Radiation source signal recognition method with online fast domain adaptation

Abstract

The application discloses an online fast domain adaptation radiation source signal identification method, comprising the following steps: pre-training a radiation source signal identification model based on a labeled source domain sample set and a small amount of unlabeled initial target domain sample set; obtaining dynamic inflow target domain sample data, and using an updated feature extractor and task classifier to perform reasoning, obtain high-dimensional feature vectors, category prediction probability distribution and pseudo labels; based on the target domain sample data and the pseudo labels, a first loss function is constructed to dynamically update the parameters of the feature extractor; the hinge loss of the target domain sample data is calculated according to the pseudo labels, and when the hinge loss is greater than zero, the parameters of the task classifier are updated online based on the kernel function; and the category prediction result of the current batch of target domain sample data is output. In this way, the continuous learning of the dynamic target domain environment and the stable and high-precision identification of the radiation source signal are realized.

Description

Technical Field

[0001] This invention relates to the field of communication technology, and in particular to an online, fast-domain-adaptive method for identifying radiation source signals. Background Technology

[0002] Radiation source identification is a key technology in the fields of electronic countermeasures, spectrum management, and wireless communication security. Its purpose is to use the signals emitted by radiation sources to achieve unique identification at the device level, thereby ensuring effective monitoring and management of communication behavior in complex electromagnetic environments.

[0003] In real-world applications, communication radiation source signals are characterized by high dimensionality, strong transients, susceptibility to interference, and difficulty in labeling. Furthermore, the electromagnetic environment is open, dynamic, and variable, with the distribution of target signals often changing with time, space, and interference conditions. This leads to a decline in the discriminative performance and insufficient generalization ability of traditional identification methods under cross-environmental conditions. Moreover, in real-world scenarios, target domain data often arrives in streaming form, or contains only a small number of labeled or even unlabeled samples, further exacerbating the difficulty of continuous model updates and adaptive learning.

[0004] To address the challenges of streaming data and continuous adaptation, some studies have explored incremental adaptation methods based on online learning. However, most online learning algorithms rely on data labels, making them difficult to update effectively when the target domain labels are scarce; and they lack robustness to sudden distribution shifts. Summary of the Invention

[0005] The purpose of this application is to provide an online, fast, domain-adaptive radiation source signal identification method that can achieve high-precision, continuous, and robust identification of individual radiation source signals in complex and dynamic target domain environments.

[0006] To achieve the above objectives: In a first aspect, embodiments of this application provide an online, fast domain-adaptive radiation source signal identification method. The method includes: pre-training a radiation source signal identification model based on a labeled source domain sample set and a small number of unlabeled initial target domain sample sets; acquiring dynamically incoming target domain sample data and using the updated feature extractor and task classifier in the radiation source signal identification model for inference to obtain a high-dimensional feature vector, category prediction probability distribution, and pseudo-label for each target domain sample; constructing a first loss function based on the target domain sample data and its pseudo-labels, and dynamically updating the parameters of the feature extractor, wherein the first loss function includes entropy loss and structural alignment loss; calculating the hinge loss of the current batch of target domain sample data based on the pseudo-labels, and updating the parameters of the task classifier online based on a kernel function when the hinge loss is greater than zero; and outputting the category prediction result of the current batch of target domain sample data based on the updated radiation source signal identification model.

