a signal sample set, and preprocess real-time collected communication signals;

[0038] A network initialization module, configured to randomly generate the activation states of each hidden layer neuron of binary type, the input weights of each hidden layer of real number type, and the hidden layer bias based on a single-hidden-layer feedforward neural network, and form multiple initial network individuals through a hybrid coding form;

[0039] The network optimization module is used to take multiple initial network individuals as the parent population, use the root mean square error and network complexity as the objective functions, and based on the signal sample set output by the data preprocessing module, adopt an improved NSGA-III algorithm for multi-objective optimization to obtain the Pareto optimal population;

[0040] The signal recognition module is used to obtain multiple network individuals as base classifiers from the Pareto optimal population obtained by the optimization network module according to the root mean square error, input the real-time collected communication signals output by the data preprocessing module into the base classifiers for ensemble learning, and after weighted summation of the output results, take the category corresponding to the maximum value as the communication signal recognition result.

[0041] Compared with the prior art, the present invention can at least achieve one of the following beneficial effects:

[0042] 1. In the learning of a single-hidden layer feedforward neural network, two conflicting factors, namely training error and network complexity, are considered simultaneously, a multi-objective learning model is established, and a hybrid coding multi-objective learning method integrated with the extreme learning machine is proposed, which improves the generalization ability and learning speed of the communication signal recognition network;

[0043] 2. The non-dominated sorting genetic algorithm III (NSGA-III) is improved to handle multi-objective models with mixed variables, making it applicable to continuous and discrete decision variables, where binary coding is used for structure learning and real number coding is used for optimizing input weights and hidden layer biases. By solving the Pareto optimal solution, a trained network model is directly obtained, overcoming the difficulty of model construction and achieving a balance between training error and network complexity;

[0044] 3. Select multiple networks from the Pareto optimal solutions for ensemble learning, which improves the recognition accuracy when the number of communication signal samples is small or the sample categories are imbalanced.

[0045] In the present invention, the above technical solutions can also be combined with each other to achieve more preferred combination schemes. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can be made obvious from the description or understood by implementing the present invention. The objectives and other advantages of the present invention can be realized and obtained from the content specifically pointed out in the description and the drawings. Description of the Drawings

[0046] The drawings are only for the purpose of showing specific embodiments and are not considered as limiting the present invention. Throughout the drawings, the same reference signs represent the same components.

[0047] Figure 1 It is a flowchart of the communication signal recognition method based on the adaptive feedforward neural network in Embodiment 1 of the present invention;

[0048] Figure 2 It is a schematic diagram of the structure of the single-hidden-layer feedforward neural network in Embodiment 1 of the present invention;

[0049] Figure 3 It is a flowchart of the use of the signal sample set in Embodiment 1 of the present invention;

[0050] Figures 4(a) and 4(b) are respectively the Pareto front graphs generated by the diabetes and glass data sets in Embodiment 1 of the present invention under different numbers of iterations. Specific Embodiments

[0051] The following will specifically describe the preferred embodiments of the present invention in conjunction with the accompanying drawings. The accompanying drawings form a part of this application and are used together with the embodiments of the present invention to explain the principles of the present invention, and are not used to limit the scope of the present invention.

[0052] A specific embodiment of the present invention discloses a communication signal recognition method based on an adaptive feedforward neural network, as Figure 1 shown, which includes the following steps:

[0053] S11: Based on a single-hidden-layer feedforward neural network, randomly generate the activation states of each hidden-layer neuron of binary type, the input weights of each hidden layer of real number type, and the hidden-layer bias, and form multiple initial network individuals through a hybrid coding form.

[0054] As Figure 2 shown, it is a schematic diagram of the structure of a single-hidden-layer feedforward neural network. The network in the figure includes N input neurons (i.e., there are N features in the input data), K hidden neurons, and L output neurons (i.e., L data categories). The activation function g(x) uses the sigmoid function. ω is the input weight connecting the input neuron and the hidden neuron, b is the bias of the hidden-layer neuron, and β is the output weight connecting the hidden neuron and the output neuron. The maximum number of hidden-layer neurons K in the single-hidden-layer feedforward neural network is 2N + 1.

