Data and knowledge enhancement dual-driven small sample equivalent model acquisition method and system and medium

Abstract

The embodiment of the invention discloses a data and knowledge enhancement dual-driven small sample equivalent model acquisition method and system and a medium. The method comprises the following steps: acquiring a simulation data set from a simulated modulation signal data set, and acquiring a small sample data set from a public data set; training according to the simulation data set to obtain a teacher model; and constructing a mixed loss objective function, and training a student model corresponding to the teacher model based on the trained teacher model, the small sample data set and the mixed loss objective function to obtain a small sample equivalent model. According to the invention, through data and knowledge dual enhancement, a deep learning model capable of effectively identifying the modulation signal under a small sample condition is constructed, and the practicability and wide applicability of a modulation signal identification technology are significantly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wireless communication technology, and particularly relates to a data and knowledge enhanced double-driven small sample equivalent model acquisition method, system and medium. BACKGROUND

[0002] The rapid development of wireless communication technology makes the electromagnetic environment increasingly complex, which brings great challenges to wireless signal monitoring, and the automatic modulation recognition technology of communication signals is particularly key. It can identify the modulation mode of the received signal without any prior information, providing important support for subsequent signal processing and demodulation. In recent years, deep learning-based methods have made significant breakthroughs in image recognition, speech recognition and other fields, and have been gradually introduced into the field of communication, especially showing great potential in automatic modulation recognition. However, the superior performance of these methods depends on large-scale labeled data sets. In practical applications, especially in complex communication and adversarial scenarios, it is difficult to obtain large-scale labeled data sets, which is difficult to meet the needs of existing deep learning models. Therefore, how to effectively use limited labeled samples has become a research hotspot.

[0003] Existing modulation recognition methods under limited labeled samples mainly include data augmentation, semi-supervised learning, transfer learning and small sample learning. The data augmentation-based method can create new training samples by transforming the original data to improve the generalization ability of the model, but its effect is limited in the case of extremely small sample size; the semi-supervised learning method can extract information from unlabeled data and train the model through pseudo-labels, but this method has high requirements for data distribution; the transfer learning method borrows knowledge from the source domain for learning in the target domain, but its effect depends on the similarity of the source domain and the target domain data; the small sample learning method relies on a specific network structure to realize feature learning and classification, but it has poor adaptability to changes in network structure. These methods have their own shortcomings in practical applications, especially in the case of limited sample size, it is difficult to provide ideal model performance. SUMMARY

[0004] Therefore, it is necessary to propose a data and knowledge enhanced double-driven small sample equivalent model acquisition method, system and medium to solve the above problems.

[0005] A data and knowledge enhanced double-driven small sample equivalent model acquisition method, the method comprising: Obtaining a simulation data set from a simulated modulation signal data set, and obtaining a small sample data set from a public data set.

[0006] Training a teacher model according to the simulation data set.

[0007] construct a mixed loss objective function, train a student model corresponding to the teacher model based on the trained teacher model, the small sample dataset and the mixed loss objective function, and obtain a small sample equivalent model.

[0008] Before the constructing the mixed loss objective function, the method further comprises: randomly enhance the small sample dataset by a signal data enhancement method, the signal data enhancement method comprising at least one of signal noise adding, overall transformation, local transformation, signal fusion, signal decomposition and masking, the overall transformation comprising rotation and flipping, and the local transformation comprising segment inversion and cyclic shift.

[0009] The method further comprises: obtaining a simulated modulation signal dataset.

[0010] extract I-channel data and Q-channel data of the modulation signal from the modulation signal dataset, and add labels to the modulation signal.

[0011] determine the mean and standard deviation of the I-channel data and Q-channel data, respectively.

[0012] normalize the I-channel data and Q-channel data based on the mean and standard deviation to obtain a simulated dataset.

[0013] The method further comprises: divide the simulated dataset into a training set and a test set.

[0014] train the teacher model by using the training set, obtain soft probability output and feature output of the teacher model, and update parameters of the teacher model by using a cross-entropy loss function.

[0015] The method further comprises: construct a classification loss based on predicted labels and real labels of the student model.

[0016] construct a distillation loss based on soft probability output of the student model and the trained teacher model.

[0017] construct a feature comparison loss based on feature output of the student model and the trained teacher model.

[0018] construct a mixed loss objective function based on the classification loss, the distillation loss and the feature comparison loss.

[0019] The mixed loss objective function is constructed according to the classification loss, the distillation loss and the feature comparison loss, and specifically includes: According to The mixed loss objective function is constructed, wherein , , is a hyperparameter, is a classification loss, is a real label of a sample, is a predicted label of a student model, is a distillation loss, is a softened probability output of a teacher model, is a softened probability output of a student model, is a feature comparison loss, is a feature output of a student model, is a feature output of a teacher model.

