An automatic modulation recognition method based on high signal-to-noise ratio pre-training
Patent Information
- Application Number
- CN202311571996.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-11-23
- Publication Date
- 2026-10-30
- Estimated Expiration
- 2043-11-23
AI Technical Summary
[0006]针对现有深度学习自动调制识别方法，将所有类型、所有信噪比条件的数据同时送入网络进行训练，导致低信噪比信号影响网络特征提取性能，识别精度下降的问题，本发明的主要目的是提供一种基于高信噪比预训练的自动调制识别方法，充分利用不同信噪比数据进行特征提取与网络训练，利用高信噪比数据重点增强网络对于不同调制方式特征的表征能力，再利用低信噪比数据增强网络的泛化能力，从而提升调制识别系统性能，并进一步用于认知无线电通信系统，为解调技术及后续处理提供便利
[0029]1、本发明公开的一种基于高信噪比预训练的自动调制识别技术，通过利用高信噪比数据预训练并加入低信噪比数据进行微调，降低训练过程中低信噪比数据对于神经网络识别精度的影响，从而提高自动调制识别技术的识别准确率。
Smart Images
Figure CN117688423B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to an automatic modulation recognition method based on high signal-to-noise ratio pre-training, belonging to the field of communication signal recognition technology. Background Technology
[0002] Modulation identification of communication signals refers to the technique of processing received signals and determining the modulation type using certain specific parameters of the signal under noisy and multi-signal-type conditions. Traditional automatic modulation identification is mainly divided into likelihood estimation-based methods and feature-based methods. Although relatively simple to operate, both have significant limitations. For example, the former theoretically guarantees that the identification result is optimal under the Bayesian minimum misjudgment cost criterion, but in practical applications, the optimal solution is limited by computational complexity, and a suboptimal solution is often sought to reduce complexity. The latter's selection of features directly affects the design and performance of the identifier and is greatly influenced by human factors.
[0003] Considering the limitations of the methods mentioned above, in recent years, with the development of deep learning methods, a large amount of research has focused on automatic modulation recognition methods based on deep learning. Modulation recognition is essentially a classification problem; therefore, the idea of using deep learning methods to train classifiers can easily be extended to solve this problem. During training on large amounts of data, deep learning methods gradually adjust network parameters through backpropagation, enhancing the network's ability to represent features and non-linearly mapping the input target from the original domain to a new dimension. Its key feature is that deep networks automatically extract significantly different features from the original input that are helpful for classification. Therefore, compared to traditional methods, this method eliminates the need for manual feature selection and complex signal preprocessing; the neural network automatically learns relevant features during training. Furthermore, because the network can learn some low-level features that are transparent to humans, deep learning methods can achieve better results than traditional methods.
[0004] Currently, there are several typical deep learning-based automatic modulation recognition methods. Convolutional neural networks (CNNs) were first applied to automatic modulation recognition tasks to extract spatial features of modulated signals. However, simple CNN models cannot fully extract representative features of various modulation signals. Therefore, improved versions of CNNs used in computer vision have been applied to modulation recognition. For example, residual neural networks trained with in-direction / orthogonal signals as input have achieved classification accuracy superior to ordinary CNNs. Another commonly used network structure is the recurrent neural network (RNN). Wireless communication modulation signals not only possess spatial correlation features but also carry temporal correlation features. RNNs can effectively extract temporal features, thus RNN-based methods have achieved better results than convolutional structures in automatic modulation recognition tasks. Based on the advantages of RNNs and CNNs in extracting features across different dimensions, related research has further proposed hybrid models combining the two to further improve the performance of modulation recognition classifiers. For example, the PET-CGDNN network structure first uses convolutional layers to extract spatial features of the modulation signal, and then inputs the obtained feature map into a gated recurrent unit for temporal feature extraction, achieving high recognition accuracy among current deep learning modulation recognition algorithms.
[0005] However, while the aforementioned deep learning-based automatic modulation recognition methods improve model accuracy from the perspectives of network structure and training parameters, they fail to consider the inherent characteristics of the modulated signal data itself. By simultaneously feeding data of all types and signal-to-noise ratios into the network for training, low signal-to-noise ratio signals negatively impact the network's feature extraction performance, leading to a decrease in recognition accuracy. How to further improve classification accuracy based on the inherent characteristics of the signal itself is a problem that requires further research. Summary of the Invention
[0006] To address the problem that existing deep learning-based automatic modulation recognition methods simultaneously feed data of all types and signal-to-noise ratios (SNR) into the network for training, resulting in low SNR signals affecting network feature extraction performance and reducing recognition accuracy, the main objective of this invention is to provide an automatic modulation recognition method based on high SNR pre-training. This method fully utilizes data with different SNRs for feature extraction and network training, leveraging high SNR data to enhance the network's ability to represent features of different modulation schemes, and then using low SNR data to enhance the network's generalization ability. This improves the performance of the modulation recognition system and can be further applied to cognitive radio communication systems, facilitating demodulation techniques and subsequent processing.
[0007] The objective of this invention is achieved through the following technical solution:
[0008] This invention discloses an automatic modulation recognition technology based on high signal-to-noise ratio (SNR) pre-training. First, signals are grouped according to prior knowledge, with easily confused signals grouped together. Then, a main classifier and sub-classifiers are trained and cascaded. The main classifier distinguishes signals from different groups, while the sub-classifiers distinguish easily confused signals within each group. For model training, signals in the training set are divided according to their SNR characteristics, and high SNR data is used for pre-training. Subsequently, lower SNR data is gradually added to the training set, and the classifier is continuously fine-tuned until all data is included, completing the training process. The cascaded method combines the outputs of the sub-classifier and the main classifier, thereby improving recognition accuracy and enhancing the performance of the modulation recognition system.
[0009] This invention discloses an automatic modulation recognition technology based on high signal-to-noise ratio pre-training, comprising the following steps:
[0010] Step 1: Based on the target task and previous research results, obtain easily confused categories, and group the modulation categories in the multi-classification task accordingly;
[0011] After obtaining the automatic modulation recognition task, based on theoretical analysis and previous research results, the easily confused categories in the task are identified. All modulation types C in the modulation recognition task are divided into k groups: {G1, G2, ..., G...} k This allows for the segmentation of difficult-to-distinguish modulation categories using subnetworks. Each group of G... k It contains one or more modulation types, groups easily confused with each other into the same group, and groups that are not easily confused with all other types into a separate group. There is no overlap between groups and the union of k groups constitutes the set of all modulation types C.
[0012] Step 2: Based on the modulation type labels of the data and the modulation type grouping obtained in Step 1, split the dataset into groups;
[0013] For each modulation type G1, G2, ..., G obtained in step one k If a group contains multiple modulation types with significant confusion, then all modulation signals corresponding to the modulation types in that group should be extracted from the dataset to construct a sub-dataset; if a group contains only one modulation type, then it is not necessary to construct a sub-dataset.
[0014] Step 3: Divide one subset of the dataset from Step 2 into a training set, a validation set, and a test set according to the specified proportions;
[0015] The automatic modulation recognition technology described in this invention is a supervised learning technique based on deep neural networks, requiring the dataset to have signal-to-noise ratio (SNR) and modulation scheme labels. The training set is used to train the deep neural network as a classifier, the validation set is used to evaluate the recognition accuracy during training, and the test set is used to evaluate the overall recognition accuracy of the classifier after training.
[0016] Step 4: Divide the training set obtained in Step 3 into several subsets according to the signal-to-noise ratio label;
[0017] The training set D obtained in step three is split into N subsets D0, D1, D2, ..., Dn according to the signal-to-noise ratio from high to low. N D0 contains data with a high signal-to-noise ratio, which is the pre-training data. Due to the influence of noise, the data quality decreases sequentially.
[0018] Step 5: Pre-train the sub-classifier using a high signal-to-noise ratio dataset;
[0019] The randomly initialized neural network classifier θ0 is pre-trained using the training set D0, and sufficient training is performed until the network reaches convergence on the dataset D0. The training hyperparameters need to be set and adjusted according to the data type and network structure to achieve optimal pre-training results.
[0020] Step 6: Add other subsets to the training set and fine-tune the previously obtained neural network using the synthesized training set to finally obtain a sub-classifier;
[0021] A subset D t (t = 1, 2, ..., N) are added to the existing training set, and the newly generated dataset is {D0, D1, ..., DN}. t Using this dataset, the neural network θ t-1 After fine-tuning and sufficient training, the neural network performs well on the dataset {D0,D1,…,D}. t The neural network θ reaches the convergence condition and is thus obtained. t Repeat the above operations until the set of all training sets D = {D0, D1, ..., D...} is obtained. N The fine-tuned neural network θ N This is a subclassifier for a group.
[0022] Step 7: Complete the training of all sub-classifiers;
[0023] For the modulation types G1, G2, ..., G obtained in step one k For groups containing multiple modulation types with severe confusion, the sub-classifier is trained using the corresponding subset dataset obtained in step two. Steps three through six are repeated to obtain all high-performance sub-classifiers to distinguish the corresponding modulation signal groups.
[0024] Step 8: Complete the training of the main classifier;
[0025] Modify the modulation type labels in the original dataset according to the grouping obtained in step one, changing the modulation type to the group number, i.e., 1, 2, ..., k, to obtain the main classifier dataset. Using this dataset, repeat steps three through six to train the group classifier θ. G The classifier takes a modulated signal as input and outputs the corresponding group G1, G2, ..., G... k .
[0026] Step 9: Cascade the main classifier and sub-classifiers, and use the test set from Step 3 to evaluate the classification accuracy of the neural network. After evaluation, an effective neural network is obtained.
[0027] It also includes step ten, applying the neural network obtained in step nine to a cognitive radio communication system, correctly identifying the modulation mode after detecting a signal, improving the recognition accuracy of automatic modulation recognition technology, enhancing the performance of the modulation recognition system, and solving specific engineering problems.
[0028] Beneficial effects:
[0029] 1. The present invention discloses an automatic modulation recognition technology based on high signal-to-noise ratio pre-training. By using high signal-to-noise ratio data for pre-training and adding low signal-to-noise ratio data for fine-tuning, the impact of low signal-to-noise ratio data on the recognition accuracy of the neural network during the training process is reduced, thereby improving the recognition accuracy of the automatic modulation recognition technology.
[0030] 2. The present invention discloses an automatic modulation recognition technology based on high signal-to-noise ratio pre-training. By dividing the task target, training sub-networks and cascaded network structures, the automatic modulation recognition problem of multi-classification is decomposed, so that the above-mentioned automatic modulation recognition method based on high signal-to-noise ratio pre-training still maintains good improvement effect in multi-classification cases. Attached Figure Description
[0031] Figure 1 This is a flowchart illustrating an automatic modulation recognition method based on high signal-to-noise ratio pre-training as described in this invention.
[0032] Figure 2 This is a schematic diagram of a method for decomposing multi-class automatic modulation recognition problems according to the present invention;
[0033] Figure 3 This is a comparison chart of the recognition accuracy of the embodiments of the present invention and the advanced method PET-CGDNN under the same conditions. Detailed Implementation
[0034] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. The technical problems solved by the present invention and its beneficial effects are also described. It should be noted that the described embodiments are only intended to facilitate understanding of the present invention and do not constitute any limitation thereof.
[0035] To make the objectives, technical solutions, and advantages of this invention more readily understood, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment describes a method for modulation identification using the RadioML 2018.01a dataset. The RadioML 2018.01a dataset is a large open-source dataset used for modulation identification and evaluation in communication systems. The dataset generation process considers the most common losses in real-world communication systems, such as delay spread, carrier frequency offset, and thermal noise. The RadioML 2018.01a dataset contains 2,555,904 frames with 24 modulation schemes, organized into 2×1024 dimensional I / Q signals.
[0036] The system parameter settings in this embodiment are shown in the table below:
[0037]
[0038]
[0039] Since the RadioML2018.01a dataset has 24 modulation schemes, to achieve better performance, this embodiment uses the automatic modulation recognition problem decomposition method based on cascaded structure disclosed in this invention to decompose the task in multi-class cases, and uses the automatic modulation recognition technology based on high signal-to-noise ratio pre-training disclosed in this invention to improve recognition accuracy. This embodiment discloses an automatic modulation recognition method based on high signal-to-noise ratio pre-training, such as... Figure 1 , Figure 2 As shown, the specific implementation steps are as follows:
[0040] Step 1: Based on the target task and previous research results, obtain easily confused categories, and group the modulation categories in the multi-classification task accordingly;
[0041] All modulation types C = {C1, C2, ..., C...} in the modulation recognition task corresponding to the RadioML2018.01a dataset are represented by C = {C1, C2, ..., C...}. 24 The modulation scheme is divided into 19 groups, with amplitude-modulated double-sideband signals forming one group G1 = {AM SSB SC, AM SSB WC}; quadrature amplitude modulation signals forming one group G2 = {16QAM, 32QAM, 64QAM, 128QAM, 256QAM}; and the remaining modulation types forming separate groups.
[0042] Step 2: Split the dataset into groups based on the modulation type labels and modulation type groupings.
[0043] Extract all amplitude-modulated double-sideband signals from the original dataset to form an amplitude-modulated double-sideband sub-dataset; extract all quadrature amplitude modulation signals from the original dataset to form a quadrature amplitude modulation sub-dataset; the other categories contain only one type of modulation signal and do not need to be split into sub-datasets.
[0044] Step 3: Divide the amplitude-modulated bilateral band sub-dataset into a training set, a validation set, and a test set.
[0045] The amplitude-modulated double-band sub-dataset was divided into training, validation, and test sets in a 6:2:2 ratio.
[0046] Step 4: Divide the training set into several subsets according to the signal-to-noise ratio label.
[0047] The dataset was divided into four subsets with signal-to-noise ratio labels of 20 to 30 dB, 10 to 20 dB, 0 to 10 dB, and -20 to 0 dB, respectively.
[0048] Step 5: Pre-train the neural network using a high signal-to-noise ratio dataset.
[0049] Using 20 to 30 dB of data, a randomly initialized MCLDNN network is pre-trained according to the hyperparameters shown in the table above. If the loss function does not decrease for 50 consecutive rounds, reaching the early stopping condition, the training is considered to have converged.
[0050] Step 6: Add other subsets to the training set and fine-tune the previously obtained neural network using the synthesized training set.
[0051] Add data subsets from 10 to 20 dB, 0 to 10 dB, and -20 to 0 dB to the training set sequentially. For each subset added, fine-tune the MCLDNN network according to the hyperparameters and settings described in step five. Once all fine-tuning is complete, the amplitude-modulated double-band subclassifier is obtained.
[0052] Step 7: Complete the training of all subclassifiers.
[0053] Replace the dataset in step three with the orthogonal amplitude modulation sub-dataset, and repeat steps three through six to complete the training of the orthogonal amplitude modulation sub-classifier.
[0054] Step 8: Train the main classifier.
[0055] Modify the modulation type labels in the original dataset according to the grouping obtained in step one, change the modulation type to the group number, replace the dataset in step three with the modified dataset, replace the network structure in step five with GRU, and repeat steps three to six to complete the training of the main classifier.
[0056] Step 9: Cascade the main classifier and the sub-classifier to obtain the modulation recognition classifier and test it.
[0057] The main classifier and sub-classifiers are cascaded as follows: The grouping classifier obtained in step eight is used to identify signal groups. If the result is an amplitude-modulated double-sideband signal or a quadrature amplitude modulation signal, the corresponding sub-classifier is used to obtain the specific modulation category of the signal. The final effect is tested using the test set obtained from the original dataset.
[0058] To demonstrate the superiority of the method disclosed in this invention, this embodiment introduces the currently advanced deep learning modulation recognition method PET-CGDNN and compares its test results with those of this embodiment under the same experimental conditions. Figure 3 As shown, since the impact of low signal-to-noise ratio data on the network's feature extraction capability is reduced during the training process, the accuracy of this embodiment is improved under high signal-to-noise ratio. Furthermore, high signal-to-noise ratio pre-training improves the network's recognition capability compared to direct training, thus its accuracy is also better than that of the PET-CGDNN network under low signal-to-noise ratio.
[0059] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is merely a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. An automatic modulation recognition method based on high signal-to-noise ratio pre-training, characterized in that: Includes the following steps, Step 1: Based on the target task and previous research results, obtain easily confused categories, and group the modulation categories in the multi-classification task accordingly; Step 2: Based on the modulation type labels of the data and the modulation type grouping obtained in Step 1, split the dataset into groups; The implementation method for step two is as follows: For each modulation type obtained in step one If a group contains multiple modulation types with significant confusion, then extract all modulation signals corresponding to the modulation types in that group from the dataset to construct a sub-dataset; if a group contains only one modulation type, then it is not necessary to construct a sub-dataset. Step 3: Divide one subset of the dataset from Step 2 into a training set, a validation set, and a test set according to the specified proportions; Step 4: Divide the training set obtained in Step 3 into several subsets according to the signal-to-noise ratio label; The implementation method for step four is as follows: The training set obtained in step three is split into subgroups according to the signal-to-noise ratio from high to low. Subset , It contains data with a high signal-to-noise ratio, which is pre-training data; due to the influence of noise, the data quality decreases sequentially. Step 5: Pre-train the sub-classifier using a high signal-to-noise ratio dataset; The implementation method for step five is as follows: Using the training set For a randomly initialized neural network classifier Pre-training and thorough training allow the network to operate on the dataset. The convergence condition has been met. Training hyperparameters need to be set and adjusted according to the data type and network structure in order to achieve better pre-training results; Step 6: Add other subsets to the training set and fine-tune the previously obtained neural network using the synthesized training set to finally obtain a sub-classifier; The implementation method for step six is as follows: a subset Add to the existing training set, The newly generated dataset is ; Using this dataset to study neural networks After fine-tuning and sufficient training, the neural network performs well on the dataset. The convergence condition is met, thus obtaining the neural network. Repeat the above steps until the entire training set is obtained. The fine-tuned neural network This is a subclassifier for a group; Step 7: Complete the training of all sub-classifiers; Step 8: Complete the training of the main classifier; The implementation method for step eight is as follows: Modify the modulation type labels in the original dataset according to the grouping obtained in step one, changing the modulation type to the group number, i.e. This yields the main classifier dataset; using this dataset, steps three through six are repeated to train the group classifier. The classifier takes a modulated signal as input and outputs the corresponding group of the signal. ; Step 9: Cascade the main classifier and sub-classifiers, and use the test set from Step 3 to evaluate the classification accuracy of the neural network. After evaluation, an effective neural network is obtained.
2. An automatic modulation recognition method based on high signal-to-noise ratio pre-training, characterized in that: It also includes step ten, applying the neural network obtained in step nine to a cognitive radio communication system, correctly identifying the modulation mode after detecting a signal, improving the recognition accuracy of automatic modulation recognition technology, enhancing the performance of the modulation recognition system, and solving specific engineering problems.
3. The automatic modulation recognition method based on high signal-to-noise ratio pre-training as described in claim 1, characterized in that: The implementation method for step one is as follows: After obtaining the automatic modulation recognition task, based on theoretical analysis and previous research results, the easily confused categories in the task are obtained; All modulation types in the modulation identification task Divided into Group: This allows for the division of difficult-to-distinguish modulation categories using subnetworks; where each group It includes one or more modulation types, groups easily confused categories together, and groups categories that are not easily confused with all other categories separately, with no overlap between groups. The union of groups constitutes the set of all modulation types. .
4. The automatic modulation recognition method based on high signal-to-noise ratio pre-training as described in claim 1, characterized in that: The implementation method for step seven is as follows: The modulation type obtained in step one For groups containing multiple modulation types with severe confusion, the sub-classifier is trained using the corresponding subset dataset obtained in step two. Steps three through six are repeated to obtain all high-performance sub-classifiers to distinguish the corresponding modulation signal groups.