[0055] It should be noted that in this embodiment, a hybrid coding of binary coding and real number coding is used to define the network structure and connection weights of the single-hidden-layer feedforward neural network. Among them, binary coding is used to represent the network structure, that is, the activation state θ of the hidden-layer neuron. The value of θ can be 1 or 0, where θ = 1 means the corresponding hidden-layer neuron is activated, and θ = 0 means the corresponding hidden-layer neuron is not activated. Real number coding is used to represent the input weight ω and the hidden-layer bias b. Since the extreme learning machine is used to calculate the output weight in this embodiment, the input weights of each hidden layer of real number type and the hidden-layer bias are randomly generated within the range of [-1, 1].

[0056] Multiple initial network individuals are formed through a hybrid coding form, as shown in the following formula:

[0057]

[0058] Among them, X i (G) represents the i-th network individual of the G-th generation, i = 1, 2,..., N pop , N pop is the population size, θ i,j (G) represents the activation state of the j-th hidden layer neuron of the i-th network individual, θ i,j (G) ∈ {0, 1}; represents the input weight of the j-th hidden layer neuron of the i-th network individual, b i,j (G) represents the bias of the j-th hidden layer neuron of the i-th network individual, b i,j (G) ∈ [-1, 1], j = 1, 2,..., K.

[0059] According to Equation (1), the total length D of the hybrid coding of a network individual is D = (N + 2) × K, and the population of the G-th generation can be expressed as:

[0060]

[0061] S12: Using multiple initial network individuals as the parent population, taking the root mean square error and network complexity as the objective functions, and based on the preprocessed signal sample set, an improved NSGA-III algorithm is used for multi-objective optimization to obtain the Pareto optimal population.

[0062] It should be noted that in this embodiment, the root mean square error and network complexity are used as the objective functions, and the smallest root mean square error and the smallest network complexity are used as two conflicting objectives to be optimized, and a two-objective optimization model is established.

[0063] The first objective function is the root mean square error, which is expressed by the following formula:

[0064]

[0065] Among them, o r1,r2 represents the network output calculated from the training data, that is, the network output value corresponding to the r1-th training data being recognized as the r2-th label, s r1,r2 represents the actual output corresponding to the training data, that is, the actual value of the r1-th training data being recognized as the r2-th label. If the training data actually belongs to the category corresponding to the l-th label, it is 1, otherwise it is 0. M is the number of training data, and L is the number of labels.

[0066] The second objective function is the network complexity, which is represented by the average value of the activation states of each hidden layer neuron in the network, as shown in the following formula:

[0067]

[0068] Among them, K is the number of hidden layer neurons, and θ j is the activation state of the j-th hidden layer neuron.

[0069] Therefore, the optimization model for the two objectives can be expressed as:

[0070]

[0071] Among them, K0 represents the minimum value of the number of activated hidden layer neurons, which can be set according to the actual situation; preferably, in this embodiment, K0 = 3.

[0072] It should be noted that in this embodiment, the non-dominated sorting genetic algorithm III (NSGA-III) is improved to enable it to handle multi-objective models with mixed variables, perform binary and real number hybrid crossover and hybrid mutation on each generation of population, and integrate with the extreme learning machine (ELM). After calculating the output weights of each network individual in each generation of population, the objective function value vector of each network individual is obtained.

[0073] Specifically, the improved NSGA-III algorithm is used for multi-objective optimization to obtain the Pareto optimal population, including:

[0074] S121: Perform binary and real number hybrid crossover and hybrid mutation on the population P of the current iteration G to generate the offspring population Q G , P G and Q G have the same number of network individuals;

[0075] It should be noted that two network individuals are randomly selected from the population P of the current iteration G to generate two network individuals in the offspring population Q G .

[0076] The binary and real number hybrid crossover performs crossover on the corresponding parts of the hybrid encoding respectively, including: binary crossover for the K-dimensional binary vector and real number crossover for the (N + 1) × K-dimensional real number vector.

[0077] The binary crossover is performed through the following formula:

[0078]

[0079] The real number crossover is performed through the following formula:

[0080]

[0081] Among them, α j∈ {0, 1} is a randomly selected binary parameter; β j is a continuous parameter randomly selected in [0, 1]; D is the total length of the mixed coding of network individuals, D = (N + 2) × K; x p,j and x q,j are respectively the j-th coding values in the p-th and q-th network individuals of the current iteration population P G , y p,j and y q,j are respectively the j-th coding values in the p-th and q-th network individuals of the generated offspring population Q G .

[0082] The mixed mutation of binary and real numbers includes: reversing the randomly assigned binary mutation, and, adding a real number mutation of a normal distribution random number;

[0083] The binary mutation is performed by the following formula:

[0084]

[0085] The real number mutation is performed by the following formula:

[0086]

[0087] where, rand(0, 1) is a random number in (0, 1), μ is a preset control parameter, N(0, σ) represents a normal distribution random number with a mean of 0 and a variance of σ, x t,j is the j-th coding value in the t-th network individual of the current iteration population P G , y t,j is the j-th coding value in the t-th network individual of the generated offspring population Q G .

[0088] Preferably, μ = 0.02 and σ = 0.2.

[0089] The population size is N pop , and in each loop, the mixed crossover is executed N pop × P c times, and the mixed mutation is executed N pop × P m times, where P c is the crossover probability, 0.7 ≤ P c ≤ 0.9, P m is the mutation probability, 0.1 ≤ P m ≤ 0.3, and N pop network individuals are generated as the offspring population Q G .

[0090] S122: Synthesize the new population R G = PG ∪Q G and input the preprocessed signal sample set into each network individual in population R G After calculating the corresponding output weights through the extreme learning machine for each network individual, calculate the two objective function values of each network individual to obtain the objective function value vector;

[0091] It should be noted that the population P of the current iteration G and the offspring population Q G are combined to obtain a new population R G where the number of network individuals is 2N pop .

[0092] The signal sample set is a set obtained by preprocessing historical communication signals, including a training set and a test set.

[0093] Specifically, through a broadband signal collector, communication signals from multiple targets are collected, such as: passenger planes, transport planes, refueling planes, and helicopters; the collected communication signals are analyzed by clustering to obtain the frequency points of interest; then, features in the time domain and frequency domain of the communication signals are extracted through methods such as Fourier transform, time-frequency analysis, ambiguity function, SPWVD, and convolutional neural network. Among them, the time-domain features include: rising edge, falling edge, top flatness, and pulse width; the frequency-domain features include: frequency point, spectrogram, fourth-order moment of instantaneous frequency, and frequency symmetry.

[0094] Normalize each feature of each communication signal data to the interval [-1, 1] using the following formula:

[0095]

[0096] where c f,h is the h-th normalized value in feature c, c b,h is the h-th value to be normalized in feature c, c b,min and c b,max are the minimum and maximum values among the values to be normalized in feature c.

[0097] Preferably, the stratified 10-fold cross-validation (10-CV) technique is adopted to evenly divide the preprocessed signal sample set into 10 parts according to the target category ratio, select 9 of them as the training set, and the rest as the test set.

[0098] As Figure 3 shown, based on the training set, according to the input weights of each hidden layer and the hidden layer biases, the generalized inverse (M-P, Moore-Penrose) is used in the extreme learning machine to calculate the corresponding output weights of each network individual by solving the least squares solution of the linear system.

[0099] Compared with the prior art, in this embodiment, the input weights and hidden biases are randomly generated values only in the first iteration. In subsequent iterations, they are all values screened according to the improved NSGA-III algorithm. When using the extreme learning machine, the stability of the network is greatly improved and the network error is reduced.

