Method and system for pattern recognition of complex time series signals

Abstract

The application relates to the technical field of pattern recognition, and specifically discloses a pattern recognition method and system for complex-valued time series, wherein the method comprises the following steps: acquiring labeled and unlabeled signals and converting the signals into complex sequences; performing complex feature extraction and multi-codebook soft quantization on the complex sequences to obtain embedding sequences, and reconstructing the embedding sequences to obtain a complex VQ-GAN loss; performing HoC feature mapping and gate factor scaling on the complex sequences, and then performing additive fusion on the complex sequences and the embedding sequences to obtain HoC fusion features; inputting the HoC fusion features into a complex domain classifier to obtain corresponding prediction probabilities, obtaining a cross-entropy loss for the labeled signals, constructing a virtual adversarial perturbation for the unlabeled signals to obtain perturbed samples, and obtaining new prediction probabilities and a consistency loss according to the perturbed samples; optimizing and training a pattern recognition model according to a total loss; and performing pattern recognition according to the trained pattern recognition model, so that the recognition accuracy and robustness are improved.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing and pattern recognition technology, and particularly to a pattern recognition method for complex-valued time-series signals, a computer-readable storage medium, a computer device, and a pattern recognition system for complex-valued time-series signals. Background Technology

[0002] In related technologies, with the development of complex signal pattern recognition and machine learning techniques, there is a need for automatic classification and recognition of a large number of complex-valued time-series signals in scenarios such as wireless communication and electronic countermeasures. Automatic modulation identification (AMR) is one of the typical applications. Its goal is to automatically determine the modulation scheme based solely on the baseband I / Q signals acquired at the receiver, providing prior information for subsequent demodulation and link evaluation. Existing AMR / complex-valued time-series signal recognition methods mainly fall into two categories: one is based on artificial features and decision rules, constructing high-order cumulants, envelope statistics, spectral features, etc., and combining them with threshold decisions or traditional classifiers for recognition; the other is an end-to-end method based on deep learning, where the I / Q time-series signals are input into models such as CNN, RNN, or Transformer, and the network automatically learns the discriminative features.

[0003] However, in real complex electromagnetic environments, especially under conditions of low to medium signal-to-noise ratio (SNR) and limited labeled samples, existing technologies still have significant shortcomings:

[0004] (1) Insufficient utilization of complex structure. Many methods still treat I / Q as two independent real-valued channels input to real-valued networks, without explicitly encoding complex algebraic structures in operators such as convolution, normalization and activation, which limits the adaptability to phase rotation, conjugate symmetry and constellation geometry.

[0005] (2) Continuous features are the main focus, and there is a lack of discrete tokenization modeling. The internal representation of existing complex or geometric perception networks is mostly continuous feature maps. There are few complex baseband signals mapped to multi-codebook discrete token sequences and combined with sequence models for structured modeling. There are still gaps in related directions.

[0006] (3) Insufficient fusion of expert high-order statistical features and deep representation. HoC has good discriminative power at low SNR, but existing methods are mostly independent links or simple splicing, lacking a deep fusion mechanism that can be trained end-to-end and can adaptively adjust the contribution according to SNR.

[0007] (4) Semi-supervised consistency constraints mostly remain in the real-valued space. Existing pseudo-label, VAT and other methods usually apply perturbations to the real-valued input or continuous feature space. The perturbation path is not consistent with the complex inference decision branch. There is a lack of a scheme to construct perturbations in the complex input plane and impose constraints along the consistent path of "encoder + expert fusion + classification head".

[0008] (5) Robustness remains limited in low SNR and limited annotation scenarios. Constellation compression and trajectory overlap make it difficult for limited annotation samples to support sufficient generalization. Existing methods still have room for improvement in terms of SNR segmentation index and annotation ratio decay curve.

[0009] Therefore, a semi-supervised method is needed that can explicitly model the I / Q complex structure, utilize multi-codebook discrete token representation and sequence modeling, deeply integrate HoC expert priors, implement virtual adversarial training consistent with the inference path in the complex input plane, and fully exploit unlabeled I / Q data, in order to improve the recognition accuracy and robustness under low to medium SNR and limited labeling conditions. Summary of the Invention

[0010] This invention aims to at least partially solve one of the technical problems in the aforementioned technologies. To this end, one objective of this invention is to propose a pattern recognition method for complex-valued time-series signals. This method utilizes complex-domain multi-codebook discrete representation, expert statistical feature (HoC) extraction and gating fusion, and complex-domain virtual adversarial training (CVAT) consistency constraints to achieve digital discrimination of complex-valued time-series signal categories (such as modulation types), thereby improving recognition accuracy and robustness under conditions of limited annotation and low to medium signal-to-noise ratio.

[0011] A second objective of this invention is to provide a computer-readable storage medium.

[0012] The third objective of this invention is to provide a computer device.

[0013] The fourth objective of this invention is to provide a pattern recognition system for complex-valued time-series signals.

[0014] To achieve the above objectives, a first aspect of the present invention proposes a pattern recognition method for complex-valued time-series signals, comprising the following steps: acquiring signal data and converting the signal data into a complex sequence, wherein the signal data includes labeled signal data and unlabeled signal data; performing complex feature extraction and multi-codebook soft quantization on the complex sequence to obtain a real-valued embedding sequence, and reconstructing based on the real-valued embedding sequence to obtain a complex VQ-GAN loss; performing HoC feature mapping and gating factor scaling on the complex sequence and then additively fusing it with the real-valued embedding sequence to obtain HoC fused features; and applying the HoC fused features... The signal is input into a complex domain classifier to obtain the predicted probability of each modulation category corresponding to the signal. For labeled signal data, cross-entropy loss is obtained based on the corresponding predicted probability. For unlabeled signal data, virtual adversarial perturbation is constructed to obtain perturbation samples, and new predicted probabilities are obtained based on the perturbation samples. Consistency loss is obtained based on the corresponding predicted probability and the new predicted probability. The model is optimized and trained based on the complex VQ-GAN loss, cross-entropy loss, and consistency loss to obtain a trained pattern recognition model. The signal to be recognized is input into the trained pattern recognition model to obtain the corresponding pattern recognition result.

[0015] The pattern recognition method for complex-valued time-series signals according to embodiments of the present invention has the following advantages: it can achieve higher recognition accuracy and robustness in complex electromagnetic environments, low to medium SNR conditions, and scenarios with limited labeled samples.

[0016] In addition, the pattern recognition method for complex-valued time-series signals proposed in the above embodiments of the present invention may also have the following additional technical features:

[0017] Optionally, complex feature extraction and multi-codebook soft quantization are performed on the complex sequence to obtain an embedded sequence, including: downsampling and feature extraction of the complex sequence using a complex encoder to obtain a multi-codebook prequantized latent representation; performing Top-k search and temperature softmax weighting on the prequantized latent representation of each time step in the corresponding codebook, and then performing a convex combination of the weighted prototype mixture and the original prequantized latent representation to obtain a stable latent vector; and concatenating all time steps and codebooks to obtain a real-valued embedded sequence of uniform width through linear mapping.

[0018] Optionally, the complex sequence is subjected to HoC feature mapping and gating factor scaling, and then additively fused with the real-valued embedding sequence. This includes: estimating multiple moments from the complex sequence and calculating higher-order cumulants accordingly, then using power normalization to obtain normalized higher-order cumulants, ultimately forming an 8-dimensional HoC feature vector; mapping the HoC feature vector through two layers of MLP to an expert embedding with the same width as the real-valued embedding sequence, then scaling it with a training round gating factor and a gating factor based on sample SNR, broadcasting it to the time dimension, and additively fused with the real-valued embedding sequence.

[0019] Optionally, the HoC fusion features are input into a complex domain classifier to obtain the predicted probabilities of each modulation category corresponding to the signal. This includes: adding the HoC fusion features to a learnable positional encoding and adding a start marker at the beginning of the sequence to obtain a preprocessed sequence; inputting the preprocessed sequence into a multi-layer Transformer encoder, wherein each Transformer encoder consists of a multi-head self-attention sublayer and a feedforward network sublayer, and uses a residual connection and layer normalization structure to ensure training stability; the sequence features output by the multi-layer Transformer encoder are time-dimensionally pooled to obtain a global representation, and then passed through a fully connected layer and a softmax output layer to obtain the predicted probabilities of each modulation category.

[0020] To achieve the above objectives, a second aspect of the present invention provides a computer-readable storage medium storing a pattern recognition program for complex-valued timing signals, which, when executed by a processor, implements the pattern recognition method for complex-valued timing signals as described above.

[0021] To achieve the above objectives, a third aspect of the present invention provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the pattern recognition method for complex-valued timing signals as described above.

[0022] To achieve the above objectives, a fourth aspect of the present invention proposes a pattern recognition system for complex-valued time-series signals, comprising: an acquisition module for acquiring signal data and converting the signal data into a complex sequence, wherein the signal data includes labeled signal data and unlabeled signal data; an encoding and multi-codebook discrete representation module for performing complex feature extraction and multi-codebook soft quantization on the complex sequence to obtain a real-valued embedding sequence, and reconstructing based on the real-valued embedding sequence to obtain a complex VQ-GAN loss; an HoC expert feature extraction and gating fusion module for performing HoC feature mapping and gating factor scaling on the complex sequence and then additively fusing it with the real-valued embedding sequence to obtain HoC fusion features; and a sequence segmentation module. The model is divided into three modules: a classifier module and a recognition module module. The classifier module is used to input the HoC fusion features into a complex domain classifier to obtain the predicted probabilities of each modulation category corresponding to the signal. For labeled signal data, cross-entropy loss is obtained based on the corresponding predicted probabilities. For unlabeled signal data, virtual adversarial perturbation is constructed to obtain perturbation samples, and new predicted probabilities are obtained based on the perturbation samples. Consistency loss is obtained based on the corresponding predicted probabilities and the new predicted probabilities. The training module is used to optimize and train the model based on the complex VQ-GAN loss, cross-entropy loss, and consistency loss to obtain a trained pattern recognition model. The recognition module is used to input the signal to be recognized into the trained pattern recognition model to obtain the corresponding pattern recognition result. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating a pattern recognition method for complex-valued time-series signals according to an embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of the overall network structure of a pattern recognition model according to an embodiment of the present invention;

[0025] Figure 3 This is a schematic diagram of the complex VQ-GAN multi-codebook soft quantization discrete feature extraction and reconstruction process according to an embodiment of the present invention;

[0026] Figure 4 A graph showing the comparison of the recognition accuracy of the method of the present invention and a comparative method according to an embodiment of the present invention on the RadioML2016.10A dataset as the signal-to-noise ratio changes;

[0027] Figure 5 This is a comparison chart of the test accuracy of the method of the present invention and a comparative method under different proportions of labeled training samples according to an embodiment of the present invention;

[0028] Figure 6 This is a block diagram of a pattern recognition system for complex-valued timing signals according to an embodiment of the present invention. Detailed Implementation

[0029] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0030] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the invention to those skilled in the art.

[0031] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0032] Figure 1 This is a flowchart illustrating a pattern recognition method for complex-valued time-series signals according to an embodiment of the present invention, as shown below. Figure 1 As shown, the pattern recognition method for this complex-valued time-series signal includes the following steps:

[0033] S101, acquire signal data and convert the signal data into a complex sequence, wherein the signal data includes labeled signal data and unlabeled signal data.

[0034] Specifically, the complex-valued baseband I / Q signal acquired by the receiver is represented as The first dimension corresponds to the I and Q components, and the second dimension represents the time sampling points. Treat it as a complex sequence .

[0035] S102, perform complex feature extraction and multi-codebook soft quantization on the complex sequence to obtain the real-valued embedding sequence, and reconstruct it based on the real-valued embedding sequence to obtain the complex VQ-GAN loss.

[0036] As one embodiment, complex feature extraction and multi-codebook soft quantization are performed on the complex sequence to obtain a real-valued embedding sequence, including: using a complex encoder to downsample and extract features from the complex sequence to obtain a multi-codebook prequantized latent representation; performing Top-k search and temperature softmax weighting on the prequantized latent representation of each time step in the corresponding codebook, and then performing a convex combination of the weighted prototype mixture and the original prequantized latent representation to obtain a stable latent vector; after concatenating all time steps and codebooks, a real-valued embedding sequence of uniform width is obtained through linear mapping.

[0037] In other words, the input complex-valued baseband I / Q signal is mapped to a multi-codebook discrete latent sequence through complex-valued convolution, complex-valued normalization, and multi-codebook soft quantization, and then reconstructed back into a time-domain signal by the decoder for pre-training and representation quality monitoring. Figure 3 The paper presents the potential representation from the original I / Q signal to multi-codebook prequantization. From there, the complete path leads to stable discrete latent sequences and reconstructed signals.

[0038] Specifically, such as Figure 3 As shown, the complex encoder consists of multiple layers of complex one-dimensional convolution, complex normalization, and phase-preserving activation units, progressively processing the complex sequence. Perform downsampling and feature extraction to output a multi-codebook pre-quantized latent representation. The multi-codebook soft quantization module performs a Top-k search and temperature-weighted softmax summation on the complex latent vector at each time step within the corresponding codebook. Then, it performs a convex combination of the weighted prototype mixture and the original latent vector to obtain a stable discrete latent vector. After concatenating all time steps and codebooks, a real-valued embedding sequence of uniform width is obtained through a linear mapping P. It is used by downstream modules and can also be based on a decoder. Or the corresponding complex latent representation is reconstructed for VQ-GAN pre-training and representation quality monitoring.

[0039] It should be noted that the complex VQ-GAN loss includes reconstruction loss and codebook loss, which is obtained according to the following formula:

[0040] ,

[0041] ,

[0042] ,

[0043]

[0044] in, For the decoder processing, This refers to the encoder process.

[0045] S103, after performing HoC feature mapping and gating factor scaling on the complex sequence, it is additively fused with the real-valued embedded sequence to obtain the HoC fused features.

[0046] As an example, the complex sequence is subjected to HoC feature mapping and gating factor scaling, and then additively fused with the real-valued embedding sequence. This includes: estimating the multi-order moments from the complex sequence and calculating the higher-order cumulants accordingly, then using power normalization to obtain the normalized higher-order cumulants, finally forming an 8-dimensional HoC feature vector; mapping the HoC feature vector through two layers of MLP to an expert embedding with the same width as the real-valued embedding sequence, then scaling it with a training round gating factor and a gating factor based on sample SNR, broadcasting it to the time dimension and additively fusing it with the real-valued embedding sequence.

[0047] Specifically, such as Figure 2 As shown, from the complex domain baseband signal Estimating multi-order moments And based on this, calculate the higher-order cumulants. Then, power normalization is used to obtain the normalized higher-order cumulants. This ultimately forms an 8-dimensional HoC feature vector. The HoC feature vectors are mapped to expert embeddings with the same width as the backbone features through two layers of MLP. After scaling by training round gating factors and sample SNR-based gating factors, the data is broadcast to the time dimension and embedded into the backbone sequence. Perform additive fusion to obtain the HoC fusion sequence. Among them, the training round gating reduces the HoC weight in the early stage of training and gradually increases it as training progresses; the SNR gating assigns greater weight when the SNR is low to medium and compresses the HoC weight when the SNR is high, so as to balance expert prior and deep representation.

[0048] S104, the HoC fusion features are input into the complex domain classifier to obtain the predicted probability of each modulation category corresponding to the signal. For labeled signal data, the cross-entropy loss is obtained based on the corresponding predicted probability. For unlabeled signal data, a virtual adversarial perturbation is constructed to obtain perturbation samples, and a new predicted probability is obtained based on the perturbation samples. The consistency loss is obtained based on the corresponding predicted probability and the new predicted probability.

[0049] As one embodiment, HoC fusion features are input into a complex domain classifier to obtain the predicted probabilities of each modulation category corresponding to the signal. This includes: adding the HoC fusion features to a learnable positional encoding and adding a start marker at the beginning of the sequence to obtain a preprocessed sequence; inputting the preprocessed sequence into a multi-layer Transformer encoder, wherein each Transformer encoder consists of a multi-head self-attention sublayer and a feedforward network sublayer, and uses residual connections and layer normalization structures to ensure training stability; the sequence features output by the multi-layer Transformer encoder are time-dimensionally pooled to obtain a global representation, and then passed through a fully connected layer and a softmax output layer to obtain the predicted probabilities of each modulation category.

[0050] Specifically, such as Figure 2 As shown, the HoC fusion sequence The sequence is added to a learnable positional encoding, and a start token is added to the beginning of the sequence before being input into a multi-layer Transformer encoder. Each Transformer encoder layer consists of a multi-head self-attention sublayer and a feedforward network sublayer, employing residual connections and layer normalization structures to ensure training stability. The sequence features output by the encoder are subjected to temporal pooling (e.g., max pooling or average pooling) to obtain a global representation, which is then passed through a fully connected layer and a softmax output layer to obtain the predicted probabilities of each modulation category, achieving supervised classification training for labeled samples.

[0051] In addition, such as Figure 2 As shown, since this application adopts a semi-supervised training strategy, the training samples are divided into labeled subsets and unlabeled subsets. Therefore, for unlabeled samples, the predicted distribution is first obtained by forward inference based on the HoC fusion features of the current model. Then, construct a function on the input plane in the complex field that satisfies... Virtual confrontation disturbance ,in, This represents the maximum allowable norm for the control disturbance and guarantees the disturbance samples. A new predicted distribution is obtained through a path identical to the original sample: "complex encoding + multi-codebook discrete representation + HoC fusion + Transformer classifier". Finally, the KL divergence is calculated as the consistency loss.

[0052] It should be noted that the cross-entropy loss is obtained from the following formula:

[0053] ,

[0054] Among them, for labeled samples The classifier outputs a probability vector for predicted categories:

[0055]

[0056] It should be noted that the consistency loss is obtained from the following formula:

[0057]

[0058] Among them, unlabeled samples First, obtain the original predicted distribution:

[0059]

[0060] Construct a perturbation in the complex input plane (Normally constrained), perturbation samples are obtained. And calculate the new predicted distribution:

[0061]

[0062] The formula for calculating disturbance constraints is:

[0063]

[0064] This is the maximum permissible norm.

[0065] S105. The model is optimized and trained using complex VQ-GAN loss, cross-entropy loss, and consistency loss to obtain a well-trained pattern recognition model.

[0066] In other words, during training, the total loss consists of the cross-entropy loss of the labeled samples. Consistency loss for complex virtual adversarial training (CVAT) of unlabeled samples and complex VQ-GAN loss Composition. Depending on the needs, weight enhancement strategies can be introduced for the supervision loss and consistency loss of low SNR samples, and a warm-up strategy can be adopted for the CVAT weights. The model parameters are updated using optimization algorithms such as AdamW, ultimately obtaining a pattern recognition model that is robust across the entire SNR range and maintains good performance under limited labeling conditions.

[0067] S106, input the signal to be identified into the trained pattern recognition model to obtain the corresponding pattern recognition result.

[0068] In other words, the signal to be identified is input into a trained pattern recognition model to obtain the predicted probability of each modulation category, and the modulation category with the highest probability is taken as the result of pattern recognition.

[0069] Furthermore, the method described in this application was validated on the publicly available wireless modulation dataset RadioML2016.10A, and the experimental results are as follows: Figure 4 and Figure 5 As shown.

[0070] Figure 4 The recognition accuracy curves of the CoMQT method in this application and several comparative methods on the RadioML2016.10A dataset as a function of signal-to-noise ratio (SNR) are presented. It can be seen that in the low to medium SNR range, the accuracy of the method in this application is significantly higher than that of baseline models such as real-valued CNNs, ordinary complex-valued networks, and multi-domain fusion. Its performance also remains at a high level in the high SNR range, indicating that complex-domain multi-codebook discrete representation, HoC expert fusion, and complex-valued virtual adversarial training jointly improve the model's robustness and discriminative ability across the entire SNR range.

[0071] Figure 5 The results show the comparison of recognition accuracy of the proposed method and the comparative method on the test set under different proportions of labeled training samples. As the labeled proportion gradually decreases from high to low, the performance degradation of the proposed method is significantly slower than that of other methods. It can still maintain a high recognition accuracy in low-labeled scenarios, which shows that the semi-supervised training framework proposed in this application can effectively utilize a large amount of unlabeled I / Q data and significantly improve labeling efficiency and generalization performance under limited labeling conditions.

[0072] As can be seen from the above embodiments, the semi-supervised pattern recognition method for complex-valued time-series signals based on complex-domain multi-codebook discrete representation and virtual adversarial training can achieve higher recognition accuracy and robustness than traditional methods in complex electromagnetic environments, low to medium SNR conditions, and scenarios with limited labeled samples.

[0073] In summary, the pattern recognition method for complex-valued time-series signals according to embodiments of the present invention first represents the wireless I / Q baseband signal in a complex form and extracts features using a complex-domain VQ-GAN encoder composed of complex convolution, complex normalization, and phase-preserving activation. Then, a multi-codebook soft-quant mechanism is introduced to generate stable multi-codebook discrete token sequences in multiple complementary latent subspaces based on complex inner product similarity, Top-k temperature weighting, and convex combination. Furthermore, HoC expert statistical features are extracted from the complex-valued signal, and deep additive fusion with the backbone token representation is achieved through training round gating and sample SNR gating. Finally, the fused sequence is input into a Transformer sequence classifier to complete supervised learning. Simultaneously, in the semi-supervised stage, CVAT virtual adversarial perturbations are directly constructed on the complex input plane, and the KL divergence consistency loss is calculated strictly along the same decision path of "complex encoding / discrete representation—HoC fusion—Transformer classifier head," jointly optimizing the model parameters with cross-entropy and VQ-GAN constraints. Therefore, by explicitly introducing complex operators and phase-preserving activation during the encoding stage, the ability to represent amplitude-phase coupling and constellation geometry is enhanced. Through multi-codebook soft-quant and convex combination, the discrete representation ability and training stability are improved without significantly increasing the size of a single codebook, enabling the Transformer to perform temporal modeling on more structured token sequences. By utilizing dual gating of training rounds and sample SNR, the contribution of expert features is adaptively amplified under medium-low SNR conditions, improving the stability of low SNR decision. Unlike the traditional VAT that applies perturbations to the real-valued space or intermediate layers, this invention constructs perturbations in the complex input plane and applies consistency regularization along the complete inference path, more effectively utilizing unlabeled I / Q data to improve robustness and generalization ability. Experimental results based on publicly available wireless modulation datasets show that, compared with baseline methods such as typical real-valued networks, ordinary complex-valued networks, and multi-domain fusion, this invention achieves significant advantages in overall accuracy across the entire SNR range, recognition performance in low-to-medium SNR subsets, and performance retention when different annotation ratios decrease. This demonstrates that the proposed complex-domain multi-codebook discrete representation + HoC gated fusion + complex-domain CVAT semi-supervised consistency learning can synergistically improve the model's recognition accuracy and robustness under complex electromagnetic environments and limited annotation conditions.

[0074] To implement the above embodiments, this invention also proposes a computer-readable storage medium storing a pattern recognition program for complex-valued timing signals. When the pattern recognition program for complex-valued timing signals is executed by a processor, it implements the pattern recognition method for complex-valued timing signals as described above.

[0075] To implement the above embodiments, this invention also proposes a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the pattern recognition method for complex-valued timing signals as described above.

[0076] To implement the above embodiments, this invention also proposes a pattern recognition system for complex-valued time-series signals, such as... Figure 6 As shown, the pattern recognition system for the complex-valued time-series signal includes: an acquisition module 10, an encoding and multi-codebook discrete representation module 20, a HoC expert feature extraction and gating fusion module 30, a sequence classification module 40, a training module 50, and a recognition module 60.

[0077] The sequence includes: an acquisition module 10 for acquiring signal data and converting it into a complex sequence; the signal data includes labeled and unlabeled signal data; an encoding and multi-codebook discrete representation module 20 for extracting complex features and performing multi-codebook soft quantization on the complex sequence to obtain a real-valued embedding sequence, and reconstructing it to obtain the complex VQ-GAN loss; an HoC expert feature extraction and gating fusion module 30 for performing HoC feature mapping and gating factor scaling on the complex sequence, and then additively fusing it with the real-valued embedding sequence to obtain HoC fused features; and a sequence classification module 40 for inputting the HoC fused features into the complex sequence. A number-domain classifier is used to obtain the predicted probability of each modulation category corresponding to the signal. For labeled signal data, the cross-entropy loss is obtained based on the corresponding predicted probability. For unlabeled signal data, a virtual adversarial perturbation is constructed to obtain perturbation samples, and a new predicted probability is obtained based on the perturbation samples. The consistency loss is obtained based on the corresponding predicted probability and the new predicted probability. The training module 50 is used to optimize and train the model based on the complex VQ-GAN loss, cross-entropy loss and consistency loss to obtain a trained pattern recognition model. The recognition module 60 is used to input the signal to be recognized into the trained pattern recognition model to obtain the corresponding pattern recognition result.

[0078] It should be noted that the above modules can be implemented through a combination of software and hardware. For example, the above functional modules can be implemented by one or more general-purpose processors, graphics processors, or dedicated acceleration chips executing program instructions stored on a computer-readable storage medium.

[0079] It should be noted that the above description of the pattern recognition method for complex-valued time-series signals also applies to the pattern recognition system for the same complex-valued time-series signals, and will not be repeated here.

[0080] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0081] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0082] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0083] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0084] It should be noted that any reference signs placed between parentheses in the claims should not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In a unit claim enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc., does not indicate any order. These words can be interpreted as names.

[0085] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0086] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0087] In the description of this invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0088] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0089] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "over," and "on top" of the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0090] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms should not be construed as necessarily referring to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0091] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A pattern recognition method for complex-valued time-series signals, characterized in that, Includes the following steps: Acquire signal data and convert the signal data into a complex sequence, wherein the signal data includes tagged signal data and untagged signal data; Complex feature extraction and multi-codebook soft quantization are performed on the complex sequence to obtain a real-valued embedding sequence, and the complex VQ-GAN loss is obtained by reconstructing the real-valued embedding sequence. The complex sequence is subjected to HoC feature mapping and gating factor scaling, and then additively fused with the real-valued embedding sequence to obtain the HoC fused features; The HoC fusion features are input into a complex domain classifier to obtain the predicted probabilities of each modulation category corresponding to the signal. For labeled signal data, cross-entropy loss is obtained based on the corresponding predicted probabilities. For unlabeled signal data, virtual adversarial perturbation is constructed to obtain perturbation samples, and new predicted probabilities are obtained based on the perturbation samples. Consistency loss is obtained based on the corresponding predicted probabilities and the new predicted probabilities. The model is optimized and trained based on the complex VQ-GAN loss, cross-entropy loss, and consistency loss to obtain a well-trained pattern recognition model. The signal to be identified is input into the trained pattern recognition model to obtain the corresponding pattern recognition result; The process of performing HoC feature mapping and gating factor scaling on the complex sequence, followed by additive fusion with the real-valued embedded sequence, includes: The multi-order moments are estimated from the complex sequence, and the higher-order cumulants are calculated accordingly. Then, power normalization is used to obtain the normalized higher-order cumulants, and finally an 8-dimensional HoC feature vector is formed. The HoC feature vector is mapped to an expert embedding with the same width as the real-valued embedding sequence through two layers of MLP. After scaling by the training round gate factor and the sample SNR-based gate factor, it is broadcast to the time dimension and additively fused with the real-valued embedding sequence.

2. The pattern recognition method for complex-valued time-series signals as described in claim 1, characterized in that, Complex feature extraction and multi-codebook soft quantization are performed on the complex sequence to obtain a real-valued embedding sequence, including: The complex sequence is downsampled and its features are extracted using a complex encoder to obtain a multi-codebook pre-quantized latent representation; For each time step, the prequantized latent representation is subjected to Top-k search and temperature softmax weighting in the corresponding codebook. The weighted prototype mixture is then combined with the original prequantized latent representation to obtain a stable latent vector. After concatenating all time steps and codebooks, a real-valued embedding sequence of uniform width is obtained through linear mapping.

3. The pattern recognition method for complex-valued time-series signals as described in claim 1, characterized in that, The HoC fusion features are input into a complex domain classifier to obtain the predicted probabilities of each modulation category corresponding to the signal, including: The HoC fusion features are added to the learnable positional codes, and a start marker is added to the beginning of the sequence to obtain the preprocessed sequence; The preprocessed sequence is input into a multi-layer Transformer encoder, where each Transformer encoder consists of a multi-head self-attention sub-layer and a feedforward network sub-layer, and uses residual connections and layer normalization structure to ensure training stability. The sequence features output by the multi-layer Transformer encoder are pooled in the time dimension to obtain a global representation, and then passed through a fully connected layer and a softmax output layer to obtain the predicted probabilities of each modulation category.

4. A computer-readable storage medium, characterized in that, It stores a pattern recognition program for complex-valued timing signals, which, when executed by a processor, implements the pattern recognition method for complex-valued timing signals as described in any one of claims 1-3.

5. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the pattern recognition method for complex-valued timing signals as described in any one of claims 1-3.

6. A pattern recognition system for complex-valued time-series signals, characterized in that, include: An acquisition module is used to acquire signal data and convert the signal data into a complex sequence, wherein the signal data includes tagged signal data and untagged signal data; The encoding and multi-codebook discrete representation module is used to perform complex feature extraction and multi-codebook soft quantization on the complex sequence to obtain a real-valued embedding sequence, and to reconstruct the complex VQ-GAN loss based on the real-valued embedding sequence. The HoC expert feature extraction and gating fusion module is used to perform HoC feature mapping and gating factor scaling on the complex sequence and then perform additive fusion with the real-valued embedded sequence to obtain HoC fused features. The sequence classification module is used to input the HoC fusion features into a complex domain classifier to obtain the predicted probability of each modulation category corresponding to the signal. For labeled signal data, cross-entropy loss is obtained based on the corresponding predicted probability. For unlabeled signal data, virtual adversarial perturbation is constructed to obtain perturbation samples, and new predicted probabilities are obtained based on the perturbation samples. Consistency loss is obtained based on the corresponding predicted probability and the new predicted probability. The training module is used to optimize the training of the model based on the complex VQ-GAN loss, cross-entropy loss and consistency loss to obtain a trained pattern recognition model. The recognition module is used to input the signal to be recognized into the trained pattern recognition model in order to obtain the corresponding pattern recognition result; The HoC expert feature extraction and gating fusion module is further used to estimate multi-order moments from the complex sequence, calculate higher-order cumulants accordingly, and then use power normalization to obtain normalized higher-order cumulants, ultimately forming an 8-dimensional HoC feature vector. The HoC feature vector is mapped to an expert embedding with the same width as the real-valued embedding sequence through two layers of MLP, and then scaled by a training round gating factor and a gating factor based on sample SNR before being broadcast to the time dimension and additively fused with the real-valued embedding sequence.

Images