[0007] In one embodiment, the radiation source signal recognition model includes: a feature extractor, a task classifier, and a domain discriminator; the pre-training of the radiation source signal recognition model includes: introducing a gradient inversion layer between the feature extractor and the domain discriminator, and training the model by jointly minimizing the task classification loss, the domain alignment loss, and the adversarial loss; wherein the domain alignment loss includes at least using multi-kernel maximum mean difference loss to align the feature distributions of the source domain and the target domain. In one embodiment, the feature extractor is a multi-scale feature alignment network, comprising at least two stacked modules, each of which includes a multi-scale convolutional module, a dynamic adaptive attention module, a one-dimensional convolutional layer, group normalization, and a ReLU activation function. In one embodiment, the multi-scale convolution module includes: multiple parallel convolution paths with different dilation rates and kernel sizes; the dynamic adaptive attention module is used to calculate channel attention weights based on input features and to perform channel weighting on the input features. In one embodiment, dynamically incoming target domain sample data is acquired, and inference is performed using the feature extractor and task classifier in the radiation source signal identification model to obtain a high-dimensional feature vector, category prediction probability distribution, and pseudo-label for each target domain sample. This includes: acquiring dynamically incoming target domain sample data in batches; inputting the target domain sample data of the current batch into the feature extractor to obtain a high-dimensional feature vector for each target domain sample in the target domain sample data; inputting the high-dimensional feature vector into the task classifier to obtain a category prediction probability distribution for the target domain samples; and assigning pseudo-labels to the target domain samples based on the category prediction probability distribution and the maximum probability criterion.

[0008] In one embodiment, constructing the first loss function includes: calculating entropy loss based on the predicted probability distribution of the target domain samples in the current batch; calculating temporary prototypes of each category of the target domain based on the pseudo-labels of the target domain samples in the current batch; and combining historical target domain prototypes to obtain updated target domain category prototypes to construct a target domain category structure matrix; calculating the difference between the target domain category structure matrix and a preset source domain category structure matrix to obtain a structure alignment loss; wherein the entropy loss and the structure alignment loss are used to optimize the parameters of the feature extractor.

[0009] In one embodiment, the hinge loss of the target domain sample data in the current batch is calculated based on the pseudo-label, and the parameters of the task classifier are updated online based on the kernel function when the hinge loss is greater than zero. This includes: incrementally updating the parameters of the task classifier based on the kernel function when the hinge loss is greater than zero; wherein, the incremental update of the parameters of the task classifier based on the kernel function includes: adding the current classifier function to a weighting term determined by the kernel function, the pseudo-label, and the hinge loss to correct the discriminant function weight vector.

[0010] In one embodiment, the class prediction result of the target domain sample data of the current batch is output based on the updated radiation source signal identification model, and then the updated feature extractor parameters and classifier parameters are saved for processing subsequent batches of target domain sample data. Secondly, embodiments of this application provide a computing device, including: a processor and a memory storing a computer program, wherein when the processor runs the computer program, the steps of the above-described method are implemented.

[0011] Thirdly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method.

[0012] The online, fast domain-adaptive radiation source signal identification method provided in this application includes: pre-training a radiation source signal identification model based on a labeled source domain sample set and a small number of unlabeled initial target domain sample sets; acquiring dynamically flowing target domain sample data and using an updated feature extractor and task classifier for inference to obtain the high-dimensional feature vector, category prediction probability distribution, and pseudo-label for each target domain sample; constructing a first loss function based on the target domain sample data and its pseudo-labels to dynamically update the parameters of the feature extractor; calculating the hinge loss of the current batch of target domain sample data based on the pseudo-labels, and updating the parameters of the task classifier online based on a kernel function when the hinge loss is greater than zero; and finally, outputting the category prediction result of the current batch of target domain sample data using the updated radiation source signal identification model. This achieves continuous learning of the dynamic target domain environment and stable identification of radiation source signals. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating an online, fast, domain-adaptive radiation source signal identification method provided in an embodiment of the present invention.

[0014] Figure 2 This invention relates to a multi-scale feature alignment network used in the feature extractor of an online fast domain-adaptive radiation source signal identification model, as provided in an embodiment of the present invention.

[0015] Figure 3 This is a comparative experimental result diagram of an embodiment of the present invention and a prior art solution.

[0016] Figure 4 This is a schematic diagram of the structure of a computing device provided in an embodiment of the present invention. Detailed Implementation

[0017] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. In the following description, when referring to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements.

[0018] It should be noted that, in this document, 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 that element. Furthermore, components, features, and elements with the same names in different embodiments of this application may have the same meaning or different meanings, the specific meaning of which must be determined by its interpretation in that specific embodiment or further in conjunction with the context of that specific embodiment.

[0019] It should be understood that although the terms first, second, third, etc., may be used herein to describe various information, such information should not be limited to these terms. These terms are used only to distinguish information of the same type from one another. For example, without departing from the scope of this document, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if," as used herein, can be interpreted as "when," "when," or "in response to determination." Furthermore, as used herein, the singular forms "a," "an," and "the" are intended to also include the plural forms unless the context indicates otherwise. It should be further understood that the terms "comprising," "including," indicate the presence of the stated feature, step, operation, element, component, item, kind, and / or group, but do not exclude the presence, occurrence, or addition of one or more other features, steps, operations, elements, components, items, kinds, and / or groups. The terms "or" and "and / or" as used herein are to be interpreted as inclusive, or mean any one or any combination thereof. Therefore, "A, B, or C" or "A, B, and / or C" means "any one of the following: A; B; C; A and B; A and C; B and C; A, B, and C". Exceptions to this definition will only occur if the combination of elements, functions, steps, or operations is inherently mutually exclusive in some way.

[0020] It should be understood that although the steps in the flowcharts of this application's embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0021] It should be noted that step designations such as S1 and S2 are used in this document for the purpose of more clearly and concisely describing the corresponding content, and do not constitute a substantial limitation on the order. In specific implementation, those skilled in the art may execute S2 first and then S1, etc., but these should all be within the protection scope of this application.

[0022] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0023] In the following description, the use of suffixes such as "module," "part," or "unit" to denote elements is solely for the purpose of illustrative purposes and has no specific meaning in itself. Therefore, "module," "part," or "unit" may be used interchangeably.

[0024] See Figure 1 This application provides an online, fast domain-adaptive radiation source signal identification method, which can be executed by a radiation source signal identification device / computing device provided in this application. This device / device can be implemented in software and / or hardware. Specifically, the online, fast domain-adaptive radiation source signal identification method provided in this embodiment includes the following steps: Step S1: Based on the labeled source domain sample set and a small number of unlabeled initial target domain sample sets, the radiation source signal identification model is pre-trained.

[0025] Specifically, a large amount of labeled multi-type radiation source signal data is obtained in advance from historical data to construct a source domain sample set; and a small amount of unlabeled radiation source signal data is obtained in advance from the target domain to construct an initial target domain sample set.

[0026] In this context, the source domain refers to the dataset or domain in which the radiation source signal identification model / model is pre-trained. It typically contains a large number of labeled samples used for training and building the initial model. The target domain, on the other hand, refers to the new dataset or domain to which the model will be applied. In the target domain, labeled samples are usually fewer (unlabeled), and the model needs to leverage the knowledge and features learned from the source domain through transfer learning to achieve good performance on the target task. In domain transfer, the source and target domains have different data distributions.

[0027] The labels are used to indicate the category of each radiation source signal in the sample set. For example, artificial radiation source signals (communication equipment: mobile phone base station signals, satellite communication signals, wireless network signals; medical equipment: X-ray machine signals, MRI machine signals), etc. Optionally, the radiation source signal identification model includes: a feature extractor, a task classifier, and a domain discriminator.

[0028] The radiation source signal identification model includes a feature extractor, a task classifier, and a domain discriminator. The feature extractor extracts the feature information of the radiation source signal; after extracting the features, these features are input into the task classifier and discriminator.

[0029] In one embodiment, the feature extractor is a multi-scale feature alignment network, comprising at least two stacked modules, each of which includes a multi-scale convolutional module, a dynamic adaptive attention module, a one-dimensional convolutional layer, group normalization, and a ReLU activation function.

[0030] In one embodiment, the multi-scale convolution module includes: multiple parallel convolution paths with different dilation rates and kernel sizes; and a dynamic adaptive attention module for calculating channel attention weights based on input features and performing channel weighting on the input features.