[0100] According to the output weights, obtain the network output values corresponding to the training data in each network individual, and substitute them into Equation (3) to obtain the first objective function value of each network individual; according to the binary coding in the hybrid coding of each network individual, substitute it into Equation (4) to obtain the second objective function value of each network individual. The objective function value vectors of each network individual are formed by the first objective function value and the second objective function value.

[0101] S123: For the population R G Perform non-dominated sorting. Based on the reference point selection mechanism, according to the objective function value vectors of each network individual, select network individuals from the population R G to obtain the population P G+1 ;

[0102] Specifically, this step is divided into the following steps:

[0103] ① Perform non-dominated sorting on the population R G to obtain multiple non-dominated layer sets (F1, F2, F3,...) with different priorities. Each non-dominated layer contains multiple network individuals;

[0104] ② Construct a new population S G , starting from the non-dominated layer with the highest priority, add them to the new population S G in turn until the number of network individuals in S G is greater than or equal to N pop . At this time, the last added non-dominated layer is denoted as F l ,

[0105] ③ If the number of network individuals in the new population S G is equal to N pop , then S G is the population P G+1 . Otherwise, first put (F1, F2, F3,... F l-1 ) into the population P G+1 , and then select the remaining number of network individuals from F l to join the population P G+1 based on the reference point selection mechanism, so that the number of network individuals in the population P G+1 is N pop individuals.

[0106] Specifically, the number of reference points is calculated using the following formula:

[0107]

[0108] where m is the number of targets, and each target is divided into v parts. In this embodiment, m = 2. Exemplarily, when v = 10, the number of reference points is pieces.

[0109] According to all the network individuals in the new population S G on each dimension target, the minimum objective function value forms the ideal point of the new population S G . Standardize the objective function value vectors of each network individual in S G according to the ideal point; calculate the perpendicular distance from each network individual to each reference line, and determine the reference point associated with each network individual according to the shortest distance. At the same time, the number of network individuals associated with each reference point is also obtained, that is, the niche number of each reference point.

[0110] Select individuals from F l according to the niche method, including:

[0111] Sort the niche numbers in ascending order. If the niche number of the reference point is 0, it means that there is no network individual in the current population associated with this reference point, and this reference point is excluded; start selecting from the reference point where the niche number is greater than or equal to 1 and the niche number is the smallest. When there are reference points with the same niche number, randomly select one. In this case, for the selected reference point, if there is one or more network individuals in F l associated with it, add the network individual with the shortest perpendicular distance to the reference line to the population P G+1 . If there is no network individual in F l associated with it, then this reference point is not considered, and the next reference point is selected until the number of network individuals in the population P G+1 is N pop pieces.

[0112] S124: Determine whether the loop has reached the maximum number of iterations. If not, return to S121, and use the population P G+1 as the new P G to perform the loop. If it has reached, exit the loop, and the population P G+1 is the Pareto optimal population.

[0113] Preferably, the maximum number of iterations is set to 50.

[0114] At the end of the iteration, the finally obtained population P G+1That is the Pareto - optimal population, where the network individuals directly correspond to the trained network models, overcoming the difficulties in model construction and achieving a balance between training error and network complexity.

[0115] Exemplarily, for the benchmark classification datasets in the UCI Machine Learning Repository (http: / / archive.ics.uci.edu), the method of this embodiment was used for training and testing. For the diabetes and glass datasets among them, the maximum number of iterations was set to 10, 20, 30, 40, and 50 generations respectively, and the generated Pareto fronts are shown in Figures 4(a) and 4(b). It can be seen from the figures that as the number of iterations increases, the convergence of the method becomes better. When the method reaches 50 generations, a set of Pareto - optimal solutions can be found, which has fewer iterations, lower computational cost, and stable performance compared with the prior art.

[0116] S13: According to the root - mean - square error, obtain multiple network individuals from the Pareto - optimal population as base classifiers, pre - process the real - time collected communication signals and then input them into the base classifiers for ensemble learning. After weighted summation of the output results, take the category corresponding to the maximum value as the communication signal recognition result.

