CN114118339A - Radio modulation signal identification and classification method based on cuckoo algorithm improved ResNet - Google Patents
Radio modulation signal identification and classification method based on cuckoo algorithm improved ResNet Download PDFInfo
- Publication number
- CN114118339A CN114118339A CN202111337282.0A CN202111337282A CN114118339A CN 114118339 A CN114118339 A CN 114118339A CN 202111337282 A CN202111337282 A CN 202111337282A CN 114118339 A CN114118339 A CN 114118339A
- Authority
- CN
- China
- Prior art keywords
- resnet
- training
- value
- cuckoo
- network
- 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.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 68
- 241000544061 Cuculus canorus Species 0.000 title claims abstract description 60
- 238000012549 training Methods 0.000 claims abstract description 110
- 238000005457 optimization Methods 0.000 claims abstract description 34
- 230000000694 effects Effects 0.000 claims abstract description 17
- 239000011159 matrix material Substances 0.000 claims abstract description 16
- 230000007547 defect Effects 0.000 claims abstract description 4
- 238000012545 processing Methods 0.000 claims abstract description 4
- 235000005770 birds nest Nutrition 0.000 claims description 44
- 235000005765 wild carrot Nutrition 0.000 claims description 44
- 238000012360 testing method Methods 0.000 claims description 39
- 230000008569 process Effects 0.000 claims description 36
- 230000006870 function Effects 0.000 claims description 31
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 claims description 6
- 238000013135 deep learning Methods 0.000 claims description 5
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 3
- 238000004458 analytical method Methods 0.000 claims description 3
- 238000013145 classification model Methods 0.000 claims description 3
- 150000001875 compounds Chemical class 0.000 claims description 3
- 238000011156 evaluation Methods 0.000 claims description 3
- 238000011423 initialization method Methods 0.000 claims description 3
- 230000007246 mechanism Effects 0.000 claims description 3
- 239000002245 particle Substances 0.000 claims description 3
- 238000011176 pooling Methods 0.000 claims description 3
- 238000013461 design Methods 0.000 claims description 2
- 238000004891 communication Methods 0.000 description 3
- 230000006872 improvement Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/004—Artificial life, i.e. computing arrangements simulating life
- G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computational Linguistics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Telephonic Communication Services (AREA)
Abstract
The invention discloses a radio modulation signal recognition and classification method for improving ResNet based on a cuckoo algorithm, which comprises the steps of firstly processing a training data set, designing a ResNet network training module, and building a ResNet training network and a connection mode; secondly, designing a CS optimizing module according to the steps related to the cuckoo algorithm; then, training a radio modulation signal model, and fusing a cuckoo algorithm optimizing module to continuously iterate and optimize to obtain a high-quality solution of the hyper-parameters of the target to be optimized, wherein a residual network improves related parameters according to the high-quality solution; finally, the trained network output accuracy, confusion matrix and the like are used as indexes for identifying and classifying effects, the CS algorithm is adopted to optimize the initial parameter setting, the condition that the network training effect and the final identification and classification accuracy are influenced when the initial value is unreasonable by manual setting through experience in the prior art is changed, the proper initial weight is obtained through iterative optimization, the defects of the traditional ResNet training are effectively improved, and meanwhile, the method has good engineering value.
Description
Technical Field
The invention belongs to the technical field of digital communication, and particularly relates to a radio modulation signal identification and classification method based on bird-brook algorithm improved ResNet.
Background
Nowadays, the demand for signal modulation identification in the military field and the civil field is higher and higher, and especially, the demand for signal modulation rises by one step along with the difficulty of the 5G communication technology.
Modern communication modulation modes are more complex, and modulation signal identification scenes are more diversified, so that how to improve the speed and accuracy of signal modulation identification classification becomes an important problem. Deep learning is used as an identification and classification method which can adapt to various signal modulation types, and the complex process of extracting features in the traditional identification and classification method is effectively solved. However, deep learning faces the problem that the initial value needs to be manually set through experience when the network initializes the hyper-parameters, and the initial value setting is unreasonable, which affects the network training effect and the final recognition and classification accuracy, so how to set the appropriate hyper-parameters as the initial values of the network is an important task at present.
Disclosure of Invention
The invention provides a radio modulation signal identification and classification method based on a cuckoo algorithm improved ResNet. When ResNet is adopted for deep learning of radio modulation signals, if inappropriate initial hyper-parameters are selected, the model convergence speed is slow, an ideal training effect is difficult to obtain, and the final recognition and classification accuracy is influenced. Therefore, the cuckoo CS algorithm is adopted to optimize the initial hyper-parameter setting, and a proper initial value is obtained through iterative optimization, so that the defects of the traditional ResNet training are effectively improved, and the method has good engineering value.
