Meteorological radar multi-form radio frequency interference classification method and device and computer equipment

Abstract

This application relates to a method, apparatus, and computer device for classifying multi-morphological radio frequency interference (RFI) from weather radar. The method includes: acquiring preprocessed polarimetric Doppler weather radar echo base data matrix samples; constructing a texture-guided attention network (TAN) to perform multi-level progressive low-level feature extraction on the preprocessed base data matrix samples after initial feature transformation and output a shallow enhanced feature map; performing multi-level progressive high-level semantic feature extraction on the shallow enhanced feature map and outputting a deep enhanced feature map; processing the deep enhanced feature map to output the RFI multi-morphological classification prediction result; training the TAN using the base data matrix samples and a pre-set loss function to obtain a trained TAN; and using the trained TAN to classify RFI multi-morphological interference from weather radar. This method enables high-precision automatic classification of RFI multi-morphological interference from weather radar.

Description

Technical Field

[0001] This application relates to the field of meteorological radar signal processing technology, and in particular to a method, apparatus and computer equipment for classifying multi-morphological radio frequency interference in meteorological radar. Background Technology

[0002] With the continuous development of technologies in meteorological observation and atmospheric research, polarimetric Doppler weather radar technology has been widely applied and deeply optimized. This technology possesses high-precision, fully polarized echo detection capabilities, enabling comprehensive capture and analysis of atmospheric meteorological targets through planar position display of echo data. It has become a core technological means for the accurate detection and early warning of extreme weather disasters, providing crucial technical support for the safety of people's lives and property and aviation flight safety. Simultaneously, with the rapid iteration and large-scale deployment of various electromagnetic equipment technologies such as 5G communication, wireless transmission, and radar, frequency band resources in the electromagnetic space are becoming increasingly congested. The radio frequency interference experienced by weather radar during operation exhibits diverse forms and complex effects. This interference can severely reduce the detection accuracy of weather radar, leading to false alarms or missed detections of meteorological targets. The classification, identification, and targeted processing of radio frequency interference have become core issues urgently needing to be addressed in the field of weather radar signal processing. Currently, traditional techniques for dealing with radio frequency interference from weather radar mainly employ spatial filtering, time-domain filtering, frequency-domain filtering, and polarization-Doppler joint filtering. In addition, interference classification and recognition methods based on deep learning image classification networks such as general convolutional neural networks and residual networks have been developed in order to effectively deal with multi-form radio frequency interference.

[0003] However, current radio frequency interference classification and identification methods, along with traditional interference processing methods, still have significant limitations in practical applications. Traditional filtering methods are mostly designed for specific single-form radio frequency interference, with limited applicability and a lack of fine classification mechanisms for radio frequency interference forms. When processing multi-form radio frequency interference, they not only fail to completely filter out interference signals but also easily cause the loss of real meteorological target information. General-purpose deep learning image classification networks can only extract local features from radar echo data and lack the ability to explicitly encode the overall directional texture features in the planar position display image of meteorological radar. They cannot fully utilize the prior directional features of various interference forms and are prone to confusion and misjudgment for highly similar ring-shaped interference. At the same time, model training often relies on a large amount of simulation data, which has a significant domain offset from the application scenario of real radar echoes. This causes a sharp drop in the generalization ability of conventional networks in real scenarios, making it impossible to meet the high-precision and strong generalization classification and identification requirements of meteorological radar for multi-form radio frequency interference in complex electromagnetic environments. Consequently, it restricts the improvement of meteorological radar detection performance and the optimization of extreme weather warning capabilities. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, and computer equipment for classifying multi-form radio frequency interference of weather radar to address the above-mentioned technical problems.

[0005] A method for classifying multi-morphological radio frequency interference from weather radar, the method comprising: Acquire a basic data matrix sample of preprocessed polarization Doppler weather radar echoes. The basic data matrix sample includes multi-morphological radio frequency interference signals and is labeled with radio frequency interference morphology. A texture-guided attention network is constructed, which includes a group of shallow feature extraction modules, a group of deep feature extraction modules, and an output module connected in sequence. The shallow feature extraction module group performs multi-level progressive low-level feature extraction on the preprocessed base data matrix samples after initial feature transformation and outputs a shallow enhanced feature map. Each shallow feature extraction module extracts the global directional texture features of radio frequency interference in the input feature map, generates a spatial saliency weight map based on the texture features, and uses the spatial saliency weight map to perform residual adaptive enhancement on the low-level feature map. The deep feature extraction module group performs multi-level progressive high-level semantic feature extraction on the shallow enhanced feature map and outputs a deep enhanced feature map. Each deep feature extraction module extracts the distance dimension structural features and orientation dimension continuity features of the circular radio frequency interference in the input feature map, generates multi-dimensional axial attention weights based on the structural features, and uses the multi-dimensional axial attention weights to perform residual adaptive enhancement on the high-level feature map. The output module processes the deep enhanced feature map and outputs the radio frequency interference multimorphic classification prediction result. The texture-guided attention network is trained using the base data matrix samples and a pre-set loss function to obtain the trained texture-guided attention network. The trained texture-guided attention network is used for the classification of multi-morphological radio frequency interference in weather radar.

[0006] A weather radar multi-morphological radio frequency interference classification device, the device comprising: The sample acquisition module is used to acquire the basic data matrix sample of the preprocessed polarization Doppler weather radar echo. The basic data matrix sample includes multi-mode radio frequency interference signals and is correspondingly labeled with radio frequency interference modes. A network construction module is used to construct a texture-guided attention network, which includes a group of shallow feature extraction modules, a group of deep feature extraction modules, and an output module connected in sequence. The shallow feature enhancement module is used to perform multi-level progressive low-level feature extraction on the preprocessed base data matrix samples after initial feature transformation and output a shallow enhanced feature map through the shallow feature extraction module group. Each shallow feature extraction module extracts the global directional texture features of radio frequency interference in the input feature map, generates a spatial saliency weight map based on the texture features, and uses the spatial saliency weight map to perform residual adaptive enhancement on the low-level feature map. The deep feature enhancement module is used to perform multi-level progressive high-level semantic feature extraction on the shallow enhanced feature map through the deep feature extraction module group and output the deep enhanced feature map. Among them, each deep feature extraction module extracts the distance dimension structural features and orientation dimension continuity features of the circular radio frequency interference in the input feature map, generates multi-dimensional axial attention weights based on the structural features, and uses the multi-dimensional axial attention weights to perform residual adaptive enhancement on the high-level feature map. The classification and prediction module is used to process the deep enhanced feature map through the output module and output the radio frequency interference multimorphic classification and prediction results. The network training module is used to train the texture-guided attention network using the base data matrix samples and a pre-set loss function to obtain the trained texture-guided attention network. The result output module is used to perform multi-morphological radio frequency interference classification for weather radar using the trained texture-guided attention network.

[0007] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps: Acquire a basic data matrix sample of preprocessed polarization Doppler weather radar echoes. The basic data matrix sample includes multi-morphological radio frequency interference signals and is labeled with radio frequency interference morphology. A texture-guided attention network is constructed, which includes a group of shallow feature extraction modules, a group of deep feature extraction modules, and an output module connected in sequence. The shallow feature extraction module group performs multi-level progressive low-level feature extraction on the preprocessed base data matrix samples after initial feature transformation and outputs a shallow enhanced feature map. Each shallow feature extraction module extracts the global directional texture features of radio frequency interference in the input feature map, generates a spatial saliency weight map based on the texture features, and uses the spatial saliency weight map to perform residual adaptive enhancement on the low-level feature map. The deep feature extraction module group performs multi-level progressive high-level semantic feature extraction on the shallow enhanced feature map and outputs a deep enhanced feature map. Each deep feature extraction module extracts the distance dimension structural features and orientation dimension continuity features of the circular radio frequency interference in the input feature map, generates multi-dimensional axial attention weights based on the structural features, and uses the multi-dimensional axial attention weights to perform residual adaptive enhancement on the high-level feature map. The output module processes the deep enhanced feature map and outputs the radio frequency interference multimorphic classification prediction result. The texture-guided attention network is trained using the base data matrix samples and a pre-set loss function to obtain the trained texture-guided attention network. The trained texture-guided attention network is used for the classification of multi-morphological radio frequency interference in weather radar.

[0008] The aforementioned meteorological radar multi-morphological radio frequency interference classification method, device, and computer equipment extract global directional texture features of radio frequency interference through a shallow feature extraction module group. Based on the texture features, a spatial saliency weight map is generated to perform residual adaptive enhancement on the low-level feature map. This explicitly encodes the global directional texture information of the interference, fully utilizes the prior directional features of different interference morphologies, and improves the discrimination accuracy of major interference categories. The deep feature extraction module group extracts the range-dimensional structural features and azimuth-dimensional continuity features of ring-type radio frequency interference. Based on the structural features, multi-dimensional axial attention weights are generated to perform residual adaptive enhancement on the high-level feature map. This accurately captures the subtle topological differences of ring-type interference, reducing the probability of confusion and misjudgment of highly similar interference morphologies. Through a hierarchical and progressive network architecture and end-to-end supervised training, the network can learn the core distinguishing features of the interference, improving the model's classification robustness. This invention enables high-precision automatic classification of multi-morphological radio frequency interference from meteorological radar, effectively adapting to the operational needs of radar interference identification in complex electromagnetic environments. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating a multi-morphological radio frequency interference classification method for weather radar in one embodiment. Figure 2 This is a schematic diagram illustrating the classification of multi-form radio frequency interference in one embodiment; Figure 3 This is a schematic diagram of multi-form radio frequency interference in one embodiment, wherein, Figure 3 (a) is a schematic diagram of the original form of the spiral radio frequency interference PPI. Figure 3 (b) is a schematic diagram of the original form of radial radio frequency interference PPI. Figure 3 (c) is a schematic diagram of the original form of point-type radio frequency interference PPI. Figure 3(d) is a schematic diagram of the original form of a real circular radio frequency interference PPI. Figure 3 (e) is a schematic diagram of the original form of the multi-ring radio frequency interference PPI. Figure 3 (f) is a schematic diagram of the original form of the virtual circular radio frequency interference PPI; Figure 4 This is a schematic diagram of the TGANet network architecture and data flow in one embodiment; Figure 5 This is a schematic diagram illustrating the texture features of various RFI (Radio Frequency Interference) types in the input matrix in one embodiment, where, Figure 5 (a) is a schematic diagram of the texture features of the spiral radio frequency interference basis data matrix. Figure 5 (b) is a schematic diagram of the texture features of the radial radio frequency interference basis data matrix. Figure 5 (c) is a schematic diagram of the texture features of the dot-type radio frequency interference basis data matrix. Figure 5 (d) is a schematic diagram of the texture features of a real circular radio frequency interference basis data matrix. Figure 5 (e) is a schematic diagram of the texture features of the multi-ring radio frequency interference basis data matrix. Figure 5 (f) is a schematic diagram of the texture features of the virtual circular ring-shaped radio frequency interference base data matrix; Figure 6 This is a schematic diagram of a Gabor filter at different angles in one embodiment, wherein, Figure 6 (a) is a schematic diagram of the Gabor filter kernel in the 0° direction. Figure 6 (b) is a schematic diagram of a Gabor filter core in the 45° direction. Figure 6 (c) is a schematic diagram of a Gabor filter kernel in the 90° direction. Figure 6 (d) is a schematic diagram of a Gabor filter core in the 135° direction; Figure 7 This is a schematic diagram illustrating the specific implementation steps of the OTE module in one embodiment; Figure 8 This is a schematic diagram of the PPI interference patterns corresponding to the simulation dataset of multi-morphological RFI in one embodiment, wherein, Figure 8 (a) is a schematic diagram of the PPI shape of the spiral radio frequency interference simulation dataset. Figure 8 (b) is a schematic diagram of the PPI shape of the radial radio frequency interference simulation dataset. Figure 8 (c) is a schematic diagram of the PPI (point-type radio frequency interference) simulation dataset. Figure 8 (d) is a schematic diagram of the PPI shape of the real circular radio frequency interference simulation dataset. Figure 8 (e) is a schematic diagram of the PPI shape of the multi-ring radio frequency interference simulation dataset. Figure 8 (f) is a schematic diagram of the PPI shape of the virtual circular radio frequency interference simulation dataset; Figure 9 This is a schematic diagram of a confusion matrix in one embodiment; Figure 10 This is a schematic diagram of the accuracy versus loss function curves in one embodiment, where, Figure 10 (a) is a schematic diagram of the total loss changes between the training set and the validation set. Figure 10 (b) is a schematic diagram of the changes in classification accuracy on the training set and validation set. Figure 10 (c) is a schematic diagram of the variation curves of each component of the combined loss; Figure 11 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0011] In one embodiment, such as Figure 1 As shown, a multi-morphological radio frequency interference classification method for weather radar is provided, including the following steps: Step 102: Obtain the basic data matrix sample of the preprocessed polarization Doppler weather radar echo. The basic data matrix sample includes multi-mode radio frequency interference signals and is labeled with radio frequency interference mode.

[0012] Polarimetric Doppler (PDO) weather radar is a core detection device for current atmospheric research and weather forecasting. It can simultaneously acquire multi-dimensional information such as the radial velocity of precipitation particles, echo intensity, and polarization scattering characteristics by emitting electromagnetic waves with different polarization directions. It is the core hardware carrier for generating meteorological echo data. The base data matrix is ​​the native two-dimensional digital matrix output by the PDO weather radar after completing low-level signal processing. The matrix dimension is the number of range gates × the number of azimuth rays. The matrix rows correspond to the radar's radial range cells, and the columns correspond to the azimuth angle of the antenna scan. The matrix element values ​​are the radar echo physical quantities corresponding to the spatial location. It is the original data source for generating radar planar position display images and the core processing object for interference classification. For example... Figure 2 The diagram illustrates the classification of multi-morphological radio frequency interference (RF interference). Multi-morphological RF interference signals are non-meteorological target echo signals emitted by external electromagnetic devices and entering the radar receiver. In the base data matrix, they exhibit six typical forms: spiral, radial, dotted, solid circular, multi-circular, and virtual circular. The RF interference morphology label is a category annotation information that corresponds one-to-one with the samples in the base data matrix, used to mark the specific morphological category of the RF interference contained in the sample.

[0013] Step 104: Construct a texture-guided attention network, which includes a group of shallow feature extraction modules, a group of deep feature extraction modules, and an output module connected in sequence.

[0014] The Texture-Guided Attention Network (TGA) is a deep learning classification network with a pre-trained residual network as its backbone, custom-designed for the texture and structural features of radio frequency interference (RFI) from weather radar. The shallow feature extraction module group, located at the front end of the classification network, is responsible for extracting low-level visual features such as edges, lines, and textures from the input data, adapting to the recognition requirements of globally directional textures in RFI. The deep feature extraction module group, located at the back end of the classification network, is responsible for extracting high-level features such as topological structure and global semantics from the input data, adapting to the detailed recognition requirements of highly similar circular RFI. The output module is the final processing unit of the classification network, responsible for mapping high-level features to specific category probability results and outputting the final interference classification judgment.

[0015] It is understandable that by building a hierarchical and progressive network architecture, it is possible to achieve progressive extraction from low-level texture features to high-level semantic features, which is conducive to capturing the differentiated features of different forms of radio frequency interference and provides a complete algorithm architecture support for the accurate classification of multi-form radio frequency interference.

[0016] Step 106: Through the shallow feature extraction module group, multi-level progressive low-level feature extraction is performed on the preprocessed base data matrix samples after initial feature transformation, and shallow enhanced feature maps are output. Among them, each shallow feature extraction module extracts the global directional texture features of radio frequency interference in the input feature map, generates a spatial saliency weight map based on the texture features, and uses the spatial saliency weight map to perform residual adaptive enhancement on the low-level feature map.

[0017] Multi-level progressive low-level feature extraction involves multiple consecutive feature extraction modules to perform multi-stage, coarse-to-fine low-level visual feature mining on the input data, gradually aggregating the texture details of radio frequency interference. The feature map is a multi-dimensional feature matrix generated after the input data undergoes network feature extraction; it is the core carrier for transmitting feature information between different layers in the network. Global directional texture features are the fixed-direction line features of different forms of radio frequency interference presented in the base data matrix, including the diagonal features of spiral interference, the vertical features of radial interference, and the horizontal features of ring-like interference; these are the core basis for distinguishing different major categories of radio frequency interference. The spatial saliency weight map is a two-dimensional weight matrix with the same size as the input feature map, generated based on the global directional texture features. The matrix element values ​​correspond to the saliency of the interference texture at that spatial location; a higher weight value indicates stronger discrimination of the interference feature at that location. Residual adaptive enhancement applies the spatial saliency weight map to the original feature map through residual connections, adaptively amplifying high-discrimination interference texture features while preserving the complete transmission of the original meteorological echo information.

[0018] It is understandable that this step can explicitly encode the global directional texture features of different types of radio frequency interference, effectively amplify the feature differences of different major types of interference, improve the differentiation accuracy of major types of interference such as spiral, radial, and point types, and at the same time, the residual enhancement method can completely preserve the original meteorological echo information and avoid the loss of effective meteorological data.

[0019] Step 108: The deep feature extraction module group performs multi-level progressive high-level semantic feature extraction on the shallow enhanced feature map and outputs the deep enhanced feature map. Among them, each deep feature extraction module extracts the distance dimension structural features and orientation dimension continuity features of the circular radio frequency interference in the input feature map, generates multi-dimensional axial attention weights based on the structural features, and uses the multi-dimensional axial attention weights to perform residual adaptive enhancement on the high-level feature map.

[0020] High-level semantic feature extraction further aggregates and abstracts the basic features extracted at the shallow level, uncovering hidden topological structures, global distribution patterns, and other high-level features in the data. This allows for the differentiation of highly similar, subdivided interference categories. For example... Figure 3 The diagram shown illustrates the various forms of radio frequency interference. Figure 3 (a) is a schematic diagram of the original form of the spiral radio frequency interference PPI. Figure 3 (b) is a schematic diagram of the original form of radial radio frequency interference PPI. Figure 3 (c) is a schematic diagram of the original form of point-type radio frequency interference PPI. Figure 3 (d) is a schematic diagram of the original form of a real circular radio frequency interference PPI. Figure 3 (e) is a schematic diagram of the original form of the multi-ring radio frequency interference PPI. Figure 3(f) is a schematic diagram of the original form of the virtual circular ring-type radio frequency interference (RFI). Circular ring-type RF interference presents a concentric ring structure in the radar echo, including three types: real circular ring, multi-circular ring, and virtual circular ring. These three types have highly similar overall shapes and are easily confused and misjudged in existing technologies. The range-dimensional structural feature refers to the peak distribution characteristics of circular ring-type interference in the radar's radial range dimension. Real circular ring interference presents a single peak in the range dimension, while multi-circular ring interference presents multiple peaks, which is the core basis for distinguishing between single-ring and multi-ring interference. The azimuth-dimensional continuity feature refers to the continuous characteristics of circular ring-type interference in the azimuth angle dimension of the radar antenna scan. Real circular ring and multi-circular ring interference are completely continuous in the azimuth dimension, while virtual circular ring interference shows discontinuous characteristics, which is the core basis for distinguishing between continuous and discontinuous rings. Multi-dimensional axial attention weights are attention weight matrices generated based on the range and azimuth-dimensional features, corresponding to the two dimensions respectively, used to specifically amplify the core distinguishing features of circular ring-type interference.

[0021] It is understandable that this step can accurately deconstruct the range-dimensional topology and orientation-dimensional continuity of circular radio frequency interference, effectively amplify the subtle differences in the three types of circular interference, significantly reduce the probability of confusion and misjudgment of circular interference with highly similar shapes, and improve the subdivision and classification accuracy of virtual circular and multi-circular interference.

[0022] Step 110: Process the deep enhanced feature map through the output module and output the radio frequency interference multimorphic classification prediction result.

[0023] The multimorphic classification prediction result of radio frequency interference is the final output of the network, which is the interference morphology classification result corresponding to the input base data matrix sample, and clearly marks the specific morphology category to which the radio frequency interference contained in the input data belongs.

[0024] Step 112: Train the texture-guided attention network using the base data matrix samples and a pre-set loss function to obtain the trained texture-guided attention network.

[0025] The loss function is used to measure the degree of deviation between the network classification prediction result and the true label. The method used here is a combined loss function composed of cross-entropy loss, supervised contrast loss and circular class focus loss, which is the core basis for guiding the optimization of network parameters.

[0026] It is understandable that this step can guide the network to autonomously learn the core distinguishing features of different forms of radio frequency interference, enabling the network to adapt to the multi-form interference identification needs in the complex electromagnetic environment of weather radar, and helping to improve the network's classification accuracy and generalization ability in real radar echo scenarios.

[0027] Step 114: Use the trained texture-guided attention network to classify multi-morphological radio frequency interference from weather radar.

[0028] The inference phase involves deploying the trained network to automatically classify unlabeled radar echo data to be identified. This is the practical application of the method in meteorological operations. The base data matrix to be identified is the raw echo base data matrix generated in real-time by meteorological radar and has not been manually labeled; it is the object processed by the network in actual operations.

[0029] Understandably, this step enables automated and high-precision classification and identification of multi-form radio frequency interference for weather radar, providing accurate category basis for the formulation of subsequent targeted interference filtering strategies. This is beneficial for improving the detection accuracy of weather radar in complex electromagnetic environments and ensuring the reliability of weather forecasts and disaster warnings.

[0030] In the aforementioned method for classifying multi-morphological radio frequency interference (RF interference) from meteorological radar, a shallow feature extraction module extracts global directional texture features of RF interference. Based on these texture features, a spatial saliency weight map is generated to perform residual adaptive enhancement on the lower-level feature map. This explicitly encodes the global directional texture information of the interference, fully utilizes the prior directional features of different morphological interferences, and improves the discrimination accuracy of major interference classes. A deep feature extraction module extracts the range-dimensional structural features and azimuth-dimensional continuity features of ring-shaped RF interference. Based on these structural features, multi-dimensional axial attention weights are generated to perform residual adaptive enhancement on the higher-level feature map. This accurately captures subtle topological differences in ring-shaped interferences, reducing the probability of confusion and misjudgment of highly similar interferences. Through a hierarchical and progressive network architecture and end-to-end supervised training, the network learns the core distinguishing features of the interference, improving the model's classification robustness. This invention enables high-precision automatic classification of multi-morphological RF interference from meteorological radar, effectively adapting to the operational needs of radar interference identification in complex electromagnetic environments.

[0031] In one embodiment, the shallow feature extraction module group includes a first shallow feature extraction module and a second shallow feature extraction module; performing multi-level progressive low-level feature extraction on the preprocessed base data matrix samples after initial feature transformation and outputting a shallow enhanced feature map includes: performing low-level feature extraction on the preprocessed base data matrix samples after initial feature transformation through the first shallow feature extraction module and outputting a first low-level feature map; and performing progressive low-level feature extraction on the first low-level feature map through the second shallow feature extraction module and outputting a shallow enhanced feature map.

[0032] In one embodiment, both the first shallow feature extraction module and the second shallow feature extraction module include a directional texture encoding submodule. The directional texture encoding submodule is used to perform channel dimensionality reduction mapping on the input feature map, perform depthwise separable convolution on the dimensionality-reduced feature map using a learnable Gabor filter bank to generate multiple directional texture response maps, perform a directional channel attention mechanism on the multiple directional texture response maps to obtain multiple attention weights, use the multiple attention weights to perform weighted fusion on the multiple directional texture response maps to generate a spatial saliency weight map, and use the spatial saliency weight map to perform residual adaptive enhancement on the input feature map.

[0033] In one embodiment, the deep feature extraction module group includes a first deep feature extraction module and a second deep feature extraction module; the deep feature extraction module group performs multi-level progressive high-level semantic feature extraction on the shallow enhanced feature map and outputs a deep enhanced feature map, which includes: performing high-level semantic feature extraction on the shallow enhanced feature map through the first deep feature extraction module and outputting a first high-level feature map; and performing progressive high-level semantic feature extraction on the first high-level feature map through the second deep feature extraction module and outputting a deep enhanced feature map.

[0034] In one embodiment, both the first deep feature extraction module and the second deep feature extraction module include an axial continuity attention submodule; the axial continuity attention submodule is used to extract the distance dimension structural features and the orientation dimension continuity features of the input feature map, generate distance dimension attention weights, orientation dimension attention weights and channel attention weights, and use the distance dimension attention weights, orientation dimension attention weights and channel attention weights to perform residual adaptive enhancement on the input feature map.

[0035] In one embodiment, the loss function is a combined loss function, which includes labeled smoothed cross-entropy loss, supervised contrastive loss, and circular focal loss.

[0036] In one embodiment, the batch normalization layers in the texture-guided attention network are all adaptive instance-batch normalization submodules; the adaptive instance-batch normalization submodules are used to adaptively mix instance normalization results and batch normalization results on the input feature map according to learnable mixing coefficients before outputting.

[0037] In this embodiment, by replacing all batch normalization layers in the texture-guided attention network with adaptive instance-batch normalization submodules, the fusion ratio of instance normalization and batch normalization results can be adaptively adjusted through learnable mixing coefficients according to the data distribution characteristics of the input feature map. This improves the network's domain invariance while preserving feature discrimination capabilities, effectively alleviating the domain offset problem between simulated training data and real radar echo data. It significantly enhances the network's generalization ability and robustness in real weather radar application scenarios, and can automatically adapt to the optimal normalization mode according to the data distribution without introducing additional performance loss.

[0038] In one embodiment, processing the deep enhanced feature map to output the radio frequency interference multimorphic classification prediction result includes: performing global average pooling and random deactivation processing on the deep enhanced feature map in sequence, inputting the pooled and deactivated features into a fully connected layer for feature mapping; and performing classification activation on the mapped features through the softmax function to output the radio frequency interference multimorphic classification prediction result.

[0039] In one specific embodiment, to address the technical problems of insufficient utilization of directional texture features, easy confusion with circular interference, and weak generalization ability in the classification of multi-morphological radio frequency interference (RFI) by existing technologies, this invention proposes a Texture-Guided Attention Network (TGANet). The input is the base data matrix used by polarimetric Doppler weather radar to generate PPI (Plane Position Indication) maps, and the output is the classification result. This method uses a pre-trained ResNet18 as the backbone network and mainly utilizes the different texture orientations of multi-morphological RFIs in the matrix. Four improved modules are added to its feature extraction process: Oriented Texture Encoding (OTE), Axis-wise Continuity Attention (ACA), Adaptive Instance-Batch Normalization (AIBN), and a combined loss function. Simultaneously, this method employs a three-stage progressive training strategy, enabling fine-tuning of network parameters, thereby achieving accurate classification of multi-morphological RFIs.

[0040] A schematic diagram of the model structure of the texture-guided attention network is shown below. Figure 4As shown, this network uses a pre-trained ResNet18 as its backbone. The input is a single-channel feature map of size 1×224×224. The input data first enters the initial feature layer of the Stem, which consists of a 1-channel input to 64-channel output, a 7×7 convolutional kernel, a convolutional layer with a stride of 2, an adaptive instance-batch normalization (AIBN) layer, a ReLU activation function, and a MaxPool max pooling layer. After the initial feature extraction and size downsampling of the input data, the feature data enters four sequentially connected feature extraction stages. The core of Stage 1 (layer 1) consists of two consecutive BasicBlock residual blocks, which output a 64-channel, 56×56 feature map. The output of the residual block is connected in series with an OTE-1 directional texture encoding module (64 input channels, Gabor filter number K=8) and a MixStyle feature-level style fusion module (activation probability p=0).5) The first explicit encoding, enhancement, and feature-level domain enhancement of the global directional texture features of radio frequency interference are completed; Stage 2 (layer 2) consists of two consecutive BasicBlock residual blocks, outputting a 128-channel, 28×28 feature map. The output of the residual block is connected in series with an OTE-2 directional texture encoding module (128 input channels, K=8 Gabor filters) to perform secondary aggregation and progressive enhancement of the underlying texture features, completing the feature decoupling of different major categories of radio frequency interference; Stage 3 (layer 3) consists of two consecutive BasicBlocks. The residual block outputs a 256-channel, 14×14 feature map. The output of the residual block is connected in series with an ACA-3 axial continuity attention module (256 input channels) to extract the range and azimuth continuity features of ring-shaped RF interference, completing the initial differentiation of ring-shaped interference. Stage 4 (layer 4) consists of two consecutive BasicBlock residual blocks, outputting a 512-channel, 7×7 feature map. The output of the residual block is connected in series with an ACA-4 axial continuity attention module (512 input channels) to further refine the axial features of ring-shaped interference. Enhancement is achieved to address the issue of misclassification caused by interference from highly similar circular ring-like structures. All native batch normalization layers in the network are globally replaced with an Adaptive Instance-Batch Normalization (AIBN) module. This module is embedded in the Stem layer and all BasicBlock residual blocks from Stage 1 to Stage 4, adaptively learning the optimal batch normalization to instance normalization mixing ratio for each feature layer to alleviate the domain offset problem between simulated training data and real radar echo data. The feature data, after full-process feature extraction and enhancement, finally enters the classification head at the end of the network. This classification head is then processed by Global Average Pooling (GA). The network consists of a P), a Dropout random deactivation layer, a fully connected layer that maps 512-dimensional features to 128-dimensional features, a ReLU activation function, and a fully connected layer that maps 128-dimensional features to a 6-dimensional output, corresponding to 6 typical types of radio frequency interference. During network training, a combined loss function, weighted by 1x labeled smooth cross-entropy loss (CE), 0.1x supervised contrastive loss (SupCon), and 0.05x ring-specific focus loss (RingFocal), is used as the optimization objective. This multi-dimensional collaborative approach guides network parameter optimization, ensuring the model's convergence stability and classification robustness. It can be understood that, besides ResNet18, networks with hierarchical progressive feature extraction capabilities, such as ResNet series variants and VGGt, can also serve as the backbone network of this network.

[0041] It is worth noting that different forms of RFI exhibit natural, directly classifiable differences in fixed-directional texture within the distance-orientation two-dimensional matrix. For example... Figure 5The diagram shows the texture features of various RFIs in the input matrix, where, Figure 5 (a) is a schematic diagram of the texture features of the spiral radio frequency interference basis data matrix. Figure 5 (b) is a schematic diagram of the texture features of the radial radio frequency interference basis data matrix. Figure 5 (c) is a schematic diagram of the texture features of the dot-type radio frequency interference basis data matrix. Figure 5 (d) is a schematic diagram of the texture features of a real circular radio frequency interference basis data matrix. Figure 5 (e) is a schematic diagram of the texture features of the multi-ring radio frequency interference basis data matrix. Figure 5 (f) is a schematic diagram of the texture features of the base data matrix for virtual circular ring-type radio frequency interference. The original dimension of the input matrix in the method of this invention is the number of range gates × the number of azimuth rays, where the range dimension corresponds to the radial distance of the radar, and the azimuth dimension corresponds to the angle direction of the antenna rotation scan. Therefore, for spiral-type RFI, the distance at which the interference occurs increases or decreases linearly with the change of azimuth angle, and is represented by a diagonal line on the matrix; for radial-type RFI, the interference occurs at all range cells at a fixed azimuth angle, and is represented by a vertical line on the matrix; point-type RFI has no specific directional features on the matrix due to its spatiotemporal regularity and low duty cycle; for solid circular ring RFI, the interference shape is a single concentric solid ring with a fixed radius and width, and is represented by a horizontal solid line on the matrix; similarly, multi-ring RFI consists of multiple concentric rings, and is represented by multiple horizontal solid lines on the matrix; virtual circular ring-type RFI is a broken concentric ring, and is represented by one or more horizontal dashed lines on the matrix. The texture features of various types of RFI on the matrix are summarized in Table 1: Table 1. Summary of texture features of various RFI types on the input matrix

[0042] The Oriented Texture Encoding (OTE) module primarily extracts the overall texture orientation features of multi-morphological RFIs by introducing a learnable Gabor filter bank and an directional attention mechanism, distinguishing spiral, radial, dotted, and various circular textures. This embodiment introduces eight Gabor filter banks with different angles into the shallow layers of ResNet18, expressed as follows: (1) in , It is the rotation angle The coordinates after Controlling the effective region size of the Gabor filter Control the thickness and spacing of the stripes. Angle Initialize to [0°, 22.5°, 45°, 67.5°, 90°, 112.5°, 135°, 157.5°]. , and As learnable parameters, they participate in backpropagation optimization, enabling the network to adaptively adjust the optimal orientation and frequency characteristics of the filters according to the texture distribution of the actual data, finding the most suitable parameter combination for resolving six types of RFI. According to the expression, a Gabor filter bank with eight different angles is equivalent to a convolution kernel rotated at different angles, such as... Figure 6 The diagrams shown depict Gabor filters at different angles. Figure 6 (a) is a schematic diagram of the Gabor filter kernel in the 0° direction. Figure 6 (b) is a schematic diagram of a Gabor filter core in the 45° direction. Figure 6 (c) is a schematic diagram of a Gabor filter kernel in the 90° direction. Figure 6 (d) is a schematic diagram of a Gabor filter kernel in the 135° direction. When the texture direction of the RFI on the matrix is ​​consistent with the angle of the Gabor filter, the convolution response value is high, thus detecting the corresponding morphological interference. For example, spiral interference has a high response value to Gabor filters with certain tilt angles such as 45° and 135°; radial interference has the highest response value to the 0° filter; point interference has little difference in the filter response at all angles; while the three types of circular interference all have a high response value to the 90° horizontal Gabor filter. Therefore, spiral, point, and radial interference can be distinguished by the difference in the results after convolution. After deep convolution of the Gabor filter bank, the texture intensity response map of the input matrix in 8 directions is obtained, representing the texture intensity of each position on the image in that direction. At this time, by introducing a directional attention mechanism, global average pooling is performed on each directional texture response map, and the average texture intensity of the 8 directions is input into a small fully connected network to calculate the weights of different directions, which are used to generate saliency maps and enhance the original features, facilitating subsequent classification learning.

[0043] The specific implementation steps of the OTE module are as follows: Figure 7 As shown: First, eight learnable Gabor filters are defined and initialized in the network. Each filter is dynamically controlled by three learnable parameters: orientation angle, Gaussian envelope scale, and cosine factor wavelength. The eight initial orientation angles are evenly distributed in the interval [0, π) to ensure balanced perception of texture in all directions during the initial stage. During model training, these parameters participate in backpropagation as part of the network weights, adaptively adjusting to the optimal filtering state that best fits the multimorphic RFI texture distribution.

[0044] During feature forward inference, the original input feature map is first reduced to 8 channels by 1×1 point convolution, realizing the projection of features onto the directional subspace. Then, depth-separable group convolution is used to independently apply the generated 8 Gabor filters to the corresponding 8 channels, resulting in 8 directional texture response maps. The absolute value of the response map is taken to eliminate the interference of positive and negative phases and obtain the pure texture energy amplitude. Then, a directional channel attention mechanism is performed on the 8 response maps: each map is compressed into a scalar by global average pooling to form an 8-dimensional directional description vector. This vector passes through two fully connected layers and activation functions in sequence, outputting 8 attention weights to evaluate the relative importance of each texture direction in the current sample. Finally, the attention weights are multiplied by the corresponding response map channel by channel and aggregated into a single two-dimensional spatial saliency map by 1×1 convolution, which is then mapped to the [0, 1] interval by the Sigmoid function.

[0045] Finally, the saliency map is applied to the original input features using residual modulation: (2) in The enhanced output feature map of the OTE module. This is the original input feature map for the module. This is the extracted spatial saliency weighted map.

[0046] This residual enhancement structure ensures the stable transmission of the underlying meteorological background and the original information flow of radar echoes, while adaptively amplifying the RFI texture energy response in specific directions such as spiral and radial based on saliency weights.

[0047] After processing by the OTE module, spiral, radial, and dot-shaped RFIs can be accurately classified. However, the three types of circular RFIs all appear as concentric ring structures centered on the radar in the PPI display and belong to the horizontal texture in the input matrix, with highly similar overall shapes, making them difficult to distinguish using OTE alone. Analysis revealed key differences between the three types of circular RFIs in the peak distribution of the range dimension and the continuity of the azimuth dimension: solid rings have only a single energy peak in the range dimension and are completely continuous in the azimuth dimension; multi-rings exhibit multiple energy peaks in the range dimension and are continuous in the azimuth dimension; while virtual rings have one or more peaks in the range dimension but exhibit discontinuous features in the azimuth dimension. These characteristics are summarized in Table 2. Table 2 Summary of Axial Characteristics of Circular RFI

[0048] Therefore, an Axial Continuity Attention (ACA) module is designed to extract global information and features of three types of circular RFIs by analyzing the number of range axis spectral peaks and azimuth axis continuity, thereby enabling accurate classification. The ACA module comprises three branches: range axis structure analysis, azimuth axis structure analysis, and inter-segment consistency analysis, which are described below: The first step is the distance axis structure analysis. This part performs global average pooling along the azimuth dimension on the input feature map, compressing the four-dimensional tensor (B,C,H,W) into a three-dimensional tensor (B,C,H), obtaining a one-dimensional structural profile of each channel in the distance dimension. Here, B is the number of samples in the batch, C is the number of channels, H is the number of distance elements, and W is the number of rays in the azimuth dimension. This profile is then processed by a two-layer one-dimensional convolutional network to generate distance-dimensional attention weights of dimension (B,C,H,1). (0~1). This branch aims to extract the peak quantity feature of the distance dimension: for a single-ring real ring or a single-ring virtual ring RFI, a single weight peak will be presented on the distance profile; for a multi-ring or multi-ring virtual ring RFI, there are multiple weight peaks. This branch can effectively decouple single-ring and multi-ring topologies.

[0049] Next is the azimuth axis structure analysis. This part performs global average pooling along the distance dimension on the input feature map, compressing it into a tensor of dimension (B, C, W), obtaining a one-dimensional structural profile of each channel in the azimuth dimension. This profile is then passed through two layers of one-dimensional convolutional networks symmetrical to the first branch structure to generate azimuth dimension attention weights of dimension (B, C, 1, W). For real circular ring RFI and multi-circular ring RFI, the weight curve is flat, while for imaginary circular ring RFI, the weight curve is undulating.

[0050] Next is the inter-segment consistency analysis. This part divides the input matrix into several sub-segments along the azimuth dimension. For each sub-segment, the average distance profile along the azimuth dimension is calculated independently, and the variance between the segments is calculated as a consistency metric to quantify the discontinuity in the azimuth dimension: for continuous real circular rings and multi-circular ring RFIs, the distance profiles of each segment are highly consistent, and the variance is minimal; for discontinuous virtual circular ring RFIs, the distance profiles of each segment differ significantly, and the variance is large. After global average pooling of the variance along the distance dimension, a channel-wise consistency descriptor (B, C) is obtained. This descriptor is then passed through two fully connected layers to generate channel attention weights of dimension (B, C, 1, 1). .

[0051] Finally, the outputs of the three branches are fused together by element-wise multiplication and applied to the original input features as residuals: (3) in This is the output feature map after modulation by the ACA module. This is the original input feature map for the module. , and These are the distance dimension attention weights, the orientation dimension attention weights, and the channel attention weights, respectively.

[0052] In practical applications of radar RFI classification, training data is typically generated from signal simulation, while deployment involves real radar echo data. These two datasets may exhibit systematic differences in signal statistical characteristics, and this domain shift can lead to a decline in model generalization performance. Conventional batch normalization-based model training methods limit the model's generalization ability. Therefore, this paper proposes adaptive instance-batch normalization to address the domain shift problem and improve the model's generalization capability.

[0053] Batch Normalization (BN) normalizes features using batch-level mean and variance statistics, preserving the relative relationships between samples and possessing strong discriminative power, but it is sensitive to domain shift. Instance Normalization (IN), on the other hand, normalizes using statistics within a single sample, eliminating style differences between individual samples and possessing strong domain invariance, but it loses some discriminative information. AIBN adaptively blends these two normalization methods to achieve an optimal balance between discriminative power and domain invariance. The specific implementation is as follows: For each channel, a learnable mixing coefficient is defined. The Sigmoid function is used to transform unconstrained parameters. Map to the [0,1] interval, and then calculate the normalized output using the following formula: (4) in =sigmoid( ), and These are the instance normalization and batch normalization results for the c-th channel, respectively. The normalized output is then subjected to a shared learnable affine transformation to restore the expressive power of the features. The initial value setting makes the initial mixing coefficient close to 0, that is, the model starts in pure BN mode and then automatically adjusts it according to the data characteristics during training.

[0054] In TGANet, all BatchNorm2d layers in the network are replaced with AIBN. This global replacement strategy allows each layer to independently learn the normalized mixing ratio best suited to its feature distribution. Experiments show that in a training scenario with a single radar data source, the mixing coefficients of all layers converge to near 0 (pure BN mode). This indicates that AIBN can automatically degenerate into standard BN based on the actual data distribution without introducing additional performance loss, while retaining its adaptive potential for handling future cross-domain scenarios.

[0055] Furthermore, to further enhance domain generalization ability, the MixStyle module is introduced as an auxiliary domain augmentation method. During the training phase, MixStyle randomly mixes the feature statistics between samples within a batch with a 50% probability, generating mixing coefficients through a Beta distribution. This forces the network to learn feature representations that are invariant to style changes, further improving the model's generalization ability.

[0056] This invention designs a combined loss function, which is composed of a weighted combination of three complementary loss components. The weight fusion method is as follows: (5) The components of the combined loss function are explained in detail below: The first component is the labeled, smoothed cross-entropy loss. We use cross-entropy loss with a label smoothing coefficient of 0.1 as the main classification loss. Label smoothing reduces the target probability of hard labels from 1.0 to 0.9, while uniformly distributing the remaining 0.1 probability to other classes. This effectively suppresses the model's overconfidence in the training data and improves generalization performance.

[0057] The second component is the supervised comparison loss. This loss is applied to the 128-dimensional L2-normalized feature vector output from the penultimate layer of the classifier. Its goal is to narrow the feature distance between samples of the same class and widen the feature distance between samples of different classes. Specifically, it first calculates the cosine similarity matrix between all sample pairs within a batch, then divides it by the temperature parameter. Scaling is applied, and then the mean of the normalized log-likelihood of each sample with all positive sample pairs is calculated as the loss. For the temperature parameter... The selection, smaller A value such as 0.07 can make the similarity distribution sharper, forcing the model to distinguish subtle differences in the feature space more precisely, which is beneficial for the model to distinguish morphologically similar ring-shaped RFIs. The weight coefficient for the auxiliary loss is 0.1, which helps to form a more compact and discriminative feature space.

[0058] The third component is the focal loss for toroidal objects. This is a loss component specifically designed for the confusion problem of circular RFI (Reference-Free Interference in Species) in this method. It only applies to circular samples; samples from other categories are ignored in this loss calculation. Its core idea is to use a focus-weighted mechanism, assigning higher loss weights to difficult-to-classify samples with lower classification confidence. (6) in γ = 2.0 represents the true class probability predicted by the model, and γ = 2.0 is the focus index. This applies when the model predicts a low probability for a certain circular RFI sample (i.e., a sample that is easily confused). The larger the value, the more significantly the sample contributes to the loss, guiding the network to focus on the most difficult-to-distinguish boundary samples within the ring family. The weight coefficient of this loss is 0.05, guiding the optimization direction with moderate force without interfering with the stable convergence of the main loss.

[0059] Since there is no readily available dataset for the various forms of radio frequency interference (RFI) from weather radar, and the probability of each type of interference occurring is unequal, a multi-morphological RFI simulation method is used to construct the dataset. For each original data file, after obtaining the base data matrix through signal processing, six simulated RFI interferences based on radar pulse regime and underlying physical timing are injected. The interference-to-noise ratio (INR) varies randomly within the range of 15-25 dB, and various forms of RFI are randomly generated at different azimuth angles, range sizes, radii, widths, and number of rings to ensure a rich and diverse dataset. Simultaneously, multiple physical-level random perturbations are superimposed on the simulated interference, such as edge broadening caused by simulated multipath effects, pixel loss caused by simulated channel fading, and interference stripe breaks, making the spatiotemporal distribution of the simulated RFI extremely close to the actual field measurement. The PPI interference forms corresponding to the simulated data are as follows: Figure 8 As shown, where, Figure 8 (a) is a schematic diagram of the PPI shape of the spiral radio frequency interference simulation dataset. Figure 8 (b) is a schematic diagram of the PPI shape of the radial radio frequency interference simulation dataset. Figure 8 (c) is a schematic diagram of the PPI (point-type radio frequency interference) simulation dataset. Figure 8 (d) is a schematic diagram of the PPI shape of the real circular radio frequency interference simulation dataset. Figure 8 (e) is a schematic diagram of the PPI shape of the multi-ring radio frequency interference simulation dataset. Figure 8 (f) is a schematic diagram of the PPI shape of the virtual circular radio frequency interference simulation dataset. After manual quality screening to remove samples with indistinct features, the final dataset contains 7229 valid samples, distributed by category as follows: Table 3. Partitioning of Multimorphic RFI Datasets

[0060] The dataset was randomly split into a training set (5783 samples), a validation set (723 samples), and a test set (723 samples) in an 8:1:1 ratio. Stratified sampling was used to ensure that the proportion of each category was consistent across the three subsets. This method employs a three-stage progressive parameter unfreezing and differential learning rate training strategy to fully utilize the pre-trained weights from ImageNet (a large-scale visual recognition dataset) while ensuring stable optimization of newly introduced modules. The core idea of ​​this strategy is to first train the new module to adapt to the data distribution, and then gradually unfreeze the pre-trained backbone network for collaborative fine-tuning, avoiding the destruction of the effectiveness of pre-trained features by the random initialization gradients of the new module in early training.

[0061] The first stage is the warm-up stage (epochs 1 to 5 of training). In this stage, all parameters of the backbone network (conv1, bn1, layers 1 to 4) are frozen, and only the newly introduced OTE module, ACA module, MixStyle module, classification heads (fc1 and fc2), and AIBN mixing coefficient parameters are trained, resulting in approximately 1.23 million trainable parameters. A uniform initial learning rate of 1e-3 is used, along with a cosine annealing scheduler. The purpose of this stage is to enable the new modules to quickly learn meaningful feature transformations, providing a good initialization for subsequent full-network collaborative training. Experiments show that after only 5 epochs of warm-up training, the validation set accuracy increases from a random level (approximately 17%) to 91.7%, indicating that the general visual features extracted by the pre-trained backbone network can quickly establish effective collaboration with the new modules.

[0062] The second phase is the partial unfreezing phase (epochs 6 to 15). While keeping layers 1 and 2 frozen, the parameters of layers 3 and 4, as well as all new modules, are unfrozen, increasing the number of trainable parameters to approximately 11.72 million. A differentiated learning rate strategy is adopted: layers 3-4 of the backbone network use a lower learning rate of 5e-5, while new modules and the classification head use a higher learning rate of 5e-4, with a learning rate ratio of 1:10. The lower backbone learning rate prevents excessive modification of pre-trained features, while the higher learning rate of new modules allows them to continue to optimize rapidly. The key role of this phase is to allow the high-level features (layers 3-4) to begin adapting to the specific distribution of radar echo data, while establishing a collaborative representation with the ACA module.

[0063] The third stage is the global fine-tuning stage (from the 16th epoch to the end of training). All parameters (approximately 12.39 million) are unfrozen, and a three-level differential learning rate is adopted: 1e-5 for layers 1-2, 3e-5 for layers 3-4, and 3e-4 for new modules and the classification head, with a learning rate ratio of 1:3:30. The shallow network uses the lowest learning rate because the low-level features extracted by the shallow layer, such as edges and textures, have good cross-domain generality and do not require significant adjustments; the deep network and new modules use an increasing learning rate to adapt to the semantic requirements of specific tasks. The entire training process uses cosine annealing learning rate scheduling, with an early stopping patience value set at 12 epochs, and a maximum training epoch count of 60 epochs.

[0064] The confusion matrix, accuracy, and loss function curves of the model trained using the above methods are shown below. Figure 9 and Figure 10 As shown, where, Figure 10 (a) is a schematic diagram of the total loss changes between the training set and the validation set. Figure 10 (b) is a schematic diagram of the changes in classification accuracy on the training set and validation set. Figure 10 (c) is a schematic diagram of the variation curves of each component of the combined loss. Figure 10 As can be seen, the TGANet proposed in this invention exhibits excellent convergence characteristics and numerical stability during a 60-epoch training process. The specific analysis is as follows: (1) Cooperative convergence of the combined loss function like Figure 10 As shown in (c), all three loss components exhibit a steady decreasing trend during training. The cross-entropy loss (CE), serving as the primary classification constraint, steadily decreased from approximately 1.5 to around 0.7, laying the foundation for the model's basic classification accuracy. The supervised contrast loss (SupCon) had a relatively high initial value of approximately 3.2, but decreased rapidly within the first 20 epochs. This indicates that the model is rapidly narrowing the feature distance between similar samples and pushing away dissimilar samples in the 128-dimensional feature space, effectively promoting the compactness of the feature space and laying the feature foundation for distinguishing morphologically similar interference. The ring focal loss (RingFocal) continuously applies pressure to difficult ring samples, with its value decreasing steadily from 0.8 to below 0.2 without significant oscillations. This demonstrates that the loss function accurately guides the network to optimize easily confused boundaries without compromising overall convergence.

[0065] (2) The phased characteristics of the three-stage progressive training Combination Figure 10(b) Obvious phased convergence characteristics can be observed: In Epochs 1-5 (warm-up period), since only the newly introduced modules and classification heads are trained, the validation set accuracy quickly climbs from a random level to over 90%, proving that the new modules can quickly establish effective cooperation with the pre-trained backbone network; In Epochs 6-15 (partial thawing period), with the thawing of the high-level semantic feature layers (Stages 3-4), the model begins to adapt to the specific distribution of radar echoes, and the accuracy further steadily climbs; After Epoch 16 (global fine-tuning period), the full parameter opening causes slight fluctuations in the early stage of training, but then enters the fine optimization stage under the guidance of the cosine annealing learning rate, and the validation set accuracy finally converges to a high level of 98.8%.

[0066] (3) Strong generalization ability brought about by regularization mechanism Depend on Figure 10 (b) It is evident that the accuracy on the validation set consistently exceeds that on the training set throughout the entire process. This phenomenon is not due to underfitting or overfitting, but rather because the present invention introduces highly targeted regularization techniques during the training phase, including MixStyle style mixing, elastic deformation simulation, CutMix (data augmentation algorithm), and Dropout (random deactivation regularization method). These mechanisms artificially increase the difficulty of the training task, forcing the network to learn the essential geometric structure of the interference signal rather than rote memorization of pixels. During the validation phase, the aforementioned interference mechanisms are turned off, and the model exhibits extremely strong generalization performance. The model's "rigorous training and robust validation" performance fully demonstrates that TGANet has extremely strong adaptability to complex radar echo backgrounds and simulation-real data domain offsets.

[0067] (4) Training stability verification Figure 10 The total loss curve in (a) shows that the validation loss decreases synchronously with the training progress and remains at an extremely low level of about 0.12 in the later stage of training without any signs of rebound. This verifies that the TGANet architecture proposed in this invention, combined with the three-stage training strategy, can effectively suppress the degradation problem of deep networks on small sample or cross-domain datasets, and ensure the reliability of the model in practical engineering applications.

[0068] Table 4 shows a comparison of the accuracy of various RFI classifications with the results of the original ResNet18: Table 4. Comparison of accuracy rates for various RFI classifications with the results of the original ResNet18.

[0069] The comparison results show that TGANet improves overall accuracy by 4.1 percentage points compared to the baseline, from 94.7% to 98.8%. Notably, the improvement is particularly significant for ring-type RFIs: the classification accuracy for virtual ring RFIs shows the largest increase, reaching 13.3 percentage points (from 85.0% to 98.3%), followed by multi-ring RFIs with an improvement of 7.3 percentage points (from 90.0% to 97.3%), and real ring RFIs with an improvement of 1.7 percentage points (from 98.3% to 100.0%). This result directly verifies the effectiveness of the ACA module and the combined loss function in distinguishing ring-type interference. The improvement in radial interference from 95.0% to 100.0% (+5.0%) also demonstrates the contribution of the OTE module to the encoding of directional texture features. From the overall macro average F1 score, TGANet significantly outperforms the baseline of 0.95 with 0.99, demonstrating that the multi-module collaborative architecture proposed in this invention has stronger classification robustness and comprehensiveness in capturing various RFI features in multi-morphological RFI recognition tasks, and can effectively take into account the refined discrimination of various RFI morphologies.

[0070] Experimental results fully demonstrate the superiority of this architecture: with almost no additional computational overhead, TGANet achieves an overall classification accuracy of 98.8%, and the macro F1 score is improved to 0.99. Particularly for the highly similar and extremely difficult-to-classify circular ring-type RFI, this invention, through the precise deconstruction of the range-azimuth axis structural continuity using the ACA module, significantly improves the recognition accuracy of virtual rings by 13.3 percentage points. Simultaneously, the combination of the AIBN module and multi-level data augmentation strategies provides built-in cross-domain generalization assurance for the smooth migration of the model from simulation training sets to real radar application scenarios.

[0071] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order in which these steps are executed, and they can be performed in other orders. Furthermore, Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0072] In one embodiment, a weather radar multi-morphological radio frequency interference classification device is provided, comprising: The sample acquisition module is used to acquire the basic data matrix sample of the preprocessed polarization Doppler weather radar echo. The basic data matrix sample includes multi-mode radio frequency interference signals and is correspondingly labeled with radio frequency interference modes. A network construction module is used to construct a texture-guided attention network, which includes a group of shallow feature extraction modules, a group of deep feature extraction modules, and an output module connected in sequence. The shallow feature enhancement module is used to perform multi-level progressive low-level feature extraction on the preprocessed base data matrix samples after initial feature transformation and output a shallow enhanced feature map through the shallow feature extraction module group. Each shallow feature extraction module extracts the global directional texture features of radio frequency interference in the input feature map, generates a spatial saliency weight map based on the texture features, and uses the spatial saliency weight map to perform residual adaptive enhancement on the low-level feature map. The deep feature enhancement module is used to perform multi-level progressive high-level semantic feature extraction on the shallow enhanced feature map through the deep feature extraction module group and output the deep enhanced feature map. Among them, each deep feature extraction module extracts the distance dimension structural features and orientation dimension continuity features of the circular radio frequency interference in the input feature map, generates multi-dimensional axial attention weights based on the structural features, and uses the multi-dimensional axial attention weights to perform residual adaptive enhancement on the high-level feature map. The classification and prediction module is used to process the deep enhanced feature map through the output module and output the radio frequency interference multimorphic classification and prediction results. The network training module is used to train the texture-guided attention network using the base data matrix samples and a pre-set loss function to obtain the trained texture-guided attention network. The result output module is used to perform multi-morphological radio frequency interference classification for weather radar using the trained texture-guided attention network.

[0073] Specific limitations regarding the multi-form radio frequency interference classification device for weather radar can be found in the limitations of the multi-form radio frequency interference classification method for weather radar mentioned above, and will not be repeated here. Each module in the aforementioned multi-form radio frequency interference classification device for weather radar can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0074] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 11As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements a multi-morphological radio frequency interference classification method for weather radar. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0075] Those skilled in the art will understand that Figure 11 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0076] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method described above.

[0077] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0078] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for classifying multi-morphological radio frequency interference in weather radar, characterized in that, The method includes: Acquire a basic data matrix sample of preprocessed polarization Doppler weather radar echoes. The basic data matrix sample includes multi-morphological radio frequency interference signals and is labeled with radio frequency interference morphology. A texture-guided attention network is constructed, which includes a group of shallow feature extraction modules, a group of deep feature extraction modules, and an output module connected in sequence. The shallow feature extraction module group performs multi-level progressive low-level feature extraction on the preprocessed base data matrix samples after initial feature transformation and outputs a shallow enhanced feature map. Each shallow feature extraction module extracts the global directional texture features of radio frequency interference in the input feature map, generates a spatial saliency weight map based on the texture features, and uses the spatial saliency weight map to perform residual adaptive enhancement on the low-level feature map. The deep feature extraction module group performs multi-level progressive high-level semantic feature extraction on the shallow enhanced feature map and outputs a deep enhanced feature map. Each deep feature extraction module extracts the distance dimension structural features and orientation dimension continuity features of the circular radio frequency interference in the input feature map, generates multi-dimensional axial attention weights based on the structural features, and uses the multi-dimensional axial attention weights to perform residual adaptive enhancement on the high-level feature map. The output module processes the deep enhanced feature map and outputs the radio frequency interference multimorphic classification prediction result. The texture-guided attention network is trained using the base data matrix samples and a pre-set loss function to obtain the trained texture-guided attention network. The trained texture-guided attention network is used for the classification of multi-morphological radio frequency interference in weather radar.

2. The method according to claim 1, characterized in that, The shallow feature extraction module group includes a first shallow feature extraction module and a second shallow feature extraction module; The step of performing multi-level progressive low-level feature extraction on the preprocessed base data matrix samples after initial feature transformation and outputting a shallow enhanced feature map includes: The first shallow feature extraction module extracts the low-level features of the preprocessed base data matrix samples after initial feature transformation and outputs the first low-level feature map. The second shallow feature extraction module performs progressive low-level feature extraction on the first low-level feature map and outputs a shallow enhanced feature map.

3. The method according to claim 2, characterized in that, Both the first shallow feature extraction module and the second shallow feature extraction module include a directional texture encoding submodule; The directional texture encoding submodule is used to perform channel dimensionality reduction mapping on the input feature map, and to perform depthwise separable convolution on the dimensionality-reduced feature map using a learnable Gabor filter bank to generate multiple directional texture response maps. A directional channel attention mechanism is performed on the multiple directional texture response maps to obtain multiple attention weights. The multiple attention weights are used to perform weighted fusion on the multiple directional texture response maps to generate a spatial saliency weight map. The spatial saliency weight map is used to perform residual adaptive enhancement on the input feature map.

4. The method according to claim 1, characterized in that, The deep feature extraction module group includes a first deep feature extraction module and a second deep feature extraction module; The deep feature extraction module group performs multi-level progressive high-level semantic feature extraction on the shallow enhanced feature map and outputs a deep enhanced feature map, including: The first deep feature extraction module extracts high-level semantic features from the shallow enhanced feature map and outputs a first high-level feature map. The second deep feature extraction module performs progressive high-level semantic feature extraction on the first high-level feature map, and outputs a deep enhanced feature map.

5. The method according to claim 4, characterized in that, Both the first deep feature extraction module and the second deep feature extraction module include an axial continuity attention submodule; The axial continuity attention submodule is used to extract the distance dimension structural features and orientation dimension continuity features of the input feature map, generate distance dimension attention weights, orientation dimension attention weights and channel attention weights, and use the distance dimension attention weights, orientation dimension attention weights and channel attention weights to perform residual adaptive enhancement on the input feature map.

6. The method according to claim 1, characterized in that, The loss function is a combined loss function, which includes labeled smoothed cross-entropy loss, supervised contrastive loss, and circular focal loss.

7. The method according to claim 1, characterized in that, The batch normalization layers in the texture-guided attention network are all adaptive instance-batch normalization submodules; The adaptive instance-batch normalization submodule is used to adaptively mix instance normalization results and batch normalization results of the input feature map according to learnable mixing coefficients before outputting the result.

8. The method according to claim 1, characterized in that, The process of processing the deep enhanced feature map and outputting the radio frequency interference multimorphic classification prediction result includes: The deep enhanced feature map is subjected to global average pooling and random deactivation in sequence. The pooled and deactivated features are then input into a fully connected layer for feature mapping. The mapped features are classified and activated using the softmax function, and the multimorphic classification prediction results of radio frequency interference are output.

9. A multi-form radio frequency interference classification device for weather radar, characterized in that, The device includes: The sample acquisition module is used to acquire the basic data matrix sample of the preprocessed polarization Doppler weather radar echo. The basic data matrix sample includes multi-mode radio frequency interference signals and is correspondingly labeled with radio frequency interference modes. A network construction module is used to construct a texture-guided attention network, which includes a group of shallow feature extraction modules, a group of deep feature extraction modules, and an output module connected in sequence. The shallow feature enhancement module is used to perform multi-level progressive low-level feature extraction on the preprocessed base data matrix samples after initial feature transformation and output a shallow enhanced feature map through the shallow feature extraction module group. Each shallow feature extraction module extracts the global directional texture features of radio frequency interference in the input feature map, generates a spatial saliency weight map based on the texture features, and uses the spatial saliency weight map to perform residual adaptive enhancement on the low-level feature map. The deep feature enhancement module is used to perform multi-level progressive high-level semantic feature extraction on the shallow enhanced feature map through the deep feature extraction module group and output the deep enhanced feature map. Among them, each deep feature extraction module extracts the distance dimension structural features and orientation dimension continuity features of the circular radio frequency interference in the input feature map, generates multi-dimensional axial attention weights based on the structural features, and uses the multi-dimensional axial attention weights to perform residual adaptive enhancement on the high-level feature map. The classification and prediction module is used to process the deep enhanced feature map through the output module and output the radio frequency interference multimorphic classification and prediction results. The network training module is used to train the texture-guided attention network using the base data matrix samples and a pre-set loss function to obtain the trained texture-guided attention network. The result output module is used to perform multi-morphological radio frequency interference classification for weather radar using the trained texture-guided attention network.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

Images