An unsupervised domain adaptive channel coding type identification method, device, and medium

Abstract

This invention discloses an unsupervised domain-adaptive channel coding type identification method, device, and medium, belonging to the field of wireless communication technology. The method includes: obtaining the log-likelihood ratio sequences of the source domain labeled and the target domain unlabeled; inputting the sequences into a model, performing low-dimensional projection on the sequence features through a self-attention mechanism, performing sequence-level domain alignment to capture codeword structure dependencies, and calculating the domain discriminant loss; maintaining class centers and using exponential moving averages for momentum updates, generating pseudo-labels for the target domain based on cosine similarity, and calculating the pseudo-label loss by selecting high-confidence samples; combining the source domain classification loss, jointly optimizing the feature extractor, classifier, and domain discriminator through a gradient inversion layer to achieve channel coding type identification. This invention effectively preserves the codeword algebraic structure and significantly improves generalization and recognition capabilities in real wireless environments.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and specifically to an unsupervised domain adaptive channel coding type identification method, device, and medium. Background Technology

[0002] The statements in this section are provided only as background information in connection with this disclosure and may not constitute prior art.

[0003] Channel coding is a crucial component of modern digital communication systems. It improves transmission reliability by introducing structured redundancy to counteract the negative effects of noise and interference. In a wide range of wireless communication applications, such as non-cooperative communication systems and adaptive modulation and coding, the receiver often cannot know in advance the coding configuration used by the transmitter. In such cases, blind identification of channel coding becomes a key step in achieving successful decoding and reliable recovery of transmitted information.

[0004] Traditional channel coding identification methods are mainly based on Gaussian-Jordanian elimination (GJETP) and rank calculation mechanisms. While these methods are effective in distinguishing macroscopic coding types, they struggle to identify different coding schemes within the same category (e.g., low-density parity-check codes (LDPC) and polar codes are both linear block codes). In recent years, deep learning has demonstrated great potential in channel coding type identification tasks, and existing methods based on convolutional neural networks (CNNs) or recurrent neural networks (RNNs) have achieved good performance in simulation environments.

[0005] However, existing deep learning methods heavily rely on idealized additive white Gaussian noise (AWGN) channels for training, making it difficult to adapt to multipath fading, hardware defects, and symbol-level non-uniform perturbations in real wireless channels. This structural distribution difference between the simulation and real domains leads to feature distribution shifts, resulting in a significant decrease in the model's recognition performance in real wireless environments.

[0006] To address the problem of cross-domain distribution shift, unsupervised domain adaptation (UDA) techniques have been proposed, with adversarial training (such as Domain Adversarial Neural Networks (DANNs) as a representative strategy. While these UDA methods have achieved success in global classification tasks such as computer vision, channel coding type identification deals with coded sequences with strict algebraic structures, where domain differences are often reflected in subtle bit-level or structurally related variations. Existing UDA methods typically perform domain alignment based on global features (i.e., pooling or averaging time-series features). This global feature aggregation operation masks local, position-dependent dependencies, which are crucial for characterizing the algebraic structure of codewords. Therefore, directly applying existing UDA strategies based on global feature alignment is not suitable, as it fails to explicitly model the structural dependencies of codewords, resulting in extracted features lacking domain discriminative power and limiting the model's generalization ability in the real target domain. Summary of the Invention

[0007] The purpose of this invention is to overcome the above-mentioned shortcomings in the prior art and to provide an unsupervised domain adaptive channel coding type identification method, device, and medium. This addresses the technical problem that existing unsupervised domain adaptive methods in channel coding identification suffer from insufficient cross-domain feature representation and low identification accuracy in real multipath fading wireless channel environments due to over-reliance on global feature alignment, which masks the local structural dependencies of codeword sequences.

