A method and system for radar signal small-sample modulation recognition based on meta-learning

Abstract

This application provides a method and system for radar signal modulation recognition based on meta-learning, relating to the field of radar signal modulation recognition technology. This application collects multi-polarized radar echo signals of a target object, preprocesses them to generate multi-polarization feature sequences and modulation feature sets; then, after analyzing the signal channel quality, multi-channel fusion is performed, and target enhanced polarization features are obtained through spatial interference suppression; subsequently, these features are aligned with the modulation feature set for multi-modal feature integration, and multi-modal feature representations are generated through joint mapping and similarity measurement, reducing intra-class differences; finally, based on a meta-learning framework, a hierarchical optimization architecture is constructed for parameter learning and meta-parameter adjustment, and normalization loss mechanism is used to standardize training, achieving adaptation and generalization of radar signal modulation recognition rules under small sample conditions. Through multi-polarization signal processing, multi-modal feature fusion, and meta-learning optimization, effective adaptation and generalization of radar signal modulation recognition rules can be achieved under small sample conditions, improving recognition performance.

Description

Technical Field

[0001] This application relates to the field of radar signal modulation recognition technology, and in particular to a radar signal small sample modulation recognition method, system, electronic device and storage medium based on meta-learning. Background Technology

[0002] During high-speed maneuvers in orbit, spacecraft operate in extremely complex electromagnetic environments, with frequent electromagnetic interference causing a sharp drop in the signal-to-noise ratio of radar echoes. Simultaneously, due to the difficulty of signal acquisition and the dynamic changes in the environment, modulation samples become very scarce. Under these circumstances, to ensure the safe operation, precise navigation, and efficient communication of spacecraft, there is an urgent need for a technology capable of rapidly and accurately identifying radar signal modulation patterns under limited sample conditions. This would ensure that the spacecraft system responds promptly and correctly to various radar signals, maintaining its reliable performance in complex electromagnetic environments.

[0003] Currently, to address this technical need, there exists a radar signal modulation recognition scheme based on deep convolutional neural networks (DCNNs) combined with time-frequency analysis. This scheme first uses time-frequency analysis to convert the radar signal into a time-frequency image, thereby highlighting the joint characteristics of the signal in the time and frequency dimensions. Then, the time-frequency image is input into a carefully constructed deep convolutional neural network, utilizing the network's powerful automatic feature extraction and classification capabilities to identify the radar signal modulation pattern. To a certain extent, this scheme can handle signal recognition problems in some complex electromagnetic environments and has a certain tolerance for sample size.

[0004] However, this approach reveals several shortcomings in scenarios involving high-speed maneuvering of spacecraft in orbit, complex electromagnetic interference, and scarce modulation samples. Firstly, deep convolutional neural networks heavily rely on large-scale, diverse training samples. In actual spacecraft operation, the scarcity of samples makes it difficult for the model to fully learn the characteristics of various signals, easily leading to overfitting and severely impacting the model's generalization ability and recognition accuracy under new samples and environments. Secondly, complex electromagnetic interference introduces a large amount of noise and interference into time-frequency images. Existing solutions have limited ability to suppress this interference, resulting in biased features extracted by the network and thus misleading recognition results. Furthermore, deep convolutional neural networks have complex structures and require massive computation. Under the limited computing resources and energy supply of spacecraft, their operating efficiency is low, making it difficult to meet real-time recognition requirements. Summary of the Invention

[0005] The purpose of this application is to provide a radar signal small sample modulation recognition method, system, electronic device and storage medium based on meta-learning, so as to solve the problems of poor recognition generalization ability and low accuracy in the existing technology in the scenario of high-speed maneuvering of spacecraft in orbit, complex electromagnetic interference and scarce modulation samples, due to the reliance on large-scale samples, which is prone to overfitting, weak anti-interference ability, large amount of computation and difficulty in meeting real-time requirements.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides a radar signal small-sample modulation recognition method based on meta-learning, comprising:

[0007] Collect the multi-polarization radar echo signal of the target object, perform signal preprocessing on the multi-polarization radar echo signal, and generate a multi-polarization feature sequence and a modulation feature set;

[0008] Channel quality analysis is performed on the signal channels corresponding to the multi-polarization feature sequence. Based on the analysis results, multi-channel fusion is performed to form channel fusion features. Spatial interference suppression processing is then applied to the channel fusion features to generate target enhanced polarization features.

[0009] The target enhanced polarization features and the modulation feature set are subjected to multimodal feature alignment processing. By jointly mapping the target enhanced polarization features and the modulation feature set and measuring similarity, a multimodal feature representation is generated. The intra-class distribution of the multimodal feature representation is constrained based on the center cosine loss function to reduce intra-class feature differences.

[0010] The multimodal feature representation is optimized and trained based on a meta-learning framework. By constructing a hierarchical optimization architecture, the multimodal feature representation is subjected to parameter learning and meta-parameter adjustment at different levels. A normalized loss mechanism is used to standardize the training process, so as to achieve the adaptation and generalization of radar signal modulation recognition rules under small sample conditions.

[0011] Optionally, the optimization training of the multimodal feature representation based on the meta-learning framework involves constructing a hierarchical optimization architecture, performing parameter learning and meta-parameter adjustment on the multimodal feature representation at different levels, and employing a normalized loss mechanism to standardize the training process, thereby achieving adaptation and generalization of radar signal modulation recognition rules under small sample conditions, including:

[0012] Based on the meta-learning framework, the multimodal feature representation is divided into a support set and a query set, and a hierarchical optimization architecture containing a basic learning layer and a meta-learning layer is constructed. The basic learning layer is used to learn local parameters of the multimodal feature representation, and the meta-learning layer is used to adjust global meta-parameters.

[0013] In the basic learning layer, the multimodal feature representation of the support set is used for forward propagation calculation to obtain preliminary recognition results and calculate preliminary loss values;

[0014] In the meta-learning layer, the preliminary identification results are verified using the multimodal feature representation of the query set, and the meta-loss value is calculated;

[0015] A normalized loss mechanism is used to scale and smooth the initial loss value and the original loss value respectively to obtain the normalized loss.

[0016] Based on the normalized loss, the local parameters of the basic learning layer and the global meta-parameters of the meta-learning layer are alternately optimized using the backpropagation algorithm;

[0017] By repeatedly executing the parameter optimization process, the multimodal feature representation is gradually adapted to small sample conditions, thereby achieving the adaptation and generalization of radar signal modulation recognition rules under small sample conditions.

[0018] Optionally, the step of alternately optimizing the local parameters of the basic learning layer and the global meta-parameters of the meta-learning layer based on the normalized loss using the backpropagation algorithm includes:

[0019] The local parameter gradients of the base learning layer are calculated based on the normalized loss, and the local parameters of the base learning layer are updated using the gradient descent algorithm.

[0020] The normalization loss is recalculated based on the updated local parameters of the base learning layer.

[0021] The gradient of the global meta-parameters of the meta-learning layer is calculated based on the recalculated normalized loss, and the global meta-parameters of the meta-learning layer are updated using the gradient descent algorithm.

[0022] The alternating update process of the local parameters and global meta-parameters is repeatedly executed to gradually reduce the normalization loss until the preset convergence condition or the maximum number of iterations is reached.

[0023] Optionally, the step of performing multimodal feature alignment processing on the target enhanced polarization features and the modulation feature set, and generating a multimodal feature representation by jointly mapping the target enhanced polarization features and the modulation feature set and measuring their similarity, includes:

[0024] The target enhanced polarization features and the modulation feature set are input into a mapping network, and the two types of features are mapped to the same feature space through nonlinear transformation to obtain a joint feature mapping;

[0025] Calculate the cosine similarity between the target enhanced polarization feature and each feature vector in the modulation feature set in the joint feature mapping, and establish a feature correlation matrix;

[0026] Based on the feature correlation matrix, the features in the target enhanced polarization features and the modulation feature set are classified into categories, and the features with cosine similarity higher than a set threshold are classified into the same category.

[0027] Based on the classification results, calculate the mean of all feature vectors within each category, and use it as the feature center point of the corresponding category;

[0028] The feature vectors, their categories, and the feature centroids corresponding to those categories in the joint feature mapping are combined to generate a multimodal feature representation.

[0029] Optionally, constraining the intra-class distribution of the multimodal feature representation based on the central cosine loss function to reduce intra-class feature differences includes:

[0030] The cosine distance between each feature vector in the multimodal feature representation and the feature center point of its class is calculated using the center cosine loss function to obtain the intra-class distribution constraint loss.

[0031] The intra-class distribution constraint loss is optimized by backpropagation algorithm, and the parameters of the mapping network are adjusted so that feature vectors of the same class gather towards the corresponding feature center point in the feature space, thereby reducing intra-class feature differences.

[0032] After iterative optimization, the output is a multimodal feature representation with reduced intra-class differences.

[0033] Optionally, the step of performing channel quality analysis on the signal channels corresponding to the multi-polarization feature sequence, performing multi-channel fusion based on the analysis results to form channel fusion features, and performing spatial interference suppression processing on the channel fusion features to generate target enhanced polarization features includes:

[0034] Quality analysis is performed on the signal channels corresponding to different polarization directions in the multi-polarization feature sequence, and the signal-to-noise ratio and signal stability index of each signal channel are calculated.

[0035] Based on the signal-to-noise ratio and signal stability index, a weighted average method is used to determine the channel quality weight of each signal channel. The higher the signal-to-noise ratio and the better the signal stability index, the higher the channel quality weight of the corresponding signal channel.

[0036] Based on the channel quality weights, a feature-level fusion method is used to weight and fuse the feature sequences of different polarization directions in the multi-polarization feature sequence, so that the features of high-quality channels occupy a larger proportion in the fusion result, in order to form channel fusion features.

[0037] Spatial distribution analysis is performed on the channel fusion features to identify abnormal feature components caused by external interference, and filtering methods are used to remove the abnormal feature components.

[0038] The channel fusion features after removing anomalous feature components are enhanced to generate target enhanced polarization features.

[0039] Optionally, the acquisition of the multi-polarization radar echo signal of the target object, and the preprocessing of the multi-polarization radar echo signal to generate a multi-polarization feature sequence and a modulation feature set, include:

[0040] The radar receiving equipment collects multi-polarization radar echo signals of the target object in multiple polarization directions, and performs segmentation processing on the multi-polarization radar echo signals to obtain multiple signal segments.

[0041] Calculate the energy value of each signal segment, filter out the corresponding signal segments whose energy values ​​meet the conditions based on a preset energy threshold, and perform spectrum analysis on the filtered signal segments to extract the peak frequency characteristics and bandwidth characteristics of each signal segment.

[0042] The peak frequency and bandwidth characteristics of the signal segments in the same polarization direction are arranged in chronological order to form a feature sequence for the corresponding polarization direction. The feature sequences of all polarization directions are integrated to generate a complete multi-polarization feature sequence.

[0043] Extract phase change information and frequency change information from each signal segment, and calculate the amount of phase change and frequency change.

[0044] Based on the phase and frequency changes of all signal segments, a set of modulation features is constructed.

[0045] Secondly, this application provides a radar signal small-sample modulation recognition system based on meta-learning, comprising:

[0046] The acquisition module is used to acquire the multi-polarization radar echo signal of the target object, perform signal preprocessing on the multi-polarization radar echo signal, and generate a multi-polarization feature sequence and a modulation feature set.

[0047] The first generation module is used to perform channel quality analysis on the signal channels corresponding to the multi-polarization feature sequence, perform multi-channel fusion based on the analysis results to form channel fusion features, and perform spatial interference suppression processing on the channel fusion features to generate target enhanced polarization features.

[0048] The second generation module is used to perform multimodal feature alignment processing on the target enhanced polarization features and the modulation feature set. By jointly mapping the target enhanced polarization features and the modulation feature set and measuring similarity, a multimodal feature representation is generated. The intra-class distribution of the multimodal feature representation is constrained based on the center cosine loss function to reduce intra-class feature differences.

[0049] The optimization module is used to optimize and train the multimodal feature representation based on the meta-learning framework. By constructing a hierarchical optimization architecture, the multimodal feature representation is subjected to parameter learning and meta-parameter adjustment at different levels. The normalized loss mechanism is used to standardize the training process so as to achieve the adaptation and generalization of radar signal modulation recognition rules under small sample conditions.

[0050] Thirdly, this application provides an electronic device, comprising:

[0051] Memory, used to store computer programs;

[0052] A processor, configured to execute the computer program to implement the steps of the radar signal small sample modulation recognition method based on meta-learning as described in the first aspect above.

[0053] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the radar signal small sample modulation recognition method based on meta-learning as described in the first aspect above.

[0054] The radar signal modulation recognition method based on meta-learning provided in this application collects and preprocesses multi-polarization radar echo signals of target objects to extract multi-polarization and modulation features, laying the foundation for recognition. Through channel quality analysis, multi-channel fusion, and spatial interference suppression, target enhanced polarization features with strong anti-interference capabilities are generated. By aligning and jointly mapping multi-modal features and combining them with the center cosine loss function, accurate fusion is achieved and the discriminative power is improved. Based on meta-learning, a hierarchical architecture is constructed to complete parameter learning and adjustment under small sample conditions. Finally, in scenarios with complex electromagnetic interference, low signal-to-noise ratio, and scarce samples, the method achieves rapid and accurate recognition of radar signal modulation, ensuring reliable system operation.

[0055] Furthermore, a hierarchical optimization architecture is constructed. The basic learning layer is responsible for learning the local parameters of multimodal features, while the meta-learning layer adjusts the global meta-parameters. In the basic learning layer, preliminary results and losses are obtained through forward propagation using support set multimodal features. The meta-learning layer verifies and calculates the meta-loss using query set features. After processing with a normalized loss mechanism, parameters are alternately optimized through backpropagation. Multiple iterations adapt the multimodal features to small samples, achieving adaptation and generalization of radar signal modulation recognition rules. This method separates local and global parameter learning, precisely controlling the direction of multimodal feature learning. Combining support sets, query sets, and a normalized loss mechanism reduces the impact of small sample data fluctuations and stabilizes loss calculation. Through alternating parameter optimization and iteration, the model's adaptability to small samples is enhanced, improving the adaptation accuracy and generalization ability of radar signal modulation recognition, and ensuring the reliability of recognition results under scarce samples. Attached Figure Description

[0056] To more clearly illustrate the technical solutions of the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below 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.

[0057] Figure 1 A flowchart illustrating a radar signal small-sample modulation recognition method based on meta-learning provided in this application embodiment;

[0058] Figure 2 A schematic diagram illustrating a specific implementation of a radar signal small-sample modulation recognition method based on meta-learning, provided in this application embodiment;

[0059] Figure 3 A schematic diagram illustrating a specific implementation of a radar signal small-sample modulation recognition method based on meta-learning, provided in this application embodiment;

[0060] Figure 4 This is a schematic diagram of the structure of a radar signal small sample modulation recognition system based on meta-learning, provided in an embodiment of this application. Detailed Implementation

[0061] In high-speed maneuvering scenarios in spacecraft orbit, complex electromagnetic interference causes a sharp drop in radar echo signal-to-noise ratio and a scarcity of modulation samples. Existing radar signal modulation recognition schemes based on deep convolutional neural networks combined with time-frequency analysis have significant shortcomings. This scheme relies on a large number of diverse training samples, which are prone to overfitting when samples are scarce, resulting in a significant decrease in generalization ability and recognition accuracy. At the same time, it has limited ability to suppress spatial interference, and interference information in time-frequency images can lead to feature extraction biases, misleading recognition results. Furthermore, the complex network structure and large computational load are difficult to adapt to the limited computing resources and energy supply of spacecraft, and cannot meet the real-time recognition requirements, seriously affecting the reliable operation of spacecraft navigation and communication systems.

[0062] To address the aforementioned issues, this application proposes a small-sample modulation recognition method for radar signals based on meta-learning. This method first acquires and preprocesses multi-polarization radar echo signals, extracting multi-polarization feature sequences and modulation feature sets. Then, it generates target-enhanced polarization features through channel quality analysis, multi-channel fusion, and spatial interference suppression, effectively reducing the impact of interference. Subsequently, it aligns these features with the modulation feature set using multi-modal features, and combines a center cosine loss function to reduce intra-class feature differences and improve feature discriminative power. Finally, it constructs a hierarchical optimization architecture based on a meta-learning framework, using a normalized loss mechanism to standardize training. This scheme does not require large-scale samples. Through multi-modal feature fusion and meta-learning optimization, it can achieve the adaptation and generalization of recognition rules under small-sample conditions, solving the overfitting problem. It can also eliminate the impact of interference on recognition through spatial interference suppression and precise feature processing. Furthermore, the hierarchical optimization architecture simplifies the training process, reduces computational burden, meets real-time requirements, fundamentally solves the shortcomings of existing technologies, and ensures the stable operation of spacecraft in complex scenarios.

[0063] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0064] The core of this application is to provide a radar signal small-sample modulation recognition method based on meta-learning, and a flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes:

[0065] S101. Acquire the multi-polarization radar echo signal of the target object, perform signal preprocessing on the multi-polarization radar echo signal, and generate a multi-polarization feature sequence and a modulation feature set.

[0066] Optionally, step S101 may specifically include the following steps:

[0067] S1011. Collect multi-polarization radar echo signals of the target object in multiple polarization directions through radar receiving equipment, and perform segmentation processing on the multi-polarization radar echo signals to obtain multiple signal segments.

[0068] S1012. Calculate the energy value of each signal segment, filter out the corresponding signal segments whose energy values ​​meet the conditions based on the preset energy threshold, and perform spectrum analysis on the filtered signal segments to extract the peak frequency characteristics and bandwidth characteristics of each signal segment.

[0069] S1013. Arrange the peak frequency characteristics and bandwidth characteristics of the signal segments in the same polarization direction in chronological order to form a feature sequence for the corresponding polarization direction, and integrate the feature sequences of all polarization directions to generate a complete multi-polarization feature sequence.

[0070] S1014. Extract phase change information and frequency change information from each signal segment, and calculate the amount of phase change and frequency change.

[0071] S1015. Based on the phase change and frequency change of all signal segments, construct a modulation feature set.

[0072] In the above scheme, the multi-polarization radar echo signal is the signal reflected back from a target object after the radar emits electromagnetic waves with different polarization directions, containing information such as the target's position and shape. Signal preprocessing involves a series of processes on the acquired raw radar echo signal, such as noise removal and filtering of valid signals, to make the signal more suitable for subsequent analysis. The multi-polarization feature sequence is a sequence formed by integrating signal features from different polarization directions and arranging them in chronological order, which can reflect the signal's variation patterns under different polarization directions. The modulation feature set is a collection of features reflecting the signal modulation mode, such as phase changes and frequency changes, used for subsequent identification of the signal's modulation type.

[0073] In this embodiment of the application, firstly, through step S1011, the radar receiving device collects multi-polarization radar echo signals of the target object in multiple polarization directions, such as horizontal polarization and vertical polarization; then, these continuous signals are divided into multiple independent signal segments according to a fixed time interval. For example, for a certain spacecraft, the radar receiving device continuously receives echo signals in both horizontal and vertical polarization directions, and the signal that lasts for 10 seconds is divided into 100 signal segments in 0.1-second intervals.

[0074] Secondly, in step S1012, the energy value of each signal segment over a certain period of time is calculated, and a preset energy threshold is set, such as a value determined based on the performance of the radar equipment and the level of environmental noise. Signal segments with energy values ​​below this threshold may be noise. Signal segments with energy values ​​greater than this threshold are then selected. Next, spectral analysis is performed on each selected signal segment to extract the frequency with the strongest signal strength in each segment, i.e., the peak frequency characteristic, as well as the frequency range occupied by the signal, i.e., the bandwidth characteristic. For example, after calculating the energy values ​​of 100 signal segments, a threshold of 5 is set, and 80 signal segments with energy values ​​greater than 5 are selected. Spectral analysis is performed on each of these 80 signal segments to obtain the peak frequency and bandwidth of each signal segment. For example, a signal segment may have a peak frequency of 1000Hz and a bandwidth of 50Hz.

[0075] Next, in step S1013, the peak frequency and bandwidth characteristics of all filtered signal segments under the same polarization direction are arranged sequentially according to the time sequence of signal segment acquisition, forming a feature sequence corresponding to that polarization direction. Then, the feature sequences of all polarization directions are integrated together to generate a complete multi-polarization feature sequence. For example, the peak frequency and bandwidth characteristics of 80 signal segments in the horizontal polarization direction are arranged in chronological order to form a horizontal polarization feature sequence, and the vertical polarization direction is similarly formed into a vertical polarization feature sequence. Integrating these two sequences yields the multi-polarization feature sequence.

[0076] Then, in step S1014, phase change information and frequency change information are extracted from each filtered signal segment. Based on this information, the amplitude of phase change over a certain period of time is calculated as the phase change amount, and the amplitude of frequency change is calculated as the frequency change amount. For example, if the phase changes from 0° to 30° within 0.1 seconds from a certain signal segment, the phase change amount is 30°, and the frequency changes from 1000Hz to 1050Hz, the frequency change amount is 50Hz.

[0077] Finally, in step S1015, the phase and frequency changes calculated from all the filtered signal segments are collected and organized into a set containing these changes, i.e., a modulation feature set. For example, the phase and frequency changes of all 80 signal segments are collected to form a modulation feature set.

[0078] In practical applications, during high-speed maneuvering scenarios of spacecraft A in orbit, radar receiver B is used to collect multi-polarization radar echo signals from spacecraft A in both horizontal and vertical polarization directions. First, the received continuous echo signals are divided into 120 signal segments at 0.1-second intervals. Then, the energy value of each signal segment is calculated, and a preset energy threshold of 6 is set. 95 signal segments with energy values ​​greater than 6 are selected, and spectral analysis is performed on these 95 segments to extract the peak frequency and bandwidth characteristics of each segment. Subsequently, the peak frequencies and bandwidths of the 95 signal segments in the horizontal polarization direction are arranged chronologically to form a horizontal polarization feature sequence, and the same applies to the vertical polarization direction to form a vertical polarization feature sequence. These two sequences are then integrated to obtain the multi-polarization feature sequence. Next, phase change and frequency change information are extracted from each selected signal segment, and the phase and frequency changes of each signal segment are calculated. Finally, the phase and frequency changes of these 95 signal segments are compiled to construct a modulation feature set.

[0079] The above-mentioned S101 overall scheme can effectively filter out useful signal segments from the multi-polarization radar echo signal of the target object, remove redundant noise signals, and extract multi-polarization feature sequences that reflect the multi-polarization characteristics of the signal and modulation feature sets that reflect the modulation mode of the signal. This provides high-quality and effective basic data for further processing and modulation identification of radar signals, making subsequent analysis and identification work more targeted and reducing the interference of invalid data on subsequent processes.

[0080] S102. Perform channel quality analysis on the signal channels corresponding to the multi-polarization feature sequence, perform multi-channel fusion based on the analysis results to form channel fusion features, and perform spatial interference suppression processing on the channel fusion features to generate target enhanced polarization features.

[0081] Optionally, step S102 may specifically include the following steps:

[0082] S1021. Perform quality analysis on the signal channels corresponding to different polarization directions in the multi-polarization feature sequence, and calculate the signal-to-noise ratio and signal stability index of each signal channel;

[0083] S1022. Based on the signal-to-noise ratio and signal stability index, the channel quality weight of each signal channel is determined by a weighted average method. The higher the signal-to-noise ratio and the better the signal stability index, the higher the channel quality weight of the corresponding signal channel.

[0084] S1023. Based on the channel quality weights, a feature-level fusion method is used to perform weighted fusion of feature sequences with different polarization directions in the multi-polarization feature sequence, so that the features of high-quality channels occupy a larger proportion in the fusion result, in order to form channel fusion features.

[0085] S1024. Perform spatial distribution analysis on the channel fusion features, identify abnormal feature components caused by external interference, and remove the abnormal feature components using a filtering method.

[0086] S1025. Enhance the channel fusion features after removing anomalous feature components to generate target enhanced polarization features.

[0087] In the above scheme, signal-to-noise ratio (SNR) and signal stability are the core indicators for evaluating signal channel quality. The former reflects the ratio of effective signal to noise, while the latter measures the degree of fluctuation using the data standard deviation. Together, they determine the clarity and stability of signal transmission. Channel quality weights are determined based on signal channel performance. They are calculated by weighted averaging of SNR and signal stability indicators; better performance results in a larger weight, used to allocate data proportions during multi-channel fusion. Channel fusion features are the core data after integrating multi-polarization signal channels. Weighted fusion based on quality weights increases the proportion of high-quality channel features, achieving spatial interference suppression and target feature enhancement. Abnormal feature components are anomalous data in the channel fusion features caused by external interference and exceeding the normal range. These can be identified through spatial analysis and removed through filtering. Target-enhanced polarization features are the final features after interference suppression and optimization. Abnormal components are removed and enhanced to amplify target feature differences, facilitating multi-modal feature alignment and radar signal identification.

[0088] In this embodiment, firstly, through step S1021, a polarization direction identifier separation algorithm is used to extract feature sequences corresponding to different polarization directions from the multi-polarization feature sequences. Each feature sequence carries data such as peak frequency and bandwidth sorted by time under that polarization direction, thereby determining the dedicated signal channel. For example, horizontal polarization and vertical polarization feature sequences containing data at multiple time points are separated, corresponding to horizontal and vertical polarization signal channels respectively. Then, a signal strength extraction tool is used to filter the effective signal strength corresponding to the peak frequency in the feature sequences of each channel, and the background noise intensity is collected to calculate the signal-to-noise ratio of each channel. Simultaneously, a sliding window is used to select the peak frequency data of multiple consecutive time points in the feature sequences of each channel. The average value is calculated first, and then the data fluctuation degree is obtained through the standard deviation calculation formula. The obtained standard deviation is used as the signal stability index.

[0089] Secondly, in step S1022, based on the radar signal identification requirements, the weights of signal-to-noise ratio (SNR) and signal stability index in the channel quality assessment are both set to 50%. Then, a feature normalization algorithm is used to normalize the SNR and signal stability index of each channel obtained in step S1021. For example, the SNR of the horizontal polarization channel is 11, and the SNR of the vertical polarization channel is 5.33. The calculated proportions of the horizontal and vertical polarization SNRs are approximately 0.674 and 0.326, respectively. The proportions of the horizontal polarization signal stability index are 7 and the vertical polarization SNR is 14, respectively, are approximately 0.333 and 0.667. Finally, using a weighted average method, based on the formula: Channel Quality Weight = SNR Proportion × SNR Weight + Signal Stability Index Proportion × Stability Weight, the weights of each channel are calculated. Because the horizontal polarization channel has a higher SNR, its weight is slightly higher than that of the vertical polarization channel, achieving the goal of allocating weights according to quality.

[0090] Next, in step S1023, a timestamp matching algorithm is used to perform time synchronization verification on the feature sequences of different polarization directions separated in step S1021 to ensure that the data at the same time point correspond. Then, based on the channel quality weights calculated in step S1022, a feature weighted fusion algorithm is used to multiply the feature data of each polarization direction feature sequence at each time point by the corresponding channel weight. For example, the horizontal polarization channel has a weight of 0.5035, and its peak frequency at the 50th time point is 305Hz and bandwidth is 82Hz. After multiplying by the weight, these are approximately 153.57Hz and 41.29Hz, respectively. The same applies to the vertical polarization channel. Then, the feature data of different polarization channels at each time point after weight reduction are added together to obtain the fused feature data. For example, the fused peak frequency at the above time point is approximately 296.69Hz and the fused bandwidth is approximately 79.78Hz. Finally, the fused data of all time points are arranged in chronological order to form a channel fusion feature that integrates the advantages of multiple channels and has a higher proportion of high-quality channel features.

[0091] Then, in step S1024, the spatial distribution of the channel fusion features is analyzed. A three-dimensional spatial modeling tool is used to construct a three-dimensional distribution model according to the spatial receiving position and time dimension. The distribution range of normal feature data is determined by the data distribution fitting algorithm, such as peak frequency 285-305Hz and bandwidth 75-85Hz. Data outside the range is marked as suspected anomalies. Then, combined with the spacecraft on-orbit environment database, the interference source tracing algorithm is used to perform interference matching on the suspected anomaly components to confirm whether they are external electromagnetic interference. Finally, the mean filtering algorithm is used to select the average value of 5 consecutive normal data before and after the abnormal data to replace the abnormal data, remove the abnormal feature components, and retain the normal data structure.

[0092] Finally, in step S1025, the channel fusion features after removing anomalous components in step S1024 are subjected to feature enhancement preprocessing. Data standardization is used to eliminate magnitude differences. A contrast enhancement algorithm is used to amplify the difference between the target and the background and suppress noise. A feature sharpening algorithm is used to enhance the temporal variation trend. Finally, feature integrity verification is performed to generate the final target enhanced polarization features that accurately reflect the characteristics of the target radar signal, providing data support for multimodal feature alignment.

[0093] In practical applications, under the radar monitoring scenario of spacecraft B's high-speed maneuvering in orbit, multi-polarization feature sequences containing horizontal and vertical polarization feature sequences are processed. First, two signal channel feature sequences are separated: the horizontal polarization channel has an effective signal strength of 25 and a noise level of 3, with a signal-to-noise ratio (SNR) of approximately 8.33; the vertical polarization channel has an effective signal strength of 18 and a noise level of 4, with a SNR of 4.5. Peak frequency data from 100 consecutive time points are taken for each channel. The average value for the horizontal polarization channel is 315Hz, with a sum of squared differences of 9000, resulting in a signal stability index of 9.5; the average value for the vertical polarization channel is 300Hz, with a sum of squared differences of 16000, resulting in a signal stability index of 12.6. Next, the weights for both SNR and signal stability index are set to 50%. After normalization, the total SNR reaches 12.83, with horizontal polarization accounting for approximately 0.65 and vertical polarization accounting for 0.35; the total stability index is 22.1, with horizontal polarization accounting for approximately 0.43 and vertical polarization accounting for 0.57. Through weighted calculation, the horizontal polarization channel weight was determined to be 0.54, and the vertical polarization channel weight to be 0.46. Then, ensuring a one-to-one correspondence between the data at 100 time points of the two feature sequences, the data at each time point of the horizontal polarization were multiplied by 0.54, and the data at each time point of the vertical polarization were multiplied by 0.46, and then summed to form the channel fusion feature. Subsequently, a three-dimensional spatial domain model was constructed, setting the normal peak frequency range to 295-325Hz and the bandwidth range to 72-88Hz, identifying three anomalous data points, such as a point with a peak frequency of 350Hz and a bandwidth of 95Hz. Based on the on-orbit environment database, these anomalies were confirmed to be caused by external interference. A mean filtering method was used, replacing the anomalous value with the average of the five normal data points before and after the anomalous point. Finally, the fused feature after removing the anomalies was standardized, the difference between the target and the background was amplified by contrast enhancement, and the time trend was enhanced by feature sharpening. After verification, the target-enhanced polarization feature was generated.

[0094] The overall scheme of S102 described above achieves reasonable fusion of multi-channel features by performing quality analysis on the signal channels and assigning different weights, allowing the effective information of high-quality channels to be fully utilized. At the same time, it removes abnormal interference components through spatial interference suppression, and then enhances the target features through enhancement processing. The generated target enhanced polarization features not only integrate the advantageous information of multi-polarization channels, but also reduce the impact of interference, making the feature data more accurate and more targeted. This provides a high-quality feature foundation for subsequent multimodal feature alignment and modulation recognition, and improves the reliability of the subsequent recognition process.

[0095] S103. Perform multimodal feature alignment processing on the target enhanced polarization feature and the modulation feature set. Generate multimodal feature representation by joint mapping and similarity measurement of the target enhanced polarization feature and the modulation feature set. Constrain the intra-class distribution of the multimodal feature representation based on the center cosine loss function to reduce intra-class feature differences.

[0096] Optionally, step S103 may specifically include the following steps:

[0097] S1031. Input the target enhanced polarization features and the modulation feature set into the mapping network, and map the two types of features to the same feature space through nonlinear transformation to obtain joint feature mapping;

[0098] S1032. Calculate the cosine similarity between the target enhanced polarization feature and each feature vector in the modulation feature set in the joint feature mapping, and establish a feature correlation matrix;

[0099] S1033. Based on the feature correlation matrix, classify the features in the target enhanced polarization features and the modulation feature set, and classify the features with cosine similarity higher than a set threshold into the same category.

[0100] S1034. Based on the classification results, calculate the mean of all feature vectors in each category, and use it as the feature center point of the corresponding category.

[0101] S1035. Combine the feature vectors, the categories to which the feature vectors belong, and the feature center points corresponding to the categories in the joint feature mapping to generate a multimodal feature representation.

[0102] S1036. Calculate the cosine distance between each feature vector in the multimodal feature representation and the feature center point of its class using the center cosine loss function to obtain the intra-class distribution constraint loss;

[0103] S1037. Optimize the intra-class distribution constraint loss through the backpropagation algorithm, adjust the parameters of the mapping network, and make the feature vectors of the same class gather towards the corresponding feature center point in the feature space to reduce the intra-class feature differences.

[0104] S1038. After iterative optimization, output the multimodal feature representation with reduced intra-class differences.

[0105] In the above scheme, joint feature mapping refers to the integrated feature data after transforming different types of features into the same space. It includes a unified dimension vector obtained by nonlinear transformation of the target enhanced polarization feature set and the modulation feature set through a mapping network, which can be used for subsequent feature similarity calculation and category classification. The feature association matrix is ​​a matrix data structure that reflects the similarity of different feature vectors. It includes the cosine similarity value between the target enhanced polarization feature vector and the modulation feature set vector, which can be used to intuitively display the correlation strength between features and assist in feature category classification. Multimodal feature representation refers to the feature data form that integrates multiple types of features and category information. It includes the feature vector in the joint feature mapping, the category label determined by the category classification, and the feature centroid information of the corresponding category, which can be used for subsequent intra-class distribution constraints and feature optimization training. Intra-class distribution constraint loss refers to the loss value that measures the distance between the feature vector and the center of its category. It includes the cosine distance result between the feature vector and the category centroid calculated by the center cosine loss function, which can be used to constrain the spatial distribution of similar features and reduce intra-class feature differences.

[0106] In the embodiments of this application, such as Figure 2 As shown, firstly, the target enhanced polarization features and modulation feature sets are input into the mapping network via S1031. This mapping network adopts a neural network structure and uses a nonlinear transformation algorithm to transform the two types of features of different dimensions into the same feature space to obtain a joint feature mapping. For example, the "peak frequency 313Hz, bandwidth 79Hz" at a certain time point in the target enhanced polarization features is converted into a 50-dimensional vector [0.21, 0.35, ..., 0.18], and the corresponding signal segment "phase change 35°, frequency change 30Hz" in the modulation feature set is converted into a 50-dimensional vector [0.23, 0.32, ..., 0.16], forming two sets of vectors in the joint feature mapping.

[0107] Secondly, in step S1032, the cosine similarity between each vector of the target-enhanced polarization feature in the joint feature mapping and each vector in the modulation feature set is calculated. The cosine similarity is calculated using the formula... ,in It is the target enhanced polarization feature vector. The more the cosine similarity is close to 1, the more similar the two features are. Then, all similarity results are arranged in matrix form to establish a feature association matrix. For example, if there are 100 vectors of target enhancement polarization features and 100 vectors of modulation feature set, a 100×100 feature association matrix is ​​calculated. The value at position (i,j) in the matrix is ​​the cosine similarity between the i-th polarization feature vector and the j-th modulation feature vector.

[0108] Then, through step S1033, based on the feature correlation matrix, a similarity threshold is set, and feature vector pairs with values ​​higher than the threshold in the matrix are determined to have strong correlation. Then, the features in the target enhanced polarization feature and modulation feature set are classified into categories. Features with cosine similarity higher than the set threshold are classified into the same category. For example, the similarity at position (5,5) in the matrix is ​​0.85, which is higher than the threshold of 0.8. Then, the 5th polarization feature vector and the 5th modulation feature vector are classified into category 1.

[0109] Next, in step S1034, based on the classification results, the mean calculation method is used to average the values ​​of each dimension for all feature vectors in each category. The average vector is used as the feature center point of the corresponding category. For example, category 1 contains 10 feature vectors, each with 50 dimensions. The 10 values ​​of each dimension are added together and then divided by 10 to obtain the 50-dimensional feature center point of category 1. Next, through S1035, the three parts of information in the joint feature map are combined: each feature vector, the category to which the vector belongs determined by the category division, and the feature centroid corresponding to the category, to generate a multimodal feature representation. For example, if a feature vector is [0.21,0.35,...,0.18], the category is 1, and the feature centroid of category 1 is [0.22,0.34,...,0.17], then the multimodal feature representation is (vector [0.21,0.35,...,0.18], category 1, centroid [0.22,0.34,...,0.17]).

[0110] Next, using S1036, the cosine distance between each feature vector in the multimodal feature representation and the center point of its respective category feature is calculated using the center cosine loss function. The formula for the center cosine loss function is as follows: in It is an eigenvector. The center point of the category y to which the feature vector belongs is the center point. The result calculated is the intra-class distribution constraint loss. The smaller the loss value, the closer the feature vector is to the category center, and the more concentrated the intra-class distribution. For example, a certain feature vector... Center point of its category The dot product is 8.5. The modulus length is 3. The modulus is 3. Substituting into the formula, we get... That is, the intra-class distribution constraint loss is 0.06. Then, through S1037, the calculated intra-class distribution constraint loss is used as the optimization objective. The backpropagation algorithm is used to deduce from the loss result and adjust the parameters of each layer in the mapping network, such as the weights and biases of the neural network, so that feature vectors of the same class gather towards the corresponding feature center point in the feature space. For example, when the loss value of a certain class feature is large, the mapping network parameters are adjusted through backpropagation so that the feature vectors of that class are closer to the class center point in the next mapping, thereby reducing the intra-class feature difference.

[0111] Finally, through step S1038, the process from step S1031 to step S1037 is repeated to perform multiple iterations of optimization. In each iteration, the mapping network parameters are adjusted according to the loss value until the loss value stabilizes at a low level. At this point, the multimodal feature representation with reduced intra-class differences is output. In this feature representation, features of the same class are more concentrated, and features of different classes are more clearly distinguished.

[0112] In practical applications, in radar monitoring scenarios where spacecraft D encounters complex electromagnetic interference during its on-orbit mission, multimodal feature alignment and optimization are performed on target enhanced polarization features containing fused peak frequency (e.g., 305-320Hz range) and bandwidth (e.g., 75-82Hz range) data at 90 time points, as well as modulation feature sets containing phase change (e.g., 28°-38° range) and frequency change (e.g., 25-35Hz range) data for 90 signal segments. First, the two types of feature inputs are transformed into a 45-dimensional feature space through a mapping network consisting of one input layer, three hidden layers (using the ReLU activation function), and one output layer, resulting in a joint feature mapping containing 90 polarization feature vectors and 90 modulation feature vectors. Then, the similarity of each pair of vectors is calculated using the cosine similarity formula, establishing a 90×90 feature association matrix. The value at position (i,j) in the matrix represents the similarity between the i-th polarization vector and the j-th modulation vector. Subsequently, a similarity threshold of 0.8 is set, and features in the matrix exceeding this threshold are classified into six categories. For example, polarization vectors 12-20 and their corresponding modulation vectors are classified into category 3 because their similarity exceeds 0.8. The mean of each dimension of all vectors within each category is calculated to obtain six 45-dimensional category feature centroids, such as [0.21, 0.33, ..., 0.17] for category 3. Then, each feature vector, its category label, and corresponding centroid are combined to generate a multimodal feature representation. The cosine distance between each vector and the category center is calculated using the center cosine loss function to obtain the intra-class distribution constraint loss, such as a loss value of 0.07 for a certain vector. The weights and biases of the mapping network are adjusted using the backpropagation algorithm to cause vectors of the same class to cluster towards the centroid. This process is iterated 90 times until the loss value stabilizes at a low level, ultimately outputting a multimodal feature representation with reduced intra-class differences, which is then used for subsequent optimization training of the meta-learning framework.

[0113] The overall scheme of S103 described above maps the target-enhanced polarization features and modulation feature sets to the same space through multimodal feature alignment processing, achieving effective association between different types of features. Similarity measurement and category classification clarify the category affiliation of features. By leveraging the center cosine loss function and backpropagation optimization, intra-class feature differences are reduced, resulting in more concentrated and discriminative features of the same type in the generated multimodal feature representation. This process fully integrates the advantages of both types of features, providing high-quality, highly discriminative feature data for subsequent optimization training based on the meta-learning framework, which helps improve the accuracy and generalization ability of radar signal modulation recognition under small sample conditions.

[0114] S104. The multimodal feature representation is optimized and trained based on the meta-learning framework. By constructing a hierarchical optimization architecture, the multimodal feature representation is subjected to parameter learning and meta-parameter adjustment at different levels. The normalization loss mechanism is used to standardize the training process so as to achieve the adaptation and generalization of radar signal modulation recognition rules under small sample conditions.

[0115] Optionally, step S104 may specifically include the following steps:

[0116] S1041. Based on the meta-learning framework, the multimodal feature representation is divided into a support set and a query set, and a hierarchical optimization architecture containing a basic learning layer and a meta-learning layer is constructed. The basic learning layer is used to learn local parameters of the multimodal feature representation, and the meta-learning layer is used to adjust global meta-parameters.

[0117] S1042. In the basic learning layer, the multimodal feature representation of the support set is used to perform forward propagation calculation to obtain preliminary recognition results and calculate preliminary loss values.

[0118] S1043. In the meta-learning layer, the multimodal feature representation of the query set is used to verify the preliminary identification result, and the meta-loss value is calculated.

[0119] S1044. Using a normalized loss mechanism, the initial loss value and the original loss value are scaled and smoothed respectively to obtain a normalized loss.

[0120] S1045. Based on the normalized loss, the local parameters of the basic learning layer and the global meta-parameters of the meta-learning layer are alternately optimized using the backpropagation algorithm.

[0121] S1046. By repeatedly executing the parameter optimization process, the multimodal feature representation is gradually adapted to small sample conditions, so as to realize the adaptation and generalization of radar signal modulation recognition rules under small sample conditions.

[0122] Specifically, step S1045 includes the following process: calculating the local parameter gradient of the base learning layer based on the normalized loss, and updating the local parameters of the base learning layer using the gradient descent algorithm; recalculating the normalized loss based on the updated local parameters of the base learning layer; calculating the global meta-parameter gradient of the meta-learning layer based on the recalculated normalized loss, and updating the global meta-parameter of the meta-learning layer using the gradient descent algorithm; repeatedly executing the alternating update process of the local parameters and global meta-parameters, so that the normalized loss gradually decreases until the preset convergence condition or the maximum number of iterations is reached.

[0123] In the above scheme, the hierarchical optimization architecture refers to a functionally layered structure used for meta-learning training. It includes a basic learning layer responsible for learning local parameters of multimodal feature representations and a meta-learning layer responsible for controlling the overall training direction. This can be used to hierarchically optimize feature training parameters under small sample conditions, improving training efficiency and adaptability. Local parameters are parameters in the basic learning layer used to process multimodal feature representations and are used to adapt feature learning to specific categories. Global meta-parameters are parameters in the meta-learning layer that control the overall training direction and are used to regulate the commonalities in training different categories of features. The preliminary loss value measures the preliminary training error of the basic learning layer, including the difference between the preliminary identification results of the support set calculated using the loss function and the actual category labels. It can be used to reflect the learning effect of the basic learning layer on the support set features. The meta-loss value evaluates the model's generalization ability on non-training data, including the difference between the query set identification results calculated using the same loss function and the actual category labels. It can be used to verify the reliability of the training effect of the basic learning layer. Normalized loss refers to the standardized loss data after processing the loss values ​​during training. It includes the results of scaling and smoothing the initial loss value and the meta-loss value using a normalized loss mechanism. It can be used to standardize the training process, avoid abnormal fluctuations in loss affecting the direction of parameter optimization, and ensure training stability.

[0124] In this embodiment, firstly, through step S1041, based on the meta-learning framework, a data partitioning algorithm is used to divide the multimodal feature representation into a support set (for initial training) and a query set (for verifying training effectiveness) according to a certain ratio. During this process, the selection of the data partitioning ratio needs to comprehensively consider the data scale and model complexity. If the support set accounts for too high a proportion, it may lead to insufficient validation; if the proportion is too low, there will be insufficient training samples. Simultaneously, a hierarchical optimization architecture is constructed. This architecture includes a basic learning layer and a meta-learning layer. The basic learning layer consists of local parameters, used for targeted local parameter learning of the multimodal feature representation. The meta-learning layer consists of global meta-parameters, used for adjusting the global meta-parameters of the overall training. The purpose of this hierarchical optimization architecture design is to allow the model to grasp the global training trend while learning local features, thereby improving training efficiency and model generalization ability.

[0125] Secondly, in step S1042, in the basic learning layer, the multimodal feature representations of the support set are input into the training model. Using the forward propagation algorithm, the feature data sequentially passes through the input layer and hidden layer of the model, finally obtaining a preliminary recognition result from the output layer, such as the predicted label of the category to which each feature belongs. For example, if a feature in the support set represents category 3, the model outputs a preliminary recognition result with a probability of 0.85 for category 3. Then, a loss function is used to calculate the difference between the preliminary recognition result and the actual category label, obtaining a preliminary loss value. For example, the difference between the predicted feature probability and the actual label yields a preliminary loss value of 0.15. The cross-entropy loss function here effectively measures the difference between the predicted distribution and the true distribution, providing a crucial basis for subsequent parameter adjustments.

[0126] Then, in step S1043, in the meta-learning layer, the multimodal feature representation of the query set is input into the model trained by the basic learning layer. Similarly, the recognition result of the query set is obtained through forward propagation. This result is compared with the actual category label of the query set to verify the accuracy of the preliminary recognition result. The difference between the query set recognition result and the actual label is then calculated using the same loss function to obtain the meta-loss value. For example, if the actual category of a feature in the query set is 2, and the probability of the model outputting the recognition result as category 2 is 0.78, the calculated meta-loss value is 0.22. This step, through validation on the query set, can examine the model's performance on unseen data and identify potential overfitting or underfitting issues.

[0127] Next, in step S1044, a normalized loss mechanism is adopted. First, the initial loss value and the original loss value are scaled. The min-max scaling algorithm is used to map the loss value to the [0,1] interval. For example, the initial loss value of 0.15 is scaled to 0.15÷(0.15+0.22)≈0.41, and the original loss value of 0.22 is scaled to 0.22÷(0.15+0.22)≈0.59. Then, smoothing is performed. The moving average algorithm is used to weight the scaled loss value with the normalized loss value of the previous round to reduce loss fluctuations. For example, the normalized loss of the previous round is 0.45, and the initial loss after scaling in this round is 0.41 and the original loss is 0.59. The weighted average normalized loss is 0.41×0.5+0.59×0.5=0.5. Normalization loss mechanism makes loss values ​​of different magnitudes comparable, while moving average can avoid parameter adjustment deviations caused by data fluctuations, making the training process more stable.

[0128] Next, in step S1045, based on the normalized loss, the backpropagation algorithm is used to adjust the local parameters of the basic learning layer and the global meta-parameters of the meta-learning layer through alternating optimization. The specific process is as follows: First, the gradient of the local parameters of the basic learning layer is calculated, that is, the derivative of the loss value with respect to the local parameters. Then, the gradient descent algorithm is used to update the local parameters according to the gradient direction. For example, when the gradient of the local parameters is -0.03 and the learning rate is 0.01, the local parameters will be updated to the original parameters + (-0.03) × 0.01. Based on the updated local parameters, the normalized loss is recalculated, assuming that the updated loss value is reduced to 0.48. Subsequently, the gradient of the global meta-parameters of the meta-learning layer is calculated and updated using the same gradient descent algorithm. For example, if the gradient of the global meta-parameters is 0.02, the learning rate is adjusted from 0.01 to 0.01 + 0.02 × 0.005 = 0.0101. The above alternating update process is repeated until the loss value tends to stabilize. The backpropagation algorithm uses the chain rule to achieve efficient transmission of gradient information; while the alternating optimization strategy effectively balances local feature learning and global model optimization.

[0129] Finally, through step S1046, the parameter optimization process of steps S1041-S1045 is repeated multiple times. Each iteration adjusts local and global meta-parameters, allowing the multimodal feature representation to gradually adapt to data patterns under small sample conditions during training. For example, after multiple iterations, the model's recognition error for small sample data gradually decreases, ultimately achieving adaptation and generalization of radar signal modulation recognition rules under small sample conditions. This enables the model to accurately identify radar signal modulation types in new small sample scenarios. In practical applications, the number of iterations is not fixed and needs to be dynamically adjusted based on model convergence. Simultaneously, an early stopping strategy should be used to avoid overfitting, ensuring the model achieves an optimal balance between training efficiency and performance.

[0130] In practical applications, under the scenario of monitoring small-sample radar signals from spacecraft E encountering complex electromagnetic interference in orbit, for data containing 85 multimodal feature representations (each feature representation containing a 40-dimensional feature vector, its category label, and category centroid), the A-brand meta-learning framework was used for S104 optimization training. First, the data was randomly partitioned into 59 support sets and 26 query sets in a 7:3 ratio. A hierarchical optimization architecture was constructed: the basic learning layer consisted of an input layer receiving 40-dimensional features and a hidden layer containing a ReLU activation function with a 40×25 weight matrix; the meta-learning layer included a learning rate adjustment module with an initial learning rate of 0.01 and a global weight module. Next, in the basic learning layer, the support set feature vectors were input into the model for forward propagation, outputting the category probability of each feature. The overall initial loss value was calculated to be 0.19 using the cross-entropy loss function. Then, in the meta-learning layer, the query sets were input into the trained model, and the recognition results were obtained through forward propagation, resulting in a meta-loss value of 0.26. Then, a normalized loss mechanism is employed. First, min-max scaling is used to map the initial loss and meta-loss to 0.42 and 0.58 respectively. Then, a moving average (smoothing coefficient 0.5) is used to weight the previous normalized loss (0.46) to obtain a current normalized loss of 0.5. Based on this loss value, backpropagation is used for alternating optimization: first, the gradients of local parameters in the basic learning layer are calculated. For example, if a weight gradient is -0.04, gradient descent is used to update the parameters. After the loss is recalculated and reduced to 0.47, the gradients of global meta-parameters in the meta-learning layer are calculated. For example, if the learning rate gradient is 0.02, the learning rate is updated to 0.0101. This optimization process is repeated 90 times, allowing the multimodal feature representation to gradually adapt to the patterns of small sample data. Ultimately, this achieves the adaptation and generalization of radar signal modulation recognition rules under small sample conditions, enabling the model to accurately identify signal modulation types in new small sample scenarios.

[0131] The overall solution of S104 described above is based on a hierarchical optimization architecture built on a meta-learning framework. It achieves feature training and validation by dividing the work between the support set and the query set, and combines the normalization loss mechanism to standardize the training process, effectively avoiding the training instability problem caused by small sample data. By alternately optimizing local parameters and global meta-parameters, the multimodal feature representation gradually adapts to small sample conditions. This not only achieves accurate adaptation of radar signal modulation recognition rules to small sample data, but also improves the generalization ability of rules in new small sample scenarios. It ensures that radar signal modulation recognition can still be completed stably and accurately when samples are scarce, providing reliable model support for subsequent practical applications.

[0132] The following is a complete example for steps 101-104, such as Figure 3As shown, in the radar monitoring scenario of spacecraft F undergoing high-speed maneuvering in orbit and subject to complex electromagnetic interference, the B-type radar receiving equipment performs step S101. First, it acquires multi-polarized radar echo signals from spacecraft F in the horizontal and vertical polarization directions, dividing them into 100 signal segments at 0.1-second intervals. The energy value of each segment is calculated, and 82 effective signal segments are selected using a threshold of 6. Spectral analysis is performed to extract the peak frequency and bandwidth, which range from 290-320Hz and 70-85Hz, respectively. These two values ​​are integrated to form a multi-polarization feature sequence. Simultaneously, phase and frequency change information is extracted from the effective signal segments, calculating the phase change and frequency change ranges to be 25°-40° and 20-35Hz, respectively, thus constructing a modulation feature set.

[0133] Based on the results of step S101, step S102, channel optimization, is performed. Quality analysis of the horizontal and vertical polarization signal channels yields a horizontal channel signal-to-noise ratio (SNR) of 13 and a signal stability index of 0.85, while the vertical channel SNR is 8 and the signal stability index is 0.6. A weighted average method (50% weight each) is used to determine the horizontal channel weight as 0.62 and the vertical channel weight as 0.38. Feature-level fusion is performed on the multi-polarization feature sequences according to these weights. For example, the peak frequency of the horizontal channel at a certain moment (310Hz) multiplied by 0.62 yields 192.2Hz, and the peak frequency of the vertical channel at the same time (295Hz) multiplied by 0.38 yields 112.1Hz. The two are added together to form the channel fusion feature. Spatial distribution analysis is performed on the fusion feature, identifying four outliers caused by external interference. These outliers are replaced with the average value of surrounding normal data (308Hz) using mean filtering, and then contrast enhancement is applied to generate the target enhanced polarization feature.

[0134] Next, step S103, multimodal alignment, is performed. The target enhanced polarization features and modulation feature sets are input into the mapping network and transformed into a 42-dimensional feature space through ReLU nonlinear transformation to obtain a joint feature map. The cosine similarity of each pair of feature vectors is calculated to establish an 82×82 feature association matrix. The features are divided into 5 categories with a threshold of 0.78. The mean of the feature vector of each category is calculated as the feature center point. For example, the center point of category 2 is [0.23, 0.31, ..., 0.16]. The feature vector, its category, and the center point are combined to generate a multimodal feature representation. The center cosine loss function is used to calculate the intra-class distribution constraint loss. The mapping network parameters are adjusted through backpropagation. After 80 iterations, a multimodal feature representation with reduced intra-class differences is output.

[0135] Finally, step S104 of the meta-learning framework was executed to optimize the training. The 82 multimodal feature representations were hierarchically divided into 57 support sets and 25 query sets in a 7:3 ratio, constructing a hierarchical optimization architecture: the basic learning layer contains an input layer (receiving 42-dimensional features) and a ReLU hidden layer (with local parameters of a 42×28 weight matrix), while the meta-learning layer contains an adjustment module with an initial learning rate of 0.01. The basic learning layer uses forward propagation with the support sets to obtain preliminary recognition results, calculating an initial loss value of 0.21. The meta-learning layer uses the query sets to verify and obtain a meta-loss value of 0.28. Min-max scaling maps these values ​​to 0.43 and 0.57, respectively, and a normalized loss of 0.5 is obtained through moving average (smoothing coefficient 0.5). Backpropagation is used to alternately optimize local and global meta-parameters. After 90 iterations, the multimodal feature representations adapt to small sample patterns, achieving adaptation and generalization of radar signal modulation recognition rules. The model can accurately identify the signal modulation type in new small sample scenarios for spacecraft F.

[0136] Figure 4 This is a schematic diagram illustrating a specific implementation of a radar signal small-sample modulation recognition system based on meta-learning, as provided in this application. (Refer to...) Figure 4 The system may include:

[0137] The acquisition module is used to acquire the multi-polarization radar echo signal of the target object, perform signal preprocessing on the multi-polarization radar echo signal, and generate a multi-polarization feature sequence and a modulation feature set.

[0138] The first generation module is used to perform channel quality analysis on the signal channels corresponding to the multi-polarization feature sequence, perform multi-channel fusion based on the analysis results to form channel fusion features, and perform spatial interference suppression processing on the channel fusion features to generate target enhanced polarization features.

[0139] The second generation module is used to perform multimodal feature alignment processing on the target enhanced polarization features and the modulation feature set. By jointly mapping the target enhanced polarization features and the modulation feature set and measuring similarity, a multimodal feature representation is generated. The intra-class distribution of the multimodal feature representation is constrained based on the center cosine loss function to reduce intra-class feature differences.

[0140] The optimization module is used to optimize and train the multimodal feature representation based on the meta-learning framework. By constructing a hierarchical optimization architecture, the multimodal feature representation is subjected to parameter learning and meta-parameter adjustment at different levels. The normalized loss mechanism is used to standardize the training process so as to achieve the adaptation and generalization of radar signal modulation recognition rules under small sample conditions.

[0141] The radar signal small sample modulation recognition system based on meta-learning in this application is used to implement the aforementioned radar signal small sample modulation recognition method based on meta-learning. Therefore, the specific implementation of the radar signal small sample modulation recognition system based on meta-learning can be found in the embodiment section of the radar signal small sample modulation recognition method based on meta-learning mentioned above. The specific implementation can be referred to the description of the corresponding embodiment, and will not be repeated here.

[0142] This application also provides an electronic device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of any of the above-described meta-learning-based radar signal small sample modulation recognition methods.

[0143] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described methods for radar signal small-sample modulation recognition based on meta-learning.

[0144] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0145] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the embodiments of the radar signal small sample modulation recognition method based on meta-learning.

[0146] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0147] The foregoing has provided a detailed description of a radar signal small-sample modulation recognition method, system, electronic device, and storage medium based on meta-learning provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A radar signal small sample modulation recognition method based on meta learning, characterized in that, The method comprises the following steps: Collecting a multi-polarization radar echo signal of a target object, performing signal preprocessing on the multi-polarization radar echo signal, generating a multi-polarization feature sequence and a modulation feature set; Performing channel quality analysis on the signal channels corresponding to the multi-polarization feature sequence, performing multi-channel fusion according to the analysis result to form a channel fusion feature, and performing spatial domain interference suppression processing on the channel fusion feature to generate a target enhanced polarization feature; Performing multi-modal feature alignment processing on the target enhanced polarization feature and the modulation feature set, generating a multi-modal feature representation through joint mapping and similarity measurement of the target enhanced polarization feature and the modulation feature set, and constraining the intra-class distribution of the multi-modal feature representation based on a center cosine loss function to reduce the intra-class feature difference; Optimizing and training the multi-modal feature representation based on a meta-learning framework, learning parameters and adjusting meta-parameters of the multi-modal feature representation in different levels through a hierarchical optimization architecture, and normalizing the training process by using a normalized loss mechanism to realize the adaptation and generalization of radar signal modulation recognition rules under small sample conditions.

2. The method of claim 1, wherein, The optimization and training of the multi-modal feature representation based on the meta-learning framework, the construction of the hierarchical optimization architecture, the parameter learning and meta-parameter adjustment of the multi-modal feature representation in different levels, and the normalization of the training process by using the normalized loss mechanism to realize the adaptation and generalization of radar signal modulation recognition rules under small sample conditions, comprising: Based on the meta-learning framework, the multi-modal feature representation is divided into a support set and a query set, and a hierarchical optimization architecture including a basic learning layer and a meta-learning layer is constructed, the basic learning layer is used for local parameter learning of the multi-modal feature representation, and the meta-learning layer is used for adjusting global meta-parameters; In the basic learning layer, forward propagation calculation is performed using the multi-modal feature representation of the support set to obtain a preliminary recognition result, and a preliminary loss value is calculated; In the meta-learning layer, the preliminary recognition result is verified using the multi-modal feature representation of the query set, and a meta-loss value is calculated; Using a normalized loss mechanism, the preliminary loss value and the meta-loss value are scaled and smoothed respectively to obtain a normalized loss; Based on the normalized loss, the local parameters of the basic learning layer and the global meta-parameters of the meta-learning layer are alternately optimized by using a back propagation algorithm; By repeatedly executing the parameter optimization process for multiple times, the multi-modal feature representation gradually adapts to the small sample condition, so as to realize the adaptation and generalization of radar signal modulation recognition rules under small sample conditions.

3. The method of claim 2, wherein, The optimization and training of the multi-modal feature representation based on the meta-learning framework, the construction of the hierarchical optimization architecture, the parameter learning and meta-parameter adjustment of the multi-modal feature representation in different levels, and the normalization of the training process by using the normalized loss mechanism to realize the adaptation and generalization of radar signal modulation recognition rules under small sample conditions, comprising: Based on the normalized loss, the local parameters of the basic learning layer and the global meta-parameters of the meta-learning layer are alternately optimized by using a back propagation algorithm; Based on the normalized loss, the local parameters of the basic learning layer and the global meta-parameters of the meta-learning layer are alternately optimized by using a back propagation algorithm; Based on the updated local parameters of the basic learning layer, the normalized loss is recalculated; The global meta-parameter gradient of the meta-learning layer is calculated based on the re-calculated normalized loss, and the global meta-parameter of the meta-learning layer is updated by using a gradient descent algorithm; The alternating updating process of the local parameters and the global meta-parameters is repeatedly performed, so that the normalized loss is gradually reduced until a preset convergence condition or a maximum iteration number is reached.

4. The method of claim 1, wherein, The multi-modal feature alignment processing of the target enhanced polarization feature and the modulation feature set includes: The target enhanced polarization feature and the modulation feature set are input into a mapping network, and the two types of features are mapped to the same feature space through nonlinear transformation to obtain joint feature mapping; The cosine similarity between the target enhanced polarization feature and each feature vector in the modulation feature set in the joint feature mapping is calculated to establish a feature correlation matrix; Based on the feature correlation matrix, the features in the target enhanced polarization feature and the modulation feature set are classified, and the features with a cosine similarity higher than a set threshold are classified into the same class; According to the classification result, the mean value of all feature vectors in each class is calculated as the feature center point of the corresponding class; The feature vectors in the joint feature mapping, the categories to which the feature vectors belong, and the feature center points corresponding to the categories are combined to generate multi-modal feature representations.

5. The method of claim 1, wherein, The intra-class distribution of the multi-modal feature representation is constrained based on the center cosine loss function to reduce the intra-class feature difference, including: The center cosine loss function is used to calculate the cosine distance between each feature vector in the multi-modal feature representation and the feature center point of the category to which the feature vector belongs to obtain an intra-class distribution constraint loss; The intra-class distribution constraint loss is optimized by a back propagation algorithm to adjust the parameters of the mapping network, so that feature vectors of the same class gather around the corresponding feature center point in the feature space to reduce the intra-class feature difference; After iterative optimization, the multi-modal feature representation with reduced intra-class difference is output.

6. The method of claim 1, wherein, The channel quality analysis of the signal channels corresponding to the multi-polarization feature sequence is performed, and the analysis result is used to perform multi-channel fusion to form channel fusion features, and the channel fusion features are subjected to spatial domain interference suppression processing to generate target enhanced polarization features, including: The quality of the signal channels corresponding to different polarization directions in the multi-polarization feature sequence is analyzed, and the signal-to-noise ratio and signal stability index of each signal channel are calculated; Based on the signal-to-noise ratio and signal stability index, the channel quality weight of each signal channel is determined by using a weighted average method, and the higher the signal-to-noise ratio and the better the signal stability index, the higher the channel quality weight of the signal channel corresponding to them; According to the channel quality weight, a feature-level fusion method is used to weight and fuse the feature sequences of different polarization directions in the multi-polarization feature sequence, so that the features of high-quality channels occupy a larger proportion in the fusion result to form channel fusion features; The spatial distribution of the channel fusion feature is analyzed, abnormal feature components caused by external interference are identified, and a filtering processing method is used to remove the abnormal feature components; The channel fusion feature after removing the abnormal feature components is enhanced to generate a target enhanced polarization feature.

7. The method of claim 1, wherein, The method comprises the following steps: Collecting the multi-polarization radar echo signal of the target object, and performing signal preprocessing on the multi-polarization radar echo signal to generate a multi-polarization feature sequence and a modulation feature set, comprising: Collecting the multi-polarization radar echo signal of the target object in multiple polarization directions through a radar receiving device, and performing segmentation processing on the multi-polarization radar echo signal to obtain multiple signal segments; Calculating the energy value of each signal segment, selecting the corresponding signal segment whose energy value meets the condition based on a preset energy threshold, and performing frequency spectrum analysis on the selected signal segment to extract the peak frequency feature and bandwidth feature of each signal segment; Arranging the peak frequency feature and bandwidth feature of the signal segment in the same polarization direction in time sequence to form a feature sequence of the corresponding polarization direction, and integrating the feature sequences of all polarization directions to generate a complete multi-polarization feature sequence; Extracting phase change information and frequency change information from each signal segment, and calculating the phase change amount and frequency change amount; 8. A radar signal small sample modulation recognition system based on meta-learning, characterized in that, Based on the phase change amount and frequency change amount of all signal segments, a modulation feature set is constructed. Comprising: The acquisition module is configured to collect the multi-polarization radar echo signal of the target object, and perform signal preprocessing on the multi-polarization radar echo signal to generate a multi-polarization feature sequence and a modulation feature set; The first generation module is configured to perform channel quality analysis on the signal channel corresponding to the multi-polarization feature sequence, perform multi-channel fusion according to the analysis result to form a channel fusion feature, and perform spatial interference suppression processing on the channel fusion feature to generate a target enhanced polarization feature; The second generation module is configured to perform multi-modal feature alignment processing on the target enhanced polarization feature and the modulation feature set, generate a multi-modal feature representation by joint mapping and similarity measurement of the target enhanced polarization feature and the modulation feature set, and constrain the intra-class distribution of the multi-modal feature representation based on a center cosine loss function to reduce the intra-class feature difference; 9. An electronic device, comprising: The optimization module is configured to optimize and train the multi-modal feature representation based on a meta-learning framework, perform parameter learning and meta-parameter adjustment on the multi-modal feature representation in different levels by constructing a hierarchical optimization architecture, and normalize the training process by using a normalization loss mechanism to realize the adaptation and generalization of radar signal modulation recognition rules under a small sample condition. Comprising: The memory is configured to store a computer program; 10. A computer-readable storage medium, characterized in that, The processor is configured to execute the computer program to implement the steps of the meta-learning based radar signal small sample modulation recognition method according to any one of claims 1 to 7. The computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the meta-learning based radar signal small sample modulation recognition method according to any one of claims 1 to 7.

Images