A radar working mode small sample recognition method based on a multi-modal model-independent meta-learning

Abstract

The application relates to the technical field of radars, in particular to a radar working mode small sample identification method based on a multi-modal model independent meta-learning, which comprises the following steps: simulating multiple radar working mode waveforms to extract corresponding pulse description word sequences and taking the pulse description word sequences as small sample tasks; converting the pulse description word sequences into multi-channel feature images and extracting sequence features and image features; mapping the sequence features and the image features to a shared feature space and obtaining fused features through weighted fusion; based on the fused features, training an initial multi-modal double-flow model independent network on the small sample tasks by adopting a model independent meta-learning framework to obtain a trained multi-modal double-flow model independent network; and classifying the working modes of query samples by using the trained multi-modal double-flow model independent network. The method can improve the rapid adaptability and reliability of a radar system in a dynamic electromagnetic environment.

Description

Technical Field

[0001] The embodiments of this application relate to the field of radar technology, and in particular to a small-sample recognition method for radar operating modes based on multimodal model-independent element learning. Background Technology

[0002] Radar operating mode recognition is one of the key tasks in electronic warfare and radar signal processing. With the development of modern radar systems, their operating modes are becoming increasingly complex, posing challenges to traditional pattern recognition methods, especially in small sample scenarios where recognition performance is often limited.

[0003] Few-Shot Learning (FSL) aims to simulate human cognitive mechanisms, enabling the rapid development of understanding of new concepts with a very small support set and generalization to unknown query sets. Current technologies propose a graph neural network-based recognition method, specifically by constructing a pulse parameter similarity graph and using a multi-relation graph neural network for feature aggregation. This method can capture local correlations between pulses, but it relies on only a small number of features and does not consider the impact of special samples (such as noise or lost pulses) on the global model, resulting in insufficient generalization ability in complex electromagnetic environments. Existing technologies also propose a graph kernel method for feature extraction and transfer learning. While this can handle inter-domain differences, in small-sample scenarios, insufficient data can lead to inadequate network learning and significant biases in the recognition results. While existing technologies are effective under ideal data conditions, they face three major problems: First, they have poor generalization ability, with the model's performance dropping sharply in new categories or scenarios; second, they lack robustness, being sensitive to packet loss, measurement errors, and parameter overlap in pulse descriptor sequences; and third, they cannot effectively model the complex relationships between radar modes, resulting in unstable classification boundaries with few samples. Summary of the Invention

[0004] In view of this, embodiments of this application propose a radar operating mode small-sample identification method based on multimodal model-independent meta-learning. By using the multimodal model-independent meta-learning framework, the method overcomes the bottleneck of data scarcity and solves the technical problems of poor generalization ability, easy failure under complex interference, and inability to effectively model complex relationships between modes in existing small-sample learning methods for radar operating mode identification. This improves the rapid adaptability and reliability of radar systems in dynamic electromagnetic environments.

[0005] To achieve the above objectives, embodiments of this application propose a small-sample identification method for radar operating modes based on multimodal model-independent meta-learning, the method comprising: Simulate waveforms of various radar operating modes to extract their corresponding pulse descriptor sequences, and use this as a small sample task; the pulse descriptor sequences include various physical parameters such as arrival time, carrier frequency, pulse amplitude, and pulse width; The pulse descriptor sequence is converted into a multi-channel feature image, and the sequence features of the pulse descriptor sequence and the image features of the multi-channel feature image are extracted by an initial multimodal two-stream model-independent network. Sequence features and image features are mapped to a shared feature space, and fused features are obtained through weighted fusion; where the fusion weights are dynamically learned by a task-adaptive mechanism. Based on fusion features, a model-independent meta-learning framework is adopted to train an initial multimodal two-stream model-independent network on few-shot tasks, resulting in a trained multimodal two-stream model-independent network. The initial multimodal two-stream meta-learning network achieves rapid adaptation to few-shot tasks through task sampling, inner loop adaptation, and outer loop meta-optimization. The few-shot tasks include support sets and query sets. The trained multimodal dual-stream model-independent network is used to classify the working modes of the query samples, and the probability vector is output. The maximum probability vector is determined as the predicted radar working mode of the query sample.

[0006] To achieve the above objectives, embodiments of this application also propose a radar operating mode small-sample identification system based on multimodal model-independent meta-learning, the system comprising: The parameter extraction module is used to simulate waveforms of various radar operating modes to extract their corresponding pulse descriptor sequences and treat them as a small sample task; the pulse descriptor sequence includes various physical parameters such as arrival time, carrier frequency, pulse amplitude and pulse width; The feature extraction module is used to convert the pulse descriptor sequence into a multi-channel feature image, and extract the sequence features of the pulse descriptor sequence and the image features of the multi-channel feature image through an initial multimodal two-stream model-independent network. The weighted fusion module maps sequence features and image features to a shared feature space and obtains fused features through weighted fusion; the fusion weights are dynamically learned by a task adaptive mechanism. The model training module is used to train an initial multimodal two-stream model-independent network on few-sample tasks based on fused features and employing a model-independent meta-learning framework, resulting in a trained multimodal two-stream model-independent network. The initial multimodal two-stream meta-learning network achieves rapid adaptation to few-sample tasks through task sampling, inner loop adaptation, and outer loop meta-optimization. The few-sample tasks include a support set and a query set. The pattern prediction module is used to classify the working modes of query samples using a trained multimodal two-stream model-independent network, and outputs a probability vector, with the maximum probability vector being determined as the predicted working mode.

[0007] To achieve the above objectives, embodiments of this application also propose an electronic device, including: a processor and a memory, wherein the memory stores instructions executable by the processor, and the processor is configured to execute the instructions such that the electronic device can implement a radar operating mode small sample recognition method based on multimodal model-independent element learning as described above.

[0008] To achieve the above objectives, embodiments of this application also propose a computer-readable storage medium storing a computer program that, when executed by a processor, can implement a radar operating mode small sample recognition method based on multimodal model-independent element learning as described above.

[0009] This application proposes a few-shot radar operating mode identification method based on multimodal model-independent meta-learning. First, it simulates waveforms of various radar operating modes to extract their corresponding pulse descriptor sequences, using this as a few-shot task. Then, the pulse descriptor sequences are converted into multi-channel feature images, and sequence features of the pulse descriptor sequences and image features of the multi-channel feature images are extracted using an initial multimodal two-stream model-independent network. Next, the sequence features and image features are mapped to a shared feature space and fused using weighted fusion to obtain fused features. Then, based on the fused features, the initial multimodal two-stream model-independent network is trained on the few-shot task using a model-independent meta-learning framework, resulting in a trained multimodal two-stream model-independent network. Finally, the trained multimodal two-stream model-independent network is used to classify the operating modes of query samples, outputting probability vectors, and determining the maximum probability vector as the predicted radar operating mode of the query sample. Since the pulse descriptor sequence includes multiple physical parameters such as arrival time, carrier frequency, pulse amplitude, and pulse width, this provides training for few-shot learning. This approach addresses the scarcity of real-world data by strengthening the foundation of data. It transforms pulse descriptor sequences into multi-channel feature images, converting one-dimensional data into two-dimensional multi-channel images to enhance the spatial texture features of the signal. Feature alignment is achieved through weighted fusion of extracted sequence and image features, avoiding single-mode failure. The initial multi-modal dual-stream learning network rapidly adapts to small-sample tasks via task sampling, inner-loop adaptation, and outer-loop optimization. Since small-sample tasks include support and query sets, a model-independent meta-learning framework can be used to train the network, enabling it to quickly adapt to new tasks under limited sample conditions. Finally, the trained model is applied to classify the operating modes of real radar signals. Based on this, this scheme overcomes the data scarcity bottleneck through a multi-modal model-independent meta-learning framework, addressing the shortcomings of existing small-sample learning methods in radar operating mode recognition, such as poor generalization ability, susceptibility to failure under complex interference, and inability to effectively model complex relationships between modes. This improves the rapid adaptability and reliability of radar systems in dynamic electromagnetic environments.

[0010] Optionally, the pulse descriptor sequence is converted into a multi-channel feature image, and the sequence features of the pulse descriptor sequence and the image features of the multi-channel feature image are extracted by an initial multimodal two-stream model-independent network, including: mapping different physical parameters in the pulse descriptor sequence to a two-dimensional plane with the pulse arrival time as the time reference axis, and generating a multi-channel image of fixed size on a discrete two-dimensional image grid; in the initial multimodal two-stream model-independent network, the sequence features of the pulse descriptor sequence are extracted by a sequence coding stream, and the image features of the multi-channel feature image are extracted by an image coding stream; wherein, the sequence coding stream uses a one-dimensional convolutional network to capture temporal dependencies, and the image coding stream uses a two-dimensional encoder to capture spatial texture features.

[0011] Optionally, record the first The pulse descriptor sequence of pulses is as follows: , , Indicates arrival time. Indicates radio frequency, Indicates amplitude, Indicates the pulse width; the image spatial resolution is... , and It is a preset integer, and Indicates the number of pixels in the height direction. The number of pixels in the width direction is indicated; the mapping of different physical parameters in the pulse descriptor sequence to a two-dimensional plane using the pulse arrival time as the time reference axis includes: Normalized relative time Mapped to x-coordinate carrier frequency Mapped to ordinate Grid coordinates The calculation is as follows: ; ; in, Indicates the first The arrival time corresponding to each pulse and These correspond to the lower and upper bounds of the time axis, respectively, and are used to represent the start and end points of the image's time range. Indicates the first The carrier frequency corresponding to each pulse and These correspond to the lower and upper bounds of the frequency axis, respectively, and are used to represent the start and end points of the image frequency range.

[0012] Optionally, generating a fixed-size multi-channel image on a discrete two-dimensional image grid includes: In the first feature channel, the data falls within a two-dimensional image grid through statistical analysis. Number of pulses The logarithmic pulse density was calculated. The calculation formula is as follows: ; In the second feature channel, by calculating the area falling into the two-dimensional image grid The amplitude-weighted sum of all pulses, based on the number of pulses. The average amplitude distribution is obtained. The calculation formula is as follows: ; in, Indicates falling into a two-dimensional image grid The pulse set, It is a constant used to prevent the denominator from being zero; In the third feature channel, the calculation is performed by falling into the two-dimensional image grid. The pulse width-weighted sum of all pulses, based on the number of pulses. The average pulse width distribution was obtained. The calculation formula is as follows: ; Normalize each of the three feature channels, linearly mapping their numerical ranges to... The mapped feature channels are then stacked in the depth dimension to form a multi-channel image of a preset size. .

[0013] Optionally, the pulse descriptor sequence is denoted as The step of extracting sequence features of pulse descriptor sequences through sequence coding streams and extracting image features of multi-channel feature images through image coding streams in the initial multimodal two-stream model-independent network includes: obtaining all weights corresponding to the sequence coding streams. All weights corresponding to the image encoded stream ; through the sequence-encoded stream and all the weights corresponding to the sequence-encoded stream For pulse descriptor sequence Encode to obtain sequence features ; through the image encoding stream and all the weights corresponding to the image encoding stream For multi-channel images Encode to obtain image features .

[0014] Optionally, sequence features and image features are mapped to a shared feature space, and fused features are obtained through weighted fusion, including: Obtain all weights corresponding to the fused projection layer; wherein, all weights corresponding to the fused projection layer include the first weights of the independent linear layers. Second weight ; Image features are extracted using independent linear layers. and sequence features Mapped to a shared feature space, and based on a nonlinear activation function and a first weight. Second weight To obtain the image projection and sequence projection ; Image features and sequence features We perform a weighted summation to obtain the aligned fusion features. ; The fusion features are obtained through weighted fusion calculation. The calculation formula is as follows: ; in, and Indicates the fusion weight, and The fusion weights are dynamically learned through a task-adaptive mechanism, and their calculation formula is as follows: ; ; in, and To use a small neural network The importance score obtained from learning, i.e. , .

[0015] Optionally, based on fusion features, a model-independent meta-learning framework is used to train an initial multimodal two-stream model-independent network on a few-shot task, resulting in a trained multimodal two-stream model-independent network, including: Obtain the base learner for the initial multimodal two-stream model-independent network. ;in, Represents the initial global parameters, the basic learner. The input data includes fusion features The output data includes a predicted probability vector, which is calculated using the following formula: ; in, and These represent the weights and biases of the classification layer in a multimodal two-stream unrelated network. This represents the normalized exponential function; From the distribution of identification tasks Small sample acquisition task and small sample task Divided into support sets and query set Among them, support sets For task adaptation, query set Used for performance evaluation and metadata updates; In support set An inner loop is used for adaptive processing, updating parameters through gradient descent to obtain task-specific parameters. The calculation formula is as follows: ; in, Indicates step size, Represents the gradient operator, This represents the loss function of the inner loop. The calculation formula is as follows: ; in, This represents the true label corresponding to the small sample task; Indicates in the parameter The following are the fusion features The prediction result is the true label. The probability of; In query set Optimize the outer loop by evaluating task-specific parameters. The performance of this is used to optimize the initial global parameters. To obtain the optimal global parameters The trained multimodal two-stream independent network is obtained. Optimal global parameters The calculation formula is as follows: ; in, This represents the loss function of the outer loop. The calculation formula is as follows: ; in, Indicates in the parameter The following are the fusion features The prediction result is the true label. The probability of.

[0016] Compared with the prior art, this application has the following technical effects: First, it exhibits stronger and faster adaptability when faced with small sample sizes and the arrival of new categories. Employing a model-independent meta-learning framework, it learns a set of rapidly adaptable initialization parameters through numerous tasks during the meta-training phase; in the application phase, adaptation to new categories requires only a small number of support samples and can be completed through a few gradient updates.

[0017] Second, in response to the challenges of numerous interferences and unstable signals in real electromagnetic environments, this application significantly improves the reliability and stability of the system through a dynamic fusion mechanism.

[0018] Third, feature representations have clear physical interpretability, which enhances the ability to distinguish patterns. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies of this application will be briefly introduced below. Obviously, the following drawings are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. The drawings described herein are only used to explain this application and are not intended to limit this application.

[0020] Figure 1 This is a flowchart of a radar operating mode small sample identification method based on multimodal model-independent element learning provided in one embodiment of this application; Figure 2 This is a simplified flowchart of a radar operating mode small sample identification method based on multimodal model-independent element learning provided in one embodiment of this application; Figure 3 This is a schematic diagram of the structure of a radar operating mode small sample recognition system based on multimodal model-independent meta-learning provided in another embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in another embodiment of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the various embodiments of this application will be described in detail below with reference to the accompanying drawings. Those skilled in the art will understand that many technical details have been presented in the embodiments of this application to facilitate better understanding. However, the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments. The division of the following embodiments is for ease of description and should not constitute any limitation on the specific implementation of this application. The following embodiments can be combined with and referenced by each other without contradiction.

[0022] Radar operating mode recognition is one of the key tasks in electronic warfare and radar signal processing. With the development of modern radar systems, their operating modes are becoming increasingly complex, posing challenges to traditional pattern recognition methods, especially in small sample scenarios where recognition performance is often limited.

[0023] For example, in practical applications, the emergence of new radar systems or operating modes often leads to a scarcity of labeled data, creating small-sample scenarios. This causes traditional deep learning models to experience a significant performance degradation due to insufficient training. Few-Shot Learning (FSL) technology has emerged to address this issue. Its core principle is to simulate the rapid cognitive mechanism of humans, quickly adapting to new categories using a very small number of support samples (such as K-Shot) and generalizing them to unknown query sets. However, the characteristics of radar signals, such as the high-dimensional sparsity, temporal dynamics, and weak correlations between pulse features of the Pulse Description Word (PDW) sequence, pose challenges to the direct application of FSL, necessitating the development of more robust feature representation and learning frameworks.

[0024] The most similar existing technologies mainly include radar operating mode recognition schemes based on graph neural networks and graph kernel methods. Patent 1 (CN2025110158062) proposes a recognition method based on graph neural networks: by constructing a pulse parameter similarity graph and using a multi-relation graph neural network for feature aggregation, this method can capture local correlations between pulses, but it only relies on a small number of features and does not consider the impact of special samples (such as noise or lost pulses) on the global model, resulting in insufficient generalization ability in complex electromagnetic environments. Patent 2 (CN2025111345371) uses graph kernel methods for feature extraction and transfer learning. Although it can handle inter-domain differences, in small sample scenarios (such as supporting sample number K≤5), insufficient data will lead to insufficient network learning and significant deviations in recognition results. The common defects of these methods are: first, they rely too much on single-modal data (e.g., using only PDW sequences or fixed image features), and cannot cope with real interference such as pulse loss and parameter drift; second, they lack a dynamic fusion mechanism, making it difficult to balance "structural features" and "statistical differences" under small sample conditions, resulting in blurred and confused category boundaries.

[0025] In summary, while existing technologies show some effectiveness under ideal data conditions, they face three core problems: First, poor generalization ability, with model performance plummeting in new categories or scenarios; second, insufficient robustness, sensitive to packet loss, measurement errors, and parameter overlap in PDW sequences; and third, inability to effectively model complex relationships between radar modes, leading to unstable classification boundaries with limited samples. These problems stem from the fact that existing methods often require full retraining or fixed fusion strategies, making it difficult to meet the rapid response requirements of dynamic electromagnetic environments.

[0026] In view of this, embodiments of this application propose a radar operating mode few-shot identification method based on multimodal model-agnostic meta-learning. By using the multimodal model-agnostic meta-learning (MAML) framework, it overcomes the bottleneck of data scarcity and solves the technical problems of poor generalization ability, easy failure under complex interference, and inability to effectively model complex relationships between modes in existing few-shot learning methods for radar operating mode identification. This improves the rapid adaptability and reliability of radar systems in dynamic electromagnetic environments.

[0027] One embodiment of this application proposes a radar operating mode small-sample identification method based on multimodal model-independent meta-learning, applied to an electronic device, wherein the electronic device can be a terminal or a server. This embodiment and the following embodiments will use a server as an example for description. The implementation details of the radar operating mode small-sample identification method based on multimodal model-independent meta-learning proposed in this embodiment are described in detail below. The following implementation details are provided for ease of understanding and are not necessary for implementing this solution.

[0028] It should be noted that the method provided in the embodiments of this application is not only applicable to radar operating pattern recognition, but can also be extended to other fields with limited data but complex patterns, such as intelligent traffic prediction, personalized recommendation systems, computer vision, speech recognition, and natural language processing. In these applications, traditional methods typically require a large amount of labeled data, while the method provided in the embodiments of this application, and the multimodal two-stream model-independent network constructed using this method, can efficiently learn category features under conditions of limited samples, thereby improving the generalization ability and practicality of the task.

[0029] The specific process of the radar operating mode small sample identification method based on multimodal model irrelevant element learning proposed in this embodiment can be described as follows: Figure 1 As shown, it includes: Step 101: Simulate waveforms of various radar operating modes to extract their corresponding pulse descriptor sequences, and use them as a small sample task.

[0030] The pulse descriptor sequence includes various physical parameters such as time of arrival (TOA), carrier frequency (RF), pulse amplitude (PA), and pulse width (PW).

[0031] For example, radar operating modes can be representative modes such as search mode, track mode, imaging mode, low probability of intercept mode, and multi-target tracking mode.

[0032] For example, the radar waveform simulation toolbox in MATLAB can be used to simulate waveforms for ten representative radar operating modes. Each mode generates a time-domain signal based on its operating principle (e.g., linear frequency modulation, frequency agility). For instance, for the search mode, the simulation parameters are set to: pulse repetition interval varying from 100μs to 500μs, and carrier frequency ranging from 8GHz to 12GHz; for the tracking mode, the pulse repetition interval is fixed but the carrier frequency is agile. After the simulation, the PDW sequence of each waveform sample is extracted, for example, by obtaining the pulse descriptor sequence of each pulse using a pulse detection algorithm.

[0033] The PDW sequences are organized into few-shot tasks; for example, a few-shot task can be in the form of N-Way K-Shot. For instance, in a 5-Way 1-Shot task, five radar operating modes are randomly selected from multiple radar operating modes. One sample from each mode is randomly selected as the support set, and another five samples are selected as the query set. Task sampling follows a meta-learning framework to simulate the need for rapid adaptation of new modes in real-world scenarios.

[0034] Step 102: Convert the pulse descriptor sequence into a multi-channel feature image, and extract the sequence features of the pulse descriptor sequence and the image features of the multi-channel feature image through the initial multimodal two-stream model-independent network.

[0035] In one possible embodiment, the denoted number is... The pulse descriptor sequence of pulses is as follows: , .in, Indicates arrival time. Indicates radio frequency, Indicates amplitude, This represents the pulse width. The image spatial resolution is... , and It is a preset integer, and Indicates the number of pixels in the height direction. This indicates the number of pixels in the width direction. For example, and It can be set to 128, meaning the image resolution is set to 128×128 pixels.

[0036] In one possible embodiment, step 102 above includes: mapping different physical parameters in the pulse descriptor sequence to a two-dimensional plane with the pulse arrival time as the time reference axis, and generating a multi-channel image of fixed size on a discrete two-dimensional image grid; in the initial multimodal two-stream model-independent network, extracting the sequence features of the pulse descriptor sequence through a sequence coding stream, and extracting the image features of the multi-channel feature image through an image coding stream.

[0037] The sequence coding stream uses a one-dimensional convolutional network to capture temporal dependencies, while the image coding stream uses a two-dimensional encoder to capture spatial texture features.

[0038] For example, the pulse arrival time can be used as a unified time reference axis, and different physical parameters (carrier frequency, pulse amplitude and pulse width) can be combined with the arrival time to form a two-dimensional plane. The pulse distribution and parameter intensity can be statistically analyzed on a discrete grid, thereby transforming the sequence structure into a spatial texture structure, and finally obtaining a multi-channel image input with a fixed size.

[0039] To map non-uniformly sampled pulses onto a uniform pixel grid, normalized relative time can be mapped to the horizontal axis of the image, and radio frequency can be mapped to the vertical axis. See the following examples for details.

[0040] In one possible embodiment, mapping different physical parameters in the pulse descriptor sequence to a two-dimensional plane using the pulse arrival time as a time reference axis includes: Normalized relative time Mapped to x-coordinate carrier frequency Mapped to ordinate Grid coordinates The calculation is as follows: ; ; in, Indicates the first The arrival time corresponding to each pulse and These correspond to the lower and upper bounds of the time axis, respectively, and are used to represent the start and end points of the image's time range. Indicates the first The carrier frequency corresponding to each pulse and These correspond to the lower and upper bounds of the frequency axis, respectively, and are used to represent the start and end points of the image frequency range.

[0041] Because radar signals have a high sampling rate, multiple pulses may fall within a single pixel grid (i.e., pulse overlap). To address this issue and retain more dimensional information, three feature channels can be constructed. See the following example for details.

[0042] In one possible embodiment, generating a fixed-size multi-channel image on a discrete two-dimensional image grid includes: The first feature channel outputs the log-density pulse, designed to capture the time-frequency distribution profile and repetition rate pattern of the signal. Considering the extremely large dynamic range of radar signal pulse density (e.g., the density of continuous wave signals is much higher than that of low-repetition-rate signals), a logarithmic function is used for dynamic range compression to prevent high-density regions from masking low-density features. Specifically, this is achieved by statistically analyzing the pulse density falling within a two-dimensional image grid. Number of pulses The logarithmic pulse density was calculated. This is to balance the dynamic range. The calculation formula is as follows: ; Understandably, the first feature channel enables the image modality to simultaneously and clearly present the main outline of the signal and the sparse background clutter features.

[0043] The second feature channel outputs the average amplitude distribution (Channel 1: Average Amplitude); this channel reflects the distribution of signal energy in the time-frequency domain. The interference of pulse aggregation on the amplitude characteristics is eliminated by calculating the amplitude-weighted sum of all pulses falling within the grid and dividing by the number of pulses in that grid (i.e., calculating the local mean). Specifically, this is achieved by calculating the amplitude of pulses falling within the two-dimensional image grid. The amplitude-weighted sum of all pulses, based on the number of pulses. The average amplitude distribution is obtained. The calculation formula is as follows: ; in, Indicates falling into a two-dimensional image grid The pulse set, It is a constant used to prevent the denominator from being zero, for example .

[0044] The average pulse width distribution is output in the third feature channel. This channel is used to help distinguish radar patterns with similar time-frequency trajectories but different time-domain modulations (e.g., distinguishing signals with different duty cycles). Specifically, this is achieved by calculating the average pulse width distribution falling within a two-dimensional image grid. The pulse width-weighted sum of all pulses, based on the number of pulses. The average pulse width distribution was obtained. The calculation formula is as follows: ; Data normalization and tensor generation are performed to adapt to the input distribution requirements of neural networks and accelerate model convergence. After completing the above statistics, the three feature channels are normalized (Min-Max normalization) to linearly map their numerical ranges to... The mapped feature channels are then stacked in the depth dimension to form a multi-channel image of a preset size (e.g., a 3×128×128 tensor). .

[0045] Understandably, in this representation, the first feature channel outlines the radar's frequency agility and repetition rate irregularity (geometric texture), while the second and third feature channels fill in the amplitude and temporal details within the texture in a manner similar to RGB color components.

[0046] In one possible embodiment, the pulse descriptor sequence is denoted as... The step of extracting sequence features of pulse descriptor sequences through a sequence coding stream and extracting image features of multi-channel feature images through an image coding stream in the initial multimodal two-stream model-independent network includes: Obtain all weights corresponding to the sequence encoded stream All weights corresponding to the image encoded stream ; through the sequence-encoded stream and all the weights corresponding to the sequence-encoded stream For pulse descriptor sequence Encode to obtain sequence features ; through the image encoding stream and all the weights corresponding to the image encoding stream For multi-channel images Encode to obtain image features .

[0047] Understandably, when constructing a multimodal dual-stream model-independent network, two independent feature processing paths can be constructed first to model radar signal representations of different dimensions.

[0048] For example, a sequence encoder can employ a one-dimensional residual convolutional network (1D-CNN) containing four convolutional layers (filter size 3, stride 1) to capture the temporal dependencies of the PDW sequence. The input sequence length... Zero-padding alignment results in a 256-dimensional embedding vector, i.e., sequence features. .

[0049] For example, the image encoder can use SE-ResNet as a two-dimensional encoder, including residual blocks and a squeeze excitation module, to extract spatial texture features from the image. The input is a 3×128×128 image, and the output is a 256-dimensional embedding vector, i.e., the image features. .

[0050] It should be noted that the two streams are processed independently to ensure that modality-specific features are not confused.

[0051] For example: The initial multimodal two-stream model independent network includes the base learner. Basic learner global parameters in The union of the following three subsets, namely: ; in, This represents all weights corresponding to the image encoded stream. This represents all weights corresponding to the sequence-encoded stream. This includes all weights corresponding to the fusion projection layer and all parameters of the classification head.

[0052] In the image encoding stream: ; In a sequence-encoded stream, ; in, is the encoder mapping function, which corresponds to the 1D-CNN architecture for sequence feature encoding and the SE-ResNet architecture for image feature encoding.

[0053] Step 103: Map sequence features and image features to a shared feature space, and obtain fused features through weighted fusion.

[0054] The fusion weights are dynamically learned by the task adaptive mechanism.

[0055] In one possible embodiment, step 103 above includes: obtaining all weights corresponding to the fusion projection layer; and using independent linear layers to process image features. and sequence features Mapped to a shared feature space, and based on a nonlinear activation function and a first weight. Second weight To obtain the image projection and sequence projection .

[0056] Among them, all weights corresponding to the fusion projection layer include the first weights of the independent linear layers. Second weight .

[0057] Image features and sequence features We perform a weighted summation to obtain the aligned fusion features. ; The fusion features are obtained through weighted fusion calculation. The calculation formula is as follows: ; in, and Indicates the fusion weight, and ; For example, two independent linear layers are used to map the extracted raw features to a shared fusion space; Image projection The calculation formula is as follows: ; Sequence Projection The calculation formula is as follows: ; Understandably, the first weight Second weight It can be used to map image features and sequence features of different dimensions to the same shared fusion feature space. Non-linear activation function, i.e. This is used to introduce nonlinearity after feature projection, thereby enhancing the model's representational capabilities.

[0058] The fusion weights are dynamically learned through a task-adaptive mechanism, and their calculation formula is as follows: ; ; in, and To use a small neural network The importance score obtained from learning, i.e. , .

[0059] Step 104: Using the model-independent meta-learning framework, train the initial multimodal two-stream model-independent network on a few-sample task to obtain the trained multimodal two-stream model-independent network.

[0060] Among them, the initial network multimodal dual-stream learning network achieves rapid adaptation to few-sample tasks through task sampling, inner loop adaptation, and outer loop optimization; the few-sample tasks include support sets and query sets.

[0061] In one possible embodiment, step 104 includes: obtaining the base learner of the initial multimodal two-stream model-independent network. ; in, Represents the initial global parameters, the basic learner. The input data includes fusion features The output data is a predicted probability vector. The calculation formula is as follows: ; in, and These represent the weights and biases of the classification layer in a multimodal two-stream unrelated network. This represents the normalized exponential function; After expanding the relationships between each layer: ; It should be noted that, The output of the classification layer can be mapped to a normalized interval to ensure that the sum of the probabilities of all classes is 1, thus obtaining the final predicted probability vector. .

[0062] For example, It can include the first weight Second weight , , , and .

[0063] For example, during the training and optimization of the initial multimodal two-stream model-independent network, It can be the union of the following subsets: ; From the distribution of identification tasks Small sample acquisition task and small sample task Divided into support sets and query set Among them, support sets For task adaptation, query set Used for performance evaluation and metadata updates.

[0064] For example, from task distribution Multiple 5-way K-Shot tasks are obtained, each with a few samples. It is divided into support set and query set.

[0065] In support set An inner loop is used for adaptive processing, updating parameters through gradient descent to obtain task-specific parameters. The calculation formula is as follows: ; in, Indicate step size (e.g., step size) ), Represents the gradient operator, This represents the loss function of the inner loop. The calculation formula is as follows: ; in, This represents the true label corresponding to the small sample task; Indicates in the parameter The following are the fusion features The prediction result is the true label. The probability of; In query set Optimize the outer loop by evaluating task-specific parameters. The performance of this is used to optimize the initial global parameters. To obtain the optimal global parameters The trained multimodal two-stream independent network is obtained. Optimal global parameters The calculation formula is as follows: ; in, This represents the loss function of the outer loop. The calculation formula is as follows: ; in, Indicates in the parameter The following are the fusion features The prediction result is the true label. The probability of.

[0066] It should be noted that the training cycle can have 1000 meta-tasks, and each task supports a certain number of sets of samples. At that time, the update step count is set to 5 steps.

[0067] For example, the network weights are initialized using the Xavier method, and early stopping is applied during training to prevent overfitting.

[0068] Step 105: Use the trained multimodal dual-stream model-independent network to classify the working modes of the query samples, and output probability vectors. The maximum probability vector is determined as the predicted radar working mode of the query sample.

[0069] For example, for query samples (such as unseen samples in a 5-Way task), input the trained multimodal two-stream model-agnostic network. The multimodal two-stream model-independent network outputs an N-dimensional probability vector (e.g., N=5), which is normalized using the Softmax function. Then, the index corresponding to the highest probability is taken as the predicted radar operating mode.

[0070] like Figure 2 As shown, Figure 2This is a simplified flowchart of a radar operating mode small-sample recognition method based on multimodal model-independent meta-learning (MAML) provided in an embodiment of this application. First, simulated radar operating mode waveforms are generated using computer simulation software to simulate radio frequency pulse signals for various typical radar operating modes. Then, descriptor feature extraction is performed, processing the simulated radar pulse sequence to extract key pulse descriptor sequences. Next, PDW image dimensionality transformation is performed to convert the one-dimensional, temporal PDW sequence into a two-dimensional, spatial multi-channel feature image. Afterward, bimodal fusion alignment is performed to establish and fuse two complementary feature streams: a sequence modal stream and an image modal stream. Then, the parameters are optimized using a model-independent meta-learning (MAML) algorithm, and model training and parameter fine-tuning are conducted to finally obtain a pre-trained model-independent meta-learning framework training network with strong cross-task generalization capabilities for subsequent operating mode recognition, i.e., inputting the unknown radar query samples to be identified into the trained model-independent meta-learning framework training network. The network extracts and fuses dual-modal features, and finally the classifier outputs a probability vector, which represents the probability that the sample belongs to each predefined radar working mode. The category corresponding to the highest probability is the final recognition result.

[0071] This application proposes a few-shot radar operating mode identification method based on multimodal model-independent meta-learning. First, it simulates waveforms of various radar operating modes to extract their corresponding pulse descriptor sequences, using this as a few-shot task. Then, the pulse descriptor sequences are converted into multi-channel feature images, and sequence features of the pulse descriptor sequences and image features of the multi-channel feature images are extracted using an initial multimodal two-stream model-independent network. Next, the sequence features and image features are mapped to a shared feature space and fused using weighted fusion to obtain fused features. Then, based on the fused features, the initial multimodal two-stream model-independent network is trained on the few-shot task using a model-independent meta-learning framework, resulting in a trained multimodal two-stream model-independent network. Finally, the trained multimodal two-stream model-independent network is used to classify the operating modes of query samples, outputting probability vectors, and determining the maximum probability vector as the predicted radar operating mode of the query sample. Since the pulse descriptor sequence includes multiple physical parameters such as arrival time, carrier frequency, pulse amplitude, and pulse width, this provides training for few-shot learning. This approach addresses the scarcity of real-world data by strengthening the foundation of data. It transforms pulse descriptor sequences into multi-channel feature images, converting one-dimensional data into two-dimensional multi-channel images to enhance the spatial texture features of the signal. Feature alignment is achieved through weighted fusion of extracted sequence and image features, avoiding single-mode failure. The initial multi-modal dual-stream learning network rapidly adapts to small-sample tasks via task sampling, inner-loop adaptation, and outer-loop optimization. Since small-sample tasks include support and query sets, a model-independent meta-learning framework can be used to train the network, enabling it to quickly adapt to new tasks under limited sample conditions. Finally, the trained model is applied to classify the operating modes of real radar signals. Based on this, this scheme overcomes the data scarcity bottleneck through a multi-modal model-independent meta-learning framework, addressing the shortcomings of existing small-sample learning methods in radar operating mode recognition, such as poor generalization ability, susceptibility to failure under complex interference, and inability to effectively model complex relationships between modes. This improves the rapid adaptability and reliability of radar systems in dynamic electromagnetic environments.

[0072] In summary, compared with the prior art, this application has the following technical effects: First, it exhibits stronger and faster adaptability when faced with small sample sizes and the arrival of new categories. Employing a model-independent meta-learning framework, it learns a set of rapidly adaptable initialization parameters through numerous tasks during the meta-training phase; in the application phase, adaptation to new categories requires only a small number of support samples and can be completed through a few gradient updates.

[0073] Second, in response to the challenges of numerous interferences and unstable signals in real electromagnetic environments, this application significantly improves the reliability and stability of the system through a dynamic fusion mechanism.

[0074] Third, feature representations have clear physical interpretability, which enhances the ability to distinguish patterns.

[0075] It should be noted that this embodiment is based on the simulation experiments in the document, aiming to verify the effectiveness of the present invention in small-sample recognition of radar operating modes based on multimodal model-independent meta-learning. The simulation platform uses an Intel(R) Core i5-13600KF processor, 32GB of memory, and a Windows 11 operating system. The software tools include MATLAB R2023b and PyTorch. The radar radiation source samples to be identified are generated through MATLAB simulation, covering ten representative radar operating modes, with 50 samples generated for each mode. For each sample, a pulse descriptor sequence is extracted, including parameters such as arrival time, carrier frequency, pulse amplitude, and pulse width, and converted into a multi-channel feature image as input.

[0076] The radar radiation source identification simulation experiment in the embodiments of this application uses the method provided in the embodiments of this application to identify the working mode of each signal sample of ten simulated radar radiation source signals with different working modes. The network model is trained and tested under different small sample task conditions to verify the robustness of the model. The accuracy of radar working mode identification under each task is obtained, and all calculation results are plotted in Table 1.

[0077] For example, the network model is trained and tested under different signal-to-noise ratios to verify its robustness. Three datasets for different tasks can be used: 5way-1shot, 5way-3shot, and 5way-5shot. The experimental results are as follows: Table 1. Recognition accuracy under different tasks

[0078] As can be seen from Table 1: First, under the above task conditions, as the number of training samples increases from 1 shot to 5 shots, the recognition accuracy eventually stabilizes at around 98.5%, which demonstrates the effectiveness of the network. Second, with the improvement of signal conditions, the network model can fully learn the patterns between signal features of different working modes in a shorter training time, and the weights of the network model hardly need to be adjusted anymore, resulting in increasingly stable recognition performance.

[0079] The steps described above are for clarity only. In implementation, they can be combined into one step, or some steps can be broken down into multiple steps, as long as they involve the same logical relationship, they are all within the scope of protection of this application. Adding insignificant modifications or introducing insignificant designs to the algorithm or process, without changing the core design of the algorithm and process, are also within the scope of protection of this application.

[0080] Another embodiment of this application proposes a radar operating mode small-sample recognition system based on multimodal model-independent meta-learning. The details of this radar operating mode small-sample recognition system based on multimodal model-independent meta-learning are described below. The following implementation details are provided for ease of understanding and are not essential for implementing this example. Figure 3 This is a schematic diagram of the structure of a radar operating mode small sample recognition system based on multimodal model-independent meta-learning proposed in this embodiment, including: The parameter extraction module 210 is used to simulate waveforms of various radar operating modes to extract their corresponding pulse descriptor sequences and to treat them as small sample tasks; wherein, the pulse descriptor sequence includes various physical parameters such as arrival time, carrier frequency, pulse amplitude and pulse width; The feature extraction module 220 is used to convert the pulse descriptor sequence into a multi-channel feature image, and extract the sequence features of the pulse descriptor sequence and the image features of the multi-channel feature image through an initial multimodal two-stream model-independent network, respectively. The weighted fusion module 230 is used to map sequence features and image features to a shared feature space and obtain fused features through weighted fusion; wherein, the fusion weights are dynamically learned by a task adaptive mechanism. The model training module 240 is used to train an initial multimodal two-stream model-independent network on a few-shot task based on fused features and using a model-independent meta-learning framework, resulting in a trained multimodal two-stream model-independent network. The initial multimodal two-stream meta-learning network achieves rapid adaptation to few-shot tasks through task sampling, inner loop adaptation, and outer loop meta-optimization. The few-shot task includes a support set and a query set. The pattern prediction module 250 is used to classify the working modes of the query samples using the trained multimodal two-stream model-independent network, and outputs a probability vector and determines the predicted working mode by the maximum probability vector.

[0081] It is not difficult to see that this embodiment is a system embodiment corresponding to the above method embodiments, and this embodiment can be implemented in conjunction with the above method embodiments. The relevant technical details and technical effects mentioned in the above method embodiments are still valid in this embodiment, and will not be repeated here to reduce repetition. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the above method embodiments.

[0082] It is worth mentioning that all modules and units involved in this embodiment are logical modules. In practical applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. Furthermore, to highlight the innovative aspects of this application, this embodiment does not introduce units that are not closely related to solving the technical problems proposed in this application; however, this does not mean that other units do not exist in this embodiment.

[0083] Another embodiment of this application provides an electronic device, such as Figure 4 As shown, it includes a processor 31 and a memory 32. The memory 32 stores instructions that the processor 31 can execute. When the processor 31 is configured to execute the instructions, the electronic device can realize a radar operating mode small sample recognition method based on multimodal model-independent element learning as described in the above method embodiment.

[0084] The memory and processor are connected via a bus, which includes any number of interconnecting buses and bridges, connecting various circuits of one or more processors and the memory. The bus can also connect various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and will not be described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single component or multiple components, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by the processor is transmitted over the wireless medium via an antenna, which further receives data and transmits it to the processor.

[0085] The processor manages the bus and general processing, and also provides various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory is used to store data used by the processor during operation.

[0086] Another embodiment of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, can implement a radar operating mode small sample recognition method based on multimodal model-independent element learning as described in the above method embodiments.

[0087] That is, those skilled in the art will understand that all or part of the steps in the above method embodiments can be implemented by a program instructing related hardware. The program is stored in a storage medium and includes several instructions to cause a device (such as a microcontroller, chip, etc.) or processor to execute all or part of the steps of the method described in the method embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0088] Those skilled in the art will understand that the above embodiments are specific implementations of this application, and in practical applications, various changes can be made in form and detail without departing from the spirit and scope of this application. For those skilled in the art, several improvements and modifications can be made without departing from the principles of this application, and these improvements and modifications are also considered to be within the scope of protection of this application.

Claims

1. A method for small-sample recognition of radar operating modes based on multimodal model-independent meta-learning, characterized in that, The method includes: Simulate waveforms of various radar operating modes to extract their corresponding pulse descriptor sequences, and use this as a small sample task; the pulse descriptor sequences include various physical parameters such as arrival time, carrier frequency, pulse amplitude, and pulse width; The pulse descriptor sequence is converted into a multi-channel feature image, and the sequence features of the pulse descriptor sequence and the image features of the multi-channel feature image are extracted by an initial multimodal two-stream model-independent network. Sequence features and image features are mapped to a shared feature space, and fused features are obtained through weighted fusion; where the fusion weights are dynamically learned by a task-adaptive mechanism. Based on fusion features, a model-independent meta-learning framework is adopted to train an initial multimodal two-stream model-independent network on few-shot tasks, resulting in a trained multimodal two-stream model-independent network. The initial multimodal two-stream model-independent network achieves rapid adaptation to few-shot tasks through task sampling, inner loop adaptation, and outer loop meta-optimization. The few-shot tasks include support sets and query sets. The trained multimodal dual-stream model-independent network is used to classify the working modes of the query samples, and the probability vector is output. The maximum probability vector is determined as the predicted radar working mode of the query sample.

2. The method according to claim 1, characterized in that, The process of converting the pulse descriptor sequence into a multi-channel feature image and extracting the sequence features of the pulse descriptor sequence and the image features of the multi-channel feature image through an initial multimodal two-stream model-independent network includes: Using the pulse arrival time as the time reference axis, different physical parameters in the pulse descriptor sequence are mapped to a two-dimensional plane, and a multi-channel image of fixed size is generated on a discrete two-dimensional image grid. In the initial multimodal dual-stream model-independent network, sequence features of pulse descriptor sequences are extracted through sequence coding stream, and image features of multi-channel feature images are extracted through image coding stream. The sequence coding stream uses a one-dimensional convolutional network to capture temporal dependencies, while the image coding stream uses a two-dimensional encoder to capture spatial texture features.

3. The method according to claim 2, characterized in that, Record No. The pulse descriptor sequence of pulses is as follows: , , Indicates arrival time. Indicates radio frequency, Indicates amplitude, Indicates the pulse width; the image spatial resolution is... , and It is a preset integer, and Indicates the number of pixels in the height direction. Indicates the number of pixels in the width direction; The process of mapping different physical parameters in the pulse descriptor sequence to a two-dimensional plane using the pulse arrival time as a time reference axis includes: Normalized relative time Mapped to x-coordinate carrier frequency Mapped to ordinate Grid coordinates The calculation is as follows: ； ； in, Indicates the first The arrival time corresponding to each pulse and These correspond to the lower and upper bounds of the time axis, respectively, and are used to represent the start and end points of the image's time range. Indicates the first The carrier frequency corresponding to each pulse and These correspond to the lower and upper bounds of the frequency axis, respectively, and are used to represent the start and end points of the image frequency range.

4. The method according to claim 3, characterized in that, The process of generating a fixed-size multi-channel image on a discrete two-dimensional image grid includes: In the first feature channel, the data falls within a two-dimensional image grid through statistical analysis. Number of pulses The logarithmic pulse density was calculated. The calculation formula is as follows: ； In the second feature channel, by calculating the area falling into the two-dimensional image grid The amplitude-weighted sum of all pulses, based on the number of pulses. The average amplitude distribution is obtained. The calculation formula is as follows: ； in, Indicates falling into a two-dimensional image grid The pulse set, It is a constant used to prevent the denominator from being zero; In the third feature channel, the calculation is performed by falling into the two-dimensional image grid. The pulse width-weighted sum of all pulses, based on the number of pulses. The average pulse width distribution was obtained. The calculation formula is as follows: ； Normalize each of the three feature channels, linearly mapping their numerical ranges to... The mapped feature channels are then stacked in the depth dimension to form a multi-channel image of a preset size. .

5. The method according to claim 4, characterized in that, Let the pulse descriptor sequence be... The step of extracting sequence features of pulse descriptor sequences through a sequence coding stream and extracting image features of multi-channel feature images through an image coding stream in the initial multimodal two-stream model-independent network includes: Obtain all weights corresponding to the sequence encoded stream All weights corresponding to the image encoded stream ; Through the sequence-encoded stream and all the weights corresponding to the sequence-encoded stream For pulse descriptor sequence Encode to obtain sequence features ; Through the image encoding stream and all the weights corresponding to the image encoding stream For multi-channel images Encode to obtain image features .

6. The method according to claim 5, characterized in that, The process of mapping sequence features and image features to a shared feature space and obtaining fused features through weighted fusion includes: Obtain all weights corresponding to the fused projection layer; wherein, all weights corresponding to the fused projection layer include the first weights of the independent linear layers. Second weight ; Image features are extracted using independent linear layers. and sequence features Mapped to a shared feature space, and based on a nonlinear activation function and a first weight. Second weight To obtain the image projection and sequence projection ; Image features and sequence features We perform a weighted summation to obtain the aligned fusion features. ; The fusion features are obtained through weighted fusion calculation. The calculation formula is as follows: ； in, and Indicates the fusion weight, and The fusion weights are dynamically learned through a task-adaptive mechanism, and their calculation formula is as follows: ； ； in, and To use a small neural network The importance score obtained from learning, i.e. , .

7. The method according to claim 6, characterized in that, The method, based on fusion features and employing a model-independent meta-learning framework, trains an initial multimodal two-stream model-independent network on a few-sample task, resulting in a trained multimodal two-stream model-independent network, including: Obtain the base learner for the initial multimodal two-stream model-independent network. ;in, Represents the initial global parameters, the basic learner. The input data includes fusion features The output data is a predicted probability vector. The calculation formula is as follows: ； in, and These represent the weights and biases of the classification layer in a multimodal two-stream unrelated network. This represents the normalized exponential function; From the distribution of identification tasks Small sample acquisition task and small sample task Divided into support sets and query set Among them, support sets For task adaptation, query set Used for performance evaluation and metadata updates; In support set An inner loop is used for adaptive processing, updating parameters through gradient descent to obtain task-specific parameters. The calculation formula is as follows: ； in, Indicates step size, Represents the gradient operator. This represents the loss function of the inner loop. The calculation formula is as follows: ； in, This represents the true label corresponding to the small sample task; Indicates in the parameter The following are the fusion features The prediction result is the true label. The probability of; In query set Optimize the outer loop by evaluating task-specific parameters. The performance of this is used to optimize the initial global parameters. To obtain the optimal global parameters The trained multimodal two-stream independent network is obtained. Optimal global parameters The calculation formula is as follows: ； in, This represents the loss function of the outer loop. The calculation formula is as follows: ； in, Indicates in the parameter The following are the fusion features The prediction result is the true label. The probability of.

8. A radar operating mode small-sample recognition system based on multimodal model-independent meta-learning, characterized in that, The system includes: The parameter extraction module is used to simulate waveforms of various radar operating modes to extract their corresponding pulse descriptor sequences and treat them as a small sample task; the pulse descriptor sequence includes various physical parameters such as arrival time, carrier frequency, pulse amplitude and pulse width; The feature extraction module is used to convert the pulse descriptor sequence into a multi-channel feature image, and extract the sequence features of the pulse descriptor sequence and the image features of the multi-channel feature image through an initial multimodal two-stream model-independent network. The weighted fusion module maps sequence features and image features to a shared feature space and obtains fused features through weighted fusion; the fusion weights are dynamically learned by a task adaptive mechanism. The model training module is used to train an initial multimodal two-stream model-independent network on few-shot tasks based on fused features and employing a model-independent meta-learning framework, resulting in a trained multimodal two-stream model-independent network. The initial multimodal two-stream model-independent network achieves rapid adaptation to few-shot tasks through task sampling, inner loop adaptation, and outer loop meta-optimization. The few-shot tasks include a support set and a query set. The pattern prediction module is used to classify the working modes of query samples using a trained multimodal two-stream model-independent network, and outputs a probability vector, with the maximum probability vector being determined as the predicted working mode.

9. An electronic device, characterized in that, include: The processor and memory, wherein the memory stores instructions executable by the processor, and the processor is configured to, when executing the instructions, enable the electronic device to implement the radar operating mode small sample recognition method based on multimodal model-independent element learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it can implement the radar operating mode small sample recognition method based on multimodal model-independent element learning as described in any one of claims 1 to 7.

Images