The invention is realized by the following technical scheme:
when a deep learning ResNet network is adopted to train the radio modulation signal identification and classification, the initial hyper-parameter setting of ResNet is optimized by adopting the cuckoo algorithm, and a proper initial value is obtained through iterative optimization, so that the defects of the traditional ResNet training are effectively improved, and the method comprises the following steps:
the method comprises the following steps: processing a training data set of a radio modulation signal, designing a ResNet network training model, and building a ResNet residual training network and a connection mode;
step two: designing a CS optimizing module according to the ResNet network training model and the cuckoo algorithm set up in the first step; performing radio modulation signal model training, continuously iterating and optimizing by a CS optimizing module fused with a cuckoo algorithm to obtain a high-quality solution of the target hyper-parameter to be optimized, and improving the target hyper-parameter to be optimized by a ResNet residual training network according to the high-quality solution;
step three: and the trained ResNet residual training network outputs the accuracy and the confusion matrix of the radio modulation signal recognition classification as indexes of the recognition classification effect.
The further technical scheme comprises the following steps:
the specific process of the step one is as follows:
(1) the 2018.01.osc.0001_1024 dataset of the radio modulation signal was read with the h5py library:
the 2018.01.osc.0001_1024 data set of the radio modulation signal comprises 24 different modulation modes, each modulation mode comprises 26 different signal-to-noise ratios; the data set folder comprises 3 groups, namely an X group, a Y group and a Z group, wherein the X group comprises I and Q modulation signals to be identified and classified, and each modulation signal comprises 1024 points; a one-hot label of a modulation mode corresponding to the sample in the X group is in the Y group; the Z group is a label of the signal-to-noise ratio corresponding to the sample in the X group;
(2) randomly extracting a modulation signal in an X group by using a random function random, simultaneously reading a label of a modulation mode corresponding to a sample in the X group from a Y group, simultaneously reading a label of a signal-to-noise ratio corresponding to the sample in the X group from a Z group, and dividing a 2018.01.osc.0001_1024 data set into small data sets with required data size for analysis and comparison;
(3) reading the partitioned 2018.01.osc.0001_1024 data set, adopting a random mode to extract samples, and dividing the 2018.01.osc.0001_1024 data set into a training set and a test set according to the proportion of 0.7:0.3, wherein 70% of data in the 2018.01.osc.0001_1024 data set is used as the training set, and the rest 30% of data is used as the test set;
(4) designing a ResNet network training model, and building a ResNet residual training network; in a residual block comprising: the system comprises a convolution kernel of 1 multiplied by 1, two residual error units and a maximum pooling layer, wherein the residual error units are residual error blocks in a residual error network; the ResNet residual training network layout comprises an input layer, six residual blocks and three full-connection layers, and is specifically shown in the ResNet residual training network layout of the table 1;
TABLE 1 ResNet residual training network topology
(5) By taking the idea of picture classification as reference, the dimension of a picture is generally width × height × channel, two paths of modulation signals corresponding to radio I and Q are 1024 × 2 × 1, that is, data of the two paths of modulation signals I and Q are defined as "pictures" with the size of 1024 × 2 of a single channel, and the pictures are used as input of a residual error network;
(6) an optimizer of the selected model is an Adam algorithm, so that the robustness is good; selecting a model loss function as a coordinated _ cross entropy loss function, wherein the model loss function is suitable for a multi-classification model; setting an initialization weight mode for a Keras layer as a kernel _ initializer ═ Glorot _ normal', namely a Glorot normal distribution initialization method;
the specific process of the second step is as follows:
designing a CS optimizing model by combining the ResNet network training model in the first step, and performing cuckoo algorithm iterative optimizing design on the learning rate and the Dropout rate initial value of the ResNet network training model:
(1) determining hyper-parameters to be optimized, namely the learning rate and Dropout rate of a ResNet network training model; defining an initial bird nest group as a learning rate and a Dropout rate, and respectively setting a value interval [ bound ] for the learning rate and the Dropout rate1,bound2]To obtain the best training effect, the learning rate and Dropout rate are defined as continuous intervals [ bound1,bound2]Any value within; randomly generating p initial bird nest initial positions
Wherein i is 1,2, …, p, let p be 150; setting relevant parameters including iteration number m and probability P of finding the cuckoo by the host birda∈[0,1]And a termination threshold h;
(2) in order to achieve the best identification and classification effect, the iterative optimization goal of the cuckoo algorithm is set to maximize the accuracy rate test _ acc of the test set, so that an objective function f (X) is defined as the maximum value test _ acc of the modelmaxI.e. by
f(X)＝test_accmax；
(3) Setting an early-stopping mechanism for the ResNet network training model, and stopping training when val _ loss continuous iteration is not reduced any more and the number of times reaches a set value probability is 10 or the total iteration number reaches a set value epoch, namely the epoch is 50;
(4) firstly, selecting a random value as an initial value in a learning rate and Dropout rate value interval by the ResNet network training model, then taking the initial value as an initial hyper-parameter of the ResNet network training model training, carrying out one-round training, outputting train _ loss, train _ acc, test _ loss and test _ acc after the training, and storing the model of the training;
(5) if the accuracy after training reaches a set termination threshold h, namely test _ acc > h, outputting that the optimization target is reached, ending the cuckoo optimization algorithm, and entering the next step; otherwise, carrying out one-time iteration optimization according to the cuckoo optimization module, updating the initial values of the learning rate and the Dropout rate, and carrying out next training by adopting the updated initial values;
(6) and (5) repeating the judgment process, and enabling the ResNet network training model to optimize the test _ acc to the maximum direction through multiple rounds of cuckoo algorithm iteration.
