A method and apparatus for pruning communication signal datasets based on supporting data.

Abstract

This invention relates to the field of communication warfare, specifically a method and apparatus for pruning communication signal datasets based on support data. The invention constructs a boundary sample selector based on a support vector machine (SVM) on top of a deep neural network-based signal classifier. This SVM approximates an explicit representation of the signal classifier's decision boundary, solving the problem of difficulty in finding boundary samples due to the implicit nature of the signal classifier's decision boundary, and achieving rapid identification of boundary samples. Simultaneously, this invention proposes a method for evaluating the importance of training samples based on support data. This method accurately selects the most representative data subset early in the signal classifier's training process, significantly reducing the training cost of the signal classifier while minimizing its impact on the classifier's recognition accuracy. This invention is applicable to signal recognition tasks with limited memory and computing resources.

Description

Technical Field

[0001] This invention relates to the field of communication countermeasures, and in particular to a method and apparatus for pruning communication signal datasets based on supporting data. Background Technology

[0002] With the rapid development of satellite, communication, and Internet of Things (IoT) technologies, the number and types of wireless signals in the electromagnetic environment have increased dramatically, seriously threatening electromagnetic environment safety. Therefore, accurate identification of various wireless signals is crucial. In recent years, deep learning-based signal recognition technology has received considerable attention and has become an important means of maintaining electromagnetic environment safety. However, deep learning-based signal classifiers typically require training on massive amounts of signal samples to achieve advanced performance, and large datasets composed of these samples consume significant amounts of memory and computing resources. Large datasets often contain redundant samples. Pruning the signal dataset, retaining only valuable samples for classifier training, would effectively reduce training costs. However, identifying which samples are valuable and how to implement dataset pruning are pressing research questions.

[0003] Existing research indicates that training examples located near the decision boundary of a signal classifier are crucial for model training. However, the decision boundary of a neural network is implicitly represented, making it difficult to find examples close to the boundary. Existing methods filter boundary examples by calculating their geometric properties or distance from the decision boundary; however, most of these methods require a well-trained classifier and have certain limitations. Summary of the Invention

[0004] The purpose of this invention is to overcome the problem of data redundancy in large signal datasets in the prior art, and to provide a communication signal dataset pruning method based on support data. By finding a representative subset that is important for training the signal classifier in the early stage of training, the size of the signal training set is reduced while ensuring the model recognition accuracy, thus reducing the memory and computation requirements of the signal classifier.

[0005] To achieve the aforementioned objective, the present invention provides the following technical solution:

[0006] A method for pruning a communication signal dataset based on supporting data, the method for communication signals, includes the following steps:

[0007] S1: Acquire communication signals, and use a deep neural network-based signal classifier to extract and classify the latent features of the communication signals;

[0008] S2: Establish a boundary sample selector based on the one-to-many rule, train the boundary sample selector with the output vector set of the signal classifier, give an explicit representation of the decision boundary of the signal classifier, and then filter out the supporting data located near the decision boundary;

[0009] S3: The importance of training examples to model training is evaluated based on the frequency with which each training sample is labeled as supporting data in the early stage of signal classifier training. The importance of training examples is positively correlated with the number of times they are labeled as supporting data.

[0010] S4: Arrange the training examples in descending order of importance, select a representative subset of the communication signal dataset, and use the representative subset for subsequent training of the signal classifier.

[0011] Compared to existing methods that prune large datasets by calculating the geometric properties of training examples or the distance from examples to the decision boundary, this approach has lower requirements for the signal classifier, fewer limitations, and improves both the efficiency and feasibility of filtering communication signals.

[0012] Step S1 includes the following steps:

[0013] S11: Acquire communication signals through a signal receiving device, and construct a signal classifier based on a deep neural network. The last layer of the signal classifier uses Softmax as the activation function to transform the output vector of the signal classifier into a discrete probability distribution.

[0014] S12: On a K-class signal dataset, cross-entropy loss is used to optimize the signal classifier, while gradient descent is used to update the model parameters and obtain the output vector set.

[0015] Step S1 involves initial processing of the original signal samples, mapping the signal sample set to a potential output vector set. The information provided by this transformation will help in the subsequent search and labeling of support data.

[0016] Furthermore, step S2 specifically includes the following steps:

[0017] S21: Construct a boundary sample selector, which is an ensemble of K support vector machine models according to a one-to-many rule, and train the boundary sample selector using the output vector set of the signal classifier;

[0018] S22: Using the hyperplane equation of the boundary sample selector, give an explicit expression for the approximate decision boundary of the signal classifier;

[0019] S23: Use the support vector coordinates labeled by the boundary sample selector to quickly find the output vectors located near the decision boundary in the latent feature space. The original signal samples corresponding to these output vectors are the boundary samples.

[0020] Optionally, step S3 includes the following steps:

[0021] S31: Record the number of times each communication signal sample is labeled as supporting data during the first 20 pre-training processes of the signal classifier, and use it to characterize the importance of training samples to model training.

[0022] S32: Arrange the training set samples in descending order of the number of times they are labeled;

[0023] S33: Select the first M original signal samples to form a representative subset to accelerate the subsequent training of the model. The value of M is the product of the total number of training samples and the proportion of training samples retained.

[0024] This invention addresses the issue of reducing the size of the training set in large signal datasets. First, it designs a boundary sample selector to directly label the samples at the decision boundary. Second, based on the frequency of each training sample appearing at the decision boundary in the early stages of training, it sorts the training samples in descending order and selects a representative subset, thereby achieving the goal of finding a data subset that is important to the model in the early stages of model training.

[0025] Based on the same inventive concept, a communication signal dataset trimming device based on supporting data is also proposed, including a computer system and at least one signal acquisition device, wherein the signal acquisition device is electrically connected to the computer processing system; the computer processing system is capable of executing any of the above-mentioned communication signal dataset trimming methods based on supporting data, and the computer processing system has instructions, which the signal acquisition device executes.

[0026] The importance of each training example to the model training is assessed by evaluating the number of times it is labeled as supporting data in the early stages of model training, thereby selecting the most representative subset of data. The method provided by this invention can reduce the training cost of the signal classifier while minimizing the impact on its recognition accuracy, making it suitable for signal recognition tasks with limited memory and computing resources.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0028] 1. Addressing the difficulty in finding boundary examples due to the implicit nature of decision boundaries in deep neural networks, this invention eliminates the need to calculate indirect parameters to reflect the relationship between examples and the decision boundary, thus finding boundary examples more directly and accurately. This invention designs a boundary example selector based on support vector machines, which approximates an explicit representation of the decision boundary, thereby quickly finding and labeling support data located near the decision boundary.

[0029] 2. This invention proposes a method for evaluating the importance of training examples based on support data. It utilizes information from the early stages of signal classifier training to assess the importance of examples to the classifier's training. Compared to most existing methods, this method identifies training examples important to the model early in the training process, without requiring model convergence, thus possessing significant practical value and being simple and easy to implement.

[0030] 3. This invention utilizes the number of times each training sample is labeled to filter out a subset of data, which can effectively reduce the training cost of the classifier while minimizing the impact on the classifier's recognition accuracy and accelerating the training speed.

[0031] 4. Through simulation experiments on the experimental dataset RML2016.04C, the performance of the method provided by this invention is significantly better than other dataset pruning methods, and it is closest to the optimal recognition accuracy when the proportion of training samples retained is relatively small.

[0032] 5. Through simulation experiments on the experimental dataset RML2016.04C, the performance of the method provided by this invention is significantly better than other dataset pruning methods. When the proportion of training samples retained is less than or equal to 0.3, it is closest to the optimal recognition accuracy. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating a communication signal dataset trimming method based on supporting data provided in Embodiment 1 of the present invention;

[0034] Figure 2 This is a schematic diagram of a boundary example of a communication signal dataset pruning method based on support data provided in Embodiment 1 of the present invention;

[0035] Figure 3 This is a schematic diagram showing the boundary sample annotation comparison results of a communication signal dataset pruning method based on support data as described in Embodiment 1 of the present invention.

[0036] Figure 4 This is a schematic diagram illustrating the number of times each sample is labeled as a support vector during the first 20 pre-training processes of the communication signal dataset pruning method based on support data described in Embodiment 1 of the present invention.

[0037] Figure 5 This is a schematic diagram comparing the performance advantages of the communication signal dataset pruning method based on support data described in Embodiment 1 of the present invention on the RML2016.04C signal dataset.

[0038] Figure 6 This is a schematic diagram of the apparatus for a communication signal dataset trimming method based on support data, as described in Embodiment 1 of the present invention. Detailed Implementation

[0039] The present invention will be further described in detail below with reference to experimental examples and specific embodiments. However, this should not be construed as limiting the scope of the above-mentioned subject matter of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.

[0040] Example

[0041] like Figure 1 As shown in this embodiment, a method for pruning a communication signal dataset based on supporting data is provided. The method is used to identify communication signals and includes the following steps:

[0042] S1: Acquire communication signals, and use a deep neural network-based signal classifier to extract and classify the latent features of the communication signals;

[0043] S2: Establish a boundary sample selector based on the one-to-many rule, train the boundary sample selector with the output vector set of the signal classifier, give an explicit representation of the decision boundary of the signal classifier, and then filter and label the supporting data located near the decision boundary;

[0044] S3: The importance of training examples to model training is evaluated based on the frequency with which each training sample is labeled as supporting data in the early stage of signal classifier training. The importance of training examples is positively correlated with the number of times they are labeled as supporting data.

[0045] S4: Arrange the training examples in descending order of importance, select a representative subset of the communication signal dataset, and use the representative subset for subsequent training of the signal classifier.

[0046] It should be noted that the deep neural network mentioned is the VT_CNN2 convolutional neural network model specifically designed for signal recognition. This model consists of two convolutional layers and two fully connected layers. The two convolutional layers contain 256 and 80 convolutional kernels, respectively, and a ReLU activation function is added at the end of each convolutional layer. The two fully connected layers contain 256 and K neurons, respectively, where K is the number of signal categories. The two fully connected layers use ReLU and Softmax as activation functions, respectively.

[0047] First, the signal classifier is pre-trained 20 times on a given signal dataset. Based on information provided by the boundary sample selector, the support data located at the decision boundary after each training iteration are recorded. Then, the number of times each training sample is labeled as support data is counted and used as an evaluation parameter for the importance of model training; the more times a sample is labeled, the more important it is for model training. Finally, all samples in the training set are sorted in descending order of the number of times they are labeled, and the top M samples are selected to form a representative subset for training the signal classifier. M is the product of the total number of training samples and the proportion of training samples retained. Compared to existing methods that select boundary samples by calculating the geometric properties of training samples or the distance of samples to the decision boundary, this approach has lower requirements for the training level of the classifier, fewer limitations, a more representative data subset, and high feasibility.

[0048] Step S1 includes the following steps:

[0049] Construct a signal classifier based on a deep neural network. After one stochastic gradient descent update, the signal samples in the training set... Mapped to output vector The last layer of the signal classifier uses Softmax as the activation function to transform the output vector of the signal classifier into a discrete probability distribution, as shown in the following expression;

[0050]

[0051] Where, p i,k Represents signal sample x i The probability of being predicted as class k. z i,k This represents the output value of the k-th node in the last layer. and These are the input vector and weight parameters of the last layer of the signal classifier, respectively. k This represents the weight of the k-th node in the last layer;

[0052] On a K-class signal dataset, cross-entropy loss is used to optimize the signal classifier, while gradient descent is employed to update the model parameters and obtain the output vector set. Specifically, the signal classifier is trained using cross-entropy loss to construct the objective function, formally expressed as:

[0053]

[0054] Where N and K represent the total number of samples and the number of classes in the signal dataset, respectively; y i,k For a symbolic function, when y i =1 when =k, otherwise 0; p i,k Represents signal sample x i The probability of being predicted as class k.

[0055] Signal classification is performed on the publicly available signal dataset RML2016.04C. The signal classifier consists of two convolutional layers and two fully connected layers. This embodiment uses a gradient descent optimizer with a learning rate of 0.001, a momentum of 0.9, and a weight decay of 5e-4. It performs 20 pre-training iterations and 350 optimization training iterations on RML2016.04C, with each batch size set to 64.

[0056] Step S1 involves building and training the signal classifier, while simultaneously performing initial processing on the original signal samples, mapping them to potential output vectors. The information provided by this transformation will aid in the subsequent search and labeling of boundary examples.

[0057] First, K support vector machines are integrated into a boundary sample selector according to the one-to-many rule. The boundary sample selector outputs vector sets... Training proceeds. The aforementioned one-to-many rule refers to the process of training K support vector machines by sequentially setting the label of one class of training set signals to 1, and setting the label of the remaining training set signals to 0. A boundary sample selector composed of K support vector machines is then used to partition the signal feature space of the K classes of signal datasets.

[0058] Secondly, using the hyperplane of the boundary sample selector, the decision boundary between every two classes of vectors in the signal feature space is represented by the following linear equation:

[0059] a T z i +b=0

[0060] Where a represents the normal vector, which determines the direction of the hyperplane; b is the displacement, which determines the output vector z corresponding to the training sample and the hyperplane. i The distance between them. Calculating the optimal solutions for a and b yields the equation of the hyperplane described above. The problem of finding this optimal solution can be transformed into the following formula:

[0061]

[0062] sty i (a T z i +b)≥1,i=1,...,N

[0063] Where C is the penalty parameter and ξ is the slack variable.

[0064] Therefore, the support data located near the decision boundary satisfies the following conditions:

[0065] y i (a T zi +b)≥1

[0066] According to the definition of support vectors, the "distance" between each training sample labeled as support data and the decision boundary in the latent feature space is denoted as .

[0067] Finally, the support vector coordinates labeled by the boundary sample selector are used to quickly find the output vectors located near the decision boundary in the signal feature space. The original signal samples corresponding to these output vectors are the boundary samples.

[0068] Based on the information about the number of times each training sample was labeled as support data in the early stages of signal classifier training, step S3 introduces a method for evaluating the importance of training samples based on support data, which includes the following steps:

[0069] First, the number of times each signal sample was labeled as supporting data during the first 20 pre-training processes of the signal classifier was recorded, and this was used as an evaluation parameter to characterize the importance of the training samples to the model training.

[0070] Then, the training set samples are sorted in descending order according to the number of times they are labeled, with the earlier the sample appears, the more important it is for model training;

[0071] Finally, the top M original signal samples are selected to form a representative subset for training the signal classifier, minimizing the impact on recognition accuracy while reducing model training costs. M is the product of the total number of training samples and the proportion of training samples retained.

[0072] Simulation experiments were conducted using the publicly available modulation signal dataset RML2016.04C, which selected two signal types: 8PSK and AM-DSB. The RML2016.04C dataset includes 3 analog modulation types and 8 digital modulation types, totaling 51,040 samples. The signal samples are in 1×2×128 format, with a signal-to-noise ratio ranging from -20dB to 18dB, in 2dB increments.

[0073] like Figure 3 As shown, the boundary sample selector represents the decision boundary of the signal classifier as a linear function and finds the signal feature vector closest to the decision boundary, i.e., the triangular data point. The original signal sample corresponding to the feature vector is the boundary sample in the training set.

[0074] Figure 4The number of times each training sample in the RML2016.04C dataset was labeled by the boundary example selector during the first 20 pre-training iterations was counted. In the early stages of model training, the number of times each training example was labeled varied, indicating different frequencies of its appearance at the decision boundary. The importance of training samples to model training was positively correlated with these frequencies; the training samples with the most labeling frequencies were considered the most important training samples for model training.

[0075] In the RML2016.04C dataset, 70% of the samples for each type of modulation signal were used as the training set, 10% as the validation set, and the remaining 20% ​​as the test set for simulation experiments. The results are as follows: Figure 5 As shown, the method provided in this embodiment is compared with the performance of various existing dataset pruning methods such as Herding, DeepFool, Cal, Craig, and Forgetting. Here, "No pruning" refers to the method of training the signal classifier on a large, unpruned signal training set. This result reflects the best recognition accuracy of the signal classifier on the current signal dataset and represents the performance ceiling of other dataset pruning methods.

[0076] like Figure 5 As shown, on the RML2016.04C signal dataset, when the proportion of training samples retained in the representative subset is small, the performance curve of the method provided in this invention is closest to the benchmark No pruning result curve. This is because during dataset pruning, the fewer signal samples retained, the worse the recognition accuracy of the signal classifier. Therefore, compared to other methods, the method provided in this embodiment achieves a better trade-off between training set size and model recognition accuracy. The performance of the method provided in this embodiment is significantly better than other dataset pruning methods.

[0077] The signal recognition dataset pruning method based on supporting data provided in this embodiment reduces the training set size by removing redundant samples from the communication signal dataset, thereby lowering the cost of model training. First, a boundary sample selector is designed to directly label boundary samples near the decision boundary. Second, the importance of each training sample to model training is evaluated by assessing the number of times each training sample is labeled as supporting data in the early stages of signal classifier training, thus selecting the most representative data subset. This reduces the memory and computing resource requirements for training the signal classifier, i.e., lowers the training cost of signal recognition, while accelerating the training speed of the signal recognition model. The method provided by this invention can effectively complete communication signal tasks with limited memory and computing resources.

[0078] Based on the same inventive concept, a communication signal dataset trimming device based on supporting data is also proposed, including a computer system and at least one signal acquisition device, wherein the signal acquisition device is electrically connected to the computer processing system; the computer processing system is capable of executing any of the above-mentioned communication signal dataset trimming methods based on supporting data, and the computer processing system has instructions, which the signal acquisition device executes.

Claims

1. A method for pruning a data set of a communication signal based on support data, characterized by, The method is used for identifying a communication signal, and comprises the following steps: S1: obtaining a communication signal, and realizing extraction and classification of potential features of the communication signal based on a signal classifier of a deep neural network; S2: establishing a boundary sample selector according to a one-to-many rule, training the boundary sample selector by using an output vector set of the signal classifier, giving an explicit expression of a decision boundary of the signal classifier, and further screening support data located near the decision boundary; S3: taking a frequency of labeling each training sample as the support data during early training of the signal classifier as a measurement index, and evaluating importance of the training sample to the model training, wherein the importance of the training sample is positively correlated with the number of times of labeling the training sample as the support data; S4: arranging the training samples in descending order of importance, screening a representative subset of the communication signal data set, and using the representative subset for subsequent training of the signal classifier.

2. The method of claim 1, wherein the method is a method of pruning a data set of a communication signal based on support data. Step S1 comprises the following steps: S11: obtaining the communication signal through signal receiving, and converting an output vector of the signal classifier into a discrete probability distribution by using Softmax as an activation function of a last layer of the signal classifier; S12: performing optimization of the signal classifier by using cross-entropy loss on a K-class signal data set, and updating model parameters by using a gradient descent method and obtaining an output vector set.

3. The method of claim 1, wherein the method further comprises: Step S2 specifically comprises the following steps: S21: constructing a boundary sample selector, the selector being integrated by K support vector machine models according to a one-to-many rule, and training the boundary sample selector by using the output vector set of the signal classifier; S22: giving an approximate explicit expression of a decision boundary of the signal classifier by using a hyperplane equation of the boundary sample selector; S23: quickly finding output vectors located near the decision boundary in a potential feature space by using support vector coordinates labeled by the boundary sample selector, and taking original signal samples corresponding to the output vectors as boundary samples.

4. The method of claim 1, wherein the method is a method of pruning a data set of a communication signal based on support data. Step S3 comprises recording a number of times of labeling each signal sample during previous 20 pre-training processes of the signal classifier as the support data, and using the number of times for evaluating importance of the training sample to the model training.

5. A data set pruning apparatus for a communication signal based on support data, characterized by, The system comprises a computer system and at least one signal acquisition device, the signal acquisition device is electrically connected with the computer processing system; the computer processing system can execute a communication signal data set pruning method based on support data according to any one of claims 1 to 4, and the computer processing system is provided with instructions, and the signal acquisition device executes the instructions.

Images