Communication station individual identification method, device and equipment based on automatic deep learning

Abstract

The present application relates to a communication station individual identification method, device and equipment based on automatic deep learning, by developing a set of radio station individual identification method using automatic deep learning framework AutoKeras, through its automatic implementation mode significantly reduces the cost of manual participation, simplifies the learning and identification workflow, effectively greatly improves the performance of communication station individual identification. Through experiment verification, compared with traditional neural network, more suitable neural architecture for communication station individual identification can be obtained, so that more representative radio frequency features can be effectively extracted, and finally higher communication station individual identification accuracy is realized, and greater operation convenience and efficiency advantage can be brought to the communication station identification in practical application.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radiation source target recognition, and relates to a communication radio individual recognition method, device and equipment based on automatic deep learning. BACKGROUND

[0002] Radio individual recognition has important research significance in the communication field. It can obtain important information such as information characteristics, signal formats and target individual characteristics by studying the subtle characteristics of radio signals. These information is crucial for data fusion, deciphering and decoding, pattern and target recognition analysis and other technologies. For a long time, the recognition of communication signal fingerprint characteristics has been of great significance. However, due to the small individual difference of radio, the feature extraction is difficult, and the research on communication radio individual characteristics is still in the preliminary stage, and there is a big gap between the recognition effect and the actual application.

[0003] At present, radio individual recognition has mature recognition methods based on statistics and neural networks, such as taking the measured signal characteristic parameters as pattern vectors, and matching the pattern with the established radar model database to realize the recognition of radar radiation sources. Another method based on statistics is the expert system method, that is, according to the professional knowledge provided by experts to construct the inference rules of radar signal recognition, and to infer and recognize the radar radiation source data. For example, the time domain characteristics of radar radiation source signals such as instantaneous amplitude, instantaneous frequency and instantaneous phase are extracted, and it is found that these characteristics are rich in individual difference information and can be used as unintentional features for recognition. For example, wavelet transform is used to extract the individual characteristics of communication radiation sources. This method can capture the local changes and nonlinear characteristics of the signal. In recent years, deep learning methods have been widely used in radiation source individual recognition. These methods realize recognition through signal preprocessing, feature extraction and classification, and use models such as convolutional neural network (CNN) and deep belief network (DBN). The above recognition methods based on traditional neural network architecture are mostly based on image structure, have poor adaptability and complex model parameter optimization, and have the technical problem of insufficient radio individual recognition accuracy. SUMMARY

[0004] In view of the problems existing in the above-mentioned traditional method, the application provides a communication radio individual recognition method based on automatic deep learning, a communication radio individual recognition device based on automatic deep learning and a computer device, which can greatly improve the radio individual recognition accuracy.

[0005] In order to achieve the above-mentioned purpose, the embodiments of the application adopt the following technical solutions:

[0006] On the one hand, a communication radio individual recognition method based on automatic deep learning is provided, comprising the steps of:

[0007] Obtaining radio signal data;

[0008] Data preprocessing is performed on the obtained radio signal data to generate a data set; the data set includes a training set and a test set;

[0009] An open-source AutoKeras architecture is used for network building processing to obtain an automatically constructed deep learning neural network model; the network building processing includes defining a supervised learning channel, neural architecture search, neural network verification, and feature extraction;

[0010] The automatically constructed deep learning neural network model is trained on the training set;

[0011] Bayesian optimization is used to optimize the hyperparameters of the automatically constructed deep learning neural network model;

[0012] After model evaluation of the automatically constructed deep learning neural network model on the test set, the corresponding communication radio individual of the radio signal data is detected and output.

[0013] In another aspect, a communication radio individual identification device based on automatic deep learning is also provided, comprising:

[0014] A data acquisition module is configured to obtain radio signal data;

[0015] A preprocessing module is configured to perform data preprocessing on the obtained radio signal data to generate a data set; the data set includes a training set and a test set;

[0016] A network building module is configured to use an open-source AutoKeras architecture for network building processing to obtain an automatically constructed deep learning neural network model; the network building processing includes defining a supervised learning channel, neural architecture search, neural network verification, and feature extraction;

[0017] A network training module is configured to train the automatically constructed deep learning neural network model on the training set;

[0018] A parameter optimization module is configured to use Bayesian optimization to optimize the hyperparameters of the automatically constructed deep learning neural network model;

[0019] An evaluation and output module is configured to detect and output the corresponding communication radio individual of the radio signal data after model evaluation of the automatically constructed deep learning neural network model on the test set.

[0020] In yet another aspect, a computer device is also provided, comprising a memory and a processor; the memory stores a computer program; and the processor implements the steps of the above-mentioned communication radio individual identification method based on automatic deep learning when executing the computer program.

[0021] One of the above technical solutions has the following advantages and beneficial effects:

[0022] The communication radio station individual identification method, device and equipment based on automatic deep learning have the advantages that a method for solving the radio station individual identification problem by using an automatic deep learning framework AutoKeras is developed, the artificial participation cost is significantly reduced by the automatic implementation mode, the learning and identification workflow is simplified, the performance of the communication radio station individual identification is effectively and greatly improved. Compared with the traditional neural network, the neural architecture more suitable for the communication radio station individual identification can be obtained through the experiment verification, the more representative radio frequency features can be effectively extracted, the higher communication radio station individual identification accuracy is finally realized, and greater operation convenience and efficiency advantages can be brought to the communication radio station identification in actual application. BRIEF DESCRIPTION OF DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0024] Figure 1 A flowchart of the communication radio station individual identification method based on automatic deep learning in an embodiment;

[0025] Figure 2 A core flow framework diagram of the automatic deep learning neural network model in an embodiment;

[0026] Figure 3 A time domain image after demodulation in an embodiment;

[0027] Figure 4 A network structure diagram of the automatic deep learning neural network model in an embodiment;

[0028] Figure 5 A confusion matrix diagram in an embodiment;

[0029] Figure 6 A comparison diagram of the accuracy rates of three models after training in an embodiment;

[0030] Figure 7 A visualization result diagram of the network automatic feature extraction and dimension reduction in an embodiment;

[0031] Figure 8 A visualization result diagram of the manual feature extraction and dimension reduction in an embodiment;

[0032] Figure 9A schematic diagram of a module structure of an automatic deep learning-based communication station individual identification device in an embodiment. DETAILED DESCRIPTION

[0033] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not intended to limit the present application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. The terms used in the specification of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application.

[0034] It should be noted that the term "embodiment" mentioned herein means that the specific features, structures or characteristics described in conjunction with the embodiment can be included in at least one embodiment of the present application. The phrase is shown at various places in the specification does not necessarily refer to the same embodiment, nor is it independent or alternative to other embodiments. Those skilled in the art can understand that the embodiments described herein can be combined with other embodiments. The term "and / or" used in the specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes these combinations.

[0035] The embodiments of the present application will be described in detail below with reference to the accompanying drawings of the embodiments of the present application.

[0036] Automated Machine Learning (AutoML) is a concept proposed in recent years, which can automatically implement data preprocessing, feature construction, feature selection, algorithm selection and hyperparameter tuning processes. Compared with traditional neural network architecture, the automatic deep learning model has great advantages.

[0037] Specifically, the automatic deep learning model can automatically clean up data, remove abnormal data and perform preprocessing; it can select and construct appropriate features, automatically explore how to automatically extract, select and convert data features to enhance the performance of the model; automatic neural network structure exploration and optimization, search and evaluation in a large number of existing network structures, similar to the way of building blocks, automatically discover efficient neural network architecture; in the process of automatic hyperparameter optimization, techniques such as Bayesian optimization, grid search and random search can be used to automatically adjust hyperparameters to improve the performance and generalization ability of the model. For some specific scenarios, automatic deep learning can almost automatically implement data cleaning, feature selection, modeling, hyperparameter optimization and model evaluation processes.

[0038] In one embodiment, as shown inFigure 1 As shown, a method for identifying individual communication stations based on automatic deep learning is provided, which may include the following processing steps S10 to S20:

[0039] S10, obtaining radio station signal data;

[0040] S12, performing data preprocessing on the acquired radio signal data to generate a data set; the data set includes a training set and a test set;

[0041] S14 uses the open source AutoKeras architecture to build a network and obtain an automated deep learning neural network model. The network building process includes defining supervised learning channels, neural architecture search, neural network verification, and feature extraction.

