An electromagnetic signal recognition method based on distributed training of an electromagnetic recognition model

Abstract

The application provides an electromagnetic signal recognition method based on distributed training of an electromagnetic recognition model, and belongs to the technical field of radio signal processing, and solves the problem of a computing resource bottleneck faced when large-scale electromagnetic signal data is processed. The electromagnetic signal recognition method comprises the following steps: constructing an electromagnetic signal sample set and a distributed architecture; wherein the electromagnetic signal sample set comprises key features and signal type labels of each historical electromagnetic signal; the distributed architecture comprises a plurality of computing nodes and a central node, each computing node maintains an electromagnetic recognition model, and the central node is used for scheduling the computing nodes; the electromagnetic recognition model is trained based on the electromagnetic signal sample set in a distributed manner by using the distributed architecture, so that a trained electromagnetic recognition model is obtained; key features of a to-be-recognized electromagnetic signal are extracted and input into the trained electromagnetic recognition model, and the electromagnetic recognition model is used for predicting and outputting a signal type of the to-be-recognized electromagnetic signal.

Description

Technical Field

[0001] This invention relates to the field of radio signal processing technology, and in particular to an electromagnetic signal recognition method based on a distributed training electromagnetic recognition model. Background Technology

[0002] With the widespread application of electromagnetic signals in many fields, accurately identifying the types of electromagnetic signals is of great significance for ensuring information security, improving communication efficiency, and coping with complex electromagnetic environments. Traditional electromagnetic signal identification methods often employ centralized training models, processing and training all electromagnetic signal sample data on a single computer. However, this approach faces numerous challenges when dealing with massive amounts of electromagnetic signal data. Centralized training requires powerful computing resources to process the large datasets, resulting in high hardware costs and low computational efficiency, making it difficult to meet the real-time requirements of electromagnetic signal identification scenarios.

[0003] Therefore, how to solve the computational resource bottleneck problem faced by existing electromagnetic signal recognition methods when processing large-scale electromagnetic signal data is a key technical challenge that urgently needs to be overcome. Summary of the Invention

[0004] Based on the above analysis, the present invention aims to provide an electromagnetic signal recognition method based on a distributed training electromagnetic recognition model, in order to solve the problem of computational resource bottleneck when processing large-scale electromagnetic signal data.

[0005] This invention discloses an electromagnetic signal recognition method based on a distributed training electromagnetic recognition model, the electromagnetic signal recognition method comprising:

[0006] An electromagnetic signal sample set and a distributed architecture are constructed. The electromagnetic signal sample set includes key features and signal type labels for each historical electromagnetic signal. The distributed architecture includes multiple computing nodes and a central node. Each computing node maintains an electromagnetic identification model, and the central node is used to schedule the computing nodes.

[0007] A distributed architecture is adopted to perform distributed training on the electromagnetic identification model based on the electromagnetic signal sample set, so as to obtain the trained electromagnetic identification model.

[0008] The key features of the electromagnetic signal to be identified are extracted and input into the trained electromagnetic identification model. The electromagnetic identification model then predicts and outputs the signal type of the electromagnetic signal to be identified.

[0009] Based on the above solution, the present invention also makes the following improvements:

[0010] Furthermore, the trained electromagnetic recognition model is obtained through the following operations:

[0011] In each round of training, each computing node initializes the model parameters of the electromagnetic identification model using the global training parameters issued by the central node, and performs multiple training optimizations on the electromagnetic identification model using the adaptive comprehensive loss function to obtain the local training parameters for the current round. The central node updates the global training parameters based on the local training parameters of each computing node for the current round and jumps to the next round of training. This process is repeated multiple times until the training termination condition is met, resulting in a trained electromagnetic identification model.

[0012] Furthermore, during each training round, the computing node performs the following operations to obtain the local training parameters for the current round:

[0013] Each time training is performed, the feature extraction vector of the electromagnetic recognition model for this training is obtained, and the softened vector for this training is obtained after selective softening.

[0014] Based on the softening vector from the first training iteration of the current round and the softening vector from this training iteration, the imbalance constraint loss function for this training iteration is obtained.

[0015] Based on the imbalance constraint loss function and cross-entropy loss function of this training, the adaptive comprehensive loss function of this training is obtained, which is used to update the model parameters for the next training.

[0016] The training is repeated multiple times, and the model parameters updated in the last training iteration are used as the local training parameters for the current iteration.

[0017] Furthermore, during each training process, the computing node selectively softens the feature extraction vector for this training based on the imbalanced set of classification categories in the training sample set according to the electromagnetic signal of the computing node, thus obtaining the softened vector for this training.

[0018] Furthermore, the adaptive synthesis loss function L is expressed as:

[0019] L=λL ce +(1-λ)L asc (1)

[0020] Where λ represents the weighting coefficient, λ∈[0,1], L ce L represents the cross-entropy loss function. asc This represents the unbalanced constraint loss function.

[0021] Furthermore, the imbalance constraint loss function L for the k-th training iteration... asc (k) is represented as:

[0022]

[0023] Where T represents the softening temperature, C imbThis represents the set of classification categories in the training sample set of electromagnetic signals for computing nodes that suffer from imbalanced sample sizes. This represents the j-th element in the softened vector of the k-th training iteration; k ranges from 1 to K, where K represents the number of training iterations in each round.

[0024] further,

[0025]

[0026] Where, p j (k) indicates that the feature extraction vector from the kth training iteration matches C. imb The j-th element that needs to be softened.

[0027] Furthermore, the central node updates the global training parameters based on the local training parameters of each computing node in the current round, and executes:

[0028] The central node uses a distributed aggregation strategy to aggregate the local training parameters of each computing node in the current round to obtain the updated global training parameters.

[0029] Furthermore, in each round of training, the central node verifies the prediction accuracy of the electromagnetic recognition model with the updated global training parameters as model parameters on the electromagnetic signal verification sample set. If the prediction accuracy is better than the historical best accuracy, the corresponding global training parameters are stored as the global best parameters.

[0030] Furthermore, if the fluctuation of the prediction accuracy in multiple consecutive training rounds is less than the preset deviation, or if the preset number of training rounds is reached, then the training termination condition is met, and the electromagnetic recognition model with the globally optimal parameters as the model parameters is taken as the completed electromagnetic recognition model.

[0031] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0032] The electromagnetic signal recognition method based on a distributed training electromagnetic recognition model provided by this invention has the following beneficial effects.

[0033] (1) Improve training efficiency: Through the distributed architecture, the training task of the electromagnetic recognition model is distributed to multiple computing nodes, making full use of the computing resources of each node, greatly shortening the training time, improving training efficiency, and enabling rapid response to electromagnetic signal recognition needs, adapting to application scenarios with high real-time requirements.

[0034] (2) Improve recognition accuracy: To address the problem of imbalanced sample classes, the computing nodes selectively soften the feature extraction vectors during training and construct an adaptive comprehensive loss function by combining the imbalance constraint loss function and the cross-entropy loss function. This effectively alleviates the impact of class imbalance on model training, enabling the model to more accurately identify different types of electromagnetic signals and improve recognition accuracy.

[0035] (3) Enhance the generalization ability of the model: During the distributed training process, each computing node trains and updates the model parameters independently, and the central node then updates the global parameters. This mechanism enables the model to learn the characteristics and patterns of data from different nodes, enhances the model's adaptability to different electromagnetic environments and signal characteristics, improves the model's generalization ability, and enables it to more accurately identify new or unknown electromagnetic signals, thereby improving the overall performance and reliability of electromagnetic signal recognition methods.

[0036] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0037] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0038] Figure 1 A flowchart of an electromagnetic signal recognition method based on a distributed training electromagnetic recognition model provided in an embodiment of the present invention. Detailed Implementation

[0039] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0040] A specific embodiment of the present invention discloses an electromagnetic signal recognition method based on a distributed trained electromagnetic recognition model, the flowchart of which is shown below. Figure 1 As shown, the electromagnetic signal identification method includes the following steps.

[0041] Step S1: Construct an electromagnetic signal sample set and a distributed architecture; wherein, the electromagnetic signal sample set includes the key features and signal type labels of each historical electromagnetic signal; the distributed architecture includes multiple computing nodes and a central node, each computing node maintains an electromagnetic identification model, and the central node is used to schedule each computing node.

[0042] In constructing the electromagnetic signal sample set, key features are extracted for each historical electromagnetic signal to obtain the corresponding key features. Preferably, depending on the characteristics of the electromagnetic signal, key features such as time-domain features and frequency-domain features can be extracted. Common time-domain features include the amplitude characteristics of the electromagnetic signal, such as maximum, minimum, and average values. Furthermore, after converting the electromagnetic signal from the time domain to the frequency domain using methods such as Fourier transform, the frequency distribution characteristics of the electromagnetic signal can be extracted. For example, the signal's center frequency and bandwidth.

[0043] In a distributed architecture, computing nodes, as the core execution units, are numerous and widely distributed. Each computing node possesses certain computing capabilities and can independently run and maintain an electromagnetic identification model. They can be physical servers, virtual machines, or dedicated computing devices, with hardware resources such as processor performance, memory capacity, and storage space configured according to the actual application scenario and performance requirements. The distribution of computing nodes can be geographically dispersed, such as being located in data centers in different geographical locations, or logically dispersed, such as being located in different racks or server clusters within the same data center. The central node plays a crucial coordinating and scheduling role in the entire architecture. It typically has high performance and stability requirements because it needs to handle information interaction and task allocation from multiple computing nodes. The hardware configuration of the central node also needs to be determined based on the number of computing nodes and the complexity of the tasks to ensure efficient scheduling. The central node can be a single server or a highly available cluster of multiple servers to improve system reliability and fault tolerance.

[0044] Step S2: Using a distributed architecture, perform distributed training on the electromagnetic identification model based on the electromagnetic signal sample set to obtain the trained electromagnetic identification model.

[0045] In this embodiment, the training process of the electromagnetic identification model is conducted independently on each computing node. Model training on different computing nodes may produce different results and performance due to factors such as data differences. Furthermore, since the model on each computing node is maintained independently, regular model updates and synchronization operations are required to ensure the identification performance and consistency of the entire distributed architecture. The central node can be responsible for coordinating this process to achieve model updates. Specifically, in step S2, the trained electromagnetic identification model is obtained through the following operations.

[0046] Step S21: In each round of training, each computing node initializes the model parameters of the electromagnetic identification model using the global training parameters issued by the central node, and performs multiple training optimizations on the electromagnetic identification model using the adaptive comprehensive loss function to obtain the local training parameters for the current round.

[0047] Step S22: The central node updates the global training parameters based on the local training parameters of each computing node in the current round, and jumps to the next round of training; the training is repeated for multiple rounds until the training termination condition is met, and the trained electromagnetic recognition model is obtained.

[0048] It should be noted that during the first round of training, the central node can pre-set global training parameters according to the actual situation and distribute them to each computing node. In subsequent rounds of training, the central node distributes the updated global training parameters from the previous round to each computing node, so that each computing node can execute the next round of training accordingly. This means that the starting point of each round of training is the combined result of the training of all computing nodes in the previous round, thus ensuring that the training parameters of the electromagnetic intelligent model are updated and optimized in each round of training.

[0049] Specifically, in step S21, during each round of training, the computing node performs the following operations to obtain the local training parameters for the current round.

[0050] Step S211: For each training session, obtain the feature extraction vector of the electromagnetic recognition model for this training session, and selectively soften it to obtain the softened vector for this training session.

[0051] Preferably, the electromagnetic recognition model selected in this embodiment includes a feature extraction network and a classification part, such as a convolutional neural network (CNN). Common CNNs include LeNet and ResNet (residual network). In specific implementation, the feature extraction network is used to obtain the feature extraction vector for each training iteration of the electromagnetic recognition model. The feature extraction vector for the k-th training iteration is denoted as p(k), and its dimension is equal to the number of classification categories. Generally, a softmax layer is used as the classification part to obtain the probability that the training sample belongs to each classification category.

[0052] Specifically, during each training process, the computing node selectively softens the feature extraction vector for this training based on the imbalanced set of classification categories in the training sample set according to the electromagnetic signal of the computing node, thus obtaining the softened vector for this training.

[0053]

[0054] Among them, C imb p represents the set of classification categories in the training sample set of electromagnetic signals of the computing node that suffers from imbalanced sample size; j (k) indicates that the feature extraction vector from the kth training iteration matches C. imb The j-th element that needs to be softened (i.e., selecting elements from p(k) that meet the set C) imb (the j-th element among all elements); This represents the j-th element in the softening vector of the k-th training iteration; k ranges from 1 to K, where K represents the number of training iterations in each round; T represents the softening temperature.

[0055] For example, suppose there are a total of 10 categories, and the labels for each category are {0, 1, 2, ..., 9}. If the number of samples in categories 2 and 4 is small, i.e., imbalanced, then C imb ={2,4}.

[0056] It should be noted that the dimension of the feature extraction vector of the electromagnetic identification model is equal to the number of classification categories. In the specific implementation, considering that the electromagnetic signal training sample set allocated to the computing nodes may have an imbalance in sample size, which can easily lead to low classification accuracy of the electromagnetic identification model for such samples, this embodiment selectively softens the feature extraction vector for this training to solve the performance bottleneck problem caused by the inconsistent distribution of sample size for different categories during the training of the electromagnetic identification model.

[0057] Step S212: Based on the softening vector from the first training iteration of the current round and the softening vector from this training iteration, obtain the imbalance constraint loss function for this training iteration.

[0058] The imbalance constraint loss function L for the k-th training iteration asc (k) is represented as:

[0059]

[0060] When k is 1 Therefore, the imbalance constraint loss function L in the first training iteration asc (1) Based on The calculation yields the result. And when k takes values ​​from 1 to K, Therefore, the imbalance constraint loss function L for the k-th training iteration asc (k) based on and It is obtained. It should be emphasized that the model parameters of the electromagnetic identification model are initialized using the global training parameters issued by the central node, and the first training is performed. Therefore, Directly related to global training parameters; and (k ranges from 1 to K) The electromagnetic recognition model is trained based on the training parameters obtained from the previous training, and is directly related to the training parameters from the previous training.

[0061] Step S213: Based on the imbalance constraint loss function and cross-entropy loss function of this training, obtain the adaptive comprehensive loss function of this training, which is used to update the model parameters for the next training.

[0062] The adaptive synthesis loss function L is expressed as:

[0063] L=λL ce +(1-λ)L asc (1)

[0064] Where λ represents the weighting coefficient, λ∈[0,1], L ce L represents the cross-entropy loss function. asc Let L represent the unbalanced constraint loss function. As shown above, the cross-entropy loss function L... ce Labels are required; the unbalanced constraint loss function L asc No labels are required; constraints are applied based on the training output of the electromagnetic recognition model.

[0065] Step S214: Repeat the training multiple times, and use the model parameters updated in the last training iteration of the current round as the local training parameters for the current round.

[0066] In step S22, the central node updates the global training parameters based on the local training parameters of each computing node in the current round. The central node uses a distributed aggregation strategy to aggregate the local training parameters of each computing node in the current round to obtain the updated global training parameters.

[0067] Preferably, the central node can employ a distributed aggregation strategy of weighted average aggregation to aggregate the local training parameters of each computing node in the current round. Specifically, the central node verifies the classification accuracy of the electromagnetic recognition model using the local parameters of each computing node in the current round as model parameters on the electromagnetic signal validation set, determines the weighting ratio of the local training parameters of each computing node in the current round based on the classification accuracy, and performs weighted average aggregation of the local training parameters of each computing node in the current round according to the determined weighting ratio, thereby obtaining the updated global training parameters.

[0068] To achieve better training results, in each training round, the central node updates the electromagnetic signal sample sets of each computing node and simultaneously distributes the updated electromagnetic signal sample sets and updated global training parameters to each computing node. For example, the processing capability of each computing node can be evaluated based on its hardware configuration (CPU, memory, GPU, etc.) and historical performance data (processing time and efficiency of each computing node in previous training rounds) to calculate and update the sample size processed by each computing node, thereby updating the electromagnetic signal sample sets of each computing node. In specific implementation, appropriate evaluation criteria can be selected according to the specific circumstances of the distributed architecture; this embodiment does not limit this selection.

[0069] Furthermore, in each training round, the central node verifies the prediction accuracy of the electromagnetic recognition model, using the updated global training parameters as model parameters, on the electromagnetic signal verification sample set. If the prediction accuracy is better than the historical best accuracy, the corresponding global training parameters are stored as the global optimal parameters. If the fluctuation of the prediction accuracy over multiple consecutive training rounds is less than a preset deviation, or if a preset number of training rounds is reached, the training termination condition is met, and the electromagnetic recognition model using the global optimal parameters as model parameters is taken as the completed electromagnetic recognition model.

[0070] Step S3: Extract the key features of the electromagnetic signal to be identified and input them into the trained electromagnetic identification model. The electromagnetic identification model then predicts and outputs the signal type of the electromagnetic signal to be identified.

[0071] In the specific implementation process, the key features of the electromagnetic signal to be identified are extracted in the same way as the historical electromagnetic signals in step S1, and the key features are input into the trained electromagnetic recognition model. Then, the trained electromagnetic recognition model can predict and output the signal type of the electromagnetic signal to be identified.

[0072] In summary, the electromagnetic signal recognition method based on a distributed training electromagnetic recognition model provided in this embodiment realizes an electromagnetic signal recognition model based on distributed training by constructing an electromagnetic signal sample set and a distributed architecture. In the distributed architecture, multiple computing nodes work collaboratively, each maintaining an electromagnetic recognition model, which is scheduled by a central node. This architecture fully utilizes the advantages of distributed computing, distributing electromagnetic signal sample data across various computing nodes for processing, greatly reducing the computational burden on individual nodes, significantly improving the efficiency of model training, and enabling rapid response to the training needs of large-scale electromagnetic signal data.

[0073] During training, each computing node employs an adaptive comprehensive loss function to optimize the electromagnetic recognition model multiple times. This adaptive comprehensive loss function integrates the imbalance constraint loss function and the cross-entropy loss function. The imbalance constraint loss function is calculated based on the softening vector obtained from each training iteration, which in turn is obtained by selectively softening the feature extraction vector. This selective softening operation fully considers the imbalanced classification sets in the electromagnetic signal training sample sets of the computing nodes, effectively mitigating the impact of class imbalance on model training. This allows the model to focus more on the signal features of a few classes during the learning process, thereby improving the overall recognition accuracy of various electromagnetic signals.

[0074] Ultimately, the electromagnetic identification model obtained through multiple rounds of distributed training can accurately predict and output the signal type after extracting the key features of the electromagnetic signal to be identified. This provides an efficient and reliable solution for the rapid and accurate identification of electromagnetic signals, and has good application prospects and practical value.

[0075] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0076] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. An electromagnetic signal recognition method based on a distributed training electromagnetic recognition model, characterized in that, The electromagnetic signal identification method includes: An electromagnetic signal sample set and a distributed architecture are constructed. The electromagnetic signal sample set includes key features and signal type labels for each historical electromagnetic signal. The distributed architecture includes multiple computing nodes and a central node. Each computing node maintains an electromagnetic identification model, and the central node is used to schedule the computing nodes. A distributed architecture is adopted to perform distributed training on the electromagnetic identification model based on the electromagnetic signal sample set, so as to obtain the trained electromagnetic identification model. Extract the key features of the electromagnetic signal to be identified and input them into the trained electromagnetic identification model. The electromagnetic identification model then predicts and outputs the signal type of the electromagnetic signal to be identified. The trained electromagnetic identification model is obtained through the following operations: During each training round, each computing node initializes the model parameters of the electromagnetic recognition model using the global training parameters issued by the central node, and performs multiple training optimizations on the electromagnetic recognition model using the adaptive comprehensive loss function to obtain the local training parameters for the current round. The central node updates the global training parameters based on the local training parameters of each computing node for the current round and jumps to the next round of training. This process is repeated multiple times until the training termination condition is met, resulting in a trained electromagnetic recognition model. During each training round, the computing node performs the following operations to obtain the local training parameters for the current round: Each time training is performed, the feature extraction vector of the electromagnetic recognition model for this training is obtained, and the softened vector for this training is obtained after selective softening. Based on the softening vector from the first training iteration of the current round and the softening vector from this training iteration, the imbalance constraint loss function for this training iteration is obtained. Based on the imbalance constraint loss function and cross-entropy loss function of this training, the adaptive comprehensive loss function of this training is obtained, which is used to update the model parameters for the next training. The training is repeated multiple times, and the model parameters updated in the last training iteration are used as the local training parameters for the current iteration.

2. The electromagnetic signal recognition method based on a distributed training electromagnetic recognition model according to claim 1, characterized in that, During each training process, the computing node selectively softens the feature extraction vector of this training based on the imbalanced set of classification categories in the training sample set according to the electromagnetic signal of the computing node, thus obtaining the softened vector of this training.

3. The electromagnetic signal recognition method based on a distributed training electromagnetic recognition model according to claim 2, characterized in that, Adaptive integrated loss function Represented as: (1) in, Indicates the weighting coefficient. , Represents the cross-entropy loss function. This represents the unbalanced constraint loss function.

4. The electromagnetic signal recognition method based on a distributed training electromagnetic recognition model according to claim 3, characterized in that, No. Imbalanced loss function during training Represented as: (2) in, Indicates the softening temperature. This represents the set of classification categories in the training sample set of electromagnetic signals for computing nodes that suffer from imbalanced sample sizes. Indicates the first The softening vector in the training iteration is the first one. One element; Take 1 to , This indicates the number of training sessions in each round.

5. The electromagnetic signal recognition method based on a distributed training electromagnetic recognition model according to claim 4, characterized in that, (3) in, Indicates the first The feature extraction vectors from the second training session were matched to The first step that needs to be softened Each element.

6. The electromagnetic signal recognition method based on a distributed training electromagnetic recognition model according to any one of claims 1-5, characterized in that, The central node updates the global training parameters based on the local training parameters of each computing node in the current round, and executes: The central node uses a distributed aggregation strategy to aggregate the local training parameters of each computing node in the current round to obtain the updated global training parameters.

7. The electromagnetic signal recognition method based on a distributed training electromagnetic recognition model according to claim 6, characterized in that, In each round of training, the central node verifies the prediction accuracy of the electromagnetic recognition model, which uses the updated global training parameters as model parameters, on the electromagnetic signal verification sample set. If the prediction accuracy is better than the historical best accuracy, the corresponding global training parameters are stored as the global best parameters.

8. The electromagnetic signal recognition method based on a distributed training electromagnetic recognition model according to claim 7, characterized in that, If the fluctuation of the prediction accuracy in multiple consecutive training rounds is less than the preset deviation, or if the preset number of training rounds is reached, then the training termination condition is met, and the electromagnetic recognition model with the globally optimal parameters as the model parameters is taken as the completed electromagnetic recognition model.

Images