The concrete process of the third step is as follows:
(1) in the process of identifying and classifying, the ResNet network training model judges whether the signal is classified correctly according to the judgment that whether the signal modulation mode of each to-be-identified and classified predicted by the ResNet network training model is consistent with the label of the modulation mode read from the Y group;
(2) since the radio signal modulation identification classification method is a 24-classification problem and belongs to a multi-classification problem, the classification is defined to be separately regarded as positive when the accuracy of each modulation mode is calculated, and the other 23 types are regarded as negative at the moment; the ResNet network training model automatically calculates and outputs the number of TP, FN, FP and TN, wherein TP is the number of positive judgment, FN is the number of negative judgment, FP is the number of positive judgment, and TN is the number of negative judgment;
(3) according to the formula
Calculating the total accuracy under each modulation mode, each signal-to-noise ratio and each modulation mode;
(4) defining each column N of the confusion matrixjkThe number of samples belonging to class k in the reference class for classification into class j, where j, k is 1, 2.. said, 24, since each class to be identified and classified is a uniform class, the total number of samples num under each class is equal, and therefore, for evaluation convenience, the confusion matrix is normalized:
normalized confusion matrix each column confnormjkIs [0,1 ]]A value in between;
(5) and outputting a confusion matrix and an accuracy result as an index for identifying the classification effect.
The process of one-time iteration optimization according to the cuckoo optimization module in the second step (5) is as follows:
a. according to the objective function f (X), judging the objective function value of each bird nest and comparing the objective function values to obtain the optimal value in the p bird nests in the current generation, namely the highest value test _ acc in the test _ accmax；
b. And (b) reserving the optimal bird nest in the process a, and updating the positions of other bird nests according to a cuckoo nest-searching path updating formula, wherein the cuckoo nest-searching path updating formula is as follows:
wherein s represents the random search step length obtained by the Lave flight, and is a Lave distribution random number with a obedient parameter of lambda, wherein u-N (0, sigma)2)，v～N(0,1)，
Beta is 1.5, and the value range of lambda is (1, 3)]，
Denotes the position of the ith bird nest in the t-th generation, where i 1, 2.. and p, t 1, 2.. and m,
is the particle with the best fitness since the algorithm began;
the cuckoo position updating formula is as follows:
in the formula (I), the compound is shown in the specification,
indicates the position of the ith bird nest in the t +1 th generation, i is 1,2, and p is the number of available bird nests,
indicating the position of the ith bird nest in the t generation,
representing point-to-point multiplication, wherein alpha is a step factor, the value alpha is 1, and L (s, lambda) is a random search path;
c. calculating an objective function value of each existing bird nest position, namely test _ acc at the moment, comparing the objective function value with the optimal value recorded in the process a, and updating the optimal value if the objective function value is better than the optimal value, namely a solution with higher test _ acc;
d. generating a random number r, r ∈ [0,1 ]]And is associated with the probability PaMaking a comparison if r > PaIf not, the position of the bird nest after the execution of the process c is kept unchanged, thereby obtaining the position of the bird nest of the next generation
e. If the iteration number m is reached or the termination threshold value h is reached, executing the next step, otherwise, turning to the process a and repeating the iteration process;
f. and outputting a final optimal bird nest position, namely a global optimal solution, wherein the optimal bird nest position is a value obtained by performing one-time iterative optimization through the cuckoo optimization module, updating the initial values of the learning rate and the Dropout rate, and performing next training by adopting the value.
The method has the advantages that the initial hyper-parameters of the residual error network are optimized by adopting the cuckoo algorithm, because the initial values need to be manually set by experience when the network is initialized in the prior art, the network training effect and the final recognition and classification accuracy rate can be influenced when the initial values are unreasonably set, the learning rate and the Dropout rate in the initial hyper-parameters of the residual error network are iteratively optimized by adopting the cuckoo algorithm, and the dynamic parameter adjusting form after the algorithm optimization can better adapt to different network models, so that the training convergence speed and the recognition and classification accuracy rate are improved. The method has good practicability and real-time performance.
Drawings
Fig. 1 is a general flow chart of a system for improving a radio modulation signal identification classification method of ResNet based on a cuckoo algorithm according to the present invention.
Fig. 2 is an example of reading IQ two-path modulation signals of a radio modulation signal identification and classification method based on cuckoo algorithm improvement ResNet provided by the present invention.
Fig. 3 is a flowchart of a one-time iterative optimization process performed by a cuckoo optimization module of the radio modulation signal identification and classification method for improving ResNet based on the cuckoo algorithm provided by the invention.
Fig. 4 is a comparison graph of the recognition accuracy of 4 ten thousand data sets of the radio modulation signal recognition classification method based on the cuckoo algorithm improved ResNet provided by the invention under different signal-to-noise ratios.
Detailed Description
The general flow chart of the system for improving the radio modulation signal identification and classification method of ResNet based on the cuckoo algorithm provided by the invention is shown in FIG. 1.
The method comprises the following steps: processing a training data set of a radio modulation signal, designing a ResNet network training model, and building a ResNet residual training network and a connection mode;
(1) the 2018.01.osc.0001_1024 data set of the radio modulated signal was read using the h5py library. The 2018.01.osc.0001_1024 data set of the radio modulation signal comprises 24 different modulation modes, namely 32PSK, 16ASK, 32QAM, FM, GMSK, 32APSK, 0QPSK, 8ASK, BPSK, 8PSK, AM-SSB-SC, 4ASK, 16PSK, 64APSK, 128QAM, 128APSK, AM-DSB-SC, AM-SSB-WC, 64QAM, QPSK, 256QAM, AM-DSB-WC, OOK and 16QAM, and each modulation mode comprises 26 different signal-to-noise ratios. The data set folder contains 3 groups, which are X, Y and Z groups. The X group comprises I and Q paths of modulation signals to be identified and classified, and each path of modulation signal comprises 1024 points. The Y group is a one-hot label of the modulation scheme corresponding to the sample in the X group. In the Z group is a label of the signal-to-noise ratio corresponding to the samples in the X group.
(2) And randomly extracting a modulation signal in the X group by using a random function random, simultaneously reading a label of a modulation mode corresponding to the sample in the X group from the Y group, simultaneously reading a label of a signal-to-noise ratio corresponding to the sample in the X group from the Z group, and dividing the 2018.01.osc.0001_1024 data set into small data sets with required data size (such as 2 ten thousand data, 4 ten thousand data and 8 ten thousand data … …) for analysis and comparison.
(3) Reading the segmented 2018.01.osc.0001_1024 data set, adopting a random mode to extract samples, and dividing the 2018.01.osc.0001_1024 data set into a training set and a test set according to the proportion of 0.7:0.3, wherein 70% of data in the 2018.01.osc.0001_1024 data set is used as the training set, and the rest 30% of data is used as the test set. Fig. 2 shows an example of one of the IQ two-path modulation signals in the read divided data set.
(4) Designing a ResNet network training model and building a ResNet residual training network. In a residual block comprising: a convolution kernel of 1 × 1, two residual units, namely residual blocks in a residual network, and a max-pooling layer. The ResNet residual training network layout comprises an input layer, six residual modules and three full-connection layers, and is specifically shown in the ResNet residual training network layout of the table 1.
TABLE 1 ResNet residual training network topology
(5) By taking the idea of picture classification as reference, the dimension of a picture is generally width × height × channel, and the two paths of modulation signals corresponding to the I and Q of the radio are 1024 × 2 × 1, that is, the data of the two paths of modulation signals I and Q are defined as "pictures" with the size of 1024 × 2 of a single channel, and the "pictures" are used as the input of the residual error network.
(6) An optimizer of the selected model is an Adam algorithm, and robustness is good. And selecting the model loss function as a coordinated _ cross entropy loss function, which is suitable for the multi-classification model. The method for setting the initialization weight for the Keras layer is kernel _ initializer ═ glort _ normal', namely a Glorot normal distribution initialization method.
Step two: designing a cuckoo CS optimizing module according to the ResNet network training model established in the first step and the related steps of the cuckoo algorithm, wherein a flow chart of the CS optimizing module is shown in FIG. 3. And training a radio modulation signal model, continuously iterating and optimizing by a CS optimizing module fused with a cuckoo algorithm to obtain a high-quality solution of the target hyper-parameter to be optimized, and improving the target hyper-parameter to be optimized by a ResNet residual training network according to the high-quality solution. The specific implementation process is as follows:
(1) and determining the hyper-parameters to be optimized, namely the learning rate and the Dropout rate of the ResNet network training model. Defining an initial bird nest group as a learning rate and a Dropout rate, and respectively setting a value interval [ bound ] for the learning rate and the Dropout rate1,bound2]To obtain the best training effect, the learning rate and Dropout rate may be defined as a continuous interval [ bound1,bound2]Any value within. Randomly generating p initial bird nest initial positions
Where i is 1,2, …, p, let p be 150. Setting relevant parameters including iteration number m and probability P of finding the cuckoo by the host birda∈[0,1]And a termination threshold h.
(2) In order to achieve the best recognition and classification effect, the iterative optimization goal of the cuckoo algorithm is set to maximize the accuracy rate test _ acc of the test set, so that an objective function f (X) is defined as the maximum value test of test _ acc of the ResNet network training model_accmaxI.e. by
f(X)＝test_accmax。
(3) And setting an early-stopping mechanism for the ResNet network training model, and stopping training when the number of continuous iterations of val _ loss is not reduced any more and reaches a set value probability equal to 10 or the total number of iterations reaches a set value epoch, namely epoch equal to 50.
(4) The ResNet network training model firstly selects a random value as an initial value in the value range of the learning rate and the Dropout rate, then takes the initial value as an initial hyper-parameter of the ResNet network training model training, carries out one round of training, outputs train _ loss, train _ acc, test _ loss and test _ acc after the training, and stores the model of the training.