[0042] S16, training the automatic deep learning neural network model on the training set;

[0043] S18, Hyperparameter optimization of automated deep learning neural network models using Bayesian optimization;

[0044] S20, after performing model evaluation on the automatic deep learning neural network model on the test set, detect and output the communication radio station individual corresponding to the radio station signal data.

[0045] It can be understood that the open source AutoKeras architecture is used in this embodiment to build and train the neural network, and its core process framework can be as follows: Figure 2 shown.

[0046] First, when using the open source AutoKeras architecture to build the network, we first define the supervised learning pipeline: In this embodiment, creating a supervised learning pipeline means finding a pipeline structure, as well as the algorithms and hyperparameter combinations that make up the pipeline, to minimize the loss. The formula for the loss function is as follows:

[0047]

[0048] in, Indicates that in the training set The pipeline created above In the test set The loss value for validation on represents the weight parameter of the model, represents the model hyperparameters, represents the regularization parameter, k Indicates the number of training sets or test sets.

[0049] For a machine learning pipeline , represents the feature vector, denotes the corresponding target value, . Given a dataset and a loss metric function L , the performance of can be obtained as:

[0050]

[0051] where denotes the pipeline for a given sample .

[0052] After defining the machine learning pipeline, the network to be built is searched for neural architecture (NAS), the first step is to define the search space, determine the range of neural architecture search, which can be set according to the actual search requirements according to the guidance of AutoKeras architecture. The second step is to select the appropriate search strategy, and use the strategy of efficient neural network architecture search (ENAS) through AutoKeras architecture to automatically design the neural network model. ENAS is a technology that improves the efficiency of neural architecture search through parameter sharing. The formula for training shared parameters is as follows:

[0053]

[0054] where m is a model sampled from a probability distribution controlled by parameters , and the right side is the unbiased formula of the gradient, denotes the shared parameters, denotes the architecture parameters, M denotes the number of sampled models, denotes the gradient of the parameter w , and denotes the loss function of the model under the parameter . First, update the weight w , then fix the weight w , so that the expectation is maximized, where is the reward function of the model under the parameter . Repeat the above two steps to make the accuracy of the model on the test set reach the best, and achieve the optimal neural architecture search.

[0055] After the neural architecture search is completed, i.e. the neural network model is built, the neural network model needs to be verified. Robustness is usually used as an evaluation index of the anti-interference performance of the neural network model. Robustness refers to the ability of the neural network model to maintain good performance when facing input data perturbations, noise or outliers. Robustness boundary is usually used to quantify the robustness of the model, and the corresponding formula is as follows:

[0056]

[0057] wherein, represents a sample space, represents a selected sample, represents a model with as a parameter, represents a loss function. The smaller the value, the lower the sensitivity of the neural network model to input perturbations, i.e. the higher the robustness, and the better the network performance. After the neural network model is verified to meet the set performance requirement, it can be used as an automatic deep learning neural network model. The data of the training set is input into the automatic deep learning neural network model to train the model.

[0058] There are usually two types of parameters in a machine learning model. One type of parameter can be estimated from data. The other type of parameter is difficult to estimate from data and needs to be designed and specified by expert experience. The latter is called a hyperparameter, such as a learning rate, a number of layers of a neural network, a number of neurons, a regularization parameter, a selection of an activation function, a selection of an optimization algorithm, and a dropout ratio (which refers to randomly deleting a certain proportion of neurons in a neural network to prevent overfitting and improve the generalization ability of the model).

[0059] The search strategy needs to quickly and accurately find the optimal hyperparameter configuration in the search space using an algorithm. It involves a classic exploration-exploitation trade-off. On the one hand, it is necessary to quickly find a hyperparameter configuration with good performance. On the other hand, it should avoid premature convergence to a suboptimal region.

[0060] The main methods for hyperparameter search optimization in AutoML are grid search and Bayesian optimization. Grid search automatically defines a complete parameter combination grid in the network according to the classification task. For each parameter combination in the grid, cross-validation is used to evaluate the performance of the model. For each hyperparameter combination, the performance index (such as accuracy) is calculated to evaluate the stability and effect of the model. Bayesian optimization uses a statistical model (usually a Gaussian process) to approximate the objective function. Based on the existing hyperparameter configurations and the corresponding performance evaluation results, the proxy model automatically trains in the network, uses historical information to guide the search, and more intelligently selects new hyperparameter combinations.

[0061] In comparison, Bayesian optimization automatic search is more commonly used in automatic deep learning, which regards the objective function as a random process and models it with a Gaussian process. In the optimization process, the Gaussian process is first defined:

[0062]

[0063] where, is the function prior mean, is the kernel function, which can be selected as the square exponential kernel and the radial basis function kernel, and the calculation formula of the kernel function is as follows:

[0064]

[0065] where, l is the length scale parameter.

[0066] By minimizing the negative improvement of the posterior mean NI find the parameters that minimize the objective function x , the corresponding formula is as follows:

[0067]

[0068] where, represents the posterior mean at point x , which is the estimate of the expected value of the objective function at this point, represents the posterior standard deviation at point x , which measures the uncertainty of the objective function at this point, is a hyperparameter that balances exploration and exploitation. The optimization of each hyperparameter can be completed through the above Bayesian optimization, thereby obtaining a more perfect automatic deep learning neural network model, which effectively enhances the effect of individual identification of communication stations.

[0069] Finally, the model evaluation is also carried out: the confusion matrix is a table used to evaluate the performance of a classification model in supervised learning, suitable for classification tasks. It compares the predicted results of the model with the actual class labels to show the classification of the model on each class. The rows of the confusion matrix represent the true labels, and the columns represent the predicted labels.

[0070] In this embodiment, the confusion matrix output by the automatic deep learning neural network model on the test set and the known class label distribution can be used to directly calculate the recall rate of each class label predicted by the model, and the overall accuracy of the model can be obtained. The recall rate refers to the proportion of samples correctly identified by the system among all relevant samples, reflecting the probability of correct identification of each class by the model. The accuracy rate refers to the proportion of correct samples in the total samples, which measures the overall identification performance of the model on the radio station. Further, the calculation formulas of the recall rate and the accuracy rate of each communication radio station are as follows:

[0071]

[0072]

[0073] wherein, TP represents the number of samples i predicted correctly, TN represents the number of other samples predicted as sample i , FN represents the number of samples i predicted as other samples, FP represents the number of other samples predicted correctly.

[0074] The above communication radio station individual identification method based on automatic deep learning develops a method for solving the radio station individual identification problem using the automatic deep learning framework AutoKeras. The automatic implementation significantly reduces the cost of human participation, simplifies the learning and identification workflow, and effectively improves the performance of communication radio station individual identification. Through experiments, it is found that compared with traditional neural networks, the method can obtain a more suitable neural architecture for communication radio station individual identification, thereby effectively extracting more representative radio frequency features, and ultimately achieving higher communication radio station individual identification accuracy. It also brings greater operational convenience and efficiency advantage to the identification of communication radio stations in practical applications.

[0075] In one embodiment, in order to more intuitively show the implementation process and effect of the above communication radio station individual identification method based on automatic deep learning, one experimental example is given. Those skilled in the art can understand that this experimental example is only an illustrative aid and is not the only limitation of the implementation of the communication radio station individual identification method based on automatic deep learning.

[0076] Experimental setup: An experimental environment is set up by deploying 4 same models of a certain communication radio station to complete the collection of experimental data sets. The details of the radio station data are shown in Table 1:

[0077] Table 1

[0078]

[0079] The time-domain image obtained after 8PSK demodulation is as shown in Figure 3 FIG. 1, where the horizontal axis X is time, in seconds; the vertical axis Y is amplitude, in volts or dimensionless. The received radio signal is divided into I channel (co-directional) and Q channel (orthogonal) through 8PSK demodulation technology, and the modulus of the two channels is extracted. The IQ two channels are spliced together to form a signal matrix similar to a picture, to facilitate subsequent network processing. In the training stage, 2000 training samples and 400 test samples are divided, and AutoKeras architecture is used for automatic construction and optimization of deep learning network.

[0080] In the process of building an automatic deep learning neural network model using AutoKeras architecture, traditional ResNet (residual) network and CNN network are used as experimental groups. The training set is input into the built neural network, and after updating the network parameters, the test set is used for model verification. The test set output accuracy is compared with that of the automatic deep learning neural network model. The network structure of the automatic deep learning neural network model can be as shown in Figure 4 FIG. 2. As can be seen from the network structure diagram, the automatic deep learning neural network model first performs dimensionality elevation and standardization on the input picture, converts the data into a form that can be processed by the network, and ensures that the input data is in the appropriate numerical range; the RandomFlip (random flip) layer randomly flips the input data with a certain probability during training to enhance data diversity and reduce overfitting; the Reisizing (size adjustment) layer adjusts the size of the input data to meet the input requirements of the neural network model; the Concatenate (connection) layer combines the features output by the previous layer into a tensor for further processing by the subsequent layer; then the residual layer (ResNet-50 network) and the pooling layer are used, and the output tensor is generated through the fully connected Dense layer and the appropriate activation function (such as the relu function), which increases the expression ability of the network, so that the neural network can learn and capture more complex patterns and relationships.

[0081] In addition, in order to save training time and resources, the AutoKeras architecture implements an early stopping strategy, that is, some training tasks are terminated in advance when it is observed that the model performance no longer improves. This strategy can be dynamically adjusted according to the training curve of the model to improve efficiency and performance. The confusion matrix obtained after model training is as shown in Figure 5 FIG. 3, where the rows represent the true labels of the radio stations, and the columns represent the predicted labels of the radio stations by the model.

[0082] The training set is input into the automatic deep learning neural network model, and after several rounds of training, the updated weights are used to verify the test set. The residual network and the convolutional network are built in a similar way. Figure 6For the comparison of test accuracy after training three models, it can be found from the figure that the test accuracy of the CNN network is the lowest, and it can be considered that the ordinary convolutional network is not suitable for the task of individual identification of radio stations; the test accuracy of the residual network Resnet is relatively high, but it is difficult to find suitable hyperparameters through manual parameter tuning; the automatic deep learning neural network model based on AutoKeras architecture automatically returns the most suitable hyperparameters for the task through continuous parameter tuning, so the accuracy is the highest.

[0083] After calculation, the recall rate of each sample in the original data is shown in Table 2:

[0084] Table 2

[0085]

[0086] It can be found from Table 2 that the recall rates of radio station 0 and radio station 1 are 1.00, indicating that the model has well identified all samples of radio station 0 and radio station 1.

[0087] The basic principle of communication station fingerprint identification is to extract the equipment differences of the communication station in the communication signal to complete the identification of the transmitter of the communication radiation source or the target individual associated therewith. In this embodiment, the output of the second-to-last layer of the trained automatic deep learning neural network model is selected as the fingerprint feature, and the dimensionality reduction algorithm t-SNE is used for dimensionality reduction and display. t-SNE is a nonlinear technique for data dimensionality reduction and visualization, which realizes the dimensionality reduction visualization of high-dimensional data by calculating the similarity between each pair of data points. The visualization result of the radio station features extracted by the neural network after dimensionality reduction is shown in Figure 7 The visualization result obtained after dimensionality reduction of another group of radio frequency features based on wavelet matrix by t-SNE is shown in Figure 8 .

[0088] Figure 7 and Figure 8 The two commonly used feature extraction methods are shown in Figure 7 The fingerprint formed by the output features of the second-to-last layer of the neural network based on the trained neural network, Figure 8 Corresponding to the manually extracted radio frequency features. The theory of extracting the output features of the second-to-last layer of the neural network as the fingerprint is based on the hierarchical structure of feature representation and the principle of information coding in the deep learning model. According to the graphical effect, it can be concluded that the features automatically extracted by the network can well identify the slight differences between each radio station, which can be used as the radio frequency fingerprint of each radio station; the manually extracted features represented by wavelet transform lose more original data information, Figure 8 The scattered points are more dispersed and cannot well distinguish each radio station.

[0089] It should be understood that although Figure 1The steps in the flowchart are shown in the order indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. In addition, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0090] In one embodiment, Figure 9 As shown, a device 100 for identifying individual communication stations based on automatic deep learning is also provided, comprising a data acquisition module 11, a preprocessing module 13, a network construction module 15, a network training module 17, a parameter optimization module 19, and an evaluation and output module 21. The data acquisition module 11 is used to acquire station signal data. The preprocessing module 13 is used to preprocess the acquired station signal data to generate a dataset; the dataset includes a training set and a test set. The network construction module 15 is used to perform network construction using the open-source AutoKeras architecture to acquire an automatic deep learning neural network model. The network construction process includes defining supervised learning channels, neural architecture search, neural network verification, and feature extraction. The network training module 17 is used to train the automatic deep learning neural network model on the training set. The parameter optimization module 19 is used to optimize the hyperparameters of the automatic deep learning neural network model using Bayesian optimization. The evaluation and output module 21 is used to evaluate the automatic deep learning neural network model on the test set and then detect and output the individual stations corresponding to the station signal data.

[0091] The aforementioned automatic deep learning-based communication station individual identification device 100 develops a method for solving the problem of individual station identification using the automatic deep learning framework AutoKeras. Its automated implementation significantly reduces manual effort, simplifies the learning and identification workflow, and effectively and significantly improves the performance of individual station identification. Experimental verification has shown that compared to traditional neural networks, this method can achieve a neural architecture more suitable for individual station identification, effectively extracting more representative RF features, ultimately achieving higher accuracy in individual station identification. It also brings greater operational convenience and efficiency advantages to station identification in practical applications.

[0092] In one embodiment, the recall rate of each communication station is:

[0093]

[0094] The accuracy of each communication station is:

[0095]

[0096] wherein, TP represents the number of samples i predicted correctly, TN represents the number of other samples predicted as sample i , FN represents the number of samples i predicted as other samples, FP represents the number of other samples predicted correctly.

[0097] The specific limitations of the automatic deep learning-based communication station individual identification device 100 can be referred to the limitations of the automatic deep learning-based communication station individual identification method in the foregoing, which will not be repeated here. Each module in the automatic deep learning-based communication station individual identification device 100 described above can be realized by software, hardware, and a combination thereof, in whole or in part. The above-mentioned modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the operations corresponding to each of the above modules.

[0098] In one embodiment, a computer device is provided, comprising a memory and a processor, the memory storing a computer program, and the processor implementing the following processing steps when executing the computer program: obtaining radio signal data; performing data preprocessing on the obtained radio signal data to generate a data set; the data set comprising a training set and a test set; performing network building processing using an open-source AutoKeras architecture to obtain a constructed automatic deep learning neural network model; the network building processing comprising defining a supervised learning channel, neural architecture search, neural network verification, and feature extraction; training the automatic deep learning neural network model on the training set; performing hyperparameter optimization on the automatic deep learning neural network model using Bayesian optimization; after model evaluation of the automatic deep learning neural network model on the test set, detecting and outputting the communication station individual corresponding to the radio signal data.

[0099] It can be understood that, in addition to the memory and the processor described above, the computer device described above can also include other software and hardware components not listed in the present specification, which can be determined according to the specific communication identification device model in different application scenarios, and the present specification will not be listed in detail.

[0100] In one embodiment, the processor, when executing the computer program, can also implement the steps or sub-steps added in each embodiment of the automatic deep learning-based communication station individual identification method described above.

[0101] In one embodiment, a computer readable storage medium is provided, and the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the following processing steps: obtaining radio signal data; performing data preprocessing on the obtained radio signal data to generate a data set; the data set includes a training set and a test set; performing network building processing using an open-source AutoKeras architecture to obtain a constructed automatic deep learning neural network model; the network building processing includes defining a supervised learning channel, neural architecture search, neural network verification, and feature extraction; training the automatic deep learning neural network model on the training set; performing hyperparameter optimization on the automatic deep learning neural network model using Bayesian optimization; after model evaluation on the test set, detecting and outputting a communication radio individual corresponding to the radio signal data.

[0102] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware, and the computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, the computer program can include the processes of the above-mentioned embodiments. Any reference to memory, storage, database or other medium used in the embodiments provided by the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus dynamic random access memory (Rambus DRAM, RDRAM for short) and interface dynamic random access memory (DRDRAM).

[0103] Any combination of the technical features of the above embodiments can be made, and in order to make the description concise, all possible combinations of the technical features in the above embodiments are not described, however, as long as the combination of the technical features does not exist, it should be considered as the scope of the present application.

[0104] The above embodiments only express several implementation manners of the present application, and the description is relatively specific and detailed, but cannot be understood as the limitation of the protection scope of the present application. It should be pointed out that, for ordinary skilled in the art, several modifications and improvements can be made without departing from the concept of the present application, and all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A method for identifying individual communication stations based on automatic deep learning, characterized in that: Including steps: Get radio signal data; Performing data preprocessing on the acquired radio signal data to generate a data set; the data set includes a training set and a test set; The open-source AutoKeras framework is used to build a network and obtain an automated deep learning neural network model. The network building process includes defining supervised learning channels, neural architecture search, neural network verification, and feature extraction. Training the automatic deep learning neural network model on the training set; Optimizing hyperparameters of the automated deep learning neural network model using Bayesian optimization; After performing model evaluation on the automatic deep learning neural network model on the test set, detecting and outputting individual communication stations corresponding to the station signal data; When using the open-source AutoKeras architecture for network construction, we first define a supervised learning pipeline. Creating a supervised learning pipeline means finding a pipeline structure, as well as a combination of algorithms and algorithm hyperparameters that minimizes loss. The loss function formula is as follows: ； in, Indicates that in the training set The pipeline created above In the test set The loss value for validation on represents the weight parameter of the model, represents the model hyperparameters, represents the regularization parameter, k Indicates the number of training sets or test sets; For a machine learning pipeline , represents the feature vector, Indicates the corresponding target value, Y represents the set of output labels, , subscript g represents the parameter weight, subscript Denotes a collection of algorithms and neural architectures, subscript Represents a set of hyperparameters; given a dataset And the loss metric function L ,get Performance: ； in, Represents a pipeline For a given sample The predicted output.

2. The method for identifying individual communication stations based on automatic deep learning according to claim 1, characterized in that: The recall rates of each communication station are: ； The accuracy of each communication station is: ； in, TP Represents a sample i The number of correct predictions, TN Indicates that other samples are predicted as samples i The number of FN Indicates that the sample i The number of other samples predicted, FP Indicates the number of correct predictions for other samples.

3. A communication station individual identification device based on automatic deep learning, characterized in that: include: A data acquisition module, used to acquire radio station signal data; A preprocessing module is used to perform data preprocessing on the acquired radio signal data to generate a data set; The data set includes a training set and a test set; A network building module, which uses the open source AutoKeras architecture to build a network and obtain an automated deep learning neural network model. The network building process includes defining supervised learning channels, neural architecture search, neural network verification, and feature extraction. A network training module, configured to train the automatic deep learning neural network model on the training set; A parameter optimization module, configured to optimize hyperparameters of the automatic deep learning neural network model using Bayesian optimization; An evaluation and output module, configured to detect and output individual radio stations corresponding to the radio station signal data after performing a model evaluation on the automatic deep learning neural network model on the test set; When using the open-source AutoKeras architecture for network construction, we first define a supervised learning pipeline. Creating a supervised learning pipeline means finding a pipeline structure, as well as a combination of algorithms and algorithm hyperparameters that minimizes loss. The loss function formula is as follows: ； in, Indicates that in the training set The pipeline created above In the test set The loss value for validation on represents the weight parameter of the model, represents the model hyperparameters, represents the regularization parameter, k Indicates the number of training sets or test sets; For a machine learning pipeline , represents the feature vector, Indicates the corresponding target value, Y represents the set of output labels, , subscript g represents the parameter weight, subscript Denotes a collection of algorithms and neural architectures, subscript Represents a set of hyperparameters; given a dataset And the loss metric function L ,get Performance: ； in, Represents a pipeline For a given sample The predicted output.

4. The communication station individual identification device based on automatic deep learning according to claim 3, characterized in that: The recall rates of each communication station are: ； The accuracy of each communication station is: ； in, TP Represents a sample i The number of correct predictions, TN Indicates that other samples are predicted as samples i The number of FN Indicates that the sample i The number of other samples predicted, FP Indicates the number of correct predictions for other samples.

5. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the communication station individual identification method based on automatic deep learning according to claim 1 or 2 are implemented.

Images