CN114595729A - Communication signal modulation identification method based on residual error neural network and meta-learning fusion - Google Patents
Communication signal modulation identification method based on residual error neural network and meta-learning fusion Download PDFInfo
- Publication number
- CN114595729A CN114595729A CN202210336623.0A CN202210336623A CN114595729A CN 114595729 A CN114595729 A CN 114595729A CN 202210336623 A CN202210336623 A CN 202210336623A CN 114595729 A CN114595729 A CN 114595729A
- Authority
- CN
- China
- Prior art keywords
- neural network
- residual
- residual error
- error neural
- signal
- 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
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 149
- 238000004891 communication Methods 0.000 title claims abstract description 85
- 238000000034 method Methods 0.000 title claims abstract description 52
- 230000004927 fusion Effects 0.000 title claims abstract description 37
- 238000012360 testing method Methods 0.000 claims abstract description 36
- 238000013100 final test Methods 0.000 claims abstract description 9
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 claims description 57
- 230000006870 function Effects 0.000 claims description 50
- 238000012549 training Methods 0.000 claims description 34
- 238000011176 pooling Methods 0.000 claims description 15
- 239000000654 additive Substances 0.000 claims description 10
- 230000000996 additive effect Effects 0.000 claims description 10
- 238000005070 sampling Methods 0.000 claims description 7
- 210000002569 neuron Anatomy 0.000 claims description 6
- 238000000605 extraction Methods 0.000 claims description 4
- 230000004913 activation Effects 0.000 claims description 3
- 238000010586 diagram Methods 0.000 description 16
- 230000015556 catabolic process Effects 0.000 description 6
- 238000006731 degradation reaction Methods 0.000 description 6
- 238000010801 machine learning Methods 0.000 description 6
- 230000000694 effects Effects 0.000 description 3
- 238000013145 classification model Methods 0.000 description 2
- 230000007850 degeneration Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000011423 initialization method Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
- 238000004804 winding Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Signal Processing (AREA)
- Probability & Statistics with Applications (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Digital Transmission Methods That Use Modulated Carrier Waves (AREA)
Abstract
The invention provides a communication signal modulation identification method based on residual error neural network and meta-learning fusion, belonging to the technical field of communication and comprising the following steps: acquiring a communication modulation signal; dividing subtasks; constructing a residual error neural network; extracting a sample; updating residual error neural network parameters; updating a meta-learning loss function; updating initial parameters of a residual error neural network; iteration is carried out; determining residual neural network parameters for the test signal; and inputting the final test communication signal into a residual error neural network to obtain a final classification result. The method integrates the residual error neural network with the meta-learning method, can effectively learn the good characteristics of the signals, solves the problem of weak generalization capability of a network model, and can complete the identification of the signals only by a small amount of samples.
Description
Technical Field
The invention belongs to the technical field of communication, and particularly relates to a communication signal modulation identification method based on residual error neural network and meta-learning fusion.
Background
Wireless signal identification (WSR) has broad prospects in military and civilian applications, which may include signal reconnaissance and interception, interference resistance, and device identification. In general, WSR mainly includes Modulation Recognition (MR) and Wireless Technology Recognition (WTR). MR, also known as Automatic Modulation Classification (AMC), is widely used first in the military field and then extended to the civilian field. MR classifies radio signals by identifying modulation patterns, which helps in evaluating wireless transmission schemes and device types. Also, MR is able to extract digital baseband information even in situations where a priori information is limited.
Conventional modulation recognition algorithms can be mainly classified into two types: likelihood-based (LB) and feature-based (FB) methods. The LB method is based on a hypothesis testing theory; the performance based on decision theory is optimal, but the computational complexity is high. Therefore, feature-based methods have been developed as sub-optimal classifiers for practical applications. However, the feature-based method requires manual feature extraction, and is complex to implement. In recent years, deep learning and machine learning have been attracting attention for applications in the field of communications, because it is excellent in computer vision, machine translation, speech recognition, and natural language processing.
The machine learning algorithm in the communication field has the characteristics of no need of special feature extraction, flexible model, end-to-end learning and the like. However, when the communication signal is greatly affected by noise, the recognition rate of the communication signal may be seriously affected. When the network depth of the deep learning model is increased to improve the recognition effect, a 'degeneration' phenomenon often occurs, that is, the model accuracy rate is greatly reduced without signs along with the continuous increase of the network depth. How to ensure the accuracy of the model while ensuring the deepening of the network depth is a problem to be solved. Secondly, the machine learning model is a data-driven model, a large amount of data is needed to ensure the realization effect of the model, and when available data is scarce, the modern deep neural network often cannot achieve the desired effect. Therefore, efficient learning of features of data through a small number of samples is also an issue to be solved. Thirdly, the current machine learning model mostly adopts a random initialization method to initialize parameters, and how to initialize the parameters by using a certain priori knowledge, so that the model knows the characteristics of data, and the specific data is identified and distinguished in a targeted manner. Finally, current machine learning algorithms are greatly affected by data quality in the case of pure data driving, and are difficult to maintain robustness in the presence of imperfect data (e.g., missing or noisy values, outliers, etc.), which results in poor model generalization capability. Machine learning algorithms need to be generalized to improve their generalization capability to result in efficient implementation of the algorithms in practical applications.
Disclosure of Invention
The invention aims to overcome the defects that the prior art has low recognition rate under the condition that signals are influenced by noise, a network model is easy to degrade, the generalization capability of the network model is weak, the influence of data quantity is large and the like, and provides a communication signal modulation recognition method based on the fusion of a residual error neural network and meta-learning, wherein the special structure of the residual error neural network is used, so that the problem of degradation is effectively solved, and the signal recognition rate is effectively improved; in addition, the model can self-adapt the initialization parameters by using different noise schools by using a meta-learning method. By effectively initializing the network parameters, the model can be quickly adapted to the modulation identification of the communication signal. The method can effectively learn the good characteristics of the signals, solve the problem of weak generalization capability of the network model, and can complete the identification of the signals only by a small amount of samples.
The technical scheme adopted by the invention for realizing the purpose is as follows: a communication signal modulation identification method based on residual error neural network and meta-learning fusion comprises the following steps:
Acquiring various communication modulation signals with different signal-to-noise ratios under additive white Gaussian noise as a training data set, wherein the expression of the communication modulation signals with different signal-to-noise ratios under the additive white Gaussian noise is as follows:
wherein, s [ t ]]And r [ t ]]Respectively representing the transmitted and received signals of the t-th time period, gamma t]For communication channel gain, ω0To shift the carrier frequency, θ0Is the phase offset between the sender and receiver, v [ t ]]Is additive white gaussian noise, and j is an imaginary unit.
The expression for the signal-to-noise ratio is:
wherein, PrIs the power of the signal, PvIs the power of the noise.
Dividing the training data set into a plurality of subtask data according to the number of noise-to-noise ratios, identifying a signal under one signal-to-noise ratio as one subtask, wherein the number of the noise-to-noise ratios is N, the number of the subtasks is N, numbering the subtasks according to the magnitude of the signal-to-noise ratios, the greater the number of the signal-to-noise ratios is, the subtask data contains r communication signals, and the types of the communication signals contained in the subtask data are the same. The number of the communication signal types contained in each subtask data is the same as that of the communication modulation signal types used for final testing, and the number of samples in each subtask is the same.
Constructing a residual neural network structure for signal classification, wherein the residual neural network consists of 6 residual blocks, 2 full connection layers, 1 Dropout layer and 1 Softmax layer, and the full connection layers integrate the automatically extracted high-level features into local feature information;
the Softmax layer is used for classifying the multiple classes, and the expression is as follows:
wherein, ak'Is the k' th input signal of the output layer, yk'Is the output of the kth' neuron, C is the number of output nodes, i.e., the number of classes classified, acRepresenting the c-th input signal of the output layer.
Each residual block adopts a network structure of a residual block I or a residual block II, each residual block I and the residual block II consist of a convolution layer, two sub-residual blocks and a pooling layer, each sub-residual block comprises 2 convolution layers, all the convolution layers comprise 32 convolution kernels (including the convolution layers in the sub-residual blocks and the convolution layers outside the sub-residual blocks), the step length is 1, and the padding is 1; the sizes of convolution kernels of sub residual blocks of the first residual block and the second residual block are different, the size of the convolution kernel of the sub residual block of the first residual block is 3 multiplied by 2, and the size of the convolution kernel of the sub residual block of the second residual block is 3 multiplied by 1; the convolutional layer located outside the sub-residual block is activated using the Relu function; the output of each sub-residual block is also activated by the Relu function and maximally pooled, wherein the convolution layer performs convolution operation on the input to extract the characteristics of the signal, and the expression of the convolution operation is as follows:
xl′＝f(wl′*xl′-1+bl′)
wherein l' represents the number of network layers (the network layers including input layers and convolutional layers), wl′Representing the convolution kernel parameters in the current convolution layer, bl′For the bias vector in the current convolutional layer, f (-) denotes the activation function, xl′-1Output, x, representing the last network layer (the network layer comprising the input layer and the convolutional layer)l′An output representing a current convolutional layer;
the pooling layer carries out down-sampling on the convolution layer, and the expression of the pooling layer is as follows:
xl′＝down(xl′-1)
wherein down (·) represents a pooling function; x is the number ofl′-1Representing the last winding layerOutput, xl′An output representing a current pooling layer;
the Relu function expression is:
where f (x) is a function of the maximum value of Relu, and when the input is negative, the output is 0, and the neuron is not activated, x is the input to the function.
And randomly extracting samples from the subtask data, and dividing the samples into a training set and a testing set as the input of the residual error neural network. The number of samples randomly extracted from the subtask data at each time is h + m, the samples include r communication signals, wherein h samples are used as a training set for extracting the samples, and m samples are used as a test set for extracting the samples.
Input data of the residual error neural network is nin×l×dim×1，ninThe number of input samples of the residual error neural network is, and l is the signal sampling length. The signal is in complex form, taking the real and imaginary parts of the signal superimposed together, so dim is 2.
And initializing initial parameters of the residual error neural network randomly, inputting the extracted sample training set into the residual error neural network, and quickly iterating to update the parameters of the residual error neural network.
The step of updating the parameters of the residual error neural network comprises the following steps:
1) random initialization residual neural network parameter phi0；
2) Inputting the extracted sample training set into a residual error neural network;
3) defining a cross entropy loss function l (phi) by using the extracted sample label and a network output result;
the function is gradiented and the parameters are updated,
wherein k is the number of samples takenThe number of times, alpha, is the inner loop learning rate, phikFor the residual neural network parameter at the kth sample draw, l (φ)k) As a cross-entropy loss function of the network parameters,
and the residual error neural network parameters after gradient updating.
The cross entropy loss function is defined as:
l(φ)＝[y(k)logg(x(k))+(1-y(k))log(1-g(x(k)))]
wherein x is(k)For the data of the k-th training set of samples, y(k)For the kth sample of the data of the training set, g represents the residual neural network structure, and l (φ) is the cross entropy loss function.
The iterative update parameters are to ensure that the residual neural network performs well on the extracted data training set. The conditions for judging whether the performance is good are as follows: and when the cross entropy loss function l (phi) is reduced for a period of time, continuing iteration, and keeping l (phi) stable, so that the residual error neural network is proved to have good performance on the extracted data training set.
And inputting the extracted sample test set into the updated network to obtain output data corresponding to the test set. Updating the expression of the meta-learning loss function according to the label corresponding to the test set and the output data corresponding to the test set as follows:
wherein,
for the kth sample test set data draw,
for the kth sample test set data label,
for the updated meta learning loss function after k-1 samples are taken,
And solving a gradient of the meta-learning loss function, and updating initial parameters of the residual neural network by using the gradient. The expression for updating the initial parameters of the residual neural network by using the gradient is as follows:
where β is the outer loop learning rate. The inner loop learning rate α is larger than the outer loop learning rate β.
And initial parameters of the updated residual error neural network.
Repeating the steps 4 to 7 for a plurality of times. When the step 5 is repeated, the parameters of the residual error neural network are not initialized randomly any more, but are initialized to the network parameters updated by the last iteration
Step 9, determining residual error neural network parameters for testing signals
And (5) inputting the communication modulation signal with the same property as the final test signal into the residual error neural network, repeating the step 5, and adjusting the initial parameters of the residual error neural network. When the step 5 is repeated, the parameters of the residual error neural network are not initialized randomly any more, but are initialized to the network parameters updated by the last iteration
And inputting the final test communication signal into a residual error neural network to obtain a final classification result.
Through the design scheme, the invention can bring the following beneficial effects:
1. the invention provides a communication signal modulation recognition method based on a residual error neural network, which learns the difference value between a target value and the input of a residual error block, namely a residual error, by utilizing a special structure (shortcut connection mode) of the residual error block, so that the accuracy rate is not reduced any more as the network is deepened, the problem of 'degradation' when the network depth is increased can be solved, and the recognition accuracy rate of communication signal modulation is improved.
2. The invention also provides a communication signal modulation identification method based on meta-learning, which is characterized in that the priori knowledge of the communication signal is utilized to learn and initialize the initial parameters of the neural network, the initial parameters of the neural network contain the high-quality characteristics of the signal after learning, and the effective characteristics of the signal can be learned without a large amount of data and only a few samples.
3. The communication signal modulation identification method based on meta-learning provided by the invention utilizes the signal-to-noise ratio and the priori knowledge of the communication signals, and learns to initialize the communication signal classification model by extracting the signals under different signal-to-noise ratios, so that the model can adapt to the signals under different signal-to-noise ratios, can adapt to different communication signal identifications under different noises, and improves the generalization capability of the classification model.
Drawings
FIG. 1 is a block diagram of a communication signal modulation recognition method based on the fusion of residual neural network and meta-learning according to the present invention;
FIG. 2 is a signal time domain waveform diagram of 4ASK signal-5 dB in the embodiment of the communication signal modulation identification method based on residual error neural network and meta-learning fusion of the present invention;
FIG. 3 is a time domain waveform diagram of a 4FSK signal of-5 dB according to an embodiment of a communication signal modulation identification method based on a residual error neural network and meta-learning fusion in the present invention;
FIG. 4 is a signal time domain waveform diagram of-5 dB of a 4PSK signal in an embodiment of a communication signal modulation identification method based on residual error neural network and meta-learning fusion in the present invention;
FIG. 5 is a signal time domain waveform diagram of 4QAM signal-5 dB in an embodiment of the communication signal modulation identification method based on residual error neural network and meta-learning fusion of the present invention;
FIG. 6 is a signal time domain waveform diagram of 0dB of a 4ASK signal in an embodiment of a communication signal modulation identification method based on residual error neural network and meta-learning fusion in the present invention;
FIG. 7 is a time domain waveform diagram of a 4 dB FSK signal according to an embodiment of the present invention based on a communication signal modulation and identification method with a residual error neural network and meta-learning fusion;
FIG. 8 is a signal time domain waveform diagram of 0dB of a 4PSK signal in an embodiment of a communication signal modulation identification method based on residual error neural network and meta-learning fusion in the present invention;
FIG. 9 is a signal time domain waveform diagram of 0dB of a 4QAM signal in an embodiment of a communication signal modulation identification method based on residual error neural network and meta-learning fusion in the present invention;
FIG. 10 is a signal time domain waveform diagram of 5dB of a 4ASK signal in an embodiment of a communication signal modulation identification method based on residual error neural network and meta-learning fusion according to the present invention;
fig. 11 is a signal time domain waveform diagram of a 4FSK signal of 5dB in an embodiment of a communication signal modulation identification method based on residual error neural network and meta-learning fusion according to the present invention;
fig. 12 is a signal time domain waveform diagram of 5dB of a 4PSK signal in an embodiment of a communication signal modulation identification method based on residual error neural network and meta-learning fusion according to the present invention;
fig. 13 is a signal time domain waveform diagram of 5dB of a 4QAM signal in an embodiment of the method for communication signal modulation identification based on residual neural network and meta-learning fusion in accordance with the present invention;
FIG. 14 is a flowchart illustrating decomposition of a training subtask in an embodiment of a communication signal modulation recognition method based on fusion of a residual neural network and meta-learning according to the present invention;
FIG. 15 is a diagram illustrating a residual neural network structure in an embodiment of a communication signal modulation recognition method based on a fusion of a residual neural network and meta-learning according to the present invention;
FIG. 16 is a block diagram of a residual block in an embodiment of a communication signal modulation recognition method based on a fusion of a residual neural network and meta-learning according to the present invention;
FIG. 17 is a diagram of a second residual block structure in an embodiment of a communication signal modulation recognition method based on the fusion of a residual neural network and meta-learning according to the present invention;
FIG. 18 is a flowchart illustrating updating parameters of a residual neural network in an embodiment of a communication signal modulation recognition method based on fusion of the residual neural network and meta-learning according to the present invention;
FIG. 19 is a flowchart illustrating updating of initial parameters of a residual neural network in an embodiment of a communication signal modulation identification method based on fusion of a residual neural network and meta-learning according to the present invention;
fig. 20 is a flowchart of a test task in an embodiment of the communication signal modulation identification method based on the fusion of the residual neural network and the meta-learning according to the present invention.
Detailed Description
In order to more clearly illustrate the invention, the invention is further described below with reference to preferred embodiments and the accompanying drawings. It is to be understood by persons skilled in the art that the following detailed description is illustrative and not restrictive, and is not to be taken as limiting the scope of the invention. Unless defined otherwise, technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention belongs.
As shown in fig. 1, the present invention relates to a communication signal modulation identification method based on residual error neural network and meta-learning fusion, firstly, acquiring communication modulation signals with different signal-to-noise ratios under additive white gaussian noise as training objects, dividing the whole data set into a plurality of subtask data sets according to different signal-to-noise ratios of the signals, wherein the identification of the signal under one signal-to-noise ratio corresponds to one subtask, and numbering the subtasks accordingly; secondly, constructing a residual error neural network for signal classification; then, by utilizing a meta-learning theory, obtaining initial parameters of a residual error neural network through learning: taking random sampling data in each subtask data as input of a residual error neural network, and sampling each subtask until all subtasks are used, so that the network learns how to initialize parameters by using the characteristics of signals; thirdly, acquiring a part of communication modulation signals of known labels, which have the same properties as the final test signal, inputting the communication modulation signals into the residual error neural network for training, and completing parameter updating of the residual error neural network; and finally, acquiring a communication modulation signal used for final testing, and inputting the communication modulation signal into the residual error neural network to complete a communication signal modulation identification task. The invention integrates the residual error neural network and the meta-learning method, avoids the phenomenon of 'degeneration' in the training process, improves the generalization capability of the neural network and improves the accuracy of modulation recognition.
The specific implementation process is as follows:
firstly, a training task and test task data set is generated by utilizing Matlab software simulation, the data set comprises 4ASK, 4FSK, 4PSK and 4QAM signals, the signal-to-noise ratio range is-5 dB to 5dB, noise is additive white Gaussian noise, and the expression of communication modulation signals under the additive white Gaussian noise with different signal-to-noise ratios is as follows:
wherein, s [ t ]]And r [ t ]]Respectively representing the transmitted and received signals of the t-th time period, gamma t]For communication channel gain, omega0To shift the carrier frequency, θ0Is the phase offset between the sender and receiver, v [ t ]]Is additive white gaussian noise, and j is an imaginary unit.
The expression for the signal-to-noise ratio is:
wherein, PrIs the power of the signal, PvIs the power of the noise. Fig. 2-5 are graphs of simulated-5 dB time domain waveforms, fig. 6-9 are graphs of simulated 0dB time domain waveforms, and fig. 10-13 are graphs of simulated 5dB time domain waveforms.
And dividing the training task into 11 subtasks according to the number of the noise signal-to-noise ratios, wherein each subtask comprises four generated communication signals. An exploded view of the training subtask is shown in FIG. 14.
Secondly, on the premise that the neural network can be converged, as the depth of the network increases, the performance of the network gradually increases to saturation and then rapidly decreases. This is a "degradation" phenomenon, and this problem makes it difficult to improve the recognition rate of classification problems. And the residual error neural network utilizes the special structure thereof, so that when the neural network propagates forwards, the input signal can directly propagate from any lower layer to a higher layer. The network degradation problem can be solved to a certain extent due to the fact that the network degradation problem is contained in the network degradation mapping method.
Constructing a residual neural network structure for signal classification, wherein the residual neural network consists of 6 residual blocks, 2 full connection layers, 1 Dropout layer and 1 Softmax layer, and the full connection layers integrate the automatically extracted high-level features into local feature information;
the Softmax layer is used for classifying the multiple classes, and the expression is as follows:
wherein, ak'Is the k' th input signal of the output layer, yk'Is the output of the kth' neuron, C is the number of output nodes, i.e., the number of classes classified, acRepresenting the c-th input signal of the output layer.
Each residual block adopts a network structure of a residual block I or a residual block II, and as shown in fig. 15, 6 residual blocks respectively adopt a residual module I, a residual module II and a residual module II in sequence; the first residual block and the second residual block are composed of a convolution layer, two sub-residual blocks and a pooling layer, each sub-residual block comprises 2 convolution layers, all convolution layers comprise 32 convolution kernels, the step length is 1, and the padding is 1; the sizes of convolution kernels of sub residual blocks of the first residual block and the second residual block are different, the size of the convolution kernel of the sub residual block of the first residual block is 3 multiplied by 2, and the size of the convolution kernel of the sub residual block of the second residual block is 3 multiplied by 1; the convolutional layer located outside the sub-residual block is activated using the Relu function; the output of each sub-residual block is also activated by the Relu function and maximally pooled, wherein the convolution layer performs convolution operation on the input to extract the characteristics of the signal, and the expression of the convolution operation is as follows:
xl′＝f(wl′*xl′-1+bl′)
wherein l' represents the number of network layers (the network layers including input layers and convolutional layers), wl′Representing the convolution kernel parameters in the current convolution layer, bl′For the bias vector in the current convolutional layer, f (-) denotes the activation function, xl′-1Output, x, representing the last network layer (the network layer comprising the input layer and the convolutional layer)l′Representing the output of the current convolutional layer.
The pooling layer carries out down-sampling on the convolution layer, and the expression of the pooling layer is as follows:
xl′＝down(xl′-1)
where down (·) represents the pooling function. x is the number ofl′-1Representing the output, x, of the previous convolution layerl′Representing the output of the current pooling layer.
The Relu function expression is:
where f (x) is a function of the maximum value of Relu, and when the input is negative, the output is 0, and the neuron is not activated, x is the input to the function.
The structure of the constructed residual neural network is shown in fig. 15. The two residual block structures in the residual neural network are shown in fig. 16 and 17.
And thirdly, randomly extracting samples from the subtasks and dividing the samples into a training set and a testing set to be used as input of the residual error neural network.
And initializing initial parameters of the residual error neural network randomly, inputting the extracted sample training set into the residual error neural network, and quickly iterating to update the parameters of the residual error neural network.
The step of updating the residual error neural network parameters comprises the following steps:
1) random initialization residual neural network parameter phi0；
2) Inputting the extracted sample training set into a residual error neural network;
3) defining a cross entropy loss function l (phi) by using the extracted sample label and a network output result;
the function is gradiented and the parameters are updated,
where k is the number of times the sample is taken, α is the inner loop learning rate, φkFor the residual neural network parameter at the kth sample draw, l (φ)k) As a cross-entropy loss function of the network parameters,
and the residual error neural network parameters after gradient updating.
The cross entropy loss function is defined as:
l(φ)＝[y(k)logg(x(k))+(1-y(k))log(1-g(x(k)))]
wherein x is(k)For the data of the k-th training set of samples, y(k)The label of the data of the training set of samples is extracted for the first time k, and g represents the residual neural network structure.
The iterative update parameters are to ensure that the residual neural network performs well on the extracted data training set. The conditions for judging whether the performance is good are as follows: when cross entropy loss function
After a period of time of falling, the iteration is continued,
and keeping the stability, and proving that the residual error neural network has good performance on the extracted data training set at the moment.
A flow chart of the residual neural network updating parameters is shown in fig. 18.
And fourthly, inputting the extracted sample test set into the updated residual error neural network to obtain output data corresponding to the test set. Updating the expression of the meta-learning loss function according to the label corresponding to the test set and the output data corresponding to the test set as follows:
wherein,
for the kth sample test set data,
for the kth sample test set data label,
for the updated meta learning loss function after k-1 samples are taken,
and solving a gradient of the meta-learning loss function, and updating initial parameters of the residual neural network by using the gradient. The expression for updating the initial parameters of the residual neural network by using the gradient is as follows:
wherein, beta is the outer ring learning rate, and the inner ring learning rate alpha is larger than the outer ring learning rate beta;
and initial parameters of the updated residual error neural network.
And randomly taking samples for multiple times, and repeating the steps. A flowchart for updating the initial parameters of the residual neural network is shown in fig. 19. And extracting sample data of the subtasks, and dividing the data into a subtask training set and a subtask test set. And inputting the subtask training set into a residual error neural network, and updating parameters of the residual error neural network. And inputting the subtask test set into the residual error neural network after updating the parameters, updating a meta-learning loss function by utilizing the subtask test set label and the output of the subtask test set input into the residual error neural network, and updating the initial parameters of the residual error neural network by calculating the gradient of the meta-learning loss function. And (4) finishing the initialization of the residual error neural network parameters by extracting the data of the subtasks for multiple times. The initial parameters of the residual error neural network are updated, so that the residual error neural network can adapt to signals under different signal-to-noise ratios, effectively learn key characteristics of the signals, and inhibit noise interference to a certain extent. When a new task occurs, the residual neural network can adjust its parameters according to the characteristics of the new task, so that it performs well on the task.
The test flow chart is shown in fig. 20. Firstly, inputting a signal with the same property as the test signal into the residual error neural network, and updating the parameters of the residual error neural network so that the residual error neural network is suitable for the test data. And finally, inputting the test communication signal into the residual error neural network with updated parameters to obtain a final classification result.
Claims (10)
1. The communication signal modulation identification method based on the fusion of the residual error neural network and the meta-learning is characterized by comprising the following steps of:
step 1: acquiring various communication modulation signals with different signal-to-noise ratios under additive white Gaussian noise as a training data set;
step 2: dividing the training data set into N subtask data according to the number of noise-to-noise ratios, wherein the identification of a signal under one signal-to-noise ratio corresponds to one subtask, and the subtasks are numbered according to the magnitude of the signal-to-noise ratio, and each subtask comprises a plurality of communication signals if the number of the signal-to-noise ratio is larger;
and 3, step 3: constructing a residual error neural network structure for signal classification;
and 4, step 4: randomly extracting a small amount of samples which are lower than 10% of subtask data from the lowest-number subtask data in the currently unused subtasks, and dividing the samples into a training set and a test set to be used as input of a residual error neural network;
and 5: initializing initial parameters of a residual error neural network randomly, inputting an extracted sample training set into the residual error neural network, quickly iterating, and updating parameters of the residual error neural network;
step 6: inputting the extracted sample test set into the updated residual error neural network, and updating a meta-learning loss function according to the test set label and the actual output;
and 7: solving a gradient of the meta-learning loss function, and updating the residual neural network parameters by using the gradient;
and 8: repeating the steps 4 to 7 for multiple times until the extraction of the N subtask data is completed;
and step 9: inputting the communication modulation signal with the same property as the final test signal into the residual error neural network, repeating the step 5, and finishing the parameter updating of the residual error neural network;
step 10: and inputting the final test signal into a residual error neural network to obtain a final classification result.
2. The communication signal modulation identification method based on residual error neural network and meta-learning fusion of claim 1, characterized in that: in step 1, the expression of the communication modulation signals with different signal-to-noise ratios under additive white gaussian noise is as follows:
wherein, s [ t ]]And r [ t ]]Respectively representing the transmitted and received signals of the t-th time period, gamma t]For communication channel gain, omega0To shift the carrier frequency, θ0Is the phase offset between the sender and receiver, v [ t ]]Is additive white Gaussian noise, and j is an imaginary number unit;
the expression for the signal-to-noise ratio is:
wherein, PrIs the power of the signal, PvIs the power of the noise.
3. The communication signal modulation identification method based on the fusion of the residual neural network and the meta-learning according to claim 1, characterized in that: in step 2, the number of the noise-to-noise ratios is N, the number of the subtasks is N, each subtask data includes r types of communication signals, each subtask data includes the same type of communication signal, the number of the types of communication signals included in each subtask is the same as the number of the types of communication modulation signals used for final testing, and the number of samples in each subtask is the same.
4. The communication signal modulation identification method based on residual error neural network and meta-learning fusion of claim 1, characterized in that: in step 3, the residual neural network consists of 6 residual blocks, 2 full-link layers, 1 Dropout layer and 1 Softmax layer, and the full-link layers integrate the automatically extracted high-level features into local feature information;
the Softmax layer is used for classifying the multiple classes, and the expression is as follows:
wherein, ak'Is the k' th input signal of the output layer, yk'Is the output of the kth' neuron, C is the number of output nodes, i.e., the number of classes classified, acRepresenting the c-th input signal of the output layer.
5. The communication signal modulation identification method based on residual error neural network and meta-learning fusion of claim 4, characterized in that: each residual block adopts a network structure of a residual block I or a residual block II, the residual block I and the residual block II consist of a convolution layer, two sub-residual blocks and a pooling layer, each sub-residual block comprises 2 convolution layers, all the convolution layers comprise 32 convolution kernels, the step length is 1, and the padding is 1; the sizes of convolution kernels of sub residual blocks of the first residual block and the second residual block are different, the size of the convolution kernel of the sub residual block of the first residual block is 3 multiplied by 2, and the size of the convolution kernel of the sub residual block of the second residual block is 3 multiplied by 1; the convolutional layer located outside the sub-residual block is activated using the Relu function; the output of each sub-residual block is also activated by the Relu function and maximally pooled, wherein the convolution layer performs convolution operation on the input to extract the characteristics of the signal, and the expression of the convolution operation is as follows:
xl′＝f(wl′*xl′-1+bl′)
wherein l' represents the number of network layers including input layers and convolutional layers, wl′Representing the convolution kernel parameters in the current convolution layer, bl′For the bias vector in the current convolutional layer, f (-) denotes the activation function, xl′-1Representing the output of the last network layer, xl′An output representing a current convolutional layer;
the pooling layer carries out down-sampling on the convolution layer, and the expression of the pooling layer is as follows:
xl′＝down(xl′-1)
wherein down (·) represents a pooling function; x is the number ofl′-1Representing the output, x, of the previous convolution layerl′An output representing a current pooling layer;
relu function expression is:
where f (x) is a function of the maximum value of Relu, and when the input is negative, the output is 0, and the neuron is not activated, x is the input to the function.
6. The residual neural network and meta-learning fusion based communication signal modulation identification method of claim 1, whichIs characterized in that: in step 4, the number of samples randomly extracted from the subtask data each time is h + m, the samples include r communication signals, wherein h samples are used as a training set of the extracted samples, and m samples are used as a test set of the extracted samples; input data of the residual error neural network is nin×l×dim×1，ninThe number of input samples of the residual neural network is l, which is the signal sampling length, the signal is in a complex form, and the real part and the imaginary part of the signal are added together, so dim is 2.
7. The communication signal modulation identification method based on residual error neural network and meta-learning fusion of claim 1, characterized in that: in step 5, the step of updating the residual error neural network parameters is as follows:
1) random initialization residual neural network parameter phi0；
2) Inputting the extracted sample training set into a residual error neural network;
3) defining a cross entropy loss function l (phi) by using the extracted sample label and a network output result;
the function is gradiented and the parameters are updated,
wherein k is the number of times of sample extraction, alpha is the inner loop learning rate, phikFor the residual neural network parameter at the kth sample draw, l (φ)k) As a cross-entropy loss function of the network parameters,
the residual error neural network parameters after gradient updating are obtained;
the cross entropy loss function is defined as:
l(φ)＝[y(k)logg(x(k))+(1-y(k))log(1-g(x(k)))]
wherein x is(k)For the data of the k-th training set of samples, y(k)For the kth sample training set data label, g represents the residual neural network structure, l (φ) is the cross entropy lossA function.
8. The communication signal modulation identification method based on residual error neural network and meta-learning fusion of claim 1, characterized in that: in step 6, updating the expression of the meta-learning loss function according to the label corresponding to the test set and the output data corresponding to the test set as follows:
wherein,
for the kth sample test set data draw,
for the kth sample test set data label,
for the updated meta learning loss function after k-1 samples are taken,
9. the communication signal modulation identification method based on residual error neural network and meta-learning fusion of claim 1, characterized in that: step 7, solving a gradient of the meta-learning loss function, and updating initial parameters of the residual error neural network by using the gradient; the expression for updating the initial parameters of the residual neural network by using the gradient is as follows:
wherein, beta is the outer ring learning rate, and the inner ring learning rate alpha is larger than the outer ring learning rate beta;
and initial parameters of the updated residual error neural network.
10. The communication signal modulation identification method based on residual error neural network and meta-learning fusion of claim 1, characterized in that: in step 8, when step 5 is repeated, the parameters of the residual error neural network are not initialized randomly any more, but are initialized to the initial parameters of the network updated by the last iteration
In step 9, when step 5 is repeated, the parameters of the residual error neural network are not initialized randomly any more, but are initialized to the initial parameters of the network updated by the last iteration
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210336623.0A CN114595729A (en) | 2022-04-01 | 2022-04-01 | Communication signal modulation identification method based on residual error neural network and meta-learning fusion |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210336623.0A CN114595729A (en) | 2022-04-01 | 2022-04-01 | Communication signal modulation identification method based on residual error neural network and meta-learning fusion |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114595729A true CN114595729A (en) | 2022-06-07 |
Family
ID=81811858
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210336623.0A Pending CN114595729A (en) | 2022-04-01 | 2022-04-01 | Communication signal modulation identification method based on residual error neural network and meta-learning fusion |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114595729A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115994303A (en) * | 2023-03-24 | 2023-04-21 | 中国人民解放军军事科学院国防科技创新研究院 | Residual neural network model and signal modulation recognition method thereof |
| CN117131416A (en) * | 2023-08-21 | 2023-11-28 | 四川轻化工大学 | Small sample modulation identification method, system, electronic equipment and storage medium |
-
2022
- 2022-04-01 CN CN202210336623.0A patent/CN114595729A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115994303A (en) * | 2023-03-24 | 2023-04-21 | 中国人民解放军军事科学院国防科技创新研究院 | Residual neural network model and signal modulation recognition method thereof |
| CN117131416A (en) * | 2023-08-21 | 2023-11-28 | 四川轻化工大学 | Small sample modulation identification method, system, electronic equipment and storage medium |
| CN117131416B (en) * | 2023-08-21 | 2024-06-04 | 四川轻化工大学 | Small sample modulation identification method, system, electronic equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112418014B (en) | Modulated signal identification method based on wavelet transformation and convolution long-term and short-term memory neural network | |
| CN111314257B (en) | Modulation mode identification method based on complex value neural network | |
| CN110120926B (en) | Communication signal modulation mode identification method based on evolution BP neural network | |
| Hu et al. | Robust modulation classification under uncertain noise condition using recurrent neural network | |
| Chen et al. | Automatic modulation classification scheme based on LSTM with random erasing and attention mechanism | |
| CN114595729A (en) | Communication signal modulation identification method based on residual error neural network and meta-learning fusion | |
| Zhang et al. | A data preprocessing method for automatic modulation classification based on CNN | |
| CN110113288B (en) | Design and demodulation method of OFDM demodulator based on machine learning | |
| CN112039820B (en) | Communication signal modulation and identification method for quantum image group mechanism evolution BP neural network | |
| Zhang et al. | Automatic modulation classification using a deep multi-stream neural network | |
| CN114881092A (en) | Signal modulation identification method based on feature fusion | |
| CN110659684A (en) | Convolutional neural network-based STBC signal identification method | |
| CN114726692B (en) | SERESESESENet-LSTM-based radiation source modulation mode identification method | |
| Zhang et al. | Lightweight automatic modulation classification via progressive differentiable architecture search | |
| CN110233810A (en) | Msk signal demodulation method based on deep learning under a kind of mixed noise | |
| CN114615118B (en) | Modulation identification method based on multi-terminal convolution neural network | |
| Sang et al. | Application of novel architectures for modulation recognition | |
| CN113298031A (en) | Signal modulation identification method considering signal physical and time sequence characteristics and application | |
| Ma et al. | Modulation classification method based on deep learning under non-Gaussian noise | |
| CN113902095A (en) | Automatic modulation identification method, device and system for wireless communication | |
| CN114070684B (en) | Modulation signal identification method based on two-dimensional convolutional neural network | |
| CN113239788A (en) | Mask R-CNN-based wireless communication modulation mode identification method | |
| Kalade et al. | Using sequence to sequence learning for digital bpsk and qpsk demodulation | |
| CN117376066A (en) | Asynchronous sampling-oriented digital signal modulation mode identification method | |
| CN112348165A (en) | Underwater acoustic communication signal classification and identification method and system based on hybrid cycle network |
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 |