(5) If the accuracy after training reaches a set termination threshold h, namely test _ acc > h, outputting that the optimization target is reached, ending the cuckoo optimization algorithm, and entering the next step; otherwise, carrying out one-time iteration optimization according to the cuckoo optimization module, updating the initial values of the learning rate and the Dropout rate, and carrying out next training by adopting the updated initial values.
The process of one iteration optimization according to the cuckoo optimization module in the process (5) is as follows:
a. and judging the objective function value of each bird nest according to the objective function f (X), and comparing to obtain the optimal value in p bird nests in the current generation, namely the highest value in test _ acc.
b. And (c) reserving the optimal bird nest in the process a, and updating the positions of other bird nests according to an updating formula, wherein the brooding path of the cuckoo is updated according to the formula.
Wherein s represents the random search step length obtained by the Lave flight, and is a Lave distribution random number with a obedient parameter of lambda, wherein u-N (0, sigma)2)，v～N(0,1)，
Beta is 1.5, and the value range of lambda is (1, 3)]，
Denotes the position of the ith bird nest in the t-th generation, where i 1, 2.. and p, t 1, 2.. and m,
is the particle with the best fitness since the algorithm began.
The cuckoo position updating formula is as follows:
in the formula (I), the compound is shown in the specification,
indicates the position of the ith bird nest in the t +1 th generation, i is 1,2, and p is the number of available bird nests,
indicating the position of the ith bird nest in the t generation,
and expressing point-to-point multiplication, wherein alpha is a step factor, the value alpha is 1, and L (s, lambda) is a random search path.
c. And (c) calculating an objective function value of each existing bird nest position, namely the test _ acc at the moment, comparing the objective function value with the optimal value recorded in the process a, and updating the optimal value if the objective function value is better than the optimal value, namely the solution of the test _ acc is higher.
d. Generating a random number r, r ∈ [0,1 ]]And is associated with the probability PaMaking a comparison if r > PaIf not, the position of the bird nest after the execution of the process c is kept unchanged, thereby obtaining the position of the bird nest of the next generation
e. And if the iteration times m or the termination threshold value h are reached, executing the next step, otherwise, turning to the process a and repeating the iteration process.
f. And outputting a final optimal bird nest position, namely a global optimal solution, wherein the optimal bird nest position is a value obtained by performing one-time iterative optimization through the cuckoo optimization module, updating the initial values of the learning rate and the Dropout rate, and performing next training by adopting the value.
(6) Repeating the judgment process in the process (5), and enabling the ResNet network training model to optimize test _ acc to the maximum direction through multi-round cuckoo algorithm iteration;
step three: and outputting the accuracy and the confusion matrix by the trained residual error network as an index for identifying the classification effect.
(1) And in the process of identifying and classifying, the ResNet network training model judges whether the signal is classified correctly according to the judgment that whether the signal modulation mode of each to-be-identified and classified predicted by the ResNet network training model is consistent with the label of the modulation mode read from the Y group.
(2) Since the radio signal modulation identification classification method is a 24-classification problem and belongs to a multi-classification problem, the classification is defined to be regarded as positive independently when the accuracy of each modulation mode is calculated, and the other 23 types are regarded as negative at this time. The ResNet network training model automatically calculates and outputs the number of TP, FN, FP and TN, wherein TP is the number of positive judgment, FN is the number of negative judgment, FP is the number of positive judgment, and TN is the number of negative judgment.
(3) According to the formula
And calculating the total accuracy under each modulation mode, each signal-to-noise ratio and each modulation mode.
(4) Defining each column N of the confusion matrixjkThe number of samples belonging to class k in the reference class for being classified into class j, where j, k is 1, 2. Because each category to be identified and classified is a uniform category and the total number of samples num under each category is equal, the confusion matrix is normalized for convenient evaluation:
normalized confusion matrix each column confnormjkIs [0,1 ]]A value in between.
(5) And outputting a confusion matrix and an accuracy result as an index for identifying the classification effect.
As shown in fig. 4, the data volume of a 2018.01.osc.0001_1024 data set after being segmented by the process (2) of the step one is compared with the recognition accuracy of 4 ten thousand small data sets under different signal-to-noise ratios, and it can be seen that under the condition of using the same data set, the recognition and classification accuracy of the improved CS-ResNet network under different signal-to-noise ratios is higher than that of the ResNet network before improvement, and the effectiveness of the method is proved.
Claims (5)
1. The radio modulation signal recognition and classification method based on the Cuckoo algorithm improved ResNet is characterized in that when a deep learning ResNet network is adopted to train the radio modulation signal recognition and classification, the Cuckoo algorithm is adopted to optimize the ResNet initial hyper-parameter setting, a proper initial value is obtained through iterative optimization, and the defects of the traditional ResNet training are effectively improved, and the method comprises the following steps:
the method comprises the following steps: processing a training data set of a radio modulation signal, designing a ResNet network training model, and building a ResNet residual training network and a connection mode;
step two: designing a CS optimizing module according to the ResNet network training model and the cuckoo algorithm set up in the first step; performing radio modulation signal model training, continuously iterating and optimizing by a CS optimizing module fused with a cuckoo algorithm to obtain a high-quality solution of the target hyper-parameter to be optimized, and improving the target hyper-parameter to be optimized by a ResNet residual training network according to the high-quality solution;
step three: and the trained ResNet residual training network outputs the accuracy and the confusion matrix of the radio modulation signal recognition classification as indexes of the recognition classification effect.
2. The method for identifying and classifying radio modulation signals based on the cuckoo algorithm improved ResNet according to claim 1, wherein the specific process of the first step is as follows:
(1) the 2018.01.osc.0001_1024 dataset of the radio modulation signal was read with the h5py library:
the 2018.01.osc.0001_1024 data set of the radio modulation signal comprises 24 different modulation modes, each modulation mode comprises 26 different signal-to-noise ratios; the data set folder comprises 3 groups, namely an X group, a Y group and a Z group, wherein the X group comprises I and Q modulation signals to be identified and classified, and each modulation signal comprises 1024 points; a one-hot label of a modulation mode corresponding to the sample in the X group is in the Y group; the Z group is a label of the signal-to-noise ratio corresponding to the sample in the X group;
(2) randomly extracting a modulation signal in an X group by using a random function random, simultaneously reading a label of a modulation mode corresponding to a sample in the X group from a Y group, simultaneously reading a label of a signal-to-noise ratio corresponding to the sample in the X group from a Z group, and dividing a 2018.01.osc.0001_1024 data set into small data sets with required data size for analysis and comparison;
(3) reading the partitioned 2018.01.osc.0001_1024 data set, adopting a random mode to extract samples, and dividing the 2018.01.osc.0001_1024 data set into a training set and a test set according to the proportion of 0.7:0.3, wherein 70% of data in the 2018.01.osc.0001_1024 data set is used as the training set, and the rest 30% of data is used as the test set;
(4) designing a ResNet network training model, and building a ResNet residual training network; in a residual block comprising: the system comprises a convolution kernel of 1 multiplied by 1, two residual error units and a maximum pooling layer, wherein the residual error units are residual error blocks in a residual error network; the ResNet residual training network layout comprises an input layer, six residual blocks and three full-connection layers, and is specifically shown in the ResNet residual training network layout of the table 1;
TABLE 1 ResNet residual training network topology
(5) By taking the idea of picture classification as reference, the dimension of a picture is generally width × height × channel, two paths of modulation signals corresponding to radio I and Q are 1024 × 2 × 1, that is, data of the two paths of modulation signals I and Q are defined as "pictures" with the size of 1024 × 2 of a single channel, and the pictures are used as input of a residual error network;
(6) an optimizer of the selected model is an Adam algorithm, so that the robustness is good; selecting a model loss function as a coordinated _ cross entropy loss function, wherein the model loss function is suitable for a multi-classification model; the method for setting the initialization weight for the Keras layer is kernel _ initializer ═ glort _ normal', namely a Glorot normal distribution initialization method.
3. The method for identifying and classifying radio modulation signals based on the cuckoo algorithm improved ResNet according to claim 2, wherein the specific process of the second step is as follows:
designing a CS optimizing model by combining the ResNet network training model in the first step, and performing cuckoo algorithm iterative optimizing design on the learning rate and the Dropout rate initial value of the ResNet network training model:
(1) determining hyper-parameters to be optimized, namely the learning rate and Dropout rate of a ResNet network training model; defining an initial bird nest group as a learning rate and a Dropout rate, and respectively setting a value interval [ bound ] for the learning rate and the Dropout rate1,bound2]To obtain the best training effect, defining learningThe ratio and Dropout rate are taken as a continuum [ bound1,bound2]Any value within; randomly generating initial positions X of p initial bird nestsi 1Wherein i is 1,2, …, p, let p be 150; setting relevant parameters including iteration number m and probability P of finding the cuckoo by the host birda∈[0,1]And a termination threshold h;
(2) in order to achieve the best identification and classification effect, the iterative optimization goal of the cuckoo algorithm is set to maximize the accuracy rate test _ acc of the test set, so that an objective function f (X) is defined as the maximum value test _ acc of the modelmaxI.e. by
f(X)＝test_accmax；
(3) Setting an early-stopping mechanism for the ResNet network training model, and stopping training when val _ loss continuous iteration is not reduced any more and the number of times reaches a set value probability is 10 or the total iteration number reaches a set value epoch, namely the epoch is 50;
(4) firstly, selecting a random value as an initial value in a learning rate and Dropout rate value interval by the ResNet network training model, then taking the initial value as an initial hyper-parameter of the ResNet network training model training, carrying out one-round training, outputting train _ loss, train _ acc, test _ loss and test _ acc after the training, and storing the model of the training;
(5) if the accuracy after training reaches a set termination threshold h, namely test _ acc > h, outputting that the optimization target is reached, ending the cuckoo optimization algorithm, and entering the next step; otherwise, carrying out one-time iteration optimization according to the cuckoo optimization module, updating the initial values of the learning rate and the Dropout rate, and carrying out next training by adopting the updated initial values;
(6) and (5) repeating the judgment process, and enabling the ResNet network training model to optimize the test _ acc to the maximum direction through multiple rounds of cuckoo algorithm iteration.
4. The method for identifying and classifying radio modulation signals based on the cuckoo algorithm improved ResNet according to claim 3, wherein the specific process of the third step is as follows:
(1) in the process of identifying and classifying, the ResNet network training model judges whether the signal is classified correctly according to the judgment that whether the signal modulation mode of each to-be-identified and classified predicted by the ResNet network training model is consistent with the label of the modulation mode read from the Y group;
(2) since the radio signal modulation identification classification method is a 24-classification problem and belongs to a multi-classification problem, the classification is defined to be separately regarded as positive when the accuracy of each modulation mode is calculated, and the other 23 types are regarded as negative at the moment; the ResNet network training model automatically calculates and outputs the number of TP, FN, FP and TN, wherein TP is the number of positive judgment, FN is the number of negative judgment, FP is the number of positive judgment, and TN is the number of negative judgment;
(3) according to the formula
Calculating the total accuracy under each modulation mode, each signal-to-noise ratio and each modulation mode;
(4) defining each column N of the confusion matrixjkThe number of samples belonging to class k in the reference class for classification into class j, where j, k is 1, 2.. said, 24, since each class to be identified and classified is a uniform class, the total number of samples num under each class is equal, and therefore, for evaluation convenience, the confusion matrix is normalized:
normalized confusion matrix each column confnormjkIs [0,1 ]]A value in between;
(5) and outputting a confusion matrix and an accuracy result as an index for identifying the classification effect.
5. The method for identifying and classifying radio modulation signals based on the cuckoo algorithm improved ResNet according to claim 3, wherein the step two process (5) comprises the following steps of performing one-time iterative optimization according to the cuckoo optimization module:
a. according to the objective function f (X), judging the objective function value of each bird nest and comparing the objective function values to obtain the optimal value in the p bird nests in the current generation, namely the highest value test _ acc in the test _ accmax；
b. And (b) reserving the optimal bird nest in the process a, and updating the positions of other bird nests according to a cuckoo nest-searching path updating formula, wherein the cuckoo nest-searching path updating formula is as follows:
wherein s represents the random search step length obtained by the Lave flight, and is a Lave distribution random number with a obedient parameter of lambda, wherein u-N (0, sigma)2)，v～N(0,1)，
Beta is 1.5, and the value range of lambda is (1, 3)]，
Denotes the position of the ith bird nest in the t-th generation, where i 1, 2.. and p, t 1, 2.. and m,
is the particle with the best fitness since the algorithm began;
the cuckoo position updating formula is as follows:
in the formula (I), the compound is shown in the specification,
indicates the position of the ith bird nest in the t +1 th generation, i is 1,2, and p is the number of available bird nests,
indicating the position of the ith bird nest in the t generation,
representing point-to-point multiplication, wherein alpha is a step factor, the value alpha is 1, and L (s, lambda) is a random search path;
c. calculating an objective function value of each existing bird nest position, namely test _ acc at the moment, comparing the objective function value with the optimal value recorded in the process a, and updating the optimal value if the objective function value is better than the optimal value, namely a solution with higher test _ acc;
d. generating a random number r, r ∈ [0,1 ]]And is associated with the probability PaMaking a comparison if r > PaIf not, the position of the bird nest after the execution of the process c is kept unchanged, thereby obtaining the position of the bird nest of the next generation
e. If the iteration number m is reached or the termination threshold value h is reached, executing the next step, otherwise, turning to the process a and repeating the iteration process;
f. and outputting a final optimal bird nest position, namely a global optimal solution, wherein the optimal bird nest position is a value obtained by performing one-time iterative optimization through the cuckoo optimization module, updating the initial values of the learning rate and the Dropout rate, and performing next training by adopting the value.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111337282.0A CN114118339B (en) | 2021-11-12 | 2021-11-12 | Radio modulation signal identification and classification method based on improvement ResNet of cuckoo algorithm |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111337282.0A CN114118339B (en) | 2021-11-12 | 2021-11-12 | Radio modulation signal identification and classification method based on improvement ResNet of cuckoo algorithm |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114118339A true CN114118339A (en) | 2022-03-01 |
| CN114118339B CN114118339B (en) | 2024-05-14 |
Family
ID=80378868
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111337282.0A Active CN114118339B (en) | 2021-11-12 | 2021-11-12 | Radio modulation signal identification and classification method based on improvement ResNet of cuckoo algorithm |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114118339B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115720184A (en) * | 2022-10-08 | 2023-02-28 | 西安电子科技大学 | Small sample signal modulation type identification method based on characteristic distribution |
| CN117236796A (en) * | 2023-11-13 | 2023-12-15 | 天津市城市规划设计研究总院有限公司 | CS-TOPSIS algorithm-based hospital logistics operation and maintenance evaluation method and system |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108694390A (en) * | 2018-05-15 | 2018-10-23 | 南京邮电大学 | A kind of cuckoo search improves the modulated signal sorting technique of grey wolf Support Vector Machines Optimized |
| WO2020073951A1 (en) * | 2018-10-10 | 2020-04-16 | 腾讯科技（深圳）有限公司 | Method and apparatus for training image recognition model, network device, and storage medium |
| WO2021007812A1 (en) * | 2019-07-17 | 2021-01-21 | 深圳大学 | Deep neural network hyperparameter optimization method, electronic device and storage medium |
| CN113240194A (en) * | 2021-06-04 | 2021-08-10 | 河北工业大学 | Energy storage battery capacity prediction method, server and computer readable storage medium |
| CN113378910A (en) * | 2021-06-07 | 2021-09-10 | 浙江工业大学 | Poisoning attack method based on electromagnetic signal modulation type identification of pure tag |
-
2021
- 2021-11-12 CN CN202111337282.0A patent/CN114118339B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108694390A (en) * | 2018-05-15 | 2018-10-23 | 南京邮电大学 | A kind of cuckoo search improves the modulated signal sorting technique of grey wolf Support Vector Machines Optimized |
| WO2020073951A1 (en) * | 2018-10-10 | 2020-04-16 | 腾讯科技（深圳）有限公司 | Method and apparatus for training image recognition model, network device, and storage medium |
| WO2021007812A1 (en) * | 2019-07-17 | 2021-01-21 | 深圳大学 | Deep neural network hyperparameter optimization method, electronic device and storage medium |
| CN113240194A (en) * | 2021-06-04 | 2021-08-10 | 河北工业大学 | Energy storage battery capacity prediction method, server and computer readable storage medium |
| CN113378910A (en) * | 2021-06-07 | 2021-09-10 | 浙江工业大学 | Poisoning attack method based on electromagnetic signal modulation type identification of pure tag |
Non-Patent Citations (2)
| Title |
|---|
| 王宇;杨艺;王宝山;王田;卜旭辉;王传云;: "深度神经网络条件随机场高分辨率遥感图像建筑物分割", 遥感学报, no. 06, 25 November 2019 (2019-11-25) * |
| 谢丽霞;王志华: "基于布谷鸟搜索优化BP神经网络的网络安全态势评估方法", 计算机应用, vol. 37, no. 7, 31 December 2017 (2017-12-31) * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115720184A (en) * | 2022-10-08 | 2023-02-28 | 西安电子科技大学 | Small sample signal modulation type identification method based on characteristic distribution |
| CN115720184B (en) * | 2022-10-08 | 2024-04-19 | 西安电子科技大学 | Small sample signal modulation type identification method based on characteristic distribution |
| CN117236796A (en) * | 2023-11-13 | 2023-12-15 | 天津市城市规划设计研究总院有限公司 | CS-TOPSIS algorithm-based hospital logistics operation and maintenance evaluation method and system |
| CN117236796B (en) * | 2023-11-13 | 2024-02-02 | 天津市城市规划设计研究总院有限公司 | CS-TOPSIS algorithm-based hospital logistics operation and maintenance evaluation method and system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114118339B (en) | 2024-05-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN114118339A (en) | Radio modulation signal identification and classification method based on cuckoo algorithm improved ResNet | |
| CN104966104B (en) | A kind of video classification methods based on Three dimensional convolution neutral net | |
| CN103559504B (en) | Image target category identification method and device | |
| CN106599051B (en) | Automatic image annotation method based on generated image annotation library | |
| CN109978035B (en) | Pedestrian detection method based on improved k-means and loss function | |
| CN113850281B (en) | MEANSHIFT optimization-based data processing method and device | |
| CN112910811B (en) | Blind modulation identification method and device under unknown noise level condition based on joint learning | |
| CN110990784B (en) | Cigarette ventilation rate prediction method based on gradient lifting regression tree | |
| CN110705516B (en) | Sole pattern image clustering method based on collaborative network structure | |
| CN112817755A (en) | Edge cloud cooperative deep learning target detection method based on target tracking acceleration | |
| CN111241939A (en) | Rice yield estimation method based on unmanned aerial vehicle digital image | |
| CN112115967A (en) | Image increment learning method based on data protection | |
| CN110751191A (en) | Image classification method and system | |
| CN114064627A (en) | Knowledge graph link completion method and system for multiple relations | |
| CN115249313A (en) | Image classification method based on meta-module fusion incremental learning | |
| CN114201632A (en) | Label noisy data set amplification method for multi-label target detection task | |
| CN116244612B (en) | HTTP traffic clustering method and device based on self-learning parameter measurement | |
| CN111639712A (en) | Positioning method and system based on density peak clustering and gradient lifting algorithm | |
| CN108846429B (en) | Unsupervised learning-based network space resource automatic classification method and unsupervised learning-based network space resource automatic classification device | |
| CN116363469A (en) | Method, device and system for detecting infrared target with few samples | |
| CN116049712A (en) | Knowledge-assisted multi-classifier fusion ultra-narrow pulse radar target identification method | |
| CN114386697A (en) | Ship main engine spare part prediction method based on improved random forest | |
| CN112132239B (en) | Training method, device, equipment and storage medium | |
| CN113033653A (en) | Edge-cloud collaborative deep neural network model training method | |
| CN114549910A (en) | Noise identification method based on clean data set and key feature detection |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |