A dual-stream robust signal classification method and device based on granule geometry correction and depth feature complementation
By constructing a dual-stream robust signal classification method that complements particle-sphere geometric correction and deep features, the problems of single feature extraction view and insufficient robustness in wireless communication signal classification are solved, and high-precision signal recognition is achieved in low signal-to-noise ratio environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANXI UNIV
- Filing Date
- 2026-03-02
- Publication Date
- 2026-06-05
AI Technical Summary
Existing wireless communication signal classification methods suffer from limitations in feature extraction perspectives, lacking an understanding of the macroscopic geometric distribution of the sample space. This results in insufficient robustness in low signal-to-noise ratio environments, making them prone to misjudgment and overfitting.
A robust dual-stream signal classification method based on particle-sphere geometric correction and complementary deep features is adopted. By constructing IQ and AP views, macroscopic geometric features are extracted using particle-sphere computation, and deep features are extracted by combining Transformer and LSTM networks. A fuzzy regularization term is introduced to optimize the loss function and enhance the robustness of the model.
It significantly improves the accuracy and robustness of signal classification, and can effectively identify wireless signals in complex electromagnetic environments, breaking through the performance bottleneck of existing technologies.
Smart Images
Figure CN122153498A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of wireless communication signal processing technology, and specifically relates to a dual-stream robust signal classification method and device based on particle-sphere geometric correction and depth feature complementarity. Background Technology
[0002] Automatic modulation identification (AMC) in wireless communication signals is a core technology in cognitive radio, spectrum monitoring, and electronic countermeasures systems. Its main task is to automatically identify the modulation type of a signal from a noisy signal when the receiver does not know the transmission parameters. In the civilian sector, AMC is an important means of achieving dynamic spectrum access, improving spectrum utilization, and effectively monitoring illegal signal transmissions. In the military sector, it is a prerequisite for electronic support measures (ESM), possessing irreplaceable strategic value for signal demodulation, interference identification, and threat assessment of friendly and enemy targets. With the increasing complexity of the electromagnetic environment, the diversification of signal types, and the prevalence of non-Gaussian noise and multipath fading, achieving high-precision and robust signal classification in low signal-to-noise ratio and high dynamic environments has become a challenging problem to be solved in the field of communication signal processing.
[0003] Looking at the current technological development path, signal classification methods have mainly evolved from traditional likelihood ratio theory and artificial feature engineering to deep learning. Although early likelihood ratio testing methods could theoretically achieve optimal performance, their computational complexity was extremely high and they heavily relied on accurate prior channel information, making it difficult to meet real-time requirements. Artificial feature-based methods relied on expert experience to extract statistical features such as higher-order cumulants and cyclic spectra, combined with classical classifiers for identification, but their generalization ability was weak and they were difficult to adapt to complex and ever-changing channel environments.
[0004] In recent years, deep learning has gradually become the mainstream paradigm in this field due to its powerful end-to-end feature extraction capabilities. Convolutional Neural Networks (CNNs) are widely used to extract spatial texture features from signal time-frequency maps or constellation diagrams, Recurrent Neural Networks (RNNs) and their variants (such as LSTMs) focus on capturing the temporal dependencies of signal sequences, while the latest Transformer architecture has been introduced to handle long sequence signals due to its excellent global attention mechanism. These methods have, to some extent, broken through the performance bottlenecks of traditional methods and achieved significant improvements in classification accuracy.
[0005] Despite significant progress made by existing deep learning-based methods, several key issues still hinder performance improvement when facing extremely complex electromagnetic environments. The primary problem lies in the singular and one-sided nature of existing methods in their feature extraction perspective. Most models rely on only a single data format for training, such as using only in-phase-orthogonal (IQ) data or only amplitude-phase (AP) data. While IQ data preserves the original projection of the signal onto the complex plane, it struggles to capture the instantaneous changes in amplitude and phase when severe aliasing occurs at low signal-to-noise ratios. AP data, while highlighting the envelope fluctuations and phase transitions of the signal, destroys the orthogonal spatial structure of the original signal. This singular perspective cannot simultaneously consider the spatial geometric distribution and physical time-varying characteristics of the signal, leading to misjudgments when processing higher-order modulation or similar modulation types due to the lack of key discriminative features. Secondly, existing deep neural networks are essentially feature extractors based on micro-pixels or points, tending to focus on subtle textures or local temporal patterns within samples, lacking the ability to perceive the overall distribution of the data. In low signal-to-noise ratio environments, signal waveforms are severely distorted due to strong noise interference, generating a large number of outliers. Since existing networks lack a mechanism that can ignore local noise and correct biases from a macroscopic distribution perspective, they are prone to mistaking noise patterns for signal features, leading to "overfitting" and severely affecting the robustness of the model. Summary of the Invention
[0006] The purpose of this invention is to provide a robust dual-stream signal classification method and device based on particle-sphere geometric correction and complementary depth features, which can solve the problems of existing wireless signal classification technology having a single feature extraction perspective, lack of perception of the macroscopic geometric distribution of the sample space, and insufficient robustness.