[0031] like Figure 2 As shown, the feature extractor is a multi-scale feature alignment network, consisting of two stacked modules. Each module includes a multi-scale convolutional module (MDCM), a dynamic adaptive attention module (DAAM), a one-dimensional convolutional layer (Conv1D), a group normalization (GroupNorm), and a ReLU activation function.

[0032] In one embodiment, the pre-training of the radiation source signal identification model includes: introducing a gradient inversion layer between the feature extractor and the domain discriminator, and training the model by jointly minimizing the task classification loss, domain alignment loss, and adversarial loss; wherein the domain alignment loss includes at least using multi-kernel maximum mean difference loss to align the feature distributions of the source domain and the target domain.

[0033] For example, the specific training method for the radiation source signal identification model is as follows: For the given radiation source data (including the source domain sample set and the initial target domain sample set) input ;in, It is the data length. This refers to the number of channels in the input features. Multi-scale features of the radiation source data are extracted from six parallel convolutional paths with different dilation rates and kernel sizes. The outputs of each convolutional path are normalized and activated by GELU before being fused to form a preliminary multi-scale feature map. ;in, This represents the number of convolutional paths. Then, average pooling along the channel axis is performed to extract the average feature of each channel; subsequently, max pooling and average pooling are applied to generate the output vector. and Then, the input is fed into a shared-parameter multilayer perceptron (MLP) to establish the intrinsic relationships between channels. After ReLU and Sigmoid activation, multi-scale dynamic channel attention weights are generated. It can be calculated as follows:

[0034] in, It is the weight vector of the MLP layer. The dimensionality reduction coefficient. It is the ReLU activation function. It is an activation function that converts attention scores into values ​​between 0 and 1.

[0035] The multi-scale dynamic channel attention weights obtained in the above steps are combined with the input multi-scale feature map. Perform channel weighting operation This enables the multi-scale feature alignment network to adaptively adjust the feature response based on the importance of different channels. Specifically, it is calculated using the following formula:

[0036] The weighted feature maps are then subjected to Conv1D convolution to increase feature dimensionality and downsample the space, followed by GroupNorm normalization and ReLU activation. These two stacked modules ultimately output a radiation source feature representation (i.e., a high-dimensional feature vector) for subsequent classification and feature alignment tasks.

[0037] In this embodiment, to adapt to the extraction requirements of features at different scales, a multi-scale convolutional module consisting of six parallel convolutional paths is proposed. Each path focuses on capturing feature information within different receptive fields through different combinations of dilation rates and kernel sizes. The specific configurations are as follows: Path 1 uses a dilation rate of 1 and a 3×3 convolutional kernel; Path 2 uses a dilation rate of 1 and a 5×5 convolutional kernel; Path 3 uses a dilation rate of 2 and a 3×3 convolutional kernel; Path 4 uses a dilation rate of 2 and a 5×5 convolutional kernel; Path 5 uses a dilation rate of 3 and a 3×3 convolutional kernel; and Path 6 uses a dilation rate of 3 and a 5×5 convolutional kernel. Through this approach, on the one hand, local and global features of the signal under different receptive fields can be obtained through different convolutional kernel sizes and dilation rate structures, enhancing the diversity of feature representation; on the other hand, a lightweight attention mechanism is introduced into each convolutional path to adaptively adjust the feature response weights of each path's convolutional kernel with minimal network parameter overhead. Ultimately, this enables the feature extractor to robustly encode multi-scale features of the radiation source signal, improving the generalization ability of cross-domain recognition.

[0038] It should be noted that multiple loss functions are introduced during the training process to jointly train the network, thereby simultaneously ensuring task discriminativity, cross-domain alignment, and feature domain invariance. Specifically, the total loss function of the network consists of three parts: task classification loss, domain alignment loss, and adversarial loss, and its expression is as follows:

[0039] Furthermore, the task classification loss is used to ensure that source domain samples are correctly classified on the task classifier, thereby guaranteeing the discriminative power of source domain features; its source domain task classification loss is:

[0040] in, Indicates the number of samples in the source domain; For the true labels of the source domain samples, This is the category probability predicted by the radiation source signal identification model.

[0041] Furthermore, the domain alignment loss is used to minimize the difference in feature distribution between the source and target domains, and its calculation formula is as follows:

[0042] in, For feature extractor mapping, and These are samples from the source and target domains, respectively. and The number of samples in the source domain and the number of samples in the target domain, respectively. The square norm operation is used to calculate the squared distance between the mean features of the source and target domains in the reproducing kernel Hilbert space. Furthermore, adversarial loss is used to enhance the domain invariance of features by adding gradient inversion layers to the feature extractor and domain discriminator, with the following formula:

[0043] in, For sample domain labels, Predict probabilities for the domain discriminator. and These represent the number of samples in the source domain and the target domain, respectively.

[0044] Through the above methods, during training, the source and target domains are aligned in the feature space by combining multi-kernel maximum mean difference and adversarial loss. Simultaneously, task classification loss ensures the discriminative training process of source domain samples, thereby obtaining a basic feature encoding model with domain invariance and class discrimination capabilities. This provides a stable foundation of model parameters and feature space for subsequent dynamic target domain streaming recognition and gradual adaptation.

[0045] Step S2: Obtain the dynamically flowing target domain sample data, and use the updated feature extractor and task classifier in the radiation source signal recognition model to perform inference, and obtain the high-dimensional feature vector, category prediction probability distribution and pseudo-label of each target domain sample.

[0046] Among them, the updated feature extractor and task classifier in the radiation source signal recognition model are a general feature extractor and a general task classifier obtained after pre-training the radiation source signal recognition model based on a labeled source domain sample set and a small number of unlabeled initial target domain sample sets.

[0047] In one embodiment, dynamically incoming target domain sample data is acquired, and inference is performed using the feature extractor and task classifier in the radiation source signal identification model to obtain a high-dimensional feature vector, category prediction probability distribution, and pseudo-label for each target domain sample. This includes: acquiring dynamically incoming target domain sample data in batches; inputting the target domain sample data of the current batch into the feature extractor to obtain a high-dimensional feature vector for each target domain sample in the target domain sample data; inputting the high-dimensional feature vector into the task classifier to obtain a category prediction probability distribution for the target domain samples; and assigning pseudo-labels to the target domain samples based on the category prediction probability distribution and the maximum probability criterion.

[0048] For example, for the same batch of target domain sample data that flows in dynamically, the cross-domain task classifier and feature extractor in the radiation source signal recognition model obtained after pre-training in the above steps are used to perform forward inference on each target domain sample in the batch.

[0049] Specifically, the target domain sample data first needs to be input into a feature extractor to extract its deep features in a multi-scale space:

[0050] in, For batches flowing in during the online phase The One target domain sample, For a well-trained feature extractor, The extracted feature vector (i.e., high-dimensional feature vector).

[0051] Subsequently, the feature vector obtained by the feature extractor is fed into the task classifier to calculate the predicted probability of each target domain sample in the batch belonging to each candidate category:

[0052] in, For a well-trained task classifier, For the sample The predicted probability vector.

[0053] Finally, based on the predicted probability distribution (e.g., 80% for category A, 50% for category B, etc.), the maximum probability criterion is used to assign a corresponding pseudo-label to each target domain sample, that is, the category with the highest probability is regarded as the category inference result (i.e., pseudo-label) of the target domain sample:

[0054] in, For the sample pseudo-tags The sample belongs to the category The predicted probability.

[0055] Using the above method, pseudo-supervisory information of target domain samples can be automatically constructed without manual annotation, thereby providing necessary data support for subsequent online adaptive updates.

[0056] Step S3: Based on the target domain sample data and its pseudo-labels, construct the first loss function and dynamically update the parameters of the feature extractor. The first loss function includes entropy loss and structure alignment loss.

[0057] In one embodiment, constructing the first loss function includes: calculating entropy loss based on the predicted probability distribution of the target domain samples in the current batch; calculating temporary prototypes of each category of the target domain based on the pseudo-labels of the target domain samples in the current batch; and combining historical target domain prototypes to obtain updated target domain category prototypes to construct a target domain category structure matrix; calculating the difference between the target domain category structure matrix and a preset source domain category structure matrix to obtain a structure alignment loss; wherein the entropy loss and the structure alignment loss are used to optimize the parameters of the feature extractor.

[0058] For example, firstly, using the labeled sample information in the source domain sample set, a category prototype is constructed for each category of the radiation source signal to characterize the center of that category in the feature space. Let the source domain sample set be... Category number is The feature extraction function is Then the category Source domain prototype Defined as the mean of the feature vectors of all samples in this category: in, Indicates belonging to a category The source domain sample set. By calculating the prototypes of all categories of radiation source signals, the source domain prototype matrix can be obtained:

[0059] in, The feature dimension is used to characterize the geometric relationships between categories and constrain the target domain structure. A source domain category structure matrix is ​​introduced. :

[0060] in, Represents the first element in the source domain category structure matrix. Line number The elements of the column represent the source domain category. With category The similarity. It is the scaling hyperparameter of Gaussian kernel similarity, used to control the sensitivity of similarity to changes in the distance between prototypes. This represents an exponential function that converts distance into similarity. The smaller the distance, the closer the similarity is to 1; the larger the distance, the closer the similarity is to 0.

[0061] Since the dynamically flowing target domain sample data lacks label information, the radiation source signal identification model can use the parameters of the feature extractor and task classifier obtained from the last online update to identify each category in the current batch of target domain samples. Using pseudo-labeled samples, estimate the temporary target domain prototype. :

[0062] in, Indicates that the current batch belongs to the category A collection of pseudo-label samples, The number of samples in this set. For the output of the feature extractor The feature vector of the sample.

[0063] Subsequently, the temporary target domain prototype calculated in the current batch will be... Compared to the previous batch of historical prototypes Recursive fusion is used to update the target domain prototype, which smooths the update process, reduces the impact of single-batch pseudo-label noise, and makes online feature alignment and subsequent structure alignment more stable and reliable. The fusion is implemented using an exponential moving average method.

[0064] in, The attenuation coefficient controls the weight ratio between historical information (processed and identified target domain sample data) and the target domain sample data in the current batch.

[0065] After obtaining the target domain prototype for each category, the structure matrix between categories of the target domain is calculated to measure the similarity between prototypes of different categories and to provide constraints for structural alignment. The calculation formula is as follows:

[0066] in, Indicate category With category Similarity in the feature space. It is the scaling hyperparameter of Gaussian kernel similarity, used to control the sensitivity of similarity to changes in the distance between prototypes.

[0067] The above approach addresses the issues of dynamic data inflow into the target domain over time, continuous data distribution drift, and catastrophic forgetting that may occur during online learning. This implementation defines a structure preservation loss to constrain the geometric relationship between the target domain category prototype and the source domain category prototype. With each batch of target domain sample data arriving, the model minimizes this structure preservation loss to maintain the historical category discriminative structure, while allowing feature representations to dynamically adjust with new data, thereby improving the model's stability, discriminative ability, and online adaptive capability. The structure preservation loss formula is expressed as:

[0068] in, and The class relationship matrices for the source domain and the online inflow target domain batches are shown respectively. Let Frobenius be the squared norm of the matrix.

[0069] Meanwhile, for dynamically flowing target domain data batches To encourage the model to output predictions with higher confidence, an entropy minimization strategy is adopted. This makes the distribution of the obtained class predictions more explicit, keeps the target domain samples away from the decision boundary, and enhances the compactness of similar samples in the embedding space. For the dynamically flowing target domain sample data... Sample The entropy loss formula can be written as:

[0070] in, This indicates that the model predicts the sample belongs to a certain category. The probability, This indicates the total number of target domain sample data in the current batch.

[0071] Finally, by backpropagating the gradients of the structure preservation loss and entropy loss to the feature extractor and updating its parameters according to the online gradient descent rule, the feature extractor can adapt to changes in the distribution of sample data in the target domain in real time. By dynamically updating the feature extractor's parameters, it can dynamically adjust to capture newly acquired feature information in the target domain while maintaining consistency with the source domain structure, thus avoiding catastrophic forgetting. The update formula can be expressed as:

[0072] in, These are the parameters of the feature extractor in the model; The learning rate is used to control the step size of parameter updates; Indicates the target domain. Batch inflow number Represents the loss function For feature extractor parameters Find the gradient. This represents the overall loss of the target domain sample data in the current batch, including structural alignment loss and entropy loss.

[0073] Step S4: Calculate the hinge loss of the target domain sample data in the current batch based on the pseudo-labels, and update the parameters of the task classifier online based on the kernel function when the hinge loss is greater than zero.

[0074] In one embodiment, the hinge loss of the target domain sample data in the current batch is calculated based on the pseudo-label, and the parameters of the task classifier are updated online based on the kernel function when the hinge loss is greater than zero. This includes: incrementally updating the parameters of the task classifier based on the kernel function when the hinge loss is greater than zero; wherein, the incremental update of the parameters of the task classifier based on the kernel function includes: adding the current classifier function to a weighting term determined by the kernel function, the pseudo-label, and the hinge loss to correct the discriminant function weight vector.

[0075] For example, when a new batch of target domain sample data is acquired, the features of the target domain sample data are extracted using the current feature extractor, and the predicted probability of its category is generated by the trained task classifier, from which pseudo-labels of the target domain samples are obtained. Then, the hinge loss is calculated based on these pseudo-labels to measure the current classifier's discriminative ability and confidence in the new samples. If there are target domain samples with misclassification or insufficient confidence, the parameters of the task classifier are updated according to the hinge loss, realizing incremental adjustment of the classifier weights, thereby adapting to the new features of the target domain while retaining historical knowledge (i.e., the source domain sample set and the processed target domain sample data). By iteratively performing this process, the classifier state of each batch of target domain sample data can be updated and used for online recognition of the next batch of target domain sample data, thereby completing dynamic and continuous online adaptive learning.

[0076] The hinge loss is expressed as:

[0077] in, It indicates that the current batch is number 1. When samples from each target domain flow in, for all The set of discriminant function weights for each category, The weight vector corresponding to category s. This represents the feature vector of the target domain sample in the current input. Indicates the current sample Pseudo-tags. This represents the category weight vector corresponding to the pseudo-label. Indicates the feature extractor's performance on the sample The output feature vector. Indicates sample Discriminant score on pseudo-label category. This represents the most competitive error category, i.e., the category with the highest score excluding pseudo-labels. Indicates sample The discrimination score on the most competitive error category. The non-negative constraint on hinge loss ensures that a positive loss is generated only when the difference between the scores of the correct class and the most competitive incorrect class is less than 1, otherwise it is zero.

[0078] The linear discriminant function of the original classifier is expressed as the inner product of the weight vector and the feature vector. Used to calculate the category to which a sample belongs. The confidence level. This implementation introduces a kernel function. Then, this linear inner product is equivalently represented as a kernelized discriminant function. That is, the classification score is obtained by calculating the kernel function of historical samples and current samples.

[0079] In this way, the linear discriminant score is mapped to a high-dimensional feature space, realizing the nonlinear decision boundary. Therefore, when When using kernel functions The parameters of the task classifier are adjusted. The update rules for the task classifier are as follows:

[0080] in, This represents the task classifier's classifier for the current time step. The discriminant function, For the newly arrived target domain sample, The weighting coefficients are calculated based on pseudo-labels and hinge loss:

[0081] in, The Kronecker function indicates that the model only performs a function if the predicted class is... Equal to pseudo-label If the result is positive, the value is 1; otherwise, it is 0. Indicates if Equal to the most competitive error category If the result is positive, the value is 1; otherwise, it is 0. Indicates hinge loss. Represented as kernel function in samples Self-similarity on. This indicates the update step size, which determines the extent of weight adjustment in this update.

[0082] By utilizing the target domain samples of the current batch, the discriminant function of the task classifier can be incrementally corrected, and the information of the newly input target domain samples can be dynamically integrated into the model. This allows the task classifier to adjust the decision boundary in real time under the streaming input target domain sample data, and the task classifier can update parameters only when necessary, effectively preventing model oscillation caused by frequent updates, while ensuring that the model continuously adapts to and improves recognition performance under dynamically changing target domain sample data.

[0083] In one embodiment, the class prediction result of the target domain sample data of the current batch is output based on the updated radiation source signal identification model, and then the updated feature extractor parameters and classifier parameters are saved for processing subsequent batches of target domain sample data.

[0084] Specifically, the predicted recognition results of the current batch of target domain sample data are output, while the updated parameters of the feature extractor and task classifier are saved for online recognition of the next batch of target domain sample data, thus starting a new round of recognition cycle and realizing iterative optimization of the radiation source signal recognition model.

[0085] This invention achieves continuous learning and stable recognition of dynamic target domain environments through the coordinated "cross-domain feature construction" and "online domain adaptation" stages described above. In the offline classifier training stage (pre-training), a multi-scale feature extraction network based on dilated convolution is employed to mine fine-grained differences in radio frequency signals from different receptive fields. A dynamic channel attention mechanism is combined to highlight key discriminative features, thereby constructing a stable and transferable cross-domain feature foundation. Simultaneously, a multi-kernel maximum mean difference and adversarial learning strategy are used to achieve joint alignment of source and target domain features, effectively reducing inter-domain distribution shifts and improving cross-domain generalization ability. In the online stage, for practical application scenarios where target domain data is unlabeled, arrives streaming, and its distribution continuously changes over time, a pseudo-label-driven online unlabeled adaptation mechanism is proposed to enable real-time incremental learning of new samples. Furthermore, a structure-preserving update strategy is embedded in the online adaptation process. Class prototype relationship constraints and structure alignment loss suppress target domain structure drift, effectively avoiding catastrophic forgetting during continuous learning and ensuring the model maintains the stability of learned class structures while adapting to new distributions.

[0086] It should be noted that, to evaluate the effectiveness of the embodiments of the present invention, experiments were conducted on a real Wi-Fi dataset with a signal-to-noise ratio of 4dB, and five methods were selected for comparison. These methods include: methods based on discrepancy metrics (such as MMD), methods based on adversarial learning (such as DANN), and hybrid methods that combine discrepancy metrics and adversarial learning (including MDD and WDDRL). In addition, the online baseline method KMPA was introduced to evaluate the model performance in scenarios with continuous input of target domain data streams.

[0087] The process of inputting target domain sample data streams into the radiation source signal identification model was simulated batch by batch. Specifically, the target domain sample data included 2000 samples, with 50 samples randomly selected from each batch, resulting in 40 consecutive batches to reproduce the dynamic online identification process. Unlike fixed-class balanced sampling, this experiment adopted a random sampling strategy, which may lead to uneven or sparse class distribution in some batches, thus more closely resembling the arrival patterns of wireless signals in real-world scenarios. To improve statistical reliability, the random sampling process was repeated 100 times, and the average of the multiple experiments was used as the final evaluation result.

[0088] Experimental results are as follows Figure 3 As shown, this embodiment achieves the highest recognