CN115049006A - Communication signal identification method and system based on self-adaptive feedforward neural network - Google Patents
Communication signal identification method and system based on self-adaptive feedforward neural network Download PDFInfo
- Publication number
- CN115049006A CN115049006A CN202210695681.2A CN202210695681A CN115049006A CN 115049006 A CN115049006 A CN 115049006A CN 202210695681 A CN202210695681 A CN 202210695681A CN 115049006 A CN115049006 A CN 115049006A
- Authority
- CN
- China
- Prior art keywords
- network
- hidden layer
- population
- signal identification
- individuals
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
- G06N20/10—Machine learning using kernel methods, e.g. support vector machines [SVM]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Software Systems (AREA)
- Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Artificial Intelligence (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Compression, Expansion, Code Conversion, And Decoders (AREA)
Abstract
The invention relates to a communication signal identification method and system based on a self-adaptive feedforward neural network, belongs to the technical field of signal identification, and solves the problems of inaccurate signal identification and difficult model parameter adjustment of the existing small sample. Randomly generating binary hidden layer neuron activation states, real number type hidden layer input weights and hidden layer offsets based on a single hidden layer feedforward neural network, and forming a plurality of initial network individuals in a mixed coding mode to serve as a parent population; taking the root mean square error and the network complexity as target functions, and obtaining a pareto optimal population by adopting an improved NSGA-III algorithm based on the preprocessed signal sample set; and acquiring a plurality of network individuals from the pareto optimal population according to the root mean square error to serve as a base classifier, preprocessing the communication signals acquired in real time, transmitting the preprocessed communication signals into the base classifier to be integrated and learned, weighting and summing the output results, and taking the category corresponding to the maximum value as a communication signal identification result. Accurate signal identification is achieved.
Description
Technical Field
The invention relates to the technical field of signal identification, in particular to a communication signal identification method and system based on a self-adaptive feedforward neural network.
Background
Data classification and recognition is an important task in machine learning, and generally requires processing of data sets in the fields of biological information, multimedia, statistics, and the like. In the field of signal identification, target categories of transmitted signals and specific target individuals are gradually identified through extraction and analysis of signal features.
In the existing machine learning, a neural network correlation algorithm is mostly adopted, but the optimal values of the connection weight and the number of hidden layer neurons are difficult to find at the same time, so that when the neural network is used for solving the actual problem, the difficulty of building a basic model exists; the feedforward neural network is trained by using a gradient-based method, and may get into a local minimum, and the convergence is slow and depends on an initial solution, so that the error is large when the neural network model is used for solving an actual problem.
An Extreme Learning Machine (ELM) is a commonly used machine learning method for training a feedforward neural network SLFN, and is applied to many fields such as data classification and biomedical engineering. And randomly distributing an input weight and hidden layer bias by using an ELM (element-weighted matrix), then calculating an output weight by using an MP (mesh point) generalized inverse, and improving the training speed by using a parameter learning method without parameter adjustment. However, randomly assigned network parameters may cause instability in network performance.
Meanwhile, when the target class of the signal sample is unbalanced or the sample data size is small, the accuracy evaluation of the classifier is likely to become meaningless, the correct class of the signal transmitting target cannot be identified, during identification, single-target optimization only determines one optimal solution in one operation, and multiple solutions need to be solved again to obtain.
Disclosure of Invention
In view of the foregoing analysis, embodiments of the present invention aim to provide a communication signal identification method and system based on an adaptive feedforward neural network, so as to solve the problems of inaccurate signal identification of existing small samples and difficult adjustment of model parameters.
In one aspect, an embodiment of the present invention provides a communication signal identification method based on an adaptive feedforward neural network, including the following steps:
based on a single hidden layer feedforward neural network, randomly generating binary hidden layer neuron activation states, real number type hidden layer input weights and hidden layer biases, and forming a plurality of initial network individuals in a mixed coding mode;
taking a plurality of initial network individuals as parent population, taking root mean square error and network complexity as objective functions, and performing multi-objective optimization by adopting an improved NSGA-III algorithm based on a preprocessed signal sample set to obtain a pareto optimal population;
and acquiring a plurality of network individuals from the pareto optimal population as a base classifier according to the root mean square error, preprocessing the communication signals acquired in real time, transmitting the preprocessed communication signals into the base classifier for ensemble learning, weighting and summing the output results, and taking the category corresponding to the maximum value as the communication signal identification result.
Based on a further improvement of the above method, the number K of hidden layer neurons of the single hidden layer feedforward neural network is at most 2N +1, where N represents the number of input neurons.
Based on a further improvement of the above method, each hidden layer neuron activation state of the binary type includes: when the hidden layer neuron activation state is 1, the hidden layer neuron is activated, and when the hidden layer neuron activation state is 0, the hidden layer neuron is not activated; the range of the hidden layer input weight and the hidden layer bias of the real number type is [ -1,1 ].
Based on a further improvement of the above method, the initial plurality of network individuals are composed by mixing the encoded forms, as shown in the following formula:
wherein, X i (G) Denotes the ith network entity of the G-th generation, i 1,2 pop ，N pop Is the size of the population, theta i,j (G) A jth hidden layer neuron activation state, θ, representing an ith network individual i,j (G)∈{0,1}；
The input weight of the jth hidden layer neuron representing the ith network individual,
b i,j (G) bias of the jth hidden layer neuron representing the ith network individual, b i,j (G)∈[-1,1],j＝1,2,...,K。
Based on the further improvement of the method, the root mean square error and the network complexity are used as objective functions, and the minimum root mean square error and the minimum network complexity are used as two mutually conflicting targets to be optimized, wherein the network complexity is represented by the average value of the neuron activation states of all hidden layers in the network.
Based on further improvement of the method, the improved NSGA-III algorithm is adopted for multi-objective optimization to obtain the pareto optimal population, and the method comprises the following steps:
s121: for the population P of the current iteration G Performing mixed intersection and mixed variation of binary numbers and real numbers to generate a filial generation population Q G ，P G And Q G The number of the medium network individuals is the same;
s122: synthesis of a novel population R G ＝P G ∪Q G Introducing the preprocessed signal sample set into the population R G In each network individual, after calculating a corresponding output weight value through an extreme learning machine, calculating two objective function values of each network individual to obtain an objective function value vector;
s123: for population R G Performing non-dominated sorting, selecting mechanism based on reference point, and selecting from population R according to objective function value vector of each network individual G Screening out network individuals to obtain a population P G+1 ；
S124: judging whether the circulation reaches the maximum iteration times, if not, returning to S121 to carry out the population P G+1 As new P G Performing circulation, if the number of the seeds reaches the preset value, exiting the circulation, and obtaining the population P G+1 Namely the pareto optimal population.
Based on a further improvement of the above method, the hybrid intersection of binary and real numbers comprises: a binary intersection for a K-dimensional binary vector, and a real intersection for an (N +1) × K-dimensional real vector;
binary interleaving is performed by:
real interleaving is performed by:
wherein alpha is j E {0,1} is a randomly selected binary parameter; beta is a j Is a continuous parameter, in [0, 1]]Selecting randomly; d is the total length of the network individual hybrid code, D ═ N +2 × K; x is the number of p,j And x q,j Respectively, the current iteration population P G And the jth coded value, y, in the qth network entity p,j And y q,j Respectively, are generated offspring populations Q G And the jth code value in the qth network individual.
Based on a further improvement of the above method, the mixed variation of binary and real numbers includes: inverting the randomly distributed binary variation and increasing the real variation of the normally distributed random number;
binary mutation was performed by the following formula:
the real number variation is performed by the following formula:
wherein 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, x t,j Is the current iteration population P G Of the tth network entity of (1) a j-th coded value, y t,j Is the generated offspring population Q G The jth coded value in the tth network individual.
Based on further improvement of the method, a plurality of network individuals are obtained from the pareto optimal population according to the root mean square error and are used as base classifiers, the network individuals are sorted from small to large according to the root mean square error, the first network individuals are selected as the base classifiers according to the quantity threshold, and weight values from large to small are respectively set and used for weighting and summing output results.
In another aspect, an embodiment of the present invention provides a communication signal identification system based on an adaptive feedforward neural network, including:
the signal preprocessing module is used for preprocessing historical communication signals to obtain a signal sample set and preprocessing the communication signals acquired in real time;
the network initialization module is used for randomly generating a binary hidden layer neuron activation state, a real number type hidden layer input weight and hidden layer bias based on a single hidden layer feedforward neural network, and forming a plurality of initial network individuals in a mixed coding mode;
the network optimization module is used for performing multi-objective optimization by adopting an improved NSGA-III algorithm on the basis of a signal sample set output by the data preprocessing module by taking a plurality of initial network individuals as parent populations and root mean square errors and network complexity as objective functions to obtain a pareto optimal population;
and the signal identification module is used for acquiring a plurality of network individuals from the pareto optimal population obtained from the optimization network module according to the root mean square error as a base classifier, transmitting the communication signals acquired in real time and output by the data preprocessing module into the base classifier for ensemble learning, weighting and summing the output results, and then taking the category corresponding to the maximum value as the communication signal identification result.
Compared with the prior art, the invention can realize at least one of the following beneficial effects:
1. in the learning of the single hidden layer feedforward neural network, two contradictory factors of training errors and network complexity are considered at the same time, a multi-target learning model is established, a hybrid coding multi-target learning method integrated with an extreme learning machine is provided, and the generalization capability and the learning speed of a communication signal recognition network are improved;
2. the non-dominated sorting genetic method III (NSGA-III) is improved and is used for processing a multi-target model with mixed variables, so that the multi-target model is suitable for continuous and discrete decision variables, wherein binary coding is used for structure learning, real number coding is used for optimizing input weight and hidden layer bias, a trained network model is directly obtained by solving a pareto optimal solution, the difficulty of model construction is overcome, and the balance between training errors and network complexity is achieved;
3. a plurality of networks are selected from the pareto optimal solution for ensemble learning, and the identification accuracy when the number of communication signal samples is small or the types of the samples are unbalanced is improved.
In the invention, the technical schemes can be combined with each other to realize more preferable combination schemes. Additional features and advantages of the invention will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. The objectives and other advantages of the invention will be realized and attained by the structure particularly pointed out in the written description and drawings.
Drawings
The drawings are only for purposes of illustrating particular embodiments and are not to be construed as limiting the invention, wherein like reference numerals are used to designate like parts throughout.
FIG. 1 is a flowchart of a communication signal identification method based on an adaptive feedforward neural network according to embodiment 1 of the present invention;
FIG. 2 is a schematic structural diagram of a single hidden layer feedforward neural network in embodiment 1 of the present invention;
FIG. 3 is a flow chart of the usage of signal sample sets in embodiment 1 of the present invention;
fig. 4(a) and 4(b) are pareto frontier maps generated by diabetes and glass datasets at different iterations in embodiment 1 of the present invention.
Detailed Description
The accompanying drawings, which are incorporated in and constitute a part of this application, illustrate preferred embodiments of the invention and together with the description, serve to explain the principles of the invention and not to limit the scope of the invention.
A specific embodiment of the present invention discloses a communication signal identification method based on an adaptive feedforward neural network, as shown in fig. 1, including the following steps:
s11: based on a single hidden layer feedforward neural network, the binary hidden layer neuron activation states, the real number hidden layer input weights and hidden layer biases are randomly generated, and a plurality of initial network individuals are formed in a mixed coding mode.
As shown in fig. 2, the structure of a single hidden layer feedforward neural network is schematically illustrated, where the network includes N input neurons (i.e., input data has N features), K hidden neurons, L output neurons (i.e., L data classes), an activation function g (x) employs a sigmoid function, ω is an input weight connecting the input neurons and the hidden neurons, b is a bias of the hidden neurons, and β is an output weight connecting the hidden neurons and the output neurons. The maximum number K of hidden layer neurons of the single hidden layer feedforward neural network is 2N + 1.
It should be noted that, in the present embodiment, a mixed code of a binary code and a real number code is used to define the network structure and the connection weights of the single hidden layer feedforward neural network. Where binary coding is used to represent the network structure, i.e. the activation state θ of hidden layer neurons, the value of θ may be 1 or 0, where θ ═ 1 represents that the corresponding hidden layer neurons are activated, and θ ═ 0 represents that the corresponding hidden layer neurons are not activated. Real number encoding is used to represent the input weight ω and the hidden layer bias b. Because the extreme learning machine is adopted to calculate the output weight in the embodiment, the hidden layer input weight and the hidden layer bias of the real number type are randomly generated in the range of [ -1,1 ].
Forming an initial plurality of network individuals by mixing the encoded forms, as shown in the following formula:
wherein, X i (G) Denotes the ith network entity of the G-th generation, i 1,2 pop ，N pop Is the size of the population, theta i,j (G) A jth hidden layer neuron activation state, θ, representing an ith network individual i,j (G)∈{0,1}；
The input weight of the jth hidden layer neuron representing the ith network individual,
b i,j (G) bias of the jth hidden layer neuron representing the ith network individual, b i,j (G)∈[-1,1],j＝1,2,...,K。
As can be seen from equation (1), the total length D of the hybrid code of one network individual is (N +2) × K, and the G-th generation group can be represented as:
s12: and taking a plurality of initial network individuals as parent population, taking root mean square error and network complexity as objective functions, and performing multi-objective optimization by adopting an improved NSGA-III algorithm based on the preprocessed signal sample set to obtain the pareto optimal population.
It should be noted that, in the present embodiment, the root mean square error and the network complexity are used as objective functions, and the minimum root mean square error and the minimum network complexity are used as two mutually conflicting objectives to be optimized, so as to establish a two-objective optimization model.
The first objective function is the root mean square error, which is expressed by:
wherein o is r1,r2 Representing the net output calculated by the training data, i.e. the r1 th training data is identified as the net output value, s, corresponding to the r2 th label r1,r2 And representing the actual output corresponding to the training data, namely, the r1 th training data is identified as the actual value of the r2 th label, if the training data is actually the category corresponding to the L-th label, the training data is 1, otherwise, the training data is 0, M is the number of the training data, and L is the number of the labels.
The second objective function is the network complexity, which is represented by the average of the hidden neuron activation states in the network, as shown in the following formula:
where K is the number of cryptic neurons, θ j Is the activation state of the jth hidden layer neuron.
Thus, the two-objective optimization model can be expressed as:
wherein, K 0 The minimum value of the number of the activated hidden layer neurons is represented and can be set according to actual conditions; preferably, K in the present embodiment 0 ＝3。
It should be noted that, in this embodiment, a non-dominated sorting genetic method III (NSGA-III) is improved, so that it can process a multi-objective model with mixed variables, perform binary and real mixed intersection and mixed variation on each generation of population, integrate with an Extreme Learning Machine (ELM), and obtain a target function value vector of each network individual after calculating an output weight of each network individual in each generation of population.
Specifically, the method adopts an improved NSGA-III algorithm to carry out multi-objective optimization to obtain the pareto optimal population, and comprises the following steps:
s121: for the population P of the current iteration G Performing mixed intersection and mixed variation of binary numbers and real numbers to generate a filial generation population Q G ，P G And Q G The number of the medium network individuals is the same;
it should be noted that the population P at the current iteration G Randomly selecting two network individuals to generate a filial generation population Q G Two network individuals.
The mixed interleaving of binary and real numbers respectively interleaves corresponding parts in the mixed coding, and comprises the following steps: binary interleaving for K-dimensional binary vectors, and real interleaving for (N +1) × K-dimensional real vectors.
Binary interleaving is performed by:
real interleaving is performed by:
wherein alpha is j E {0,1} is a randomly selected binary parameter; beta is a j Is a continuous parameter, in [0, 1]]Selecting randomly; d is the total length of the network individual hybrid code, D ═ N +2 × K; x is the number of p,j And x q,j Respectively, the current iteration population P G And the jth coded value, y, in the qth network entity p,j And y q,j Respectively, are generated offspring populations Q G And the jth code value in the qth network individual.
Mixed variations of binary and real include: inverting the randomly distributed binary variation and increasing the real variation of the normally distributed random number;
binary mutation was performed by the following formula:
the real number variation is performed by the following formula:
wherein 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 value of 0 and a variance of σ, and x t,j Is the current iteration population P G Of the tth network entity of (1) a j-th coded value, y t,j Is the generated offspring population Q G The jth coded value in the tth network individual.
Preferably, μ ═ 0.02 and σ ═ 0.2.
Population size is N pop Mixed interleaved execution of N in each cycle pop ×P c Next, mixed mutation execution N pop ×P m Wherein P is c P is more than or equal to 0.7 and is the cross probability c ≤0.9，P m Is variation probability, P is more than or equal to 0.1 m Less than or equal to 0.3, generating N pop Using individual network as filial generation population Q G 。
S122: synthesis of a novel population R G ＝P G ∪Q G Transmitting the preprocessed signal sample set into the population R G In each network individual, after calculating a corresponding output weight value through an extreme learning machine, calculating two objective function values of each network individual to obtain an objective function value vector;
it should be noted that the population P of the current iteration is G And progeny population Q G Combining to obtain a new population R G Wherein the number of network individuals is 2N pop 。
The signal sample set is a set of preprocessed historical communication signals, including a training set and a test set.
Specifically, communication signals from various targets are collected by the broadband signal collector, such as: passenger aircraft, transport aircraft, fuel dispensers, and helicopters; the collected communication signals are subjected to cluster analysis to obtain concerned frequency points; and extracting the characteristics of the communication signal on the time domain and the frequency domain by methods such as Fourier transform, time-frequency analysis, fuzzy function, SPWVD, convolutional neural network and the like. Wherein the time domain features include: rising edge, falling edge, top unevenness and pulse width; the frequency domain features include: frequency points, frequency spectrogram, instantaneous frequency fourth moment and frequency symmetry.
Normalizing each characteristic of each piece of communication signal data to an [ -1,1] interval using the following equation:
wherein, c f,h For the h-th normalization value in feature c, c b,h For the h-th value to be normalized in feature c, c b,min And c b,max The minimum and maximum values of the values to be normalized for feature c.
Preferably, a hierarchical 10-fold cross validation (10-CV) technology is adopted, the preprocessed signal sample set is divided into 10 parts according to the target class ratio on average, 9 parts of the 10 parts are selected as a training set, and the rest are selected as a testing set.
As shown in fig. 3, based on the training set, according to the hidden layer input weights and hidden layer biases, the output weight corresponding to each network individual is calculated by solving the least square solution of the linear system using the generalized inverse (M-P, Moore-Penrose) in the extreme learning machine.
Compared with the prior art, the input weight and the hidden bias are randomly generated values only in the first iteration, and are screened according to the improved NSGA-III algorithm in the subsequent iterations, so that the stability of the network is greatly improved and the network error is reduced when the extreme learning machine is used.
Obtaining a corresponding network output value of the training data in each network individual according to the output weight, and substituting the formula (3) to obtain a first objective function value of each network individual; and (4) obtaining a second objective function value of each network individual by substituting the formula (4) according to the binary code in the mixed code of each network individual. And forming an objective function value vector of each network individual according to the first objective function value and the second objective function value.
S123: for population R G Performing non-dominated sorting, selecting mechanism based on reference point, and selecting from population R according to objective function value vector of each network individual G Screening out network individuals to obtain a population P G+1 ；
Specifically, the steps are divided into the following steps:
to population R G Performing non-dominant sorting to obtain a plurality of non-dominant layer sets (F) with different priorities 1 ,F 2 ,F 3 ,..), each non-dominated layer comprising a plurality of network individuals;
② a new population S is constructed G Adding the new population S in sequence from the non-dominant layer with the highest priority G Until S first appears G The number of medium network individuals is greater than or equal to N pop At this time, the last added non-dominant layer is denoted as F l ，
(iii) if new population S G The number of middle network individuals is equal to N pop Then S is G Namely the population P G+1 Otherwise, first (F) 1 ,F 2 ,F 3 ,...F l-1 ) Put into a population P G+1 Then from F l The selection mechanism based on the reference point selects the rest number of network individuals to join the population P G+1 In (C) making the population P G+1 The number of network units is N pop And (4) respectively.
Specifically, the number of reference points is calculated using the following formula:
where m is the number of targets, each target is divided into v parts. In the present embodiment, m is 2, and illustratively, when v is 10, the number of reference points is
And (4) respectively.
According to the new population S G The minimum objective function value of all network individuals on each dimension of the target forms a new population S G Normalized S based on the ideal point G Target function value vector of each network; and calculating the vertical distance from each network individual to each reference line, determining the reference point associated with each network individual according to the shortest distance, and simultaneously acquiring the number of the network individuals associated with each reference point, namely the number of the parasitic mirrors of each reference point.
From F according to the lenticule method l Selecting an individual comprising:
sorting the ecological niche numbers in an ascending order, if the ecological niche number of the reference point is 0, indicating that no network individual is associated with the reference point in the current population, and excluding the reference point; selecting from the reference points with the number of niches being greater than or equal to 1 and the number of niches being the smallest, selecting one at random when there are reference points with the same number of niches, in which case, for the selected reference point, if at F l One or more network individuals are associated with the group P, and the network individual with the shortest vertical distance to the reference line is added into the group P G+1 In, if atF l If no network individual is associated with it, the next reference point is selected, regardless of the reference point, until the population P G+1 The number of network units is N pop And (4) respectively.
S124: judging whether the circulation reaches the maximum iteration times, if not, returning to S121 to carry out the population P G+1 As new P G Performing circulation, if the number of the seeds reaches the preset value, exiting the circulation, and obtaining the population P G+1 Namely the pareto optimal population.
Preferably, the maximum number of iterations is set to 50.
The iteration is finished, and the finally obtained population P G+1 The method is the pareto optimal population, wherein the network individuals directly correspond to the trained network model, so that the difficulty of model construction is overcome, and the balance between training errors and network complexity is achieved.
Illustratively, the reference classification dataset of the UCI machine learning repository (http:// archive. ics. UCI. edu) was trained and tested using the method of the present embodiment. Maximum iterations for the diabetes and glass datasets set to 10, 20, 30, 40 and 50 generations respectively, resulting in pareto fronts as shown in fig. 4(a) and 4 (b). As can be seen from the figure, the convergence of the method becomes better as the number of iterations increases. When the method reaches 50 generations, a group of pareto optimal solutions can be found, and compared with the prior art, the method has the advantages of less iteration times, low calculation cost and stable performance.
S13: and acquiring a plurality of network individuals from the pareto optimal population as a base classifier according to the root mean square error, preprocessing the communication signals acquired in real time, transmitting the preprocessed communication signals into the base classifier for ensemble learning, weighting and summing the output results, and taking the category corresponding to the maximum value as the communication signal identification result.
It should be noted that, according to the root mean square error, a plurality of network individuals are obtained from the pareto optimal population and are used as base classifiers, the network individuals are sorted from small to large according to the root mean square error, the first network individuals are selected as the base classifiers according to the quantity threshold, and weight values from large to small are respectively set and used for weighting and summing output results.
Preferably, in fig. 3, the identification performance of each base classifier is tested based on the test set in the signal sample set, and the average accuracy, the standard deviation of the accuracy and the number of hidden neurons of the signal identification are obtained.
Preferably, the number threshold is set to 3, that is, 3 network individuals with the minimum root mean square error are selected as the base classifier, and the corresponding weight λ is obtained cj 0.7, 0.2 and 0.1 in sequence.
In implementation, the communication signal is collected by the broadband signal collector, such as a real-time spectrum analyzer, and is preprocessed according to the preprocessing method in step S122 and then transmitted to the finally determined base classifier, and the maximum value is obtained by using the following formula, and the corresponding target class is the recognition result:
wherein x is te JN is the number of the base classifiers for the test data or the pre-processed data collected in real time; (C) cj,1 (x te ),C cj,2 (x te ),...,C cj,L (x te ) Is the classification result of the cj-th base classifier, and after weighted summation, the maximum value C (x) is taken te ) And the corresponding category is used as a final signal identification result, namely the transmission target category of the acquired signal.
It should be noted that the method in this embodiment is not limited to the identification of the type of the transmission target, and may be configured to label a specific individual type when preprocessing the signal history data, and retrain the specific individual type to obtain a network individual structure for identifying a specific source individual for the signal acquired in real time; and the targets (airplanes, tanks, ships and the like) can be classified and identified based on image data such as visible light, infrared and the like.
Compared with the prior art, the communication signal identification method based on the self-adaptive feedforward neural network provided by the embodiment considers two contradictory factors, namely training errors and network complexity, simultaneously in the learning of the single-hidden-layer feedforward neural network, establishes a multi-target learning model, provides a hybrid coding multi-target learning method integrated with an extreme learning machine, and improves the generalization capability and the learning speed of the communication signal identification network; the non-dominated sorting genetic method III (NSGA-III) is improved and is used for processing a multi-target model with mixed variables, so that the multi-target model is suitable for continuous and discrete decision variables, wherein binary coding is used for structure learning, real number coding is used for optimizing input weight and hidden layer bias, a trained network model is directly obtained by solving a pareto optimal solution, the difficulty of model construction is overcome, and the balance between training errors and network complexity is achieved; a plurality of networks are selected from the pareto optimal solution for ensemble learning, and the identification accuracy is improved when the number of signal samples is small or the types of the samples are unbalanced.
Example 2
In another embodiment of the present invention, a communication signal identification system based on an adaptive feedforward neural network is disclosed, so as to implement the signal identification method in embodiment 1. The concrete implementation of each module refers to the corresponding description in embodiment 1. The method comprises the following steps:
the signal preprocessing module is used for preprocessing historical communication signals to obtain a signal sample set and preprocessing the communication signals acquired in real time;
the network initialization module is used for randomly generating a binary hidden layer neuron activation state, a real number type hidden layer input weight and hidden layer bias based on a single hidden layer feedforward neural network, and forming a plurality of initial network individuals in a mixed coding mode;
the network optimization module is used for performing multi-objective optimization by adopting an improved NSGA-III algorithm on the basis of a signal sample set output by the data preprocessing module by taking a plurality of initial network individuals as parent populations and root mean square errors and network complexity as objective functions to obtain a pareto optimal population;
and the signal identification module is used for acquiring a plurality of network individuals from the pareto optimal population obtained from the optimization network module according to the root mean square error as a base classifier, transmitting the communication signals acquired in real time and output by the data preprocessing module into the base classifier for ensemble learning, weighting and summing the output results, and then taking the category corresponding to the maximum value as the communication signal identification result.
Since the relevant parts of the communication signal identification system based on the adaptive feedforward neural network and the communication signal identification method in the embodiment can be referred to each other, the description is repeated here, and thus, the details are not repeated here. Since the principle of the embodiment of the system is the same as that of the embodiment of the method, the system also has the corresponding technical effect of the embodiment of the method.
Those skilled in the art will appreciate that all or part of the flow of the method implementing the above embodiments may be implemented by a computer program, which is stored in a computer readable storage medium, to instruct related hardware. The computer readable storage medium is a magnetic disk, an optical disk, a read-only memory or a random access memory.
The above description is only for the preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present invention are included in the scope of the present invention.
Claims (10)
1. A communication signal identification method based on an adaptive feedforward neural network is characterized by comprising the following steps:
based on a single hidden layer feedforward neural network, randomly generating binary hidden layer neuron activation states, real number type hidden layer input weights and hidden layer biases, and forming a plurality of initial network individuals in a mixed coding mode;
taking a plurality of initial network individuals as parent population, taking root mean square error and network complexity as objective functions, and performing multi-objective optimization by adopting an improved NSGA-III algorithm based on a preprocessed signal sample set to obtain a pareto optimal population;
and acquiring a plurality of network individuals from the pareto optimal population as a base classifier according to the root mean square error, preprocessing the communication signals acquired in real time, transmitting the preprocessed communication signals into the base classifier for ensemble learning, weighting and summing the output results, and taking the category corresponding to the maximum value as the communication signal identification result.
2. A communication signal identification method based on an adaptive feedforward neural network according to claim 1, wherein the number K of hidden layer neurons of the single hidden layer feedforward neural network is at most 2N +1, where N represents the number of input neurons.
3. The adaptive feedforward neural network-based communication signal identification method of claim 2, wherein the binary type hidden layer neuron activation states each include: when the hidden layer neuron activation state is 1, the hidden layer neuron is activated, and when the hidden layer neuron activation state is 0, the hidden layer neuron is not activated; the range of the hidden layer input weight and the hidden layer bias of the real number type is [ -1,1 ].
4. A method as claimed in claim 3, wherein the initial plurality of network individuals are composed by mixed coding, as shown in the following formula:
wherein, X i (G) Denotes the ith network entity of the G-th generation, i 1,2 pop ，N pop Is the size of the population, theta i,j (G) The jth hidden layer neuron activation state, θ, representing the ith network individual i,j (G)∈{0,1}；
The input weight of the jth hidden layer neuron representing the ith network individual,
b i,j (G) bias of the jth hidden layer neuron representing the ith network individual, b i,j (G)∈[-1,1],j＝1,2,...,K。
5. The adaptive feedforward neural network-based communication signal identification method according to claim 4, wherein the root mean square error and the network complexity are used as objective functions, and the minimum root mean square error and the minimum network complexity are used as two conflicting targets to be optimized, wherein the network complexity is represented by an average value of activation states of hidden neurons in the network.
6. The method for identifying communication signals based on the adaptive feedforward neural network as claimed in claim 4, wherein the performing multi-objective optimization by using the improved NSGA-III algorithm to obtain the pareto optimal population comprises:
s121: for the population P of the current iteration G Performing mixed intersection and mixed variation of binary numbers and real numbers to generate a filial generation population Q G ，P G And Q G The number of the medium network individuals is the same;
s122: synthesis of a novel population R G ＝P G ∪Q G Introducing the preprocessed signal sample set into the population R G In each network individual, after calculating a corresponding output weight value through an extreme learning machine, calculating two objective function values of each network individual to obtain an objective function value vector;
s123: for population R G Performing non-dominated sorting, selecting mechanism based on reference point, and selecting from population R according to objective function value vector of each network individual G Screening out network individuals to obtain a population P G+1 ；
S124: judging whether the circulation reaches the maximum iteration times, if not, returning to S121 to carry out the population P G+1 As new P G Performing circulation, if the number of the seeds reaches the preset value, exiting the circulation, and obtaining the population P G+1 Namely the pareto optimal population.
7. The adaptive feedforward neural network-based communication signal identification method of claim 6, wherein the mixed intersection of binary and real numbers comprises: a binary intersection for a K-dimensional binary vector, and a real intersection for an (N +1) × K-dimensional real vector;
the binary interleaving is performed by:
the real interleaving is performed by:
wherein alpha is j E {0,1} is a randomly selected binary parameter; beta is a j Is a continuous parameter, in [0, 1]]Selecting randomly; d is the total length of the network individual hybrid code, D ═ N +2 × K; x is the number of p,j And x q,j Respectively, the current iteration population P G And the jth coded value, y, in the qth network entity p,j And y q,j Respectively, are generated offspring populations Q G And the jth code value in the qth network entity.
8. The adaptive feedforward neural network-based communication signal identification method of claim 6, wherein the mixed variation of binary and real numbers comprises: inverting the randomly distributed binary variation and increasing the real variation of the normally distributed random number;
the binary mutation is performed by:
performing the real number variation by:
wherein 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 value of 0 and a variance of σ, and x t,j Is the current iteration population P G Of the tth network entity of (1) a j-th coded value, y t,j Is the generated offspring population Q G The jth coded value in the tth network individual.
9. The communication signal identification method based on the adaptive feedforward neural network as claimed in claim 6, wherein the plurality of network individuals obtained from the pareto optimal population according to the root mean square error are used as the basis classifiers, the network individuals are sorted from small to large according to the root mean square error, the first network individuals are selected as the basis classifiers according to a quantity threshold, and weights from large to small are respectively set for weighting and summing the output results.
10. A communication signal identification system based on an adaptive feedforward neural network, comprising:
the signal preprocessing module is used for preprocessing historical communication signals to obtain a signal sample set and preprocessing the communication signals acquired in real time;
the network initialization module is used for randomly generating a binary hidden layer neuron activation state, a real number type hidden layer input weight and hidden layer bias based on a single hidden layer feedforward neural network, and forming a plurality of initial network individuals in a mixed coding mode;
the network optimization module is used for performing multi-objective optimization by adopting an improved NSGA-III algorithm on the basis of a signal sample set output by the data preprocessing module by taking a plurality of initial network individuals as parent populations and root mean square errors and network complexity as objective functions to obtain a pareto optimal population;
and the signal identification module is used for acquiring a plurality of network individuals from the pareto optimal population obtained from the optimization network module according to the root mean square error as a base classifier, transmitting the communication signals acquired in real time and output by the data preprocessing module into the base classifier for ensemble learning, weighting and summing the output results, and then taking the category corresponding to the maximum value as the communication signal identification result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210695681.2A CN115049006A (en) | 2022-06-20 | 2022-06-20 | Communication signal identification method and system based on self-adaptive feedforward neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210695681.2A CN115049006A (en) | 2022-06-20 | 2022-06-20 | Communication signal identification method and system based on self-adaptive feedforward neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115049006A true CN115049006A (en) | 2022-09-13 |
Family
ID=83162781
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210695681.2A Pending CN115049006A (en) | 2022-06-20 | 2022-06-20 | Communication signal identification method and system based on self-adaptive feedforward neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115049006A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116578913A (en) * | 2023-03-31 | 2023-08-11 | 中国人民解放军陆军工程大学 | Reliable unmanned aerial vehicle detection and recognition method oriented to complex electromagnetic environment |
-
2022
- 2022-06-20 CN CN202210695681.2A patent/CN115049006A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116578913A (en) * | 2023-03-31 | 2023-08-11 | 中国人民解放军陆军工程大学 | Reliable unmanned aerial vehicle detection and recognition method oriented to complex electromagnetic environment |
| CN116578913B (en) * | 2023-03-31 | 2024-05-24 | 中国人民解放军陆军工程大学 | Reliable unmanned aerial vehicle detection and recognition method oriented to complex electromagnetic environment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112465040B (en) | Software defect prediction method based on class unbalance learning algorithm | |
| CN110009030B (en) | Sewage treatment fault diagnosis method based on stacking meta-learning strategy | |
| CN107992976B (en) | Hot topic early development trend prediction system and prediction method | |
| CN108921604B (en) | Advertisement click rate prediction method based on cost-sensitive classifier integration | |
| CN110363230B (en) | Stacking integrated sewage treatment fault diagnosis method based on weighted base classifier | |
| CN107506590A (en) | A kind of angiocardiopathy forecast model based on improvement depth belief network | |
| CN107292097B (en) | Chinese medicine principal symptom selection method based on feature group | |
| CN110059852A (en) | A kind of stock yield prediction technique based on improvement random forests algorithm | |
| CN110348608A (en) | A kind of prediction technique for improving LSTM based on fuzzy clustering algorithm | |
| CN112784031B (en) | Method and system for classifying customer service conversation texts based on small sample learning | |
| CN105913450A (en) | Tire rubber carbon black dispersity evaluation method and system based on neural network image processing | |
| CN116542382A (en) | Sewage treatment dissolved oxygen concentration prediction method based on mixed optimization algorithm | |
| CN115115389A (en) | Express customer loss prediction method based on value subdivision and integrated prediction | |
| CN109919236A (en) | A kind of BP neural network multi-tag classification method based on label correlation | |
| Fatima et al. | Xgboost and random forest algorithms: an in depth analysis | |
| CN117291069A (en) | LSTM sewage water quality prediction method based on improved DE and attention mechanism | |
| CN116911459A (en) | Multi-input multi-output ultra-short-term power load prediction method suitable for virtual power plant | |
| Zhao et al. | A hybrid learning method for constructing compact rule-based fuzzy models | |
| CN115392387B (en) | Low-voltage distributed photovoltaic power generation output prediction method | |
| CN115049006A (en) | Communication signal identification method and system based on self-adaptive feedforward neural network | |
| CN110177112B (en) | Network intrusion detection method based on double subspace sampling and confidence offset | |
| CN114330135B (en) | Classification model construction method and device, storage medium and electronic equipment | |
| CN109919374A (en) | Prediction of Stock Price method based on APSO-BP neural network | |
| Johansson | Obtaining accurate and comprehensible data mining models: An evolutionary approach | |
| CN116956160A (en) | Data classification prediction method based on self-adaptive tree species algorithm |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |