Identification model autonomous evolution method oriented to electromagnetic environment
By employing an autonomous evolution method for recognition models oriented towards electromagnetic environments, combined with meta-learning and domain generalization techniques, a complex deep separable convolutional neural network and a visual Transformer module are constructed. This solves the problem of declining recognition performance of traditional machine learning in dynamic and complex electromagnetic environments, and achieves efficient recognition and improved adaptability of the model in complex environments.
Patent Information
- Application Number
- CN202511051374.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-29
- Publication Date
- 2025-11-14
AI Technical Summary
Traditional machine learning struggles to adapt to changes in target information in dynamic and complex electromagnetic environments, leading to a decline in recognition performance. Existing autonomous evolution methods also lack the ability to adapt to changing environments.
An autonomous evolution method for identification models oriented towards electromagnetic environments is adopted. By representing electromagnetic signals in multiple domains and dimensions, and combining meta-learning and domain generalization techniques, a complex deep separable convolutional neural network and a visual Transformer module are constructed for feature extraction and radiation source feature pre-training. The autonomous evolution of the model is achieved by using cross-modal contrastive learning and meta-learning inner and outer loop optimization strategies.
It significantly improves the adaptability and scalability of electromagnetic signal recognition models in dynamic and complex environments, enhances recognition efficiency and robustness, and is suitable for scenarios with small samples and distribution bias.
Smart Images
Figure CN120951084A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of electromagnetic signal processing technology and is a method for the autonomous evolution of recognition models in electromagnetic environments. It is particularly suitable for the rapid recognition and processing of diverse electromagnetic signals in dynamic and complex electromagnetic environments, providing support for the autonomous evolution of recognition models. Background Technology
[0002] Machine learning is widely used in electromagnetic signal processing, especially in electromagnetic signal recognition. In recent years, machine learning-based signal processing techniques have overcome the shortcomings of traditional signal analysis techniques, helping to improve the recognition capabilities of massive signal data. However, traditional machine learning cannot effectively adapt to the changes in the distribution, behavior, scale, and complexity of target information brought about by complex electromagnetic environments, posing challenges to the extraction and recognition of target information based on machine learning.
[0003] With the rapid development of artificial intelligence theories and methods, new tools applicable to the electromagnetic field are emerging in an endless stream. For example, deep learning, by constructing models with deeper layers and more parameters, can accurately characterize and memorize complex electromagnetic big data. Deep learning can not only automatically extract target features from complex electromagnetic data, but also learn target representations in an "end-to-end" manner, realizing an integrated process of electromagnetic target detection, localization, and recognition, showing broad application prospects in the electromagnetic field.
[0004] Although deep learning technology has improved the automation of signal recognition, there are still gaps in its practical application in dynamic and complex electromagnetic environments. It is prone to learning "false" patterns when there is signal interference and distortion; it is prone to overfitting when the sample complexity is low; and its predictive performance will significantly decrease when the data distribution drifts and deviates.
[0005] Therefore, methods related to autonomous evolution have gradually become one of the research hotspots in the field of electromagnetic identification. Autonomous evolutionary learning methods mainly include incremental learning, transfer learning, and meta-learning. Incremental learning, as the earliest continuous learning method, generates a series of models on a given training data stream, suitable for continuous and real-time processing of large-scale data streams, but it often emphasizes streaming scenario settings or theoretical concept drift. Transfer learning enables continuous learning in changing scenarios, helping new environments learn tasks by transferring existing knowledge, but its effectiveness depends on the consistency of invariants such as space, instances, and features. Meta-learning, as a new research branch, can leverage previously accumulated experience to learn quickly in new environments, but typical meta-learning methods lose their adaptability to changing environments after pre-training.
[0006] Choosing a suitable autonomous evolution method can effectively mitigate signal distortion and data drift in dynamic and complex electromagnetic environments, enabling the model to comprehensively and deeply represent the original electromagnetic signal in such environments, which greatly helps improve the recognition efficiency of electromagnetic signals. Therefore, to address the technical problem of the rapid decline in recognition accuracy caused by data drift in existing recognition technologies under dynamic and complex electromagnetic environments, a self-evolution method for recognition models that is highly adaptable and scalable to electromagnetic environments is needed. Summary of the Invention
[0007] This invention addresses the shortcomings of existing technologies by proposing an autonomous evolution method for electromagnetic environment recognition models. This invention enhances the generalization ability of electromagnetic signal recognition models in multi-noise scenarios through multi-domain, multi-dimensional representation and enhancement of electromagnetic signal data; combining meta-learning and domain generalization techniques, it optimizes the model to meet the signal recognition needs of complex and variable electromagnetic environments, improving the model's ability to recognize target signals in complex and variable electromagnetic environments with high radiation source characteristic similarity.
[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0009] An autonomous evolution method for identification models in electromagnetic environments includes the following steps:
[0010] Step 1: Set up the processor to monitor the electromagnetic environment in the target area, collect data on the target electromagnetic signal, obtain the IQ data of the target electromagnetic signal and perform local buffering processing.
[0011] Step 2: Extract IQ data from the local cache, add composite noise to it, and use a method based on discrete wavelet decomposition and reconstruction to convert the IQ data into a dual-channel narrow-side two-dimensional image; then apply a random mask to the dual-channel narrow-side two-dimensional image, capture the graph characteristics of the target electromagnetic signal through a topology modeling method based on finite fixed graph mapping, and delete the edges or perturbation nodes of the graph structure with a set probability value θ1.
[0012] Step 3: Construct a complex depth separable convolutional neural network model for IQ data and extract IQ data features; extract features from the two-dimensional image using the visual Transformer module to obtain two-dimensional image features; extract features from the graph structure using the DiffPool module to obtain graph structure features; and then, based on the IQ data features, two-dimensional image features, and graph structure features, complete the construction of the recognition model using a cross-modal contrastive learning radiation source feature pre-training framework.
[0013] Step 4: Use the constructed recognition model as a meta-learner, and update the meta-learner using a metric strategy to achieve target recognition based on meta-learning in the case of small samples.
[0014] Step 5: Based on the meta-learning inner and outer loop collaborative optimization strategy, fine-tune some parameters in the recognition model to achieve unsupervised autonomous evolution of the recognition model;
[0015] A method for autonomous evolution of identification models oriented towards electromagnetic environments was developed.
[0016] Furthermore, in step 3, feature extraction is performed on the two-dimensional image using the visual Transformer module, specifically including the following steps:
[0017] Step 201: Apply a vertical asymmetric segmentation strategy to the two-dimensional image, dynamically adjust the segment size to adapt to the structural characteristics of the narrow-side image, and obtain the corresponding image segments;
[0018] Step 202: Use 2D position coding to independently encode the row and column position information of the image blocks, and add the 2D position coding result to the tile embedding vector of the image block element by element to obtain the input sequence of the visual Transformer module;
[0019] Step 203: Introduce a multi-head attention mechanism in the visual Transformer module, where the multi-head attention mechanism has M attention heads; split the input sequence of the visual Transformer module into approximate and detail components by channel, generate corresponding Query, Key, and Value matrices, and introduce positional encoding; calculate positional correlation through linear projection, use the positional correlation as a positional bias function, calculate the attention score after incorporating the positional bias function, and use it as the output of a single attention head; then concatenate the outputs of the M attention heads, perform linear projection, and perform residual connection with the input sequence;
[0020] Step 204: Set the position consistency loss, minimize the position consistency loss, obtain the parameters of the visual Transformer module, and apply the model parameters to extract the two-dimensional image features from the two-dimensional image.
[0021] Furthermore, in step 3, based on IQ data features, two-dimensional image features, and graph structure features, a pre-training framework for radiation source features based on cross-modal contrastive learning is used to complete the construction of the recognition model. This specifically includes the following steps:
[0022] Step 301: Map the IQ data features, two-dimensional image features and graph structure features to a unified metric space. Each feature processing module independently uses a projection head composed of a two-layer perceptron to perform L2 normalization and output the normalized modal feature vector.
[0023] Step 302: Establish positive and negative phase pairs for modal feature vectors, and use cosine similarity to measure the consistency of features between modalities to obtain the similarity of cross-modal features;
[0024] Step 303: Utilize the similarity between modal features to calculate the cross-entropy loss of information for all modal pairs, which is used as the total loss value of the pre-training framework for radiation source features, and perform cross-modal comparative learning of radiation source features.
[0025] Step 304: Deploy the radiation source feature pre-training framework, introduce the classification head, concatenate the three normalized modal feature vectors from step 301, use the fused features as input to the classification head, obtain the recognition result, and complete the construction of the recognition model. The recognition model specifically includes a complex depth separable convolutional neural network model, a visual Transformer module, a DiffPool module, and a classification head consisting of a fully connected layer.
[0026] Furthermore, step 4 updates the meta-learner using a metric strategy, specifically including the following steps:
[0027] Step 401: Taking target recognition based on small sample scenarios as the meta-task, randomly select data containing N radiation source types from a dataset that follows a task distribution as the meta-task dataset, where each type has K labeled instances. Divide the meta-task dataset into a support set and a query set according to a set ratio, and perform scenario training using a model-independent meta-learning strategy; where K < <N;
[0028] Step 402: Divide the entire training into multiple meta-tasks. For a single meta-task, construct a meta-learner to extract support set features, use the similarity score between support sets as the loss function of the support set, calculate one step of gradient descent to obtain the expected parameters of the meta-learner, and measure the similarity between the support set and the query set based on the support set, query set, and expected parameters to obtain the similarity loss between the support set and the query set as the loss function of a single meta-task. Then, calculate the sum of the loss functions under each meta-task as the training loss for the entire scenario training, and use stochastic gradient descent to train the meta-learner.
[0029] Step 403: Extract data from the unknown categories that follow the test task distribution as the test set, and perform meta-testing on the meta-learner that has been trained in step 402 to complete target recognition in the case of few samples.
[0030] Furthermore, in step 5, the recognition model is fine-tuned and optimized based on the meta-learning inner and outer loop collaborative optimization strategy, specifically including the following steps:
[0031] Step 501: Using low-rank adaptation and dynamic gradient freezing strategies, low-rank matrix bypasses are inserted into the visual Transformer module and DiffPool module, updating only some parameters in the module; and the bottom temporal convolutional layers of the complex depth separable convolutional neural network model are frozen, with only the top complex fully connected layer being fine-tuned.
[0032] Step 502 introduces a dynamic class reweighting mechanism to quantify sample importance. A learnable weight generator adaptively adjusts the weights of each class in the loss function. An improved model-independent meta-learning strategy is adopted, updating the recognition model in both inner and outer loops. Specifically, this includes:
[0033] The support set of the meta-task is used as the input of the inner loop. The parameters of the feature extractor are frozen, a gradient descent step is performed on the support set, the classification head parameters are updated, the model is adapted, and the recognition model in the inner loop is updated. The feature extractor includes a complex depth separable convolutional neural network model, a visual Transformer module, and a DiffPool module.
[0034] The adapted model and query set are used as input to the outer loop. A meta-optimization strategy is adopted to jointly optimize the parameters of the feature extractor and the classification head, minimize the query set loss, update the global parameters of the recognition model, and complete the update of the recognition model in the outer loop.
[0035] Compared with the prior art, the present invention has the following advantages:
[0036] The proposed autonomous evolution method for electromagnetic environment recognition models significantly improves the applicability of recognition models through lightweight deployment and enhances their scalability through a meta-learning fine-tuning framework. It boasts advantages such as wide adaptability and strong scalability, while also offering innovations in electromagnetic signal representation, modeling, and learning. In terms of representation, it explores multi-dimensional signal descriptions across space, time, and frequency, and applies different signal enhancement strategies to enhance the feature representation of signal data based on these multi-dimensional descriptions. Regarding modeling, this invention proposes an innovative complex-depth separable convolutional neural network architecture. This architecture, by inheriting a series of complex operation units with rotating gate designs, can learn richer and more robust signal representations. More importantly, by employing complex-depth separable convolution technology, this improvement significantly reduces model complexity, facilitating lightweight deployment. In terms of learning, meta-learning is introduced into the field of electromagnetic signal recognition to address the simultaneous issues of small sample sizes and distribution bias in practical applications. Multiple related meta-tasks are constructed to improve domain-adaptive classification capabilities under conditions of very few labeled instances. This invention provides a new solution for electromagnetic target recognition in dynamic and complex electromagnetic environments. Attached Figure Description
[0037] Figure 1 This is the overall flowchart of the present invention. Detailed Implementation
[0038] Reference Figure 1A method for autonomous evolution of recognition models in electromagnetic environments is proposed, comprising a series of processing steps including electromagnetic signal data reception and local caching, conversion of two-dimensional images and capture of graph features of electromagnetic signals, extraction of IQ data features, two-dimensional image features and graph structure features, construction of the recognition model, realization of target recognition based on meta-learning, and realization of unsupervised autonomous evolution of the recognition model. Specifically, it includes the following steps:
[0039] (1) Set up a processor to monitor the electromagnetic environment in the target area, collect data on the target electromagnetic signal, obtain the IQ data of the target electromagnetic signal and perform local buffering processing.
[0040] (2) Extract IQ data from the local cache, add composite noise to it, and use the method based on discrete wavelet decomposition and reconstruction to convert the IQ data into a two-dimensional image with narrow side of dual channel; then apply a random mask to the two-dimensional image with narrow side of dual channel, capture the graph features of the target electromagnetic signal by the topology modeling method based on finite fixed graph mapping, and delete the edges or disturbed nodes of the graph structure with a set probability value θ1.
[0041] (3) Construct a complex depth separable convolutional neural network model for IQ data and extract IQ data features; extract features from two-dimensional images through the visual Transformer module to obtain two-dimensional image features; extract features from graph structures through the DiffPool module to obtain graph structure features; and then, based on IQ data features, two-dimensional image features and graph structure features, use the radiation source feature pre-training framework of cross-modal contrastive learning to complete the construction of the recognition model.
[0042] The feature extraction of two-dimensional images using the visual Transformer module includes the following steps:
[0043] Step 201: Apply a vertical asymmetric segmentation strategy to the two-dimensional image, dynamically adjust the segment size to adapt to the structural characteristics of the narrow-side image, and obtain the corresponding image segments;
[0044] Step 202: Use 2D position coding to independently encode the row and column position information of the image blocks, and add the 2D position coding result to the tile embedding vector of the image block element by element to obtain the input sequence of the visual Transformer module;
[0045] Step 203: Introduce a multi-head attention mechanism in the visual Transformer module, where the multi-head attention mechanism has M attention heads; split the input sequence of the visual Transformer module into approximate and detail components by channel, generate corresponding Query, Key, and Value matrices, and introduce positional encoding; calculate positional correlation through linear projection, use the positional correlation as a positional bias function, calculate the attention score after incorporating the positional bias function, and use it as the output of a single attention head; then concatenate the outputs of the M attention heads, perform linear projection, and perform residual connection with the input sequence;
[0046] Step 204: Set the position consistency loss, minimize the position consistency loss, obtain the parameters of the visual Transformer module, and apply the model parameters to extract the two-dimensional image features from the two-dimensional image.
[0047] The recognition model is built based on IQ data features, two-dimensional image features, and graph structure features, using a cross-modal contrastive learning radiation source feature pre-training framework. The specific steps include:
[0048] Step 301: Map the IQ data features, two-dimensional image features and graph structure features to a unified metric space. Each feature processing module independently uses a projection head composed of a two-layer perceptron to perform L2 normalization and output the normalized modal feature vector.
[0049] Step 302: Establish positive and negative phase pairs for modal feature vectors, and use cosine similarity to measure the consistency of features between modalities to obtain the similarity of cross-modal features;
[0050] Step 303: Utilize the similarity between modal features to calculate the cross-entropy loss of information for all modal pairs, which is used as the total loss value of the pre-training framework for radiation source features, and perform cross-modal comparative learning of radiation source features.
[0051] Step 304: Deploy the radiation source feature pre-training framework, introduce the classification head, concatenate the three normalized modal feature vectors from step 301, use the fused features as input to the classification head, obtain the recognition result, and complete the construction of the recognition model. The recognition model specifically includes a complex depth separable convolutional neural network model, a visual Transformer module, a DiffPool module, and a classification head consisting of a fully connected layer.
[0052] (4) Use the constructed recognition model as a meta-learner and use the metric strategy to update the meta-learner to achieve target recognition based on meta-learning in the case of small sample size.
[0053] The strategy of updating the meta-learner using metrics specifically includes the following steps:
[0054] Step 401: Taking target recognition based on small sample scenarios as the meta-task, randomly select data containing N radiation source types from a dataset that follows a task distribution as the meta-task dataset, where each type has K labeled instances. Divide the meta-task dataset into a support set and a query set according to a set ratio, and perform scenario training using a model-independent meta-learning strategy; where K < <N;
[0055] Step 402: Divide the entire training into multiple meta-tasks. For a single meta-task, construct a meta-learner to extract support set features, use the similarity score between support sets as the loss function of the support set, calculate one step of gradient descent to obtain the expected parameters of the meta-learner, and measure the similarity between the support set and the query set based on the support set, query set, and expected parameters to obtain the similarity loss between the support set and the query set as the loss function of a single meta-task. Then, calculate the sum of the loss functions under each meta-task as the training loss for the entire scenario training, and use stochastic gradient descent to train the meta-learner.
[0056] Step 403: Extract data from the unknown categories that follow the test task distribution as the test set, and perform meta-testing on the meta-learner that has been trained in step 402 to complete target recognition in the case of few samples.
[0057] (5) Based on the meta-learning internal and external loop collaborative optimization strategy, some parameters in the recognition model are fine-tuned to achieve unsupervised autonomous evolution of the recognition model;
[0058] The identification model is fine-tuned and optimized based on a meta-learning internal and external loop collaborative optimization strategy, which includes the following steps:
[0059] Step 501: Using low-rank adaptation and dynamic gradient freezing strategies, low-rank matrix bypasses are inserted into the visual Transformer module and DiffPool module, updating only some parameters in the module; and the bottom temporal convolutional layers of the complex depth separable convolutional neural network model are frozen, with only the top complex fully connected layer being fine-tuned.
[0060] Step 502 introduces a dynamic class reweighting mechanism to quantify sample importance. A learnable weight generator adaptively adjusts the weights of each class in the loss function. An improved model-independent meta-learning strategy is adopted, updating the recognition model in both inner and outer loops, specifically including:
[0061] The support set of the meta-task is used as the input of the inner loop. The parameters of the feature extractor are frozen, a gradient descent step is performed on the support set, the classification head parameters are updated, the model is adapted, and the recognition model in the inner loop is updated. The feature extractor includes a complex depth separable convolutional neural network model, a visual Transformer module, and a DiffPool module.
[0062] The adapted model and query set are used as input to the outer loop. A meta-optimization strategy is adopted to jointly optimize the parameters of the feature extractor and the classification head, minimize the query set loss, update the global parameters of the recognition model, and complete the update of the recognition model in the outer loop.
[0063] In summary, this invention implements a method for the autonomous evolution of an electromagnetic environment recognition model, comprising a series of processes including electromagnetic signal data reception and local caching, conversion of two-dimensional images and capture of graph features of electromagnetic signals, extraction of IQ data features, two-dimensional image features and graph structure features, construction of the recognition model, realization of target recognition based on meta-learning, and unsupervised autonomous evolution of the recognition model. This invention combines meta-learning and domain generalization techniques to improve the recognition capability of the electromagnetic signal recognition model in dynamic and complex electromagnetic environments. The invention significantly improves the applicability of the recognition model through lightweight deployment and significantly enhances its scalability through meta-learning fine-tuning framework, exhibiting advantages such as wide adaptability and strong scalability.
Claims
1. A method for autonomous evolution of identification models for electromagnetic environments, characterized in that, Includes the following steps: Step 1: Set up the processor to monitor the electromagnetic environment in the target area, collect data on the target electromagnetic signal, obtain the IQ data of the target electromagnetic signal and perform local buffering processing. Step 2: Extract IQ data from the local cache, add composite noise to it, and use a method based on discrete wavelet decomposition and reconstruction to convert the IQ data into a dual-channel narrow-side two-dimensional image; then apply a random mask to the dual-channel narrow-side two-dimensional image, capture the graph characteristics of the target electromagnetic signal through a topology modeling method based on finite fixed graph mapping, and delete the edges or perturbation nodes of the graph structure with a set probability value θ1. Step 3: Construct a complex depth separable convolutional neural network model for IQ data and extract IQ data features; extract features from the two-dimensional image using the visual Transformer module to obtain two-dimensional image features; extract features from the graph structure using the DiffPool module to obtain graph structure features; and then, based on the IQ data features, two-dimensional image features, and graph structure features, complete the construction of the recognition model using a cross-modal contrastive learning radiation source feature pre-training framework. Step 4: Use the constructed recognition model as a meta-learner, and update the meta-learner using a metric strategy to achieve target recognition based on meta-learning in the case of small samples. Step 5: Based on the meta-learning inner and outer loop collaborative optimization strategy, fine-tune some parameters in the recognition model to achieve unsupervised autonomous evolution of the recognition model; A method for autonomous evolution of identification models oriented towards electromagnetic environments was developed.
2. The self-evolutionary method for identification models based on electromagnetic environment as described in claim 1, characterized in that, Step 3 involves feature extraction from the 2D image using the Visual Transformer module, specifically including the following steps: Step 201: Apply a vertical asymmetric segmentation strategy to the two-dimensional image, dynamically adjust the segment size to adapt to the structural characteristics of the narrow-side image, and obtain the corresponding image segments; Step 202: Use 2D position coding to independently encode the row and column position information of the image blocks, and add the 2D position coding result to the tile embedding vector of the image block element by element to obtain the input sequence of the visual Transformer module; Step 203: Introduce a multi-head attention mechanism in the visual Transformer module, where the multi-head attention mechanism has M attention heads; split the input sequence of the visual Transformer module into approximate and detail components by channel, generate corresponding Query, Key, and Value matrices, and introduce positional encoding; calculate positional correlation through linear projection, use the positional correlation as a positional bias function, calculate the attention score after incorporating the positional bias function, and use it as the output of a single attention head; then concatenate the outputs of the M attention heads, perform linear projection, and perform residual connection with the input sequence; Step 204: Set the position consistency loss, minimize the position consistency loss, obtain the parameters of the visual Transformer module, and apply the model parameters to extract the two-dimensional image features from the two-dimensional image.
3. The method for autonomous evolution of recognition models under electromagnetic environments according to claim 1, characterized in that, Step 3, based on IQ data features, two-dimensional image features, and graph structure features, utilizes a cross-modal contrastive learning-based radiation source feature pre-training framework to complete the construction of the recognition model. Specifically, it includes the following steps: Step 301: Map the IQ data features, two-dimensional image features and graph structure features to a unified metric space. Each feature processing module independently uses a projection head composed of a two-layer perceptron to perform L2 normalization and output the normalized modal feature vector. Step 302: Establish positive and negative phase pairs for modal feature vectors, and use cosine similarity to measure the consistency of features between modalities to obtain the similarity of cross-modal features; Step 303: Utilize the similarity between modal features to calculate the cross-entropy loss of information for all modal pairs, which is used as the total loss value of the pre-training framework for radiation source features, and perform cross-modal comparative learning of radiation source features. Step 304: Deploy the radiation source feature pre-training framework, introduce the classification head, concatenate the three normalized modal feature vectors from step 301, use the fused features as input to the classification head, obtain the recognition result, and complete the construction of the recognition model. The recognition model specifically includes a complex depth separable convolutional neural network model, a visual Transformer module, a DiffPool module, and a classification head consisting of a fully connected layer.
4. The method for autonomous evolution of recognition models under electromagnetic environments according to claim 1, characterized in that, Step 4 updates the meta-learner using a metric strategy, specifically including the following steps: Step 401: Taking target recognition based on small sample scenarios as the meta-task, randomly select data containing N radiation source types from a dataset that follows a task distribution as the meta-task dataset, where each type has K labeled instances. Divide the meta-task dataset into a support set and a query set according to a set ratio, and perform scenario training using a model-independent meta-learning strategy; where K < <N; Step 402: Divide the entire training into multiple meta-tasks. For a single meta-task, construct a meta-learner to extract support set features, use the similarity score between support sets as the loss function of the support set, calculate one step of gradient descent to obtain the expected parameters of the meta-learner, and measure the similarity between the support set and the query set based on the support set, query set, and expected parameters to obtain the similarity loss between the support set and the query set as the loss function of a single meta-task. Then, calculate the sum of the loss functions under each meta-task as the training loss for the entire scenario training, and use stochastic gradient descent to train the meta-learner. Step 403: Extract data from the unknown categories that follow the test task distribution as the test set, and perform meta-testing on the meta-learner that has been trained in step 402 to complete target recognition in the case of few samples.
5. The method for autonomous evolution of recognition models under electromagnetic environments according to claim 1, characterized in that, Step 5 involves fine-tuning and optimizing the recognition model based on a meta-learning inner and outer loop collaborative optimization strategy. This includes the following steps: Step 501: Using low-rank adaptation and dynamic gradient freezing strategies, low-rank matrix bypasses are inserted into the visual Transformer module and DiffPool module, updating only some parameters in the module; and the bottom temporal convolutional layers of the complex depth separable convolutional neural network model are frozen, with only the top complex fully connected layer being fine-tuned. Step 502 introduces a dynamic class reweighting mechanism to quantify sample importance. Through a learnable weight generator, the weights of each class in the loss function are adaptively adjusted. An improved model-independent meta-learning strategy is adopted, which updates the recognition model in two loops: an inner loop and an outer loop. Specifically, this includes: The support set of the meta-task is used as the input of the inner loop. The parameters of the feature extractor are frozen, a gradient descent step is performed on the support set, the classification head parameters are updated, the model is adapted, and the recognition model in the inner loop is updated. The feature extractor includes a complex depth separable convolutional neural network model, a visual Transformer module, and a DiffPool module. The adapted model and query set are used as input to the outer loop. A meta-optimization strategy is adopted to jointly optimize the parameters of the feature extractor and the classification head, minimize the query set loss, update the global parameters of the recognition model, and complete the update of the recognition model in the outer loop.