[0117] It should be noted that obtaining multiple network individuals from the Pareto - optimal population as base classifiers according to the root - mean - square error is to sort them in ascending order of the root - mean - square error, and according to the quantity threshold, select the first few network individuals as base classifiers, and set the weights from large to small respectively for the weighted summation of the output results.

[0118] Preferably, in Figure 3 Based on the test set in the signal sample set, test the recognition performance of each base classifier, and obtain the average correct rate, standard deviation of the accuracy rate, and the number of hidden neurons for signal recognition.

[0119] Preferably, set the quantity threshold to 3, that is, select the 3 network individuals with the smallest root - mean - square error as base classifiers, and the corresponding weights λ cj are 0.7, 0.2, and 0.1 in turn.

[0120] During implementation, collect the communication signals through a broadband signal collector, such as a real - time spectrum analyzer. After pre - processing the communication signals according to the pre - processing method in step S122, input them into the finally determined base classifiers, and use the following formula to obtain the maximum value, and the corresponding target category is the recognition result:

[0121]

[0122] where x te is the test data or the pre - processed data collected in real - time, JN is the number of base classifiers; (Ccj,1 (x te ), C cj,2 (x te ),..., C cj,L (x te ) is the classification result of the $c_j$-th base classifier. After weighted summation, the category corresponding to the maximum value $C(x te )$ is taken as the final signal recognition result, that is, the emission target category of the signal is obtained.

[0123] It should be noted that the method in this embodiment is not limited to the recognition of emission target categories. When preprocessing the signal historical data, specific individual categories can be labeled, and the network individual structure can be retrained to identify specific source individuals from the real-time collected signals; or based on image data such as visible light and infrared, the classification and recognition of targets (aircraft, tanks, ships, etc.) can be realized.

[0124] Compared with the prior art, the communication signal recognition method based on the adaptive feedforward neural network provided in this embodiment simultaneously considers two conflicting factors, the training error and the network complexity, in the learning of the single-hidden-layer feedforward neural network, establishes a multi-objective learning model, and proposes a hybrid coding multi-objective learning method integrated with the extreme learning machine, which improves the generalization ability and learning speed of the communication signal recognition network; improves the non-dominated sorting genetic algorithm III (NSGA-III) to handle the multi-objective model with mixed variables, making it applicable to continuous and discrete decision variables, where binary coding is used for structure learning and real-number coding is used for optimizing the input weights and hidden layer biases. By solving the Pareto optimal solution, the trained network model can be directly obtained, overcoming the difficulty of model construction and achieving a balance between the training error and the network complexity; multiple networks are selected from the Pareto optimal solutions for ensemble learning, which improves the recognition accuracy when the number of signal samples is small or the sample categories are unbalanced.

[0125] Embodiment 2

[0126] Another embodiment of the present invention discloses a communication signal recognition system based on an adaptive feedforward neural network to implement the signal recognition method in Embodiment 1. The specific implementation manners of each module refer to the corresponding descriptions in Embodiment 1. It includes:

[0127] A signal preprocessing module for preprocessing historical communication signals to obtain a signal sample set and preprocessing real-time collected communication signals;

[0128] A network initialization module for randomly generating the activation states of each hidden layer neuron of binary type, the input weights of each hidden layer of real-number type, and the hidden layer biases based on the single-hidden-layer feedforward neural network, and forming multiple initial network individuals in a hybrid coding form;

[0129] A network optimization module, which uses multiple initial network individuals as the parental population, uses the root mean square error and network complexity as the objective functions, and based on the signal sample set output by the data preprocessing module, adopts an improved NSGA-III algorithm for multi-objective optimization to obtain a Pareto optimal population;

[0130] A signal recognition module, which, according to the root mean square error, obtains multiple network individuals from the Pareto optimal population obtained by the optimization network module as base classifiers, inputs the real-time collected communication signals output by the data preprocessing module into the base classifiers for ensemble learning, and after weighted summation of the output results, takes the category corresponding to the maximum value as the communication signal recognition result.

[0131] Since the communication signal recognition system based on the adaptive feedforward neural network in this embodiment can be mutually referenced with the foregoing communication signal recognition method in related parts, and this is a repetitive description here, it will not be elaborated here. Since the principle of this system embodiment is the same as that of the above method embodiment, this system also has the corresponding technical effects of the above method embodiment.

[0132] Those skilled in the art can understand that all or part of the processes for implementing the above method embodiments can be completed by instructing relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. Among them, the computer-readable storage medium is a magnetic disk, an optical disk, a read-only memory, or a random access memory, etc.

[0133] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention.

Claims

1. A communication signal recognition method based on an adaptive feedforward neural network, characterized in that, It includes the following steps: Based on a single-hidden-layer feedforward neural network, randomly generate the activation states of each hidden-layer neuron of binary type, the input weights of each hidden layer of real number type, and the hidden-layer biases, and form multiple initial network individuals in a hybrid coding form; Use the multiple initial network individuals as the parent population, use the root mean square error and network complexity as the objective functions, and based on the preprocessed signal sample set, use the improved NSGA-III algorithm for multi-objective optimization to obtain the Pareto optimal population; The signal sample set is a set obtained by preprocessing historical communication signals, including: collecting communication signals from multiple targets through a broadband signal collector; obtaining the concerned frequency points by clustering analysis of the collected communication signals; extracting the time-domain and frequency-domain features of the corresponding communication signals according to the concerned frequency points, where the time-domain features include: rising edge, falling edge, top flatness, and pulse width; the frequency-domain features include: frequency point, spectrogram, fourth-order moment of instantaneous frequency, and frequency symmetry; According to the root mean square error, obtain multiple network individuals from the Pareto optimal population as base classifiers, preprocess the real-time collected communication signals and input them into the base classifiers for ensemble learning, and after weighted summation of the output results, take the category corresponding to the maximum value as the communication signal recognition result; the step of obtaining multiple network individuals from the Pareto optimal population as base classifiers according to the root mean square error is to sort according to the root mean square error from small to large, and according to the quantity threshold, select the first few network individuals as base classifiers, and respectively set the weights from large to small for weighted summation of the output results.

2. The communication signal recognition method based on an adaptive feedforward neural network according to claim 1, characterized in that The maximum number K of hidden-layer neurons in the single-hidden-layer feedforward neural network is 2N + 1, where N represents the number of input neurons.

3. The communication signal recognition method based on an adaptive feedforward neural network according to claim 2, characterized in that The activation states of each hidden-layer neuron of binary type include: when the activation state of this hidden-layer neuron is 1, it means this hidden-layer neuron is activated, and when the activation state of this hidden-layer neuron is 0, it means this hidden-layer neuron is not activated; the ranges of the input weights of each hidden layer and the hidden-layer biases of real number type are [-1, 1].

4. The communication signal recognition method based on an adaptive feedforward neural network according to claim 3, wherein The composition of multiple initial network individuals in a hybrid coding form is as shown in the following formula: Among them, X i (G) represents the i-th network individual in the G-th generation, where i = 1, 2, ..., N pop , N pop is the population size, and θ i,j (G) represents the activation state of the j-th hidden layer neuron of the i-th network individual, and θ i,j (G) ∈ {0, 1}; represents the input weight of the j-th hidden layer neuron of the i-th network individual, b i,j (G) represents the bias of the j-th hidden layer neuron of the i-th network individual, and b i,j (G) ∈ [-1, 1], where j = 1, 2, ..., K.

5. The communication signal recognition method based on an adaptive feedforward neural network according to claim 4, characterized in that Using the root mean square error and network complexity as the objective functions means taking the minimum root mean square error and the minimum network complexity as two conflicting objectives to be optimized, where the network complexity is represented by the average value of the activation states of each hidden-layer neuron in the network.

6. The communication signal recognition method based on an adaptive feedforward neural network according to claim 4, characterized in that The use of the improved NSGA-III algorithm for multi-objective optimization to obtain the Pareto optimal population includes: S121: Perform binary and real number hybrid crossover and hybrid mutation on the population P of the current iteration G to generate the offspring population Q G , P G and Q G have the same number of network individuals; S122: Synthesize the new population R G = P G ∪ Q G , and input the preprocessed signal sample set into each network individual in the population R G . After calculating the corresponding output weights through the extreme learning machine, calculate the two objective function values of each network individual to obtain the objective function value vector; S123: Perform non-dominated sorting on population R G and, based on the reference point selection mechanism, screen out network individuals from population R G according to the objective function value vectors of each network individual to obtain population P G+1 ; S124: Determine whether the loop has reached the maximum number of iterations. If not, return to S121, and use the population P G+1 as the new P G to perform the loop. If it has reached, exit the loop, and the population P G+1 is the Pareto optimal population.

7. The communication signal recognition method based on an adaptive feedforward neural network according to claim 6, characterized in that The hybrid crossover of binary and real numbers includes: binary crossover for K-dimensional binary vectors, and real crossover for (N + 1)×K-dimensional real vectors; The binary crossover is carried out through the following formula: The real crossover is carried out through the following formula: where, α j ∈ {0, 1} is a randomly selected binary parameter; β j is a continuous parameter randomly selected in [0, 1]; D is the total length of the network individual hybrid coding, D = (N + 2) × K; x p,j and x q,j are the j-th coding values in the p-th and q-th network individuals in the current iteration population P G respectively, and y p,j and y q,j are the j-th coding values in the p-th and q-th network individuals in the generated offspring population Q G respectively.

8. The communication signal recognition method based on an adaptive feedforward neural network according to claim 6, wherein The hybrid mutation of binary and real numbers includes: binary mutation by randomly reversing, and real mutation by adding normally distributed random numbers; The binary mutation is carried out through the following formula: The real mutation is carried out through the following formula: Among them, rand(0,1) is a random number in (0,1), μ is a preset control parameter, N(0,σ) represents a normally distributed random number with a mean of 0 and a variance of σ, and x t,j is the j-th coded value in the t-th network individual of the current iteration population P G , and y t,j is the j-th coded value in the t-th network individual of the generated offspring population Q G .

9. A communication signal recognition system based on an adaptive feedforward neural network, characterized in that, It includes: A signal preprocessing module, which is used to preprocess historical communication signals to obtain a signal sample set, and to preprocess real-time collected communication signals; The signal sample set is a set obtained by preprocessing historical communication signals, including: collecting communication signals from multiple targets through a broadband signal collector; obtaining the concerned frequency points by clustering analysis of the collected communication signals; extracting the time-domain and frequency-domain features of the corresponding communication signals according to the concerned frequency points, where the time-domain features include: rising edge, falling edge, top flatness, and pulse width; the frequency-domain features include: frequency point, spectrogram, fourth-order moment of instantaneous frequency, and frequency symmetry; A network initialization module, which is used to randomly generate the activation states of each hidden-layer neuron of binary type, the input weights of each hidden layer of real number type, and the hidden-layer bias based on a single-hidden-layer feedforward neural network, and form multiple initial network individuals in a hybrid coding form; A network optimization module, which is used to take the multiple initial network individuals as the parent population, use the root mean square error and network complexity as the objective function, and adopt an improved NSGA-III algorithm for multi-objective optimization based on the signal sample set output by the data preprocessing module to obtain a Pareto optimal population; A signal recognition module, which is used to obtain multiple network individuals as base classifiers from the Pareto optimal population obtained by the optimization network module according to the root mean square error, input the real-time collected communication signals output by the data preprocessing module into the base classifiers for ensemble learning, and after weighted summation of the output results, take the category corresponding to the maximum value as the communication signal recognition result; the step of obtaining multiple network individuals as base classifiers from the Pareto optimal population obtained by the optimization network module according to the root mean square error is to sort according to the root mean square error from small to large, and select the first few network individuals as base classifiers according to the quantity threshold, and respectively set the weights from large to small for the weighted summation of the output results.

Images