Communication modulation identification model training method and apparatus

Abstract

The application discloses a communication modulation identification model training method and device, sensitive feature sets of actually collected communication signals are extracted, the sensitive feature sets are input into a deep neural network model to obtain initial classification results, the signals are input into a naive Bayes classifier, the naive Bayes classifier is initially constructed by using the communication signals, a reward threshold thr is set, a reward function Q(k) is constructed based on the reward threshold thr, the naive Bayes classifier is optimized based on a reward value and a penalty value to obtain a naive Bayes classifier with a weight value, and the weight of the naive Bayes classifier is updated by using a reinforcement learning algorithm. The rough classification results of the deep neural network model are classified again by using the naive Bayes classifier, the naive Bayes classifier is optimized by setting the reward value and the penalty value, the weight of the naive Bayes classifier is continuously updated by using the reinforcement learning algorithm, and the accuracy of the communication modulation identification model in identifying signals is further improved.

Description

TECHNICAL FIELD

[0001] The embodiment of the present application relates to the technical field of signal processing, in particular to a communication modulation identification model training method and device. BACKGROUND

[0002] With the continuous development of wireless communication and radar technology, the application of signal recognition technology is more and more widely. The identification process of communication signal can be divided into three steps: first, pre-processing the signal; second, selecting and extracting some key features; finally, classifying and identifying through constructing a classifier. At present, the recognition and identification of communication signal mainly adopts the discriminant identification method based on maximum likelihood ratio. The discriminant identification method based on maximum likelihood ratio is to transform the identification problem of signal into a hypothesis testing problem, to obtain the feature quantity that can be used for classification by defining the likelihood function of signal and processing it, and then inputting it into the classifier for comparison, so as to obtain the identification result.

[0003] In the process of implementing the present application, the inventors have found the following technical problems: At present, AI-based communication modulation identification is generally used, and many scholars have used ResNet50, CNN, CLDNN and other deep neural networks for communication signal identification. However, in actual engineering application, due to the complex and changeable electromagnetic channel environment, communication signals will experience different degrees of Rayleigh fading, Rice fading, Doppler shift and natural noise influence, etc. in the process of transmission, and with the increase of transmission distance, the received signal amplitude will be very weak, greatly reducing the accuracy of deep learning network model for communication signal identification. SUMMARY

[0004] The embodiment of the present application provides a communication modulation identification model training method and device, which establishes a new model and trains it to solve the technical problem of poor accuracy of deep learning network model for communication signal identification in the prior art.

[0005] In a first aspect, the embodiment of the present application provides a communication modulation identification model training method, comprising:

[0006] extracting a sensitive feature set of the actually collected communication signal;

[0007] inputting the sensitive feature set into a deep neural network model to obtain a preliminary classification result;

[0008] inputting the preliminary classification result into a Naive Bayes classifier, and the Naive Bayes classifier is initially constructed using the communication signal;

[0009] setting a reward threshold thr, and constructing a reward function Q(k) based on the reward threshold thr,

[0010] Ata value is a penalty value, T value is a reward value;

[0011] optimizing the Naive Bayes classifier based on the reward value and the penalty value to obtain a Naive Bayes classifier with a weight value;

[0012] adopting a reinforcement learning algorithm to update the weight of the Naive Bayes classifier.

[0013] Further, the Naive Bayes classifier comprises:

[0014] According to the number of communication signal categories collected, the same number of Naive Bayes classifiers is constructed:

[0015]

[0016]

[0017]

[0018] …

[0019]

[0020] wherein fea1, fea2, fea3, …, feaα are respectively the electromagnetic channel environment of the collected communication signals, a total of α kinds, and P{·} represents the statistical probability of · appearing.

[0021] Further, the Naive Bayes classifier with a weight value comprises:

[0022] Class_β1=β 11 *T value +β 12 *Ata value ,

[0023] Class_β2=β 21 *T value +β 22 *Ata value ,

[0024] Class_β2=β 21 *T value +β 22 *Ata value ,

[0025] …

[0026] Class_β k =β k1 *T value +β k2Ata value

[0027] Class_β={Class_β1,Class_β2,Class_β3,…,Class_β k ,}。

[0028] Further, the extracting the sensitive feature set of the actually collected communication signal comprises:

[0029] extracting a time domain IQ feature of the actually collected communication signal;

[0030] extracting a frequency domain power spectrum feature of the actually collected communication signal;

[0031] extracting a frequency domain phase difference feature of the actually collected communication signal;

[0032] constructing a sensitive feature fusion set based on the time domain IQ feature, the frequency domain power spectrum feature and the frequency domain phase difference feature.

[0033] Further, before the extracting the time domain IQ feature of the actually collected communication signal, the method further comprises:

[0034] preprocessing the collected communication signal according to a channel environment in which the actually collected communication signal is transmitted in an actual engineering environment.

[0035] Still further, the updating the weight of the Naive Bayes classifier by using the reinforcement learning algorithm comprises:

[0036] counting the weight value;

[0037] initializing a Q-Table, wherein a row of the Q-Table corresponds to a reward value weight coefficient, a column of the Q-Table corresponds to a penalty value weight coefficient, and a value in the Q-Table corresponds to a corresponding reward function;

[0038] recursively updating each reward function value in the Q-Table by using a greedy strategy until a reward function value on a main diagonal line in the Q-Table meets a set expected value, wherein the greedy strategy is formulated according to a Naive Bayes multi-round recognition result, an update step length of the reward function value is 1, and an update step length of the weight coefficient is 0.001-0.01.

[0039] In a second aspect, an embodiment of the present application further provides a communication modulation recognition model training device, comprising:

[0040] an extracting module configured to extract a sensitive feature set of an actually collected communication signal;

[0041] a deep neural network model input module configured to input the sensitive feature set into a deep neural network model to obtain a preliminary classification result;

[0042] a Naive Bayes classifier input module, configured to input the initial classification result into a Naive Bayes classifier, wherein the Naive Bayes classifier is initially constructed by using the communication signals;

[0043] a configuration module, configured to set a reward threshold thr, and configure a reward function Q(k) based on the reward threshold thr,

[0044] Ata value is a penalty value, T value is a reward value;

[0045] an optimization module, configured to optimize the Naive Bayes classifier based on the reward value and the penalty value, to obtain a Naive Bayes classifier with a weight value;

[0046] a correction module, configured to update the weight of the Naive Bayes classifier by using a reinforcement learning algorithm.

[0047] Further, the Naive Bayes classifier comprises:

[0048] According to the number of collected communication signal categories, the same number of Naive Bayes classifiers are constructed:

[0049]

[0050]

[0051]

[0052] …

[0053]

[0054] wherein fea1, fea2, fea3, …, feaα are respectively the electromagnetic channel environments of the collected communication signals, and there are a total of α kinds, and P{·} represents the statistical probability of ·.

[0055] Further, the Naive Bayes classifier with a weight value comprises:

[0056] Class_β1=β11*T_value+β12*Ata_value,

[0057] Class_β2=β21*T_value+β22*Ata_value,

[0058] Class_β2=β21*T_value+β22*Ata_value,

[0059] …

[0060] Class_βk=βk1*T_value+βk2*Ata_value

[0061] Class_β={Class_β1,Class_β2,Class_β3,…,Class_βk,}。

[0062] Further, the extraction module comprises:

[0063] a time-domain IQ feature extraction unit configured to extract a time-domain IQ feature of the actually collected communication signal;

[0064] a frequency-domain power spectrum feature extraction unit configured to extract a frequency-domain power spectrum feature of the actually collected communication signal;

[0065] a frequency-domain phase difference feature extraction unit configured to extract a frequency-domain phase difference feature of the actually collected communication signal;

[0066] a construction unit configured to construct a sensitive feature fusion set based on the time-domain IQ feature, the frequency-domain power spectrum feature and the frequency-domain phase difference feature.

[0067] Further, the extraction module further comprises:

[0068] a preprocessing unit configured to pre-process the collected communication signal according to a channel environment in which the actually collected communication signal is transmitted in an actual engineering environment.

[0069] Still further, the correction module comprises:

[0070] a statistical unit configured to statistically process the weight values;

[0071] an initialization unit configured to initialize a Q-Table, wherein rows of the Q-Table correspond to reward value weight coefficients, columns of the Q-Table correspond to penalty value weight coefficients, and values in the Q-Table correspond to corresponding reward functions;

[0072] an updating unit configured to cyclically update each reward function value in the Q-Table by using a greedy strategy until reward function values on a main diagonal line of the Q-Table satisfy a set expected value, wherein the greedy strategy is formulated according to a naive Bayes multi-round recognition result, an updating step length of the reward function values is 1, and an updating step length of the weight coefficients is 0.001-0.01.

[0073] The communication modulation identification model training method and device provided by the embodiment of the present application, by extracting the sensitive feature set of the actually collected communication signal; inputting the sensitive feature set into the deep neural network model to obtain the initial classification result; inputting the initial classification result into the Naive Bayes classifier, the Naive Bayes classifier is initially constructed by using the communication signal; setting the reward threshold thr, and constructing the reward function Q(k) based on the reward threshold thr, optimizing the Naive Bayes classifier based on the reward value and the penalty value, and obtaining the Naive Bayes classifier with the weight value; using the reinforcement learning algorithm to update the weight of the Naive Bayes classifier. The Naive Bayes classifier is used to classify the rough classification result of the deep neural network model again, and the Naive Bayes classifier is optimized by setting the reward value and the penalty value, and the weight of the Naive Bayes classifier is constantly updated by using the reinforcement learning algorithm, which further improves the accuracy of the communication modulation identification model in identifying signals. The communication modulation identification model obtained by the training method can avoid the situation that the signal identification accuracy is reduced due to the complex and changeable electromagnetic channel environment, and further improves the accuracy of the communication signal identification. BRIEF DESCRIPTION OF DRAWINGS

[0074] Other features, objects, and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments thereof as read in conjunction with the accompanying drawings:

[0075] Figure 1 The flowchart of the communication modulation identification model training method provided for the first embodiment of the present application;

[0076] Figure 2 The structural diagram of the communication modulation identification model training device provided for the second embodiment of the present application. DETAILED DESCRIPTION

[0077] The present application will be further described in conjunction with the drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the present application, but not to limit the present application. In addition, it should be noted that, for the convenience of description, only the parts related to the present application are shown in the drawings, but not all the structures.

[0078] Embodiment one

[0079] Figure 1 The flowchart of the communication modulation identification model training method provided for the first embodiment of the present application, the present embodiment can be applied to the training of a communication modulation identification model based on a deep learning neural network model and a Naive Bayes classifier, the method can be executed by a communication modulation identification model training device and can be integrated in a server, and specifically includes the following steps:

[0080] Step 110, extracting the sensitive feature set of the actually collected communication signal.

[0081] In the embodiment, the sensitive feature set can be generated by extracting the actually collected communication signals and performing corresponding feature extraction. The communication signals can be received by an existing communication signal receiving device, and the communication signals can be processed. The sensitive feature set can be generated by extracting the time domain IQ feature of the actually collected communication signals, extracting the frequency domain power spectrum feature of the actually collected communication signals, extracting the frequency domain phase difference feature of the actually collected communication signals, and constructing the sensitive feature fusion set based on the time domain IQ feature, the frequency domain power spectrum feature, and the frequency domain phase difference feature. In addition, the signals can be preprocessed due to the influence of the channel environment of the actual engineering environment transmission. The preprocessing can be performed by collecting a certain amount of k types of communication signals x(t) according to the actual engineering environment and task conditions: x(t) = {x1(t), x2(t), x3(t), …, xk(t)}. The communication signals are evaluated in the channel environment of the actual engineering environment transmission, and the collected signals x(t) are preprocessed, for example, frequency offset correction, filtering, synchronization, signal enhancement, etc., to obtain the signals s(t). k The communication signals are evaluated in the channel environment of the actual engineering environment transmission, and the collected signals x(t) are preprocessed, for example, frequency offset correction, filtering, synchronization, signal enhancement, etc., to obtain the signals s(t).

[0082] First, orthogonal decomposition is performed to obtain the communication IQ signal. The orthogonal decomposition can be performed by the following method to obtain the time domain IQ feature.

[0083] The time domain IQ feature s I (t) is extracted. Q The preprocessed s(t) signal is a complex signal, which is specifically represented as follows:

[0084] s(t) = s I (t) + j s Q (t).

[0085] The real part and the imaginary part of s(t) are extracted to obtain the time domain features s I (t) and s Q (t).

[0086] The frequency domain power spectrum feature Ps(w) can be extracted by the following method:

[0087] Ps(w) = 20*log 10 (|S(w)|),

[0088] S(w) = fft(s(t)), where fft is a fast Fourier transform function, which is used to calculate the frequency spectrum S(w) of s(t), and |S(w)| is the absolute value of the frequency spectrum S(w), which represents the amplitude spectrum of s(t).

[0089] The frequency domain phase difference feature can also be generated by the following method:

[0090] Frequency domain phase difference feature Ψs(t) extraction:

[0091] Ψs(t) = arctan(s Q (t) / s I (t))-arctan(s Q (t-t0) / s I (t-t0)), where t0 is the interval period of sampling, and arctan is the inverse tangent function.

[0092] Based on the time domain IQ features, frequency domain power spectrum features and frequency domain phase difference features calculated above, a sensitive feature fusion set Φ is constructed, Φ = {s I (t), s Q (t), Ps(w), Ψs(t)}.

[0093] Step 120, input the sensitive feature set into a deep neural network model to obtain a preliminary classification result.

[0094] A deep neural network DeepNet is constructed, and hyperparameters (learning rate, epoch, batch size, rand state), AI model training callback functions callbacks, optimizers, etc. are set. Optionally, the deep neural network can use a ResNet50 model, and the output is a communication signal corresponding to the classification result of various signals.

[0095] Step 130, input the preliminary classification result into a Naive Bayes classifier, and the Naive Bayes classifier is initially constructed using the communication signal.

[0096] The Naive Bayes classifier is a series of simple probability classifiers based on the assumption that the features are strongly independent and the use of Bayes' theorem. The classifier model assigns a class label represented by a feature value to a problem instance, and the class label is taken from a limited set. Therefore, the Naive Bayes classifier can be initially constructed based on probability.

[0097] Since the collected communication signals have k classes, k BCBTs are constructed as follows:

[0098]

[0099]

[0100]

[0101] …,

[0102]

[0103] Wherein fea1, fea2, fea3, …, feaα are respectively the electromagnetic channel environment of the collected communication signal, a total of α, P{·} represents the statistical probability of · occurrence

[0104] Step 140, set the reward threshold thr, and construct the reward function Q(k) based on the reward threshold thr.

[0105] In this embodiment, the reward threshold thr can be set according to experience, and the reward function Q(k) is:

[0106] Ata value is a penalty value, T value is a reward value.

[0107] Step 150, based on the reward value and the penalty value, the Naive Bayes classifier is optimized to obtain a weighted Naive Bayes classifier.

[0108] Exemplarily, the weighted Naive Bayes classifier is as follows:

[0109] Class_β1=β 11 *T value +β 12 *Ata value ,

[0110] Class_β2=β 21 *T value +β 22 *Ata value ,

[0111] Class_β2=β 21 *T value +β 22 *Ata value ,

[0112] …,

[0113] Class_β k =β k1 *T value +β k2 *Ata value

[0114] Class_β={Class_β1,Class_β2,Class_β3,…,Class_β k ,}。

[0115] In the above manner, each BCPT of the Naive Bayes classifier can be rewritten in a weighted manner.

[0116] Step 160, updating the weight of the Naive Bayes classifier by using the reinforcement learning algorithm.

[0117] For example, the weight value of the Naive Bayes classifier can be updated in the following way: the weight value is counted; the Q-Table is initialized, the rows of the Q-Table correspond to the reward value weight coefficient, the columns correspond to the penalty value weight coefficient, and the values in the table correspond to the corresponding reward function; each reward function value in the table is updated in a loop by using the greedy strategy until the reward function value on the main diagonal line in the Q-Table meets the set expected value, wherein the greedy strategy is formulated according to the Naive Bayes multi-round recognition result, the update step of the reward function value is 1, and the update step of the weight coefficient is 0.001-0.01.

[0118] The counted weight value β is as follows:

[0119] β={[β 11 ,β 12 ],[β 21 ,β 22 ],[β 31 ,β 32 ],…,[β k1 ,β k2 ]}.

[0120] The Q-Table is initialized, the rows of the Q-Table correspond to the reward value weight coefficient, the columns correspond to the penalty value weight coefficient, and the values in the table correspond to the corresponding reward function, the Q-Table is filled based on the counted weight value β, and the filled Q-Table is as follows:

[0121]

[0122] Each reward function value in the table is updated in a loop by using the greedy strategy until the reward function value on the main diagonal line in the Q-Table meets the set expected value, wherein the greedy strategy is formulated according to the Naive Bayes multi-round recognition result, the update step of the reward function value is 1, and the update step of the weight coefficient is 0.001-0.01. The above method is used for iterative updating, and the appropriate weight value is finally determined. Through the above method, the weight value corresponding to each BCPT is determined, and the training of the entire model is completed.

[0123] In addition, the above method can also be used to identify and test the newly collected signals in the subsequent process, and the reinforcement learning algorithm is optimized according to the test results to continuously correct the weight β of the Naive Bayes classifier, so as to continuously optimize the accuracy of the classifier.

[0124] The embodiment extracts a sensitive feature set of an actually collected communication signal, inputs the sensitive feature set into a deep neural network model to obtain a preliminary classification result, inputs the preliminary classification result into a Naive Bayes classifier, the Naive Bayes classifier is initially constructed by using a communication signal, a reward threshold thr is set, a reward function Q(k) is constructed based on the reward threshold thr, the Naive Bayes classifier is optimized based on a reward value and a penalty value, and a Naive Bayes classifier with a weight value is obtained, and the weight of the Naive Bayes classifier is updated by using a reinforcement learning algorithm. The Naive Bayes classifier is used for re-classification of the rough classification result of the deep neural network model, the Naive Bayes classifier is optimized by setting the reward value and the penalty value, the weight of the Naive Bayes classifier is constantly updated by using the reinforcement learning algorithm, the accuracy of the communication modulation recognition model in identifying signals is further improved, and the communication modulation recognition model obtained by the training method can avoid the situation that the signal recognition accuracy is reduced due to the complex and changeable electromagnetic channel environment, and further improves the accuracy of the communication signal recognition.

[0125] Embodiment two

[0126] Figure 2 is a structural schematic diagram of a communication modulation recognition model training device provided by the embodiment two of the present application, as Figure 2 shown, the device comprises:

[0127] The extraction module 210 is configured to extract a sensitive feature set of an actually collected communication signal.

[0128] The deep neural network model input module 220 is configured to input the sensitive feature set into a deep neural network model to obtain a preliminary classification result.

[0129] The Naive Bayes classifier input module 230 is configured to input the preliminary classification result into a Naive Bayes classifier, and the Naive Bayes classifier is initially constructed by using a communication signal.

[0130] The construction module 240 is configured to set a reward threshold thr, and construct a reward function Q(k) based on the reward threshold thr.

[0131] Ata value is a penalty value, T value is a reward value.

[0132] The optimization module 250 is configured to optimize the Naive Bayes classifier based on the reward value and the penalty value, and obtain a Naive Bayes classifier with a weight value.

[0133] The correction module 260 is configured to update the weight of the Naive Bayes classifier by using a reinforcement learning algorithm.

[0134] The communication modulation identification model training device provided by the embodiment extracts a sensitive feature set of an actually collected communication signal, inputs the sensitive feature set into a deep neural network model to obtain a preliminary classification result, inputs the preliminary classification result into a naive Bayes classifier, the naive Bayes classifier is initially constructed by using a communication signal, a reward threshold thr is set, a reward function Q(k) is constructed based on the reward threshold thr, the naive Bayes classifier is optimized based on a reward value and a penalty value, and a weighted naive Bayes classifier is obtained. The weight of the naive Bayes classifier is updated by using a reinforcement learning algorithm. The rough classification result of the deep neural network model is classified again by using the naive Bayes classifier, the naive Bayes classifier is optimized by setting the reward value and the penalty value, the weight of the naive Bayes classifier is constantly updated by using the reinforcement learning algorithm, the accuracy of the communication modulation identification model in identifying a signal is further improved, and the communication modulation identification model obtained by the training method can avoid the situation that the accuracy of signal identification is reduced due to the complexity and variability of an electromagnetic channel environment, and further improves the accuracy of communication signal identification.

[0135] On the basis of the above-mentioned embodiments, the naive Bayes classifier comprises:

[0136] According to the number of collected communication signal categories, the same number of naive Bayes classifiers is constructed:

[0137]

[0138]

[0139]

[0140] …

[0141]

[0142] Wherein fea1, fea2, fea3, …, feaα are respectively electromagnetic channel environments of collected communication signals, and there are α kinds in total, and P{·} represents the statistical probability of ·.

[0143] On the basis of the above-mentioned embodiments, the weighted naive Bayes classifier comprises:

[0144] Class_β1=β11*T_value+β12*Ata_value,

[0145] Class_β2=β21*T_value+β22*Ata_value,

[0146] Class_β2=β21*T_value+β22*Ata_value,

[0147] …

[0148] Class_βk=βk1*T_value+βk2*Ata_value

[0149] Class_β={Class_β1,Class_β2,Class_β3,…,Class_βk,}。

[0150] On the basis of each of the above embodiments, the extraction module comprises:

[0151] a time-domain IQ feature extraction unit configured to extract a time-domain IQ feature of the actually collected communication signal;

[0152] a frequency-domain power spectrum feature extraction unit configured to extract a frequency-domain power spectrum feature of the actually collected communication signal;

[0153] a frequency-domain phase difference feature extraction unit configured to extract a frequency-domain phase difference feature of the actually collected communication signal;

[0154] a construction unit configured to construct a sensitive feature fusion set based on the time-domain IQ feature, the frequency-domain power spectrum feature, and the frequency-domain phase difference feature.

[0155] On the basis of each of the above embodiments, the extraction module further comprises:

[0156] a preprocessing unit configured to preprocess the collected communication signal according to a channel environment in which the actually collected communication signal is transmitted in an actual engineering environment.

[0157] On the basis of each of the above embodiments, the correction module comprises:

[0158] a statistics unit configured to perform statistics on the weight values;

[0159] an initialization unit configured to initialize a Q-Table, wherein rows of the Q-Table correspond to reward value weight coefficients, columns of the Q-Table correspond to penalty value weight coefficients, and values in the Q-Table correspond to corresponding reward functions;

[0160] an updating unit configured to cyclically update each reward function value in the Q-Table by using a greedy strategy until reward function values on a main diagonal line of the Q-Table satisfy a set expected value, wherein the greedy strategy is formulated according to a naive Bayes multi-round recognition result, an updating step length of the reward function values is 1, and an updating step length of the weight coefficients is 0.001-0.01.

[0161] The communication modulation identification model training device provided by the embodiment of the present application can execute the communication modulation identification model training method provided by any embodiment of the present application, and has the corresponding function modules and beneficial effects of the execution method.

[0162] Computer program code for carrying out operations of the present application can be written in any combination of one or more programming languages, including an object oriented programming language such as Java, Smalltalk, C++ or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code can execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider).

[0163] Note that the above only describes the preferred embodiments of the present application and the applied technical principles. Those skilled in the art will understand that the present application is not limited to the specific embodiments described herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the scope of the present application. Therefore, although the present application has been described in detail through the above embodiments, the present application is not limited to the above embodiments, and can include more other equivalent embodiments without departing from the concept of the present application, and the scope of the present application is determined by the scope of the appended claims.

Claims

1. A method for training a communication modulation recognition model, characterized in that, include: Extract the sensitive feature set of the actual acquired communication signals; The sensitive feature set is input into a deep neural network model to obtain the initial classification result; The initial classification result is input into the Naive Bayes classifier, which is initially constructed using communication signals; Set a reward threshold thr, and construct a reward function Q(k) based on the reward threshold thr. Ata value T is the penalty value. value As a reward value; The Naive Bayes classifier is optimized based on the reward and penalty values ​​to obtain a Naive Bayes classifier with weights. The weights of the Naive Bayes classifier are updated using a reinforcement learning algorithm. The Naive Bayes classifier includes: Based on the number of communication signal categories collected, construct the same number of Naive Bayes classifiers: Where fea1, fea2, fea3, ..., feaα represent the electromagnetic channel environments for collecting communication signals, totaling α types, and P{·} represents the statistical probability of the occurrence of ·.

2. The method according to claim 1, characterized in that, The weighted Naive Bayes classifier includes: Class_β1=β 11 *T value +b 12 *Ata value , Class_β2=β 21 *T value +b 22 *Ata value , Class_β3=β 31 *T value +b 32 *Ata value , …， Class_b k =b k1 *T value +b k2 *Ata value Class_β={Class_β1,Class_β2,Class_β3,…,Class_β k ,}。 3. The method according to claim 1, characterized in that, The sensitive feature set extracted from the actual acquired communication signals includes: Extract the time-domain IQ features of the actual acquired communication signals; Extract the frequency domain power spectrum characteristics of the actual acquired communication signal; Extract the frequency domain phase difference characteristics of the actual acquired communication signals; A sensitive feature fusion set is constructed based on time-domain IQ features, frequency-domain power spectrum features, and frequency-domain phase difference features.

4. The method according to claim 3, characterized in that, Before extracting the time-domain IQ features of the actual acquired communication signal, the method further includes: Based on the channel environment in which the actual acquired communication signals are transmitted in the actual engineering environment, the acquired communication signals are preprocessed.

5. The method according to claim 2, characterized in that, The step of using a reinforcement learning algorithm to update the weights of the Naive Bayes classifier includes: Statistical analysis of the weight values; Initialize the Q-Table, where rows correspond to reward value weight coefficients, columns correspond to penalty value weight coefficients, and values ​​in the table correspond to the corresponding reward functions; A greedy strategy is used to iteratively update each reward function value in the table until the reward function values ​​on the main diagonal of the Q-Table meet the set expected value. The greedy strategy is formulated based on the results of multiple rounds of Naive Bayes recognition. The update step size of the reward function value is 1, and the update step size of the weight coefficient is 0.001 to 0.

01.

6. A communication modulation recognition model training device, characterized in that, include: The extraction module is used to extract the sensitive feature set of the actual acquired communication signals; The deep neural network model input module is used to input the sensitive feature set into the deep neural network model to obtain the initial classification result; The Naive Bayes classifier input module is used to input the initial classification result into the Naive Bayes classifier, which is initially constructed using communication signals. The module is used to set the reward threshold thr and construct the reward function Q(k) based on the reward threshold thr. Ata value T is the penalty value. value As a reward value; An optimization module is used to optimize the Naive Bayes classifier based on reward and penalty values ​​to obtain a Naive Bayes classifier with weights. The correction module is used to update the weights of the Naive Bayes classifier using a reinforcement learning algorithm; The Naive Bayes classifier includes: Based on the number of communication signal categories collected, construct the same number of Naive Bayes classifiers: Where fea1, fea2, fea3, ..., feaα represent the electromagnetic channel environments for collecting communication signals, totaling α types, and P{·} represents the statistical probability of the occurrence of ·.

7. The apparatus according to claim 6, characterized in that, The weighted Naive Bayes classifier includes: Class_β1=β 11 *T value +b 12 *Ata value , Class_β2=β 21 *T value +b 22 *Ata value , Class_β3=β 31 *T value +b 32 *Ata value , …， Class_b k =b k1 *T value +b k2 *Ata value Class_β={Class_β1,Class_β2,Class_β3,…,Class_β k ,}。 8. The apparatus according to claim 6, characterized in that, The extraction module includes: The time-domain IQ feature extraction unit is used to extract the time-domain IQ features of the actual acquired communication signals; The frequency domain power spectrum feature extraction unit is used to extract the frequency domain power spectrum features of the actual acquired communication signals; The frequency domain phase difference feature extraction unit is used to extract the frequency domain phase difference features of the actual acquired communication signal; The building unit is used to construct a sensitive feature fusion set based on time-domain IQ features, frequency-domain power spectrum features, and frequency-domain phase difference features.

Images