[0008] The technical solution of the present invention is as follows: An unsupervised domain adaptive channel coding type identification method includes: Step S1: Obtain the source domain dataset and the target domain dataset; the source domain dataset contains a log-likelihood ratio sequence with channel coding type labels generated by simulation, and the target domain dataset contains an unlabeled log-likelihood ratio sequence extracted from signals received from a real wireless environment; Step S2: Construct a channel coding type identification model, which includes a feature extractor, a classifier, and a domain discriminator; pre-train the feature extractor and classifier using the source domain dataset; Step S3: Perform sequence-level domain alignment on the model based on adversarial training: The log-likelihood ratio sequences from the source domain dataset and the target domain dataset are input into a pre-trained feature extractor to extract sequence features. The sequence features are input into the domain discriminator, and the sequence features are projected into a query matrix, a key matrix, and a value matrix through a self-attention mechanism. Low-dimensional projection is performed on the key matrix and the value matrix to capture the structural dependencies of the sequence. Based on the attention output after low-dimensional projection, it is determined whether the sample comes from the source domain or the target domain, and the domain discrimination loss is calculated. Step S4: Perform self-supervised training on target domain samples based on pseudo-labels: Extract the classification features of the target domain samples, and output the predicted class probability through the classifier; Maintain a category center for each channel coding type, and update the momentum of the category center using an exponential moving average mechanism; Calculate the cosine similarity between the classification features of the target domain samples and the class centers to generate pseudo-labels for the target domain samples; use a confidence threshold to filter out target domain samples with high confidence and calculate the pseudo-label loss; Step S5: Combining the classification loss of the source domain dataset, the domain discrimination loss, and the pseudo-label loss, the feature extractor, the classifier, and the domain discriminator are jointly optimized through a gradient inversion layer to obtain the trained channel coding type recognition model; Step S6: Obtain the communication signal to be identified, calculate its log-likelihood ratio sequence and input it into the trained channel coding type identification model, and output the channel coding type identification result.

[0009] Further, obtain the log-likelihood ratio sequences from the source domain dataset and the target domain dataset, including: The receiving end receives the baseband signal modulated by orthogonal frequency division multiplexing; Symbol timing synchronization is performed using the preamble structure in the baseband signal, and carrier frequency offset correction is performed using a two-step carrier frequency offset estimation method. Channel estimation is performed in the frequency domain using the long training sequence in the preamble structure to obtain the channel frequency response on the subcarrier; The baseband signal is equalized according to the channel frequency response, the equalized symbols are demodulated into corresponding bit sequences, and the log-likelihood ratio of each bit in the bit sequence is calculated to form a log-likelihood ratio sequence.

[0010] Furthermore, the specific formula for calculating the log-likelihood ratio is as follows:

[0011] in: Indicates the first A symbol after equilibrium The demodulated first Bits The log-likelihood ratio; The symbol after the equilibrium is represented Posterior probability under given conditions.

[0012] Further, the step of projecting the sequence features into a query matrix, a key matrix, and a value matrix using a self-attention mechanism, and performing low-dimensional projection on the key matrix and value matrix, includes: sequence features Projected into query matrices respectively Key matrix Sum matrix ; The key matrix and value matrix are subjected to a low-dimensional projection using the following formula:

[0013]

[0014] in: The projected key matrix; This is the projected value matrix; and These are projection matrices used to map the key matrix and value matrix to a preset domain discriminant subspace, respectively. Calculate the attention output after low-dimensional projection:

[0015] in: The feature dimension of the sequence feature; This represents the normalized exponential function; The attention output is input into the residual multilayer perceptron for binary classification to determine whether the sample comes from the source domain or the target domain.

[0016] Furthermore, the step of updating the momentum of the category center using an exponential moving average mechanism includes: The category corresponding to the maximum predicted category probability of the target domain sample is taken as the predicted label. ; Based on the predicted class probabilities of the target domain samples and the preset temperature hyperparameters, the weighting weights are calculated. ; According to the weighted weights Calculate the weighted batch center within the predicted category. :

[0017] in: Batch size; For the first Classification features of each target domain sample; For predicting labels; For category number; For indicator functions; To prevent division by zero of constants; The category center is updated using a momentum update mechanism:

[0018] in: For category The category center; The momentum coefficient; Perform L2 normalization on the updated category centers: .

[0019] Further, the cosine similarity between the classification features of the target domain samples and the class centers is calculated to generate pseudo-labels for the target domain samples, and the pseudo-label loss is calculated, including: Classification features of target domain samples With normalized category centers Perform cosine similarity matching to generate pseudo-labels for the current batch. :

[0020] The maximum predicted category probability is greater than the pre-configured confidence threshold. The samples are used as a high-confidence sample set. Participate in training; Calculate the average cross-entropy on the high-confidence sample set as the pseudo-label loss. :

[0021] in: Represents the cross-entropy loss function; This indicates that the classifier classifies features. The output.

[0022] Furthermore, when jointly optimizing the feature extractor, the classifier, and the domain discriminator through the gradient inversion layer, the gradient inversion layer performs an identity mapping during forward propagation and flips the gradient sign during backward propagation. The flipping formula is as follows:

[0023]

[0024] in: The input gradient; The gradient after flipping; Indicates the normalized training progress; A coefficient used to control the growth rate; The inversion coefficient is dynamically adjusted according to the training progress.

[0025] Furthermore, the comprehensive loss function, which combines the classification loss of the source domain dataset, the domain discrimination loss, and the pseudo-label loss for joint optimization, is expressed as:

[0026] in: , , These represent the parameters of the feature extractor, the classifier, and the domain discriminator, respectively. The classification loss is the loss for the source domain dataset; For the domain, determine the loss; The pseudo-label loss; This is the tradeoff coefficient for the false label loss.

[0027] The present invention also proposes an electronic device, comprising: At least one processor; and a memory communicatively connected to said at least one processor; The memory stores instructions that can be executed by the at least one processor, and the at least one processor executes the instructions stored in the memory to perform the method described above.

[0028] The present invention also proposes a computer-readable storage medium for storing instructions that, when executed, cause the method described above to be implemented.

[0029] Compared with existing technologies, the advantages of this invention are: 1. This invention overcomes the limitations of traditional global feature alignment and effectively preserves the algebraic structural features of codewords. It innovatively proposes an unsupervised domain adaptive method based on sequence-level feature alignment, utilizing a self-attention mechanism to perform low-dimensional projection modeling of sequence features. This mechanism avoids the masking of local dependencies by traditional temporal pooling operations and can effectively capture the inherent structural dependencies in the encoded sequence (i.e., changes in codeword structure rather than uniform shifts in global features), thereby extracting cross-domain key structural features with strong domain discriminative capabilities.

[0030] 2. By employing a self-supervised training mechanism in the target domain, the discriminative power and stability of target domain features are significantly enhanced. Addressing the issue of missing supervision signals due to the lack of labels in the target domain, this invention maintains class centers for each channel coding type within the target domain and uses an exponential moving average (EMA) mechanism for momentum updates. This mechanism not only smooths out changes in class centers during iteration and provides stable guidance anchors throughout the training process, but also, combined with a confidence-based sample selection strategy, effectively eliminates low-confidence noise samples, ensuring extremely high reliability of pseudo-label supervision information.

[0031] 3. An end-to-end adversarial training framework was constructed, significantly improving the model's generalization and recognition capabilities in real wireless environments. This invention integrates sequence-level domain discriminant loss, target domain pseudo-label loss, and source domain classification loss, and jointly optimizes the feature extractor and classifier through a gradient inversion layer. Experimental results show that the method of this invention can effectively extract robust domain-invariant features with good generalization capabilities across both simulated and real domains; especially under harsh channel conditions such as low signal-to-noise ratio (SNR), the method of this invention significantly improves the average classification accuracy compared to existing domain adaptive baseline models (such as DANN, MCC, etc.), and exhibits highly balanced and stable recognition capabilities across various coding types (Cohen's Kappa coefficient is significantly better than existing comparative methods). Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments recorded in the embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0033] Figure 1 This is a general framework diagram of an unsupervised domain adaptive channel coding type identification method. Figure 2 This is a comparison chart of average accuracy rates.

[0034] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0035] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0036] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0037] Example 1 This embodiment provides an unsupervised domain adaptive channel coding type identification method, applicable to scenarios such as non-cooperative communication signal analysis and adaptive modulation and coding. In non-cooperative communication systems, the receiver cannot know in advance the coding configuration used by the transmitter, making blind channel coding identification a crucial step for successful decoding and reliable recovery of transmitted information. However, multipath fading, hardware defects, and symbol-level non-uniform perturbations in real wireless channels can lead to feature distribution shifts. To address this issue, the method in this embodiment aims to mitigate the structural distribution differences between the simulated and real domains.

[0038] Please see Figure 1 Specifically, the method includes the following steps: Step S1: Obtain the source domain dataset and the target domain dataset; the source domain dataset contains a log-likelihood ratio sequence with channel coding type labels generated by simulation, and the target domain dataset contains an unlabeled log-likelihood ratio sequence extracted from signals received from a real wireless environment.

[0039] Specifically, a labeled source domain dataset is defined. ,in For the data samples generated in the simulation, Its corresponding category label; Simultaneously define an unlabeled target domain dataset. ,in This refers to data received from a real wireless environment. Source domain. Based on AWGN channel generation, and target domain This refers to wireless reception from the actual environment. It is important to emphasize that the source and target domains share the same tag space, i.e. and The samples in this example all come from the same set of channel coding schemes. In this context, the objectives of this embodiment include two aspects: (1) Analyze the impact of domain differences between the source domain and the target domain on classification performance; (2) By mitigating cross-domain distribution shift, the generalization ability of the model in real-world scenarios can be improved.

[0040] To transform the underlying physical layer baseband signal into a feature sequence suitable for deep learning, the process of obtaining the log-likelihood ratio (LLR) sequence includes: Transmitter baseband signal Orthogonal Frequency Division Multiplexing (OFDM) modulation is employed. After propagation through the dynamic wireless channel, the signal received by the receiver is modeled as follows:

[0041] in: Indicates normalized carrier frequency offset (CFO); This refers to carrier phase offset; For the first The time-varying channel coefficients of the multipath components, with corresponding time delays of: , For the number of multipaths; It is additive white Gaussian noise.

[0042] Transmitter baseband signal Constructed according to the IEEE 802.11a frame format, including the Legacy ShortTraining Field (L-STF, 32) in the preamble. ), Legacy Long Training Field (L-LTF, 32 ), LegacySignal Field (L-SIG, 16 ), and including Data fields for each OFDM symbol (16 consecutive symbols per symbol) To support more types of channel coding and parameter configurations, this embodiment modifies the data field of the standard IEEE 802.11a frame, replacing its default payload with data encoded using various coding schemes, while keeping the preamble structure unchanged to maintain synchronization, channel estimation, and signaling functions.

[0043] The signal processing chain at the receiving end proceeds sequentially as follows: Cross-correlation with a known long training sequence to detect the OFDM start position (symbol timing synchronization); A two-step CFO estimation method is used to correct the carrier frequency offset (frequency offset estimation and correction) for the preamble. Channel estimation is performed in the frequency domain using L-LTF to obtain the channel frequency response on all subcarriers; Then, the baseband signal is equalized.

[0044] After balancing, the balanced symbols will be... The modulation is performed to obtain the corresponding bit sequence. The task uses the log-likelihood ratio of each bit in the modulation symbol as the input feature. The first symbol The LLR of 1 bit is defined as:

[0045] in: For symbols The demodulated first Bits; This represents the posterior probability under the condition of receiving the symbol after the equalization.

[0046] Step S2: Construct a channel coding type identification model, which includes a feature extractor, a classifier, and a domain discriminator; pre-train the feature extractor and classifier using the source domain dataset.

[0047] First, in the source domain dataset Train a supervised model. This model consists of two components: a feature extractor. Mapping the input sequence to a categorical feature representation, the classifier. Predict based on the encoding type corresponding to the feature output. After training in the source domain, the feature extractor... The learned weights will be used to initialize the shared feature extractor in the subsequent domain adaptation framework. .

[0048] During the source domain pre-training phase, the classifier With feature extractor Joint training is performed by minimizing the cross-entropy loss of the source domain samples:

[0049] in, and These are the parameters for the feature extractor and the classifier, respectively. After training in the source domain, the feature extractor... The learned weights will be used to initialize the shared feature extractor in the subsequent domain adaptation framework. Subsequent adversarial frameworks will utilize feature extractors shared between the source and target domains. and shared classifiers .

[0050] Step S3: Perform sequence-level domain alignment on the model based on adversarial training: The log-likelihood ratio sequences from the source and target domain datasets are input into the feature extractor. Extract sequence features ,in Indicates the number of time steps. For feature dimension, For the first Local feature embedding at each time step. In this embodiment, it should be noted that the adversarial adaptive framework provides a solid foundation for cross-domain feature alignment. However, traditional domain discriminators mainly make decisions based on classification-level features, i.e., the global representation obtained by averaging time-series features. This temporal pooling operation masks a large number of local positional changes and structural features, which are crucial for revealing inter-domain differences. Therefore, domain alignment relying solely on global features is often insufficient, thus limiting the model's generalization ability in the target domain. This embodiment focuses on modeling the joint dependencies of feature vectors in the temporal dimension to more effectively preserve domain-related information and capture potential structural distribution differences between the source and target domains. Formally, consider sequence features. The joint distribution is expressed by the chain rule as follows:

[0051] The sequence features are input into the domain discriminator. Low-dimensional projection is performed using a self-attention mechanism to efficiently capture key structural feature relationships in cross-domain alignment: Project sequence features onto a query ( ),key( ),value( ) matrix, where and Represents the feature set at all time steps. Indicates the current location where the context representation needs to be computed. Then, a low-dimensional projection is performed:

[0052]

[0053] in and ( (For the projection dimension). The learnable projection matrix acts as a temporal structure filter, enabling the model to highlight key structural patterns reflecting source-target domain differences while suppressing sequence fluctuations unrelated to source-target domain distinction; and reducing computational complexity from Down to It supports efficient modeling of long sequences.

[0054] Attention-based input-output residuals are processed by a multilayer perceptron (MLP) for binary classification to determine the source of samples. The domain alignment objective is formalized as:

[0055] in, The parameters of the domain discriminator.

[0056] Step S4: Perform self-supervised training on target domain samples based on pseudo-labels: To ensure in the unlabeled target domain The learned features exhibit good discriminative power. A self-supervised strategy based on pseudo-labels is employed, and its stability is enhanced by updating class centers using momentum. Class centers are updated using an exponential moving average (EMA) method, which maintains a relatively smooth class representation throughout training iterations. Specifically, a class center memory is maintained. Each of them Indicate category The normalized classification feature centers serve as reliable anchors for generating pseudo-labels, contributing to stable target domain feature learning.

[0057] Extracting classification features from target domain samples Through classifier Obtain the category probability Sample predicted labels With confidence level Defined as:

[0058] To reduce the impact of noise prediction, a temperature hyperparameter is introduced. Regulated softmax weights :

[0059] in For batch size. Calculate the weighted batch center within the predicted category. :

[0060] in, For indicator functions, To prevent small constants from being divided by zero, momentum is updated using an exponential moving average (EMA) and normalized using L2.

[0061] Then, pseudo-labels were matched using cosine similarity. .

[0062] Select those with confidence levels higher than the threshold. The samples used in training: The pseudo-label loss is defined as:

[0063] Reversal coefficient .in Indicates the normalized training progress. The growth rate is controlled. This scheduling mechanism prioritizes classification in the early stages of training and enhances domain alignment in the later stages.

[0064] Feature extractor Minimize simultaneously And maximize Learning domain-invariant features; domain discriminator minimize Improve discrimination ability. The comprehensive loss function is expressed as:

[0065] in, This is the loss balance coefficient.

[0066] Step S6: Obtain the communication signal to be identified, calculate its log-likelihood ratio sequence and input it into the trained channel coding type identification model, and output the channel coding type identification result.

[0067] During the inference phase, the trained feature extractor and classifier are directly applied to the target domain data without the need for a domain discriminator or memory. For the target domain test samples... Feature extractor First, extract domain-adaptive features, then classify. Output the final prediction:

[0068] in, This represents the softmax activation function. The predicted encoding type label.

[0069] Based on the same technical concept, embodiments of the present invention also provide an electronic device that can implement the unsupervised domain adaptive channel coding type identification method flow provided in the above embodiments of the present invention. In one embodiment, the electronic device can be a server, a terminal device, or other electronic devices. Figure 3 As shown, the electronic device may include: At least one processor and a memory connected to the at least one processor. In this embodiment of the invention, the specific connection medium between the processor and the memory is not limited. Figure 3 The example used is the connection between the processor and memory via a bus. The bus... Figure 3 The connections between other components are indicated by thick lines and are for illustrative purposes only, not as limiting information. Buses can be categorized into address buses, data buses, control buses, etc., but for ease of representation, [the specific bus type is not shown here]. Figure 3 The processor is represented by a single thick line, but this does not imply that there is only one bus or one type of bus. Alternatively, a processor can also be called a controller; there are no restrictions on the name.

[0070] In this embodiment of the invention, the memory stores instructions executable by at least one processor. By executing the instructions stored in the memory, the at least one processor can perform the unsupervised domain adaptive channel coding type identification method discussed above. The processor can implement... Figure 3 The functions of each module in the device shown.

[0071] The processor is the control center of the device. It can connect to various parts of the control device through various interfaces and lines. By running or executing instructions stored in memory and calling data stored in memory, it can monitor the device's various functions and process data, thereby enabling overall monitoring of the device.

[0072] In an alternative design, the processor may include one or more processing units. The processor may integrate an application processor and a modem processor, wherein the application processor primarily handles the operating system, user interface, and applications, while the modem processor primarily handles wireless communication. It is understood that the modem processor may also not be integrated into the processor. In some embodiments, the processor and memory may be implemented on the same chip; in some embodiments, they may also be implemented separately on separate chips.

[0073] The processor can be a general-purpose processor, such as a CPU, digital signal processor, application-specific integrated circuit, field-programmable gate array or other programmable logic device, discrete gate or transistor logic device, or discrete hardware component, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the unsupervised domain adaptive channel coding type identification method disclosed in the embodiments of this invention can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules within the processor.

[0074] Memory, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules. Memory can include at least one type of storage medium, such as flash memory, hard disk, multimedia cards, card-type memory, random access memory (RAM), static random access memory (SRAM), programmable read-only memory (PROM), read-only memory (ROM), and electrically erasable programmable read-only memory (EPROM). Only memory (EEPROM), magnetic storage, magnetic disks, optical disks, etc. A memory is any other medium capable of carrying or storing desired program code in the form of instructions or data structures, and accessible by a computer, but is not limited thereto. The memory in embodiments of this invention can also be a circuit or any other device capable of performing storage functions for storing program instructions and / or data.

[0075] By designing and programming the processor, the code corresponding to the unsupervised domain adaptive channel coding type identification method described in the foregoing embodiments can be embedded into the chip, enabling the chip to execute the steps of the method described in the foregoing embodiments during runtime. How to design and program the processor is a technique well-known to those skilled in the art and will not be elaborated upon here.

[0076] Based on the same inventive concept, embodiments of the present invention also provide a storage medium storing computer instructions that, when executed on a computer, cause the computer to perform an unsupervised domain adaptive channel coding type identification method as described above.

[0077] In some alternative embodiments, the present invention also provides aspects of an unsupervised domain adaptive channel coding type identification method that can also be implemented as a program product comprising program code that, when the program product is run on a device, causes the control device to perform the steps in an unsupervised domain adaptive channel coding type identification method according to various exemplary embodiments of the present invention as described above.

[0078] It should be noted that although several units or sub-units of the apparatus have been mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of the invention, the features and functions of two or more units described above can be embodied in one unit. Conversely, the features and functions of one unit described above can be further divided and embodied by multiple units. Furthermore, although the operation of the method of the invention is described in a specific order in the drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0079] 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 be implemented in one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs) containing computer-usable program code. The form of a computer program product implemented on ROM, optical memory, etc.

[0080] 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 server, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0081] Program code for performing the operations of this invention can be written using any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0082] In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0083] 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.

[0084] 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.

[0085] Example 2 To further verify and illustrate the effectiveness and feasibility of the unsupervised domain adaptive channel coding type identification method described in Embodiment 1, this embodiment provides experimental environment settings, hyperparameter configurations, and a benchmark model for comparison in specific applications. Those skilled in the art should understand that the specific experimental parameters below are merely a preferred embodiment of the present invention, intended to meet the requirements of sufficient disclosure under the Patent Law, and do not constitute a limitation on the scope of protection of the present invention.

[0086] In this embodiment, following the unsupervised domain adaptation method, labeled data from the source domain and unlabeled data from the target domain are used during the training phase. All experiments were conducted on an environment equipped with an NVIDIA GeForce RTX 3090 GPU and accelerated using PyTorch 1.7.1 and CUDA 11.0. Each dataset was randomly divided into a 60% training set, a 20% validation set, and a 20% test set. The hyperparameter settings used in the experiments are shown in Table 1.

[0087] Table 1 Experimental hyperparameter settings

[0088] In this embodiment, eight different methods are compared: (1) Source Only: Training and evaluation are performed using only the source domain dataset. Typical CNN is used as the backbone network to measure the best performance the model can achieve when the training distribution is exactly the same as the test distribution.

[0089] (2) Target Only: Training is performed on the source domain and testing is performed on the target domain, without any domain adaptation processing. This baseline also uses Typical CNN as the backbone network to quantify the performance degradation caused by domain shift and as a performance lower bound when evaluating the effectiveness of various adaptation methods.

[0090] (3) Domain Adversarial Neural Network, a widely used method, learns domain-invariant feature representations through adversarial training.

[0091] (4) Conditional Domain Adversarial Network improves domain alignment by simultaneously conditionalizing the input of the domain discriminator to the feature representation and the output of the classifier.

[0092] (5) Batch Spectral Penalization enhances domain adaptability by penalizing the maximum singular value of the feature covariance matrix of each batch.

[0093] (6) Minimum Class Confusion reduces class confusion by reducing off-diagonal elements in the target domain prediction class probability matrix, thereby improving the model's ability to identify the target domain.

[0094] (7) Batch Nuclear-norm Maximization improves model learning performance when labels are scarce by maximizing the nuclear norm of the batch output matrix, and encourages prediction results to maintain both high discriminativeness and high diversity.

[0095] (8) Proposed is the method proposed in this patent.

[0096] The experimental results are analyzed as follows: As shown in Table 2, there is a significant performance gap between the Source Only and Target Only baselines, mainly due to the distribution mismatch caused by domain shift. Specifically, under low signal-to-noise ratio (SNR) conditions, the average classification accuracy decreases by approximately 24%, from 54.39% in the source domain to 30.56% in the target domain; even at high SNR levels, classification performance still shows a decrease of nearly 11% (from 93.54% to 82.64%). These differences are mainly attributed to the inherent differences between the two domains, resulting in inconsistent feature distributions between the training and test domains. Consequently, the classification features extracted by the model from the target domain are difficult to align with the features learned from the source domain, ultimately weakening class separability and causing instability of the classification boundary in the target domain.

[0097] Furthermore, adversarial adaptive techniques have shown significant effectiveness in mitigating domain offset. For example... Figure 2 As shown, DANN improves the average target domain recognition accuracy by nearly 8% (from 30.56% to 38.65%) under low SNR conditions. This improvement is attributed to the alignment of source and target domain feature distributions achieved through adversarial training, enabling the model to learn more generalizable domain-invariant features, thereby sharing the same classifier and improving cross-domain recognition performance.

[0098] Among all adversarial adaptive methods, the proposed method exhibits the best classification performance. Its average classification accuracy reaches 64.31%, surpassing the strongest contrastive adaptive method, MCC, at 63.39%. Compared to the Target Only baseline, the proposed method achieves an average accuracy gain of nearly 8%, indicating its greater effectiveness in reducing domain variance and learning domain-invariant features. This performance improvement primarily stems from two key mechanisms: first, sequence-level alignment captures domain-specific structural dependencies, thus enabling more effective learning of domain-invariant representations; second, the high-confidence pseudo-label strategy provides a more reliable supervisory signal for the target domain.

[0099] In addition to classification accuracy, the effectiveness of the proposed method can be further verified by Cohen's Kappa coefficient. The Kappa values ​​in Table 2 are calculated by averaging the confusion matrix across all SNRs, reflecting the model's ability to maintain consistent classification performance under different noise conditions. Under this metric, the proposed method achieves a Kappa value of 0.5538, higher than the adaptive baseline MCC (0.5423), Target Only (0.4575), and other comparative methods. This indicates that the performance improvement of the proposed method in the target domain does not rely on predictions biased towards the dominant class, but rather achieves balanced recognition across all classes, demonstrating more stable and consistent classification capabilities.

[0100] Table 2 Recognition Performance Analysis Results

[0101] Explanation: (1) Low SNR: SNR ≤ 4 dB; High SNR: SNR ≥ 6 dB. (2) Acc. (%): Average classification accuracy within the selected SNR interval. (3) Max. (%): Maximum classification accuracy within the selected SNR interval. (4) Coeff: Cohen's Kappa coefficient. (5) std: Standard deviation of the results obtained from three independent experiments.

[0102] The embodiments described above merely illustrate specific implementation methods of this application, and while the descriptions are detailed and specific, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the technical solution of this application, and these modifications and improvements all fall within the scope of protection of this application.

[0103] This background section is provided to generally present the context of the invention. The work of the currently named inventors, the work to the extent described in this background section, and aspects described in this section that did not constitute prior art at the time of application are neither expressly nor impliedly acknowledged as prior art to the invention.

Claims

1. An unsupervised domain adaptive channel coding type recognition method, characterized in that, include: Step S1: Obtain the source domain dataset and the target domain dataset; the source domain dataset contains a log-likelihood ratio sequence with channel coding type labels generated by simulation, and the target domain dataset contains an unlabeled log-likelihood ratio sequence extracted from signals received from a real wireless environment; Step S2: Construct a channel coding type identification model, which includes a feature extractor, a classifier, and a domain discriminator; pre-train the feature extractor and classifier using the source domain dataset; Step S3: Perform sequence-level domain alignment on the model based on adversarial training: The log-likelihood ratio sequences from the source domain dataset and the target domain dataset are input into a pre-trained feature extractor to extract sequence features. The sequence features are input into the domain discriminator, and the sequence features are projected into a query matrix, a key matrix, and a value matrix through a self-attention mechanism. Low-dimensional projection is performed on the key matrix and the value matrix to capture the structural dependencies of the sequence. Based on the attention output after low-dimensional projection, it is determined whether the sample comes from the source domain or the target domain, and the domain discrimination loss is calculated. Step S4: Perform self-supervised training on target domain samples based on pseudo-labels: Extract the classification features of the target domain samples, and output the predicted class probability through the classifier; Maintain a category center for each channel coding type, and update the momentum of the category center using an exponential moving average mechanism; Calculate the cosine similarity between the classification features of the target domain samples and the class centers to generate pseudo-labels for the target domain samples; use a confidence threshold to filter out target domain samples with high confidence and calculate the pseudo-label loss; Step S5: Combining the classification loss of the source domain dataset, the domain discrimination loss, and the pseudo-label loss, the feature extractor, the classifier, and the domain discriminator are jointly optimized through a gradient inversion layer to obtain the trained channel coding type recognition model; Step S6: Obtain the communication signal to be identified, calculate its log-likelihood ratio sequence and input it into the trained channel coding type identification model, and output the channel coding type identification result.

2. The unsupervised domain adaptive channel coding type identification method of claim 1, wherein, Obtain the log-likelihood ratio sequences from the source and target domain datasets, including: The receiving end receives the baseband signal modulated by orthogonal frequency division multiplexing; Symbol timing synchronization is performed using the preamble structure in the baseband signal, and carrier frequency offset correction is performed using a two-step carrier frequency offset estimation method. Channel estimation is performed in the frequency domain using the long training sequence in the preamble structure to obtain the channel frequency response on the subcarrier; The baseband signal is equalized according to the channel frequency response, the equalized symbols are demodulated into corresponding bit sequences, and the log-likelihood ratio of each bit in the bit sequence is calculated to form a log-likelihood ratio sequence.

3. The unsupervised domain adaptive channel coding type identification method of claim 2, wherein, The specific formula for calculating the log-likelihood ratio is as follows: in: representing the equalized symbols demodulated to obtain log likelihood ratios of the bits represents the posterior probability under the condition that the equalized symbol is 4. The unsupervised domain adaptive channel coding type identification method of claim 1, wherein, The step of projecting the sequence features into a query matrix, a key matrix, and a value matrix using a self-attention mechanism, and performing low-dimensional projection on the key matrix and value matrix, includes: Projecting the sequence features as a query matrix , a key matrix and a value matrix respectively; The key matrix and value matrix are subjected to a low-dimensional projection using the following formula: in: The projected key matrix; This is the projected value matrix; and These are projection matrices used to map the key matrix and value matrix to a preset domain discriminant subspace, respectively. Calculate the attention output after low-dimensional projection: in: The feature dimension of the sequence feature; This represents the normalized exponential function; The attention output is input into the residual multilayer perceptron for binary classification to determine whether the sample comes from the source domain or the target domain.

5. The unsupervised domain adaptive channel coding type identification method according to claim 1, characterized in that, The step of updating the momentum of the category centers using an exponential moving average mechanism includes: The category corresponding to the maximum predicted category probability of the target domain sample is taken as the predicted label. ; Based on the predicted class probabilities of the target domain samples and the preset temperature hyperparameters, the weighting weights are calculated. ; According to the weighted weights Calculate the weighted batch center within the predicted category. : in: Batch size; For the first Classification features of each target domain sample; For predicting labels; For category number; For indicator functions; To prevent division by zero of constants; The category center is updated using a momentum update mechanism: in: For category The category center; The momentum coefficient; Perform L2 normalization on the updated category centers: 。 6. The unsupervised domain adaptive channel coding type identification method according to claim 5, characterized in that, Calculate the cosine similarity between the classification features of the target domain samples and the class centers, generate pseudo-labels for the target domain samples, and calculate the pseudo-label loss, including: Classification features of target domain samples With normalized category centers Perform cosine similarity matching to generate pseudo-labels for the current batch. : The maximum predicted category probability is greater than the pre-configured confidence threshold. The samples are used as a high-confidence sample set. Participate in training; Calculate the average cross-entropy on the high-confidence sample set as the pseudo-label loss. : in: Represents the cross-entropy loss function; This indicates that the classifier classifies features. The output.

7. The unsupervised domain adaptive channel coding type identification method according to claim 1, characterized in that, When jointly optimizing the feature extractor, the classifier, and the domain discriminator using a gradient inversion layer, the gradient inversion layer performs an identity mapping during forward propagation and flips the gradient sign during back propagation. The flipping formula is as follows: in: The input gradient; The gradient after flipping; Indicates the normalized training progress; A coefficient used to control the growth rate; The inversion coefficient is dynamically adjusted according to the training progress.

8. The unsupervised domain adaptive channel coding type identification method according to claim 1, characterized in that, The comprehensive loss function, which combines the classification loss of the source domain dataset, the domain discrimination loss, and the pseudo-label loss for joint optimization, is expressed as: in: , , These represent the parameters of the feature extractor, the classifier, and the domain discriminator, respectively. The classification loss is the loss for the source domain dataset; For the domain, determine the loss; The pseudo-label loss; This is the tradeoff coefficient for the false label loss.

9. An electronic device, characterized in that, include: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores instructions executable by the at least one processor, which executes the instructions stored in the memory to perform the method as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store instructions that, when executed, cause the method as described in any one of claims 1-8 to be implemented.

Images