[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0008] A robust dual-stream signal classification method based on particle-sphere geometric correction and complementary depth features includes the following steps:
[0009] Acquire in-phase and quadrature data of the wireless communication signal to be identified;
[0010] Based on the in-phase-orthogonal data, a first feature view data and a second feature view data are constructed. The method for constructing the first feature view data is to flatten the in-phase-orthogonal data into a one-dimensional vector. The method for constructing the second feature view data is to perform a nonlinear coordinate transformation on the in-phase-orthogonal data to generate amplitude-phase data.
[0011] Based on the first feature view data of the training samples, an adaptive splitting clustering algorithm is used to construct a sphere coverage set. Each sphere in the sphere coverage set is composed of sphere center, radius, purity, reliability and density characterization parameters.
[0012] For the first feature view data of the sample to be identified, calculate its Euclidean distance to the center of all spheres in the sphere coverage set, select the closest sphere as a reference, extract the geometric feature vector based on the purity, reliability and density of the sphere, and map the geometric feature vector into coarse-grained features through a multilayer perceptron;
[0013] For the second feature view data, a convolutional layer is used to downsample and superimpose position encoding to obtain sequence features. After adaptive recalibration by the channel attention module, it is input into the Transformer encoder for global dependency modeling. Then, the output sequence of the Transformer encoder is input into the Long Short-Term Memory network for local temporal modeling. The hidden state of the last time step is taken and mapped into fine-grained features through the projection layer.
[0014] The coarse-grained features and the fine-grained features are concatenated along the channel dimension to obtain a joint feature vector;
[0015] The joint feature vector is input into a classifier containing a fully connected layer and a Softmax output layer. The classifier outputs the probability distribution of each signal category and takes the category with the highest probability as the classification result.
[0016] Preferably, the step of constructing the particle-sphere cover set using the adaptive splitting clustering algorithm specifically includes the following steps:
[0017] The preset purity threshold is adaptively adjusted based on the total number of modulation categories to obtain the actual purity threshold.
[0018] If the number of samples in the current data subset is less than the preset minimum number of samples, then the data subset is directly generated into a single sphere.
[0019] Otherwise, calculate the label purity for the current data subset; if the purity is greater than or equal to the actual purity threshold, then generate the data subset into a single sphere.
[0020] If the purity is less than the actual purity threshold, the 2-Means algorithm is used to divide the data subset into two subsets, and the above judgment and splitting operations are recursively performed on each subset until all samples are covered by spheres.
[0021] Preferably, the calculation of the Euclidean distance between the first feature view data and the centers of all particles, selecting the nearest particle as a reference, and extracting the geometric feature vector specifically includes the following steps:
[0022] Let the first feature view data be denoted as The first sphere in the sphere-covered set The center of each ball is marked as ,calculate Find the nearest neighbor index of the Euclidean distance from the center of all particles. :
[0023]
[0024] index as The nearest neighbor sphere is denoted as the sphere with the highest purity. The radius is denoted as The set of samples included is denoted as The sample size is denoted as The set of labels corresponding to this sample set is denoted as Based on the nearest neighbor spheres, the following features are extracted:
[0025] Purity characteristics: ;
[0026] Reliability characteristics: ,in For a set of tags Information entropy Total number of modulation categories;
[0027] Density characteristics: ;
[0028] Affinity characteristics: ,in To prevent small positive constants with a denominator of zero;
[0029] Adversarial characteristics: , , , ;
[0030] Finally, the geometric eigenvectors are obtained. .
[0031] Preferably, the multilayer perceptron includes two fully connected layers, each followed by a BN layer and ReLU activation.
[0032] Preferably, the adaptive recalibration via the channel attention module specifically includes the following steps:
[0033] Channel statistics of sequence features are generated through global average pooling;
[0034] The channel weight vector is learned through two fully connected layers;
[0035] The channel weight vector is multiplied with the sequence features channel by channel to obtain the recalibrated sequence features.
[0036] Preferably, in the global dependency modeling of the input Transformer encoder, the Talking-Heads mechanism is introduced into the multi-head self-attention mechanism of the Transformer encoder, and a linear transformation is applied to the attention score matrix before and after Softmax to enhance the information interaction between heads.
[0037] Preferably, the classifier consists of a fully connected layer, a ReLU activation layer, and a Dropout layer. It outputs the probability distribution of each signal category through a Softmax function and takes the category with the highest probability as the classification result.
[0038] Preferably, during the training process of the classifier, the total loss function Classification loss based on cross-entropy With fuzzy regularization terms Weighted composition:
[0039]
[0040] in, The regularization coefficient is . This refers to the training batch size;
[0041] Regular terms The prediction probability is constructed using the degree of dispersion and information entropy, and its calculation formula is as follows:
[0042]
[0043] in, and To be constrained, the following conditions must be met:
[0044]
[0045] in, For training batch size, The number of similar probability categories selected. Indicates the first The sample belongs to the first The predicted probability values for each category, This is a preset fuzzy threshold constant.
[0046] In another aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps in the dual-stream robust signal classification method based on particle-sphere geometric correction and depth feature complementarity described above.
[0047] In another aspect, the present invention provides a dual-stream robust signal classification device based on particle-sphere geometric correction and complementary depth features, comprising:
[0048] Memory, used to store software applications.
[0049] A processor is configured to execute the software application, wherein each program of the software application correspondingly performs the steps in the dual-stream robust signal classification method based on particle-sphere geometric correction and depth feature complementarity described above.
[0050] This invention innovatively constructs a dual-stream feature extraction architecture: On the one hand, it introduces particle-sphere computation theory to process the IQ view, adaptively dividing the irregular signal sample space into several hyperspheres to extract macroscopic geometric features such as purity, density, and reliability. This is equivalent to adding a low-pass filter to the model, significantly enhancing the model's noise robustness in low signal-to-noise ratio environments from a macroscopic data distribution perspective. On the other hand, it utilizes a hybrid deep network to process the AP view, combining the global attention of the Transformer with the local temporal modeling capability of LSTM to extract deep semantic features of the signal. Furthermore, to address the prediction ambiguity phenomenon easily generated by similar modulation types, this invention also introduces a fuzzy regularization term. By adding a sharpening constraint on the output probability distribution in the loss function, it forces the model to tend towards determinism in predicting easily confused categories. This organic integration of macroscopic geometric view, microscopic spatiotemporal view, and fuzzy regularization strategy enables this invention to more comprehensively and accurately characterize wireless signals, effectively breaking through the performance bottleneck of signal classification in complex communication environments. Attached Figure Description
[0051] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0052] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.
[0053] This embodiment verifies the proposed dual-stream robust signal classification method based on particle-sphere geometric correction and deep feature complementarity using wireless communication signal datasets (such as RML2016.10a and RML2016.10b). This method achieves robust recognition of complex signals by constructing a dual-channel architecture of "macroscopic geometric flow" and "microscopic spatiotemporal flow".
[0054] S1. Construction and Preprocessing of Multi-View Data Space
[0055] Let the original wireless signal dataset be... ,in The total number of samples, For modulation category labels.
[0056] For a single input sample , Representing the two channels of IQ, The length of the signal sequence is given in this embodiment. This invention constructs two complementary feature views:
[0057] S1.1 First View (Geometric View): Preserves the Euclidean space distribution characteristics of the original in-phase and orthogonal components. The sample is flattened into a one-dimensional vector, which serves as the first-order input.
[0058]
[0059] S1.2 Second View (AP View): Performs a nonlinear coordinate transformation on the original IQ data to generate amplitude-phase (AP) data, which serves as the input to the second stream. Let time... The signal components are The transformed signal Defined as:
[0060]
[0061]
[0062] in To prevent numerically unstable small constants, in this embodiment... .
[0063] S2. Coarse-grained geometric feature extraction based on particle size distribution.
[0064] This step aims to extract the macroscopic geometric distribution features of the samples in the feature space.
[0065] S2.1 Particle Coverage Generation (Training Phase)
[0066] An adaptive overlay of the training set data is performed using a particle-sphere computation algorithm. Let the current dataset be... The granule generation process follows the following iterative splitting strategy:
[0067] Purity calculation: for a subset of data Calculate its label purity .like (Setting a threshold) If ), then a sphere is generated. Cover the area; otherwise, proceed to the splitting step.
[0068] Sphere splitting: If purity is insufficient, use the 2-Means algorithm to split the sphere. Divided into two subsets The purity calculation is then recursively performed on the subset until all samples are covered or the minimum sample number constraint is reached.
[0069] Particle parameterization: The final set of particles is denoted as . Each sphere From the center of the ball ,radius ,purity ,reliability and density Characteristic parameters are defined;
[0070] radius The maximum Euclidean distance from the sample inside the sphere to the center of the sphere;
[0071] purity : Percentage of samples from the dominant category within the sphere;
[0072] Sample set : The set of sample indices belonging to this ball.
[0073] S2.2 Geometric Feature Transformation (Reference Stage)
[0074] For the samples to be classified Calculate its geometric eigenvectors :
[0075] Nearest neighbor matching: Calculate samples Find the nearest neighbor index of the nearest neighbor particle, based on the distance from the center of all particles. :
[0076]
[0077] Explicit feature calculation: based on matched spheres Extract the following features:
[0078] Purity characteristics: ;
[0079] Reliability characteristics: ,in Information entropy;
[0080] Density characteristics: ;
[0081] Affinity characteristics: ;
[0082] Adversarial characteristics: ;
[0083] Finally, we obtain the vector. .
[0084] S2.3 Deep Feature Mapping
[0085] Multilayer perceptron (MLP) is used to map low-dimensional geometric features to high-dimensional coarse-grained features. This MLP contains two fully connected layers, each followed by BatchNormalization (BN) and ReLU activation:
[0086]
[0087]
[0088] in, , This is the weight matrix. In this embodiment, the output feature dimension is... .
[0089] S3. Fine-grained feature extraction based on Transformer-LSTM
[0090] This step processes the second view data. Extract deep spatiotemporal semantic features.
[0091] S3.1 Convolutional Coding and Position Embedding
[0092] First, feature downsampling and local feature extraction are performed using a two-layer one-dimensional convolutional neural network (CNN). In this embodiment, the kernel size is set to 4, the stride to 2, and the padding to 1.
[0093]
[0094] in Then, learnable positional codes are superimposed. To preserve sequence timing information:
[0095]
[0096] S3.2 SE Channel Attention Enhancement
[0097] The Squeeze-and-Excitation (SE) module is introduced to perform adaptive recalibration of feature channels.
[0098] Squeeze: Generates channel statistics through global average pooling. :
[0099]
[0100] Excitation: Channel weights are learned through two fully connected layers. :
[0101]
[0102] in For the Sigmoid function, For ReLU function, scaling ratio .
[0103] Reweighting: .
[0104] S3.3 Transformer Global Dependency Modeling
[0105] Will The input to the TransformerEncoder is transposed into a sequence. The core computation utilizes a multi-head self-attention mechanism.
[0106] Let the input sequence be Calculate query ,key ,value :
[0107]
[0108] To enhance inter-head information interaction, a Talking-Heads mechanism is introduced, performing a linear transformation on the attention score matrix before and after Softmax. The final output is:
[0109]
[0110]
[0111] S3.4 Local Timing Modeling with LSTM
[0112] The output sequence of the Transformer is fed into a two-layer LSTM network to capture local short-term dependencies:
[0113]
[0114] Take the hidden state of the last time step As the final temporal representation of the sequence.
[0115] S3.5 Feature Projection
[0116] The LSTM output is mapped to fine-grained features through a projection layer. :
[0117]
[0118] Where the output dimension .
[0119] S4. Dual-stream Feature Fusion and Classification Decision
[0120] S4.1 Feature Cascade
[0121] First-order macroscopic geometric features are combined with second-order microscopic spatiotemporal features along the feature channel dimension:
[0122]
[0123] The fused feature dimensions are dimension.
[0124] S4.2 Classification Prediction
[0125] Building a classifier It consists of a fully connected layer, a ReLU activation layer, and a Dropout layer:
[0126]
[0127]
[0128] in This represents the predicted probability distribution of a sample belonging to each modulation category.
[0129] S4.3 Model Optimization and Fuzzy Regularization
[0130] To address the problem of smoothing of prediction probability distributions (i.e., fuzzy uncertainty) that easily occurs in categories with similar modulation types, this invention introduces a fuzzy regularization term based on the statistical properties of probability distributions into the optimization objective. .
[0131] For the first in the training batch For each sample, the regularization term is constructed using the dispersion of the predicted probability and the information entropy, and its calculation formula is as follows:
[0132]
[0133] in, and To be constrained, the following conditions must be met:
[0134]
[0135] The meanings of the parameters in the above formula are as follows:
[0136] This refers to the training batch size;
[0137] The number of similar probability categories selected is used to calculate the local probability distribution characteristics;
[0138] Indicates the first The sample belongs to the first The predicted probability values for each category;
[0139] This is a preset fuzzy threshold constant used to adjust the tolerance for uncertainty;
[0140] The information entropy calculation based on the predicted probability characterizes the average uncertainty level of the current batch of samples;
[0141] The term is constructed based on the second-order moment statistics of the predicted probability, reflecting the sharpness and discreteness of the predicted distribution.
[0142] S4.4 Joint Loss Optimization
[0143] Finally, the model's total loss function The classification loss is determined by the cross-entropy of the main task. Combined with the above fuzzy regularization terms, the following is formed:
[0144]
[0145] in This is the regularization coefficient. By minimizing this total loss, the model can utilize [variables] while maintaining classification accuracy. Suppressing fuzzy predictions with high entropy values significantly improves the model's confidence in distinguishing easily confused modulation categories.
[0146] Through the above steps, this invention achieves effective complementarity between geometric priors and deep semantic information, significantly enhancing the robustness of the signal classification system in complex environments. This invention was validated on the RML2016.10a and RML2016.10b standard datasets, with the training, validation, and test sets divided in a 6:2:2 ratio for model parameter training, weight optimization, and performance evaluation, respectively. The experiments used three metrics for multi-dimensional evaluation: F1-Score, ACC, and H-ACC. F1-Score evaluates the model's overall balanced performance across different categories, ACC measures the average classification accuracy across the entire signal-to-noise ratio range, and H-ACC captures the performance peak at a single signal-to-noise ratio. The experimental results are shown in Table 1, demonstrating that the proposed method outperforms existing technologies in multiple metrics.
[0147] Table 1
[0148]
[0149] In another aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps in the dual-stream robust signal classification method based on particle-sphere geometric correction and depth feature complementarity described above.
[0150] In another aspect, the present invention provides a dual-stream robust signal classification device based on particle-sphere geometric correction and complementary depth features, comprising:
[0151] Memory, used to store software applications.
[0152] A processor is configured to execute the software application, wherein each program of the software application correspondingly performs the steps in the dual-stream robust signal classification method based on particle-sphere geometric correction and depth feature complementarity described above.
Claims
1. A robust dual-stream signal classification method based on particle-sphere geometric correction and complementary depth features, characterized in that, Includes the following steps: Acquire in-phase and quadrature data of the wireless communication signal to be identified; Based on the in-phase-orthogonal data, a first feature view data and a second feature view data are constructed. The method for constructing the first feature view data is to flatten the in-phase-orthogonal data into a one-dimensional vector. The method for constructing the second feature view data is to perform a nonlinear coordinate transformation on the in-phase-orthogonal data to generate amplitude-phase data. Based on the first feature view data of the training samples, an adaptive splitting clustering algorithm is used to construct a sphere coverage set. Each sphere in the sphere coverage set is composed of sphere center, radius, purity, reliability and density characterization parameters. For the first feature view data of the sample to be identified, calculate its Euclidean distance to the center of all spheres in the sphere coverage set, select the closest sphere as a reference, extract the geometric feature vector based on the purity, reliability and density of the sphere, and map the geometric feature vector into coarse-grained features through a multilayer perceptron; For the second feature view data, a convolutional layer is used to downsample and superimpose position encoding to obtain sequence features. After adaptive recalibration by the channel attention module, it is input into the Transformer encoder for global dependency modeling. Then, the output sequence of the Transformer encoder is input into the Long Short-Term Memory network for local temporal modeling. The hidden state of the last time step is taken and mapped into fine-grained features through the projection layer. The coarse-grained features and the fine-grained features are concatenated along the channel dimension to obtain a joint feature vector; The joint feature vector is input into a classifier containing a fully connected layer and a Softmax output layer. The classifier outputs the probability distribution of each signal category and takes the category with the highest probability as the classification result.
2. The method according to claim 1, characterized in that, The process of constructing the particle-sphere cover set using the adaptive splitting clustering algorithm specifically includes the following steps: The preset purity threshold is adaptively adjusted based on the total number of modulation categories to obtain the actual purity threshold. If the number of samples in the current data subset is less than the preset minimum number of samples, then the data subset is directly generated into a single sphere. Otherwise, calculate the label purity for the current data subset; if the purity is greater than or equal to the actual purity threshold, then generate the data subset into a single sphere. If the purity is less than the actual purity threshold, the 2-Means algorithm is used to divide the data subset into two subsets, and the above judgment and splitting operations are recursively performed on each subset until all samples are covered by spheres.
3. The method according to claim 1, characterized in that, The calculation of the Euclidean distance between the first feature view data and the centers of all particles, selecting the nearest particle as a reference, and extracting the geometric feature vector specifically includes the following steps: Let the first feature view data be denoted as The first sphere in the sphere-covered set The center of each ball is marked as ,calculate Find the nearest neighbor index of the Euclidean distance from the center of all particles. : index as The nearest neighbor sphere is denoted as the sphere with the highest purity. The radius is denoted as The set of samples included is denoted as The sample size is denoted as The set of labels corresponding to this sample set is denoted as Based on the nearest neighbor spheres, the following features are extracted: Purity characteristics: ; Reliability characteristics: ,in For a set of tags Information entropy Total number of modulation categories; Density characteristics: ; Affinity characteristics: ,in To prevent small positive constants with a denominator of zero; Adversarial characteristics: , , , ; Finally, the geometric eigenvectors are obtained. .
4. The method according to claim 1, characterized in that, The multilayer perceptron comprises two fully connected layers, each followed by a BN layer and ReLU activation.
5. The method according to claim 1, characterized in that, The adaptive recalibration via the channel attention module specifically includes the following steps: Channel statistics of sequence features are generated through global average pooling; The channel weight vector is learned through two fully connected layers; The channel weight vector is multiplied with the sequence features channel by channel to obtain the recalibrated sequence features.
6. The method according to claim 1, characterized in that, In the global dependency modeling of the input Transformer encoder, the Talking-Heads mechanism is introduced into the multi-head self-attention mechanism of the Transformer encoder. A linear transformation is applied to the attention score matrix before and after Softmax to enhance the information interaction between heads.
7. The method according to claim 1, characterized in that, The classifier consists of a fully connected layer, ReLU activation, and Dropout layer. It outputs the probability distribution of each signal category through the Softmax function and takes the category with the highest probability as the classification result.
8. The method according to claim 1, characterized in that, During the training of the classifier, the total loss function Classification loss based on cross-entropy With fuzzy regularization terms Weighted composition: in, The regularization coefficient is . This refers to the training batch size; Regular terms The prediction probability is constructed using the degree of dispersion and information entropy, and its calculation formula is as follows: in, and To be constrained, the following conditions must be met: in, For training batch size, The number of similar probability categories selected. Indicates the first The sample belongs to the first The predicted probability values for each category, This is a preset fuzzy threshold constant.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it performs the steps in the dual-stream robust signal classification method based on particle-sphere geometric correction and depth feature complementarity as described in any one of claims 1 to 8.
10. A robust dual-stream signal classification device based on particle-sphere geometric correction and complementary depth features, characterized in that, include: Memory, used to store software applications. A processor for executing the software application, wherein each program of the software application correspondingly performs the steps in the dual-stream robust signal classification method based on particle-sphere geometric correction and depth feature complementarity as described in any one of claims 1 to 8.