[0020] The student model corresponding to the teacher model is trained based on the trained teacher model, the small sample data set and the mixed loss objective function to obtain a small sample equivalent model, and specifically includes: The enhanced small sample data is input into the student model to obtain a predicted label, a softened probability output and a feature output of the student model.

[0021] The softened probability output and the feature output of the teacher model, and the predicted label, the softened probability output and the feature output of the student model are substituted into the mixed loss objective function to determine a loss value.

[0022] The gradient of the loss value on the student model is calculated by a back propagation algorithm.

[0023] The gradient is used to adjust the parameters of the student model until the loss value reaches a threshold value, and a small sample equivalent model is obtained.

[0024] The method further includes: The test set is input into the small sample equivalent model to obtain a predicted label, and the number of correctly classified samples is counted by comparing the predicted label with a real label.

[0025] The samples in the test set are grouped according to a signal-to-noise ratio, each group corresponds to a signal-to-noise ratio, and the total number of samples of each signal-to-noise ratio is recorded.

[0026] For each sample in the group, the number of correctly classified samples is counted by comparing the predicted label with a real label.

[0027] Determine the average classification accuracy and the highest recognition accuracy evaluation according to the total number of samples and the number of correctly classified samples, and evaluate the performance of the small sample equivalent model through the average classification accuracy and the highest recognition accuracy.

[0028] A data and knowledge enhanced double-driven small sample equivalent model acquisition system, comprising: A simulation data set and small sample data set acquisition module is configured to acquire a simulation data set from a simulated modulated signal data set and a small sample data set from a public data set.

[0029] A training module is configured to train a teacher model according to the simulation data set.

[0030] A small sample equivalent model acquisition module is configured to construct a hybrid loss objective function, train a student model corresponding to the teacher model based on the trained teacher model, the small sample data set and the hybrid loss objective function, and obtain a small sample equivalent model.

[0031] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to make the processor execute the steps of the method.

[0032] The embodiments of the present application have the following beneficial effects: The present application utilizes the simulation data set in the simulation data set and the small sample data in the public data set, effectively improves the performance of the small sample equivalent model through a double driving mechanism. The teacher model is trained according to the simulation data set, the trained teacher model is used as a knowledge source, and the knowledge is transmitted to the student model to guide the training of the student model. Further, a hybrid loss objective function is constructed based on the trained teacher model and the student model, and the learning process of the student model is systematically guided through three dimensions of classification loss, distillation loss and feature comparison loss, so that the student model can better generalize on the small sample data set. The student model is trained according to the small sample data set, and the model parameters of the student model are updated using the hybrid loss objective function to obtain a small sample equivalent model. In summary, the present application enhances the data and knowledge, and constructs a deep learning model that can effectively recognize the modulated signal under the condition of small sample, which significantly improves the practicality and wide applicability of the modulated signal recognition technology. BRIEF DESCRIPTION OF DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiment or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0034] in: Figure 1 A flowchart illustrating an embodiment of a method for obtaining a small sample equivalence model driven by both data and knowledge enhancement provided by the present invention; Figure 2 A flowchart illustrating another embodiment of the method for obtaining a small sample equivalence model driven by both data and knowledge enhancement provided by the present invention; Figure 3 This is a flowchart illustrating another embodiment of the method for obtaining a small sample equivalence model driven by both data and knowledge enhancement provided by the present invention. Figure 4 A schematic diagram of an embodiment of a small sample equivalence model acquisition system driven by both data and knowledge enhancement provided by the present invention; Figure 5 A schematic diagram of the structure of an embodiment of the medium provided by the present invention. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] like Figure 1 As shown, Figure 1 This is a flowchart illustrating an embodiment of a method for obtaining a small-sample equivalence model driven by both data and knowledge enhancement, as provided by the present invention. The method includes: S101: Obtain the simulation dataset from the simulated modulation signal dataset, and obtain the small sample dataset from the public dataset.

[0037] For example, first, the simulated modulation signal dataset is obtained.

[0038] The simulated modulation signal dataset was obtained by Matlab 2020a simulation. There are 11 modulation categories with signal-to-noise ratios ranging from -20dB to 18dB. Each category has 1000 samples. The details are shown in Table 1. Table 1 is the simulated modulation signal dataset.

[0039] Table 1. Simulated Modulation Signal Dataset

[0040] The modulation signal can be represented in complex baseband form as follows:

[0041] wherein, and denote the in-phase and quadrature components of the signal, is a virtual unit. By sampling the received signal of length , the signal can be represented as a discrete sequence, where L = 128.

[0042] Further, the I and Q channel data of the modulated signal are extracted, as shown in the following formula: ; The modulated signal is labeled to construct a data set, as shown in the following formula: ; wherein, is the th modulated signal, is the corresponding label, and the number of samples is N.

[0043] Further, the mean and standard deviation of the I and Q channel data are calculated, the extracted I and Q channel data are normalized to eliminate the amplitude difference and dimension influence of non-modulation essence, and the normalized modulated signal is obtained to obtain the simulation data set, as shown in the following formula: ; wherein, and are the mean and standard deviation of the th modulated signal.

[0044] Meanwhile, the public data set RML2016.10a is disclosed. The data set is composed of modulated signal examples synthesized by GNU Radio. The data set has 11 modulation categories, including 8 digital modulations and 3 analog modulations, and the specific modulation types are consistent with the types of the simulated data set. The signal-to-noise ratio ranges from -20 dB to 18 dB with an interval of 2 dB. The original data set is stratified sampled, and 10 samples are selected for each signal-to-noise ratio in each category as a training set, and 200 samples are selected as a test set. The total number of small sample signal data actually used for network training is 2200.

[0045] S102: Training the teacher model according to the simulation data set.

[0046] Exemplarily, the simulation data set is divided into a training set and a test set according to a ratio of 4:1. The teacher model is trained through the training set, the soft probability output and the feature output of the teacher model are obtained, and the parameter update of the teacher model is performed by using a cross-entropy loss function.

[0047] The teacher model is trained according to a simulation data set, and the teacher model uses VGG16. The VGG16 network comprises convolution blocks, pooling layers and full connection layers; wherein the convolution block comprises a convolution layer, a batch normalization layer and an activation layer, different convolution blocks have similar structures but different specific parameters. The learning rate of the network is dynamically adjusted by adopting a fixed step attenuation in the process of training the predetermined network.

[0048] Specifically, the teacher model takes the simulation data set as input, and outputs the prediction label, the softened probability output and the feature output of the teacher model, wherein the softened probability output and the feature output of the teacher model provide a “supervision signal” for knowledge distillation of the student model.

[0049] Moreover, the parameter updating of the teacher model is performed by using a cross-entropy loss function, and the cross-entropy loss function is shown in the following formula: ; wherein, is the prediction label of the teacher model, is the parameter of the teacher model. The model parameter is adjusted by a back propagation algorithm, M is the total number of training samples, and the optimization target is represented as: ; The prediction label of the teacher model is substituted into the cross-entropy loss function to determine the loss value. The gradient of the loss value to the teacher model is calculated by the back propagation algorithm, the teacher model parameter is adjusted by using the gradient, until the loss value is less than a threshold value, and the trained teacher model is obtained. The specific training parameter is shown in Table 2, and Table 2 is the teacher model parameter.

[0050] Table 2 Teacher model parameter

[0051] After the loss converges, the trained teacher model can be used for modulation type identification.

[0052] S103: A mixed loss objective function is constructed, and a student model corresponding to the teacher model is trained based on the trained teacher model, a small sample data set and the mixed loss objective function, to obtain a small sample equivalent model.

[0053] Exemplarily, for each sample data (modulation signal), the feature outputs of a teacher model (represented as f T (x)) and a student model (represented as f S (x)) are respectively: ; ; wherein, is the sample data, is the feature output of the student model, ​​The feature output of the teacher model.

[0054] Based on the predicted label of the student model and the true label, a classification loss is constructed using a cross-entropy loss function, which is the difference between the predicted label of the student model and the true label, as shown in the following formula: ; Wherein, is the true label of the sample, is the predicted label of the student model, is the cross-entropy loss function.

[0055] Based on the softened probability output of the student model and the trained teacher model, a distillation loss is constructed using KL divergence, which reflects the knowledge transfer effect of the teacher model to the student model, as shown in the following formula: ; Wherein, is the softened probability output of the teacher model, is the softened probability output of the student model, is the KL divergence. The softened output is calculated using the formula .

[0056] Based on the feature output of the student model and the trained teacher model, a feature comparison loss is constructed using a contrastive loss function, and the feature output of each student model must identify the corresponding feature output of the teacher model in all other teacher representations in the batch .

[0057] ; Wherein, is the cosine similarity function, is the temperature parameter, is the bias parameter, is the total number of negative samples. Parameters and provide fine control over this classification process. In network learning, these two parameters are changeable and learnable.

[0058] Finally, a hybrid loss objective function is constructed according to the classification loss, the distillation loss and the feature comparison loss, as shown in the following formula: ; Wherein, , , is a hyperparameter, which is set to = 0.3, = 0.7, = 0.2 in the experiment, is a classification loss, is a true label of a sample, is a predicted label of a student model, is a distillation loss, is a softened probability output of a teacher model, is a softened probability output of a student model, is a feature contrastive loss, is a feature output of a student model, is a feature output of a teacher model.

[0059] Further, the student model is a CLDNN network, and the CLDNN network comprises a convolutional layer, a batch normalization layer, an activation layer, an LSTM layer, and a fully connected layer. During the training of the predetermined network, a learning rate of the network is dynamically adjusted by using fixed step decay.

[0060] Specifically, each sample data in a small sample data set is randomly augmented and input into the CLDNN network; the augmented sample data is input into the student model to obtain a predicted label, a softened probability output, and a feature output of the student model; the softened probability output and the feature output of the teacher model, and the predicted label, the softened probability output, and the feature output of the student model are substituted into a mixed loss objective function to determine a loss value.

[0061] Based on the loss value, the gradient of the loss value with respect to all parameters (such as convolution kernel weights, LSTM weights, and bias terms) of the student model is calculated in reverse from the output layer to the input layer by using the chain rule, that is, the influence degree and direction of each parameter on the loss are quantified.

[0062] An optimizer (Adam) is used to adjust the parameters according to the gradient, and the parameter settings are as follows: the batch size is 200, the total training rounds are 100, the initial learning rate is 1e-2, the weight decay is 5e-4, and the learning rate is dynamically adjusted by using fixed step decay.

[0063] The above process is repeated until the loss value reaches a threshold, at which time the model parameters are optimal parameters, and the corresponding model is a trained small sample equivalent model.

[0064] As can be known from the above description, the simulation data set in the simulation data set and the small sample data in the public data set are utilized, and the performance of the small sample equivalent model is effectively improved by a double driving mechanism. The teacher model is trained according to the simulation data set, the trained teacher model is used as a knowledge source, knowledge is transmitted to the student model, and the training of the student model is guided. Further, a hybrid loss objective function is constructed based on the trained teacher model and the student model, the learning process of the student model is systematically guided through three dimensions of classification loss, distillation loss and feature comparison loss, so that the student model can better generalize on the small sample data set. The student model is trained according to the small sample data set, and the model parameters of the student model are updated by using the hybrid loss objective function to obtain the small sample equivalent model. In summary, the deep learning model capable of effectively identifying the modulated signal under the condition of small sample is constructed by the data and knowledge double enhancement, and the practicability and wide applicability of the modulated signal identification technology are significantly improved.

[0065] As Figure 2 shown, Figure 2 is a flowchart of another embodiment of a data and knowledge enhanced double driving small sample equivalent model acquisition method provided by the application. The data and knowledge enhanced double driving small sample equivalent model acquisition method comprises the following steps: S201: Obtain a simulated modulated signal data set.

[0066] Exemplarily, first, the simulated modulated signal data set is obtained.

[0067] The simulated modulated signal data set is obtained by Matlab 2020a simulation, and there are 11 modulation categories, the signal-to-noise ratio is distributed between-20dB and 18dB, there are 1000 samples in each category, and the specific introduction is shown in Table 3, and Table 3 is the simulated modulated signal data set.

[0068] Table 3 Simulated modulated signal data set

[0069] Among them, the modulated signal can be expressed in the form of complex baseband as:

[0070] Among them, and respectively represent the in-phase and quadrature components of the signal, is a virtual unit. By sampling the received signal with a length of , the signal can be expressed as a discrete sequence, and at this time L=128.

[0071] S202: Extract the I-channel data and Q-channel data of the modulation signal from the modulation signal dataset, and add a label to the modulation signal.

[0072] Exemplarily, the I and Q channel data of the modulation signal are extracted, as shown in the following formula: ; A label is added to the modulation signal, thereby constructing a dataset, as shown in the following formula: ; wherein, is the i-th modulation signal, is the corresponding label, and N is the number of samples.

[0073] S203: Determine the mean and standard deviation of the I-channel data and Q-channel data, respectively.

[0074] S204: Normalize the I-channel data and Q-channel data based on the mean and standard deviation to obtain the simulation dataset.

[0075] Exemplarily, the mean and standard deviation of the I and Q channel data are calculated, the extracted I and Q channel data are normalized to eliminate the amplitude difference and dimension influence of non-modulation essence, the normalized modulation signal is obtained, and thereby the simulation dataset is obtained. The normalized modulation signal is shown in the following formula: ; wherein, and are the mean and standard deviation of the i-th modulation signal, respectively.

[0076] S205: Obtain a small sample dataset from the public dataset.

[0077] Exemplarily, the public dataset RML2016.10a. The dataset is composed of modulation signal examples synthesized by GNU Radio. The dataset has 11 modulation categories, including 8 digital modulations and 3 analog modulations, and the specific modulation types are consistent with the types of the simulation dataset. The signal-to-noise ratio ranges from -20dB to 18dB with an interval of 2dB. The original dataset is stratified sampled, 10 samples are selected for each signal-to-noise ratio in each category as the training set, and 200 samples are selected as the test set. The total number of small sample signal data actually used for network training is 2200.

[0078] S206: Divide the simulation dataset into a training set and a test set.

[0079] ​​S207: Train the teacher model through the training set, obtain the soft probability output and feature output of the teacher model, and perform parameter updating of the teacher model by using a cross-entropy loss function.

[0080] It should be noted that step S206 has been described in detail in the embodiment scenario shown in Figure 1 and will not be described here again.

[0081] S208: Randomly enhance the small sample data set by using a signal data enhancement method, the signal data enhancement method including at least one of signal noise addition, overall transformation, local transformation, signal fusion, signal decomposition, and masking, the overall transformation including rotation and flipping, and the local transformation including segmented inversion and cyclic shift.

[0082] Exemplarily, the signal data enhancement method is that each batch of sample data sent into the network for training is randomly enhanced by using the signal data enhancement method. The specific enhancement methods include signal noise addition, overall transformation (rotation and flipping), local transformation (inversion and shift), signal fusion / decomposition, and masking.

[0083] Specifically, the signal noise addition is shown in the following formula: ; wherein, is the signal noise addition, is an IQ signal, i.e., a modulation signal, is a zero-mean Gaussian noise with a variance independently added to the in-phase and quadrature components. The parameter controls the noise intensity. In this example, = 0.002 is used, and the Gaussian white noise can be added to the modulation signal by calculation according to the above formula.

[0084] The signal rotation is shown in the following formula: ; wherein, is the signal rotation, is a rotation angle. In this example, = 90°, 180°, and 270° are used, and the signal can be rotated by calculation according to the above formula.

[0085] The signal flipping is shown in the following formula: ; ; ; wherein, is a horizontally flipped signal, is a vertically flipped signal, for horizontal and vertical flipping of a signal simultaneously, for negating the value of I, for negating the value of Q, for negating both I and Q, switching I to the opposite value to define horizontal flipping and switching Q to the opposite value to define vertical flipping. Horizontal and vertical flipping can be performed simultaneously. The flipping of a signal can be achieved by the above formula calculation.

[0086] The signal inversion is shown as follows: ; where, is the signal inversion, is a randomly selected starting position, is the length of the segment to be inverted. Both the segment length and the breakpoint are augmented hyperparameters that control the degree of signal change. The inversion of a signal can be achieved by the above formula calculation.

[0087] The signal shift is shown as follows: ; where, is the signal shift, is a randomly selected split point, dividing the signal into two sub-sequences and . The shift of a signal can be achieved by the above formula calculation.

[0088] The signal fusion is shown as follows: ; ; where, and are the two fused signals, are the two sub-signals divided from the first signal, and are the two sub-signals divided from the second signal, , is a randomly selected split point, randomly splitting and exchanging segments for a given pair of modulated signals. Since the two signals are randomly divided into two parts at , , the lengths of the two fused signals are - + - , assuming > , the length of the first fused signal is greater than the original signal length L, and the length of the second fused signal is less than the original signal length L, so the fused signal needs to be cropped and padded. If the fused signal is longer than the original signal length, it is randomly cropped to a length of L. If they are less than L, zero padding of a random padding width is applied on both sides until the target length is reached. The signal can be fused by calculating the above formula.

[0089] The signal decomposition is shown in the following formula: ; ; wherein, is the first signal of decomposition, is the second signal of decomposition, , and are randomly selected. The signal can be decomposed by calculating the above formula.

[0090] The signal mask is shown in the following formula: ; wherein, is the signal mask, is the ith signal, is the mask sequence (continuous and discrete) is a configurable hyperparameter to control the degree and randomness of the disturbance. The signal can be masked by calculating the above formula.

[0091] Each signal data is randomly transformed into any one form in a batch and then sent to the student model for training. These methods simulate various pattern transformations, losses or attenuations that may occur to the signal, which can encourage the model to infer meaningful representations from these augmented samples, thereby enhancing the generalization ability and noise tolerance.

[0092] S209: Construct a classification loss based on the predicted label of the student model and the real label.

[0093] S210: Construct a distillation loss based on the softened probability output of the student model and the trained teacher model.

[0094] S211: Construct a feature comparison loss based on the feature output of the student model and the trained teacher model.

[0095] S212: Construct a hybrid loss objective function according to the classification loss, the distillation loss and the feature comparison loss.

[0096] S213: training the student model corresponding to the teacher model based on the trained teacher model, the small sample dataset, and the mixed loss objective function, to obtain a small sample equivalent model.

[0097] It should be noted that steps S209-S213 have been described in detail in the embodiment scenario shown in Figure 1 and will not be described here again.

[0098] As described above, by combining deep knowledge distillation and data enhancement strategy, the present application realizes deep knowledge transfer from the feature layer to the output layer, and greatly enriches the diversity of training data. On the one hand, the mixed loss function containing classification loss, distillation loss and feature comparison loss enables the student model to learn the essence of the teacher model more fully, thereby significantly improving the recognition accuracy and robustness to different signal-to-noise ratio signals under small sample conditions. On the other hand, the data enhancement strategy is applied to the small sample training data, which effectively suppresses the model overfitting and fundamentally enhances the generalization ability of the model to unseen data. Therefore, the present application effectively solves the problem of dependence of deep learning modulation recognition technology on large-scale labeled data sets, making it possible to construct a high-performance recognition model in the actual scene with only a small number of labeled samples. In addition, the technical framework of the present application does not depend on a specific neural network structure, and can realize cross-model architecture knowledge transfer, and has wide applicability.

[0099] As shown in Figure 3 , Figure 3 is a flowchart of another embodiment of a data and knowledge enhancement double-driven small sample equivalent model acquisition method provided by the present application. The data and knowledge enhancement double-driven small sample equivalent model acquisition method comprises the following steps: S301: inputting the test set into the small sample equivalent model to obtain a predicted label, and comparing the predicted label with the real label to count the number of correctly classified samples.

[0100] S302: grouping the samples in the test set according to signal-to-noise ratio, each group corresponding to a signal-to-noise ratio, and recording the total number of samples for each signal-to-noise ratio.

[0101] S303: for each sample in the group, comparing the predicted label with the real label to count the number of correctly classified samples.

[0102] S304: determining the average classification accuracy and the highest recognition accuracy evaluation according to the total number of samples and the number of correctly classified samples, and performing performance evaluation of the small sample equivalent model through the average classification accuracy and the highest recognition accuracy.

[0103] The test set is input into the small sample equivalent model for prediction label, and the predicted label is compared with the real label to count the number of correctly classified samples.

[0104] Furthermore, the average classification accuracy and the highest recognition accuracy are determined based on the total number of samples and the number of correctly classified samples. The average classification accuracy is... As shown in the following formula: .

[0105] Specifically, the classification performance at a specific signal-to-noise ratio is first determined based on the total number of samples and the number of correctly classified samples. As shown in the following formula: ; in, For signal-to-noise ratio is The total number of samples, For a signal-to-noise ratio of The number of correctly classified samples at that time. The classification performance at a specific signal-to-noise ratio.

[0106] Furthermore, the highest recognition accuracy The definition is as follows: ; Finally, through calculation, it was found that = 60.336%, = 89.864%, and classification performance curves under different signal-to-noise ratios were plotted. In addition, the confusion matrix under high signal-to-noise ratio (14dB) and low signal-to-noise ratio (-4dB) were plotted, showing that the model has good classification performance and robustness.

[0107] like Figure 4 As shown, Figure 4 This is a schematic diagram of an embodiment of a small-sample equivalence model acquisition system driven by both data and knowledge enhancement provided by the present invention. The system 10 comprises: The simulation dataset and small sample dataset acquisition module 11 is used to acquire the simulation dataset from the simulated modulation signal dataset and the small sample dataset from the public dataset.

[0108] Training module 12 is used to train a teacher model based on the simulation dataset.

[0109] The small sample equivalent model acquisition module 13 is used to construct a hybrid loss objective function. Based on the trained teacher model, the small sample dataset, and the hybrid loss objective function, the student model corresponding to the teacher model is trained to obtain the small sample equivalent model.

[0110] Exemplarily, in the simulation data set and small sample data set acquisition module 11, a simulated modulation signal data set is acquired; I channel data and Q channel data of the modulation signal are extracted from the modulation signal data set, and the modulation signal is labeled; the mean and standard deviation of the I channel data and the Q channel data are determined respectively; the I channel data and the Q channel data are normalized based on the mean and the standard deviation, and the simulation data set is acquired.

[0111] In the training module 12, the simulation data set is divided into a training set and a test set; the teacher model is trained through the training set, the soft probability output and the feature output of the teacher model are acquired, and the parameter update of the teacher model is performed by using a cross-entropy loss function.

[0112] In the small sample equivalent model acquisition module 13, a classification loss is constructed based on the prediction label of the student model and the real label; a distillation loss is constructed based on the soft probability output of the student model and the trained teacher model; a feature comparison loss is constructed based on the feature output of the student model and the trained teacher model; and a mixed loss objective function is constructed according to the classification loss, the distillation loss and the feature comparison loss.

[0113] Further, the enhanced small sample data is input into the student model to acquire the prediction label, the soft probability output and the feature output of the student model; the soft probability output and the feature output of the teacher model, and the prediction label, the soft probability output and the feature output of the student model are substituted into the mixed loss objective function to determine the loss value; the gradient of the loss value to the student model is calculated by using a back propagation algorithm; the student model parameters are adjusted by using the gradient until the loss value reaches a threshold value, and the small sample equivalent model is acquired.

[0114] As shown in Figure 5 , Figure 5 An embodiment structure diagram of the medium provided by the application. The medium 20 stores at least one computer program 21, and the computer program 21 is executed by a processor to implement the method as shown in Figure 1 、 Figure 2 and Figure 3 The detailed method can be referred to the above, which will not be repeated here. In an embodiment, the medium 20 can be a storage chip, a hard disk or a mobile hard disk or an optical disc, and other readable and writable storage tools, and can also be a server and the like.

[0115] In addition, the processes depicted in the drawings do not necessarily have to be implemented in the specific order or continuous order shown in order to achieve the desired results. In some embodiments, multi-task processing and parallel processing are also possible or can be advantageous.

[0116] Each of the embodiments in the specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other. Each of the embodiments focuses on the difference from other embodiments. In particular, for the apparatus, device, and non-transitory computer-readable storage medium embodiments, since they are basically similar to the method embodiments, they are described more simply, and the relevant parts can be referred to the part of the method embodiments.

[0117] The apparatus, device, non-transitory computer-readable storage medium, and method provided by the embodiments of the specification are corresponding, and therefore, the apparatus, device, and non-transitory computer-readable storage medium also have similar beneficial technical effects as the corresponding method. Since the beneficial technical effects of the method have been described in detail above, the beneficial technical effects of the corresponding apparatus, device, and non-transitory computer-readable storage medium will not be described here.

[0118] The system, apparatus, module, or unit illustrated in the above embodiments can be specifically implemented by a computer chip or entity, or by a product with certain functions. A typical implementation device is a computer. Specifically, the computer may, for example, be a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0119] For the convenience of description, the above apparatus is described as various units divided by functions during description. Of course, the functions of each unit can be implemented in the same or multiple software and / or hardware in the implementation of the specification. Those skilled in the art should understand that the embodiments of the specification can be provided as a method, a system, or a computer program product. Therefore, the embodiments of the specification can adopt a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware. Moreover, the embodiments of the specification can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0120] The specification is described with reference to flowcharts and / or block diagrams of the method, device (system), and computer program product according to the embodiments of the specification. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to produce a machine, so that the instructions executed by the computer or other programmable data processing devices produce a machine for implementing the functions described in the flowcharts and / or block diagrams. Figure 1one or more processes and / or blocks Figure 1 an apparatus with the functionality to achieve the specified function of a block or blocks.

[0121] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the Figure 1 one or more processes and / or blocks Figure 1 an apparatus with the functionality to achieve the specified function of a block or blocks.

[0122] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the Figure 1 one or more processes and / or blocks Figure 1 an apparatus with the functionality to achieve the specified function of a block or blocks.

[0123] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0124] The memory can include non-persistent memory and / or volatile memory, such as a random access memory (RAM) including a cache area for the temporary storage of data. A

[0125] Computer-readable media includes permanent and non-permanent, moveable and non- moveable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory (ROM), electrically erasable programmable read only memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), digital versatile disks (DVDs) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information that is accessible to a computing device. According to the definition provided herein, a computer readable medium does not include transitory media, such as modulated data signals and carrier waves.

[0126] It is also to be noted that the terms "comprising", "including", and any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises a... " does not, without more constraints, exclude the existence of additional identical elements in the process, method, article, or apparatus that comprises the recited element.

[0127] The specification can be described in the general context of computer-executable instructions, such as program modules, being executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. The specification can also be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules can be located in both local and remote computer storage media including memory storage devices.

[0128] The various embodiments in the specification are described in progressive manner, and the same or similar parts between the embodiments can be mutually referred to, and each embodiment focuses on the difference from other embodiments. In particular, the system embodiments are described more simply because they are basically similar to the method embodiments, and the relevant parts can be referred to the part of the method embodiments.

[0129] The above disclosure is merely preferred embodiments of the present application and is not intended to limit the scope of the present application. Therefore, any equivalent changes made according to the claims of the present application are still within the scope of the present application.

Claims

1. A method for obtaining a small-sample equivalence model driven by both data and knowledge enhancement, characterized in that, The method includes: The simulation dataset is obtained from the simulated modulation signal dataset, and the small sample dataset is obtained from the public dataset. The teacher model was trained based on the simulation dataset; A hybrid loss objective function is constructed, and the student model corresponding to the teacher model is trained based on the trained teacher model, the small sample dataset, and the hybrid loss objective function to obtain a small sample equivalent model.

2. The method for obtaining a small-sample equivalence model driven by both data and knowledge enhancement as described in claim 1, characterized in that, Before constructing the hybrid loss objective function and training the student model corresponding to the teacher model based on the trained teacher model, the small sample dataset, and the hybrid loss objective function to obtain the small sample equivalent model, the method further includes: The small sample dataset is randomly augmented using signal data augmentation methods, which include at least one of signal noise addition, global transformation, local transformation, signal fusion, signal decomposition, and masking. The global transformation includes rotation and flipping, and the local transformation includes segmented inversion and cyclic shifting.

3. The method for obtaining a small-sample equivalence model driven by both data and knowledge enhancement as described in claim 1, characterized in that, The step of obtaining the simulation dataset from the simulated modulation signal dataset specifically includes: Obtain the simulation modulation signal dataset; Extract the I-channel and Q-channel data of the modulation signal from the modulation signal dataset, and add tags to the modulation signal; Determine the mean and standard deviation of the I-channel data and the Q-channel data respectively; The I-channel and Q-channel data are normalized based on the mean and standard deviation to obtain the simulation dataset.

4. The method for obtaining a small sample equivalence model driven by both data and knowledge enhancement as described in claim 3, characterized in that, The teacher model trained based on the simulation dataset specifically includes: The simulation dataset is divided into a training set and a test set; The teacher model is trained using the training set to obtain the softening probability output and feature output of the teacher model, and the parameters of the teacher model are updated using the cross-entropy loss function.

5. The method for obtaining a small-sample equivalence model driven by both data and knowledge enhancement as described in claim 4, characterized in that, The construction of the hybrid loss objective function specifically includes: A classification loss is constructed based on the predicted labels and the true labels of the student model. The distillation loss is constructed based on the softening probability output of the student model and the trained teacher model. A feature contrast loss is constructed based on the feature outputs of the student model and the trained teacher model. A hybrid loss objective function is constructed based on the classification loss, distillation loss, and feature contrast loss.

6. The method for obtaining a small-sample equivalence model driven by both data and knowledge enhancement as described in claim 5, characterized in that, The construction of the hybrid loss objective function based on the classification loss, distillation loss, and feature contrast loss specifically includes: according to Construct a hybrid loss objective function, where α, β, and λ are hyperparameters. For classification loss, y i For the true label of the sample, The predicted labels for the student model. For distillation losses, This is the softening probability output for the teacher model. The softening probability output for the student model. For feature contrast loss, For the feature output of the student model, This is the feature output for the teacher model.

7. The method for obtaining a small-sample equivalence model driven by both data and knowledge enhancement as described in claim 6, characterized in that, The process of training the student model corresponding to the teacher model based on the trained teacher model, the small sample dataset, and the hybrid loss objective function to obtain a small sample equivalent model specifically includes: The enhanced small sample data is input into the student model to obtain the student model's predicted label, softening probability output, and feature output; Substitute the softening probability output and feature output of the teacher model, and the predicted label, softening probability output and feature output of the student model into the hybrid loss objective function to determine the loss value; The gradient of the loss value with respect to the student model is calculated using the backpropagation algorithm; The student model parameters are adjusted using the gradient until the loss value reaches a threshold, thus obtaining a small-sample equivalent model.

8. The method for obtaining a small-sample equivalence model driven by both data and knowledge enhancement as described in claim 6, characterized in that, The method further includes: The test set is input into the small sample equivalence model to obtain the predicted labels. The number of correctly classified samples is counted by comparing the predicted labels with the true labels. The samples in the test set are grouped according to the signal-to-noise ratio, with each group corresponding to a signal-to-noise ratio, and the total number of samples for each signal-to-noise ratio is recorded. For each sample in the group, the number of correctly classified samples is counted by comparing the predicted label with the true label; The average classification accuracy and the highest recognition accuracy are determined based on the total number of samples and the number of correctly classified samples. The performance of the small sample equivalent model is then evaluated using the average classification accuracy and the highest recognition accuracy.

9. A system for acquiring small-sample equivalence models driven by both data and knowledge enhancement, characterized in that, The system includes: The simulation dataset and small sample dataset acquisition module is used to acquire the simulation dataset from the simulated modulation signal dataset and the small sample dataset from the public dataset. The training module is used to train a teacher model based on the simulation dataset; The small sample equivalent model acquisition module is used to construct a hybrid loss objective function, and to train the student model corresponding to the teacher model based on the trained teacher model, the small sample dataset, and the hybrid loss objective function to obtain the small sample equivalent model.

10. A computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the steps of the method as claimed in any one of claims 1 to 8.

Images