A robust modulation signal recognition method and system based on sample selection and semi-supervised learning
Patent Information
- Application Number
- CN202410354385.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-03-27
- Publication Date
- 2026-09-08
- Estimated Expiration
- 2044-03-27
AI Technical Summary
[0005]本发明的目的在于提供一种基于样本选择与半监督学习的鲁棒调制信号识别方法及系统，以解决传统的自动调制识别方法通常依赖于人工设计的特征提取器和分类器，但面对复杂无线信号环境时效果有限的问题
[0048] This invention creatively introduces sparse representation and proximal gradient descent algorithm into the backpropagation process of convolutional neural networks, which can effectively remove redundant features, select key features, and improve the interpretability of the network model. Moreover, this invention is applicable to radio frequency signals of various IoT devices and has high accuracy and robustness.
Smart Images
Figure CN118171073B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a robust modulated signal recognition method and system based on sample selection and semi-supervised learning, belonging to the field of signal processing technology. Background Technology
[0002] Automatic modulation identification (AMI) is a technology that analyzes and identifies the modulation patterns of wireless communication signals. It aims to automatically classify the modulation type of received signals to improve demodulation and decoding efficiency. This technology plays a crucial role in telecommunications, electronic warfare, unmanned systems, and the Internet of Things (IoT). Specifically, in the military field, it is used for electronic reconnaissance, unmanned system communications, and electronic warfare, aiding in intelligence gathering and spectrum management. In the civilian field, it optimizes wireless communication systems, conducts spectrum monitoring, and supports the development of IoT and smart cities.
[0003] Traditional automatic modulation identification methods typically rely on manually designed feature extractors and classifiers, but their effectiveness is limited when faced with complex wireless signal environments. Deep learning, by training deep neural networks, can automatically learn efficient feature representations of signals, improving the accuracy and robustness of identification. By training on large-scale datasets, deep learning models can adapt to various modulation types and environmental conditions, providing better adaptability for practical applications.
[0004] However, in non-cooperative military scenarios, it is difficult to obtain signal labels, while in cooperative civilian scenarios, manual labeling errors are prone to occur, making it difficult to obtain enough high-quality signals and their corresponding labels to train deep neural network models, thus reducing the robustness and recognition performance of the models. Summary of the Invention
[0005] The purpose of this invention is to provide a robust modulation signal recognition method and system based on sample selection and semi-supervised learning, in order to solve the problem that traditional automatic modulation recognition methods usually rely on manually designed feature extractors and classifiers, but have limited effectiveness in complex wireless signal environments.
[0006] First aspect:
[0007] A robust modulation signal recognition method based on sample selection and semi-supervised learning includes:
[0008] Acquire the modulation signal to be tested;
[0009] The modulated signal to be tested is input into a pre-trained network model with weights W2 to identify the type of modulated signal;
[0010] The network model with weight W2 is obtained by semi-supervised training, where correctly labeled samples and incorrectly labeled samples are treated as labeled samples and unlabeled samples, respectively, and label smoothing regularization constraints and entropy minimization constraints are applied during training.
[0011] Optionally, training the network model with weights W2 includes:
[0012] Acquire modulated signals with noise labels;
[0013] The modulated signal with noise label is input into the convolutional neural network for initial training, resulting in a network model with weights W1.
[0014] The modulated signal with noise label is input into the network model with weight W1 to obtain the predicted probability, and the confidence level of each category is calculated based on the predicted probability.
[0015] The confidence-based sample selection method compares the obtained predicted probabilities with the confidence scores of each category to select the correct-labeled samples and the incorrect-labeled samples.
[0016] The correctly labeled samples and incorrectly labeled samples are treated as labeled samples and unlabeled samples, respectively, and input into the initial network model for semi-supervised training. During training, label smoothing regularization constraints and entropy minimization constraints are applied to obtain a network model with weights W2.
[0017] Optionally, the modulation signal is N is the number of signal samples, x i For the modulated signal sample, y i ∈{0,1,2……,K-1}, where K is the number of signal categories.
[0018] Optionally, the initial input of the modulated signal with noise labels into the convolutional neural network to obtain a network model with weights W1 includes:
[0019] A signal with dimension 2×d Input the signal sample length into the CNN network for initial training, and obtain a network model with weights W1.
[0020] Optionally, the confidence-based sample selection method compares the obtained predicted probabilities with the confidence scores of each category to filter out correctly labeled samples and incorrectly labeled samples, including:
[0021] Acquired modulated signals with noise tags The predicted probability is obtained after the final Softmax activation function operation of the CNN network. With predictive labels For each category, based on the true label y i With predictive labels Calculate the confidence level C for each category. j ;
[0022] Based on the calculated confidence level, a confidence-based sample selection method is used. When the true label of a sample is the same as the predicted label, and its predicted probability is greater than the confidence level of the corresponding category, the sample is a correctly labeled sample; otherwise, it is an incorrect sample.
[0023] Optionally, the confidence level C j The expression is:
[0024]
[0025] Where I(·) is an indicator function, with a value of 1 when the condition in parentheses is true and a value of 0 when the condition is false. This represents the sum of predicted probabilities when both the true label and the predicted label of a sample belong to that category. This indicates the number of samples whose true or predicted label is this category.
[0026] Optionally, the step of treating correctly labeled samples and incorrectly labeled samples as labeled samples and unlabeled samples, respectively, inputting them into the initial network model for semi-supervised training, and applying label smoothing regularization constraints and entropy minimization constraints during training to obtain a network model with weights W2 includes:
[0027] The correctly labeled samples selected are considered as labeled signal samples, and a label smoothing regularization constraint is added. Label smoothing refers to one-hot encoding of the true label of the sample. Transform into:
[0028]
[0029] Where ε is often set to a small constant, the label smoothing regularization loss term is obtained as follows:
[0030]
[0031] Convert the actual binary hard labels into soft labels with continuous values between 0 and 1;
[0032] Treating the selected incorrectly labeled samples as unlabeled signal samples, an entropy minimization constraint is added, specifically:
[0033]
[0034] Will and The two terms are combined using a balancing factor α to form the loss function of the CNN:
[0035]
[0036] After calculating the loss function value through the forward propagation process, backpropagation is performed to update the network weights W2, resulting in a network model with weights W2, specifically:
[0037]
[0038] in, Let be the parameters at training times t+1 and t respectively, η be the learning rate of the CNN, and t = 1, 2, ..., T, where T is the training period.
[0039] Optionally, the backpropagation update of network weights W2 is performed using the stochastic gradient descent algorithm.
[0040] Optionally, the structure of the CNN network includes:
[0041] An encoder consisting of two-dimensional convolution operations, ReLU linear function operations, two-dimensional max pooling operations, Dropout operations, and flattening operations; and a decoder consisting of Dense operations, ReLU linear function operations, and Softmax activation function operations.
[0042] The second aspect:
[0043] A robust modulation signal recognition system based on sample selection and semi-supervised learning, the system comprising:
[0044] The signal acquisition module is used to acquire the modulated signal to be tested.
[0045] The signal recognition module is used to input the modulated signal to be tested into a pre-trained network model with weights W2 to identify the signal;
[0046] The network model with weight W2 is obtained by semi-supervised training, where correctly labeled samples and incorrectly labeled samples are treated as labeled samples and unlabeled samples, respectively, and label smoothing regularization constraints and entropy minimization constraints are applied during training.
[0047] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:
[0048] This invention creatively introduces sparse representation and proximal gradient descent algorithm into the backpropagation process of convolutional neural networks, which can effectively remove redundant features, select key features, and improve the interpretability of the network model. Moreover, this invention is applicable to radio frequency signals of various IoT devices and has high accuracy and robustness. Attached Figure Description
[0049] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0050] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.
[0051] Example 1:
[0052] like Figure 1 As shown, the present invention provides a robust modulation signal recognition method based on sample selection and semi-supervised learning, comprising:
[0053] Acquire the modulation signal to be tested;
[0054] The modulated signal to be tested is input into a pre-trained network model with weights W2 to identify the type of modulated signal;
[0055] The network model with weight W2 is obtained by semi-supervised training, where correctly labeled samples and incorrectly labeled samples are treated as labeled samples and unlabeled samples, respectively, and label smoothing regularization constraints and entropy minimization constraints are applied during training.
[0056] The training of the network model with weight W2 includes:
[0057] Acquire modulated signals with noise labels;
[0058] The modulated signal with noise label is input into the convolutional neural network for initial training, resulting in a network model with weights W1.
[0059] The modulated signal with noise label is input into the network model with weight W1 to obtain the predicted probability, and the confidence level of each category is calculated based on the predicted probability.
[0060] The confidence-based sample selection method compares the obtained predicted probabilities with the confidence scores of each category to select the correct-labeled samples and the incorrect-labeled samples.
[0061] The correctly labeled samples and incorrectly labeled samples are treated as labeled samples and unlabeled samples, respectively, and input into the initial network model for semi-supervised training. During training, label smoothing regularization constraints and entropy minimization constraints are applied to obtain a network model with weights W2.
[0062] This embodiment elaborates on the above method:
[0063] Step 1: Acquire the modulated signal, which may contain noise labels, and denote it as...
[0064] Step 2: Input the signal into the convolutional neural network for initial training to obtain a network model with weights W1;
[0065] Step 3: Design a confidence-based sample selection method. Compare the predicted probability obtained by inputting the signal into the network model with weight W1 with the confidence level to select the correct label samples and the incorrect label samples.
[0066] Step 4: Treat the correctly labeled samples and incorrectly labeled samples selected in the previous step as labeled samples and unlabeled samples respectively, and perform unsupervised training. During training, apply label smoothing regularization constraints and entropy minimization constraints respectively to obtain a network model with weight W2.
[0067] Step 5: Input the modulated signal to be tested into the network model with weight W2 to complete the signal recognition task.
[0068] In step 1, the modulation signal is N is the number of signal samples, y i ∈{0,1,2……,K-1}, where K is the number of signal categories and the label y is... i Quality cannot be guaranteed and errors may occur.
[0069] In step 2, the signal with dimension 2×d is... The input is used for initial training of the CNN, resulting in a network model with weights W1. The CNN network consists of an encoder composed of two-dimensional convolutional operations, ReLU linear function operations, two-dimensional max pooling operations, Dropout operations, and flattening operations, and a decoder composed of Dense operations, ReLU linear function operations, and Softmax activation function operations.
[0070]
[0071]
[0072] In step 3, a confidence-based sample selection method is designed. The predicted probability obtained by inputting the modulated signal into a network model with weights W1 is compared with the confidence level to filter out correctly labeled samples and incorrectly labeled samples. Specific steps include:
[0073] Step 3-1, Modulated signal with noise The predicted probability is obtained after the final Softmax activation function operation of the CNN. With predictive labels For each category, based on the true label y i With predictive labels Calculate the confidence level C j :
[0074]
[0075] Where I(·) is an indicator function, with a value of 1 when the condition in parentheses is true and a value of 0 when the condition is false. This represents the sum of predicted probabilities when both the true label and the predicted label of a sample belong to that category. This indicates the number of samples whose true or predicted label is this category.
[0076] Step 3-2: After calculating the confidence level, a sample selection method based on the confidence level is used. This method selects samples when the true label and the predicted label are the same. And when its predicted probability is greater than the confidence level (threshold) for the corresponding category This sample is correctly labeled; otherwise, it is an incorrect sample.
[0077] In step 4, the correctly labeled samples and incorrectly labeled samples selected in the previous step are treated as labeled samples and unlabeled samples, respectively, for unsupervised training. During training, label smoothing regularization constraints and entropy minimization constraints are applied respectively. The specific steps include:
[0078] Step 4-1: Treat the selected correctly labeled samples as labeled signal samples and add label smoothing regularization constraints. Label smoothing refers to one-hot encoding of the true labels of the samples. Transform into:
[0079]
[0080] Here, ε is often set to a small constant. Therefore, the label smoothing regularization loss term is:
[0081]
[0082] Converting actual binary hard labels into soft labels with continuous values between 0 and 1 helps reduce the impact of noisy labels on model training, thereby mitigating overfitting. Robust features can be learned from signal samples with noisy labels for identification.
[0083] Step 4-2: Treat the selected incorrectly labeled samples as unlabeled signal samples and add an entropy minimization constraint, specifically:
[0084]
[0085] Entropy minimization constraints can learn as much information as possible from erroneous samples, concentrate the output distribution of CNNs, further improve the reliability of predictions, and enhance robustness.
[0086] Step 4-3, and The two terms are combined using a balancing factor α to form the loss function of the CNN:
[0087]
[0088] Step 4-4: After calculating the loss function value during the forward propagation process, the network weights W2 are updated using the stochastic gradient descent algorithm for backpropagation. Specifically:
[0089]
[0090] in, Let be the parameters at training times t+1 and t respectively, η be the learning rate of the CNN, and t = 1, 2, ..., T, where T is the training period.
[0091] In step 5, after training is complete, the modulated signal to be tested is input into the network model with weight W2 to complete the signal recognition task.
[0092] This invention creatively introduces a confidence-based sample selection method to achieve adaptive sample selection of modulated signals with noise labels. Furthermore, it applies different constraints to the selected samples for semi-supervised training, improving the model's robustness. Moreover, this invention is applicable to signals with different modulation types and exhibits high accuracy and robustness.
[0093] Example 2:
[0094] This invention discloses a robust modulation signal recognition system based on sample selection and semi-supervised learning, the system comprising:
[0095] The signal acquisition module is used to acquire the modulated signal to be tested.
[0096] The signal recognition module is used to input the modulated signal to be tested into a pre-trained network model with weights W2 to identify the signal;
[0097] The network model with weight W2 is obtained by semi-supervised training, where correctly labeled samples and incorrectly labeled samples are treated as labeled samples and unlabeled samples, respectively, and label smoothing regularization constraints and entropy minimization constraints are applied during training.
[0098] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A robust modulation signal recognition method based on sample selection and semi-supervised learning, characterized in that, include: Acquire the modulation signal to be tested; The modulated signal to be tested is input into the pre-trained weights. The network model identifies the type of modulation signal; Wherein, the weight is The network model is obtained by semi-supervised training, where correctly labeled samples and incorrectly labeled samples are treated as labeled samples and unlabeled samples, respectively, and label smoothing regularization constraints and entropy minimization constraints are applied during training. The weight is The training of the network model includes: Acquire modulated signals with noise labels; The modulated signal with noise labels is input into the convolutional neural network for initial training, resulting in weights of... The network model; The modulated signal with noise label is input with weights of The network model is used to obtain the predicted probability, and the confidence level of each category is calculated based on the predicted probability. The confidence-based sample selection method compares the obtained predicted probabilities with the confidence scores of each category to select the correct-labeled samples and the incorrect-labeled samples. Correctly labeled samples and incorrectly labeled samples are treated as labeled samples and unlabeled samples, respectively, and input into the initial network model for semi-supervised training. During training, label smoothing regularization constraints and entropy minimization constraints are applied, resulting in weights of... The network model; The modulation signal is , The number of signal samples. For modulated signal samples, , Number of signal categories; The confidence-based sample selection method compares the obtained predicted probabilities with the confidence scores of each category to filter out correctly labeled samples and incorrectly labeled samples, including: Acquired modulated signals with noise tags The predicted probability is obtained after the final Softmax activation function operation of the CNN network. With predictive labels For each category, based on the actual label With predictive labels Calculate the confidence level for each category ; Based on the calculated confidence level, a confidence-based sample selection method is used. When the true label of a sample is the same as the predicted label and its predicted probability is greater than the confidence level of the corresponding category, the sample is a correctly labeled sample; otherwise, it is an incorrect sample.
2. The robust modulation signal recognition method based on sample selection and semi-supervised learning according to claim 1, characterized in that, The modulated signal with noise labels is input into the convolutional neural network to initially obtain weights. The network models include: Dimension is signal Input the CNN network for initial training. Given the signal sample length, the weights are obtained as follows: The network model.
3. The robust modulation signal recognition method based on sample selection and semi-supervised learning according to claim 1, characterized in that, The confidence level The expression is: ； in This is an indicator function; the value is 1 when the condition in parentheses is true and 0 when the condition is false. This represents the sum of predicted probabilities when both the true label and the predicted label of a sample belong to that category. This indicates the number of samples whose true or predicted label is this category.
4. The robust modulation signal recognition method based on sample selection and semi-supervised learning according to claim 1, characterized in that, The process involves treating correctly labeled samples and incorrectly labeled samples as labeled and unlabeled samples, respectively, and inputting them into the initial network model for semi-supervised training. During training, label smoothing regularization constraints and entropy minimization constraints are applied, resulting in weights... The network models include: The correctly labeled samples selected are considered as labeled signal samples, and a label smoothing regularization constraint is added. Label smoothing refers to one-hot encoding of the true label of the sample. Transform into: ； in As a relatively small constant, the label smoothing regularization loss term is obtained as follows: ； Convert the actual binary hard labels into soft labels with continuous values between 0 and 1; Treating the selected incorrectly labeled samples as unlabeled signal samples, an entropy minimization constraint is added, specifically: ； Will and Two items using a balance factor Combined, these form the loss function of a CNN: ； After calculating the loss function value through forward propagation, backpropagation is performed to update the network weights. The weights are obtained as follows The network model is as follows: ； in, , The first , The network weights during the second training iteration. The learning rate of the CNN. ,in For training cycles.
5. The robust modulation signal recognition method based on sample selection and semi-supervised learning according to claim 4, characterized in that, The backpropagation update of network weights is performed. The stochastic gradient descent algorithm is used.
6. The robust modulation signal recognition method based on sample selection and semi-supervised learning according to claim 1, characterized in that, The structure of the CNN network includes: An encoder consisting of two-dimensional convolution operations, ReLU linear function operations, two-dimensional max pooling operations, Dropout operations, and flattening operations; and a decoder consisting of Dense operations, ReLU linear function operations, and Softmax activation function operations.
7. A system for implementing the robust modulation signal recognition method based on sample selection and semi-supervised learning as described in any one of claims 1-6, characterized in that, The system includes: The signal acquisition module is used to acquire the modulated signal to be tested. The signal recognition module is used to input the modulated signal to be tested into a pre-trained weighted array. The network model identifies the signal; Wherein, the weight is The network model is obtained by semi-supervised training, where correctly labeled samples and incorrectly labeled samples are treated as labeled samples and unlabeled samples, respectively, and label smoothing regularization constraints and entropy minimization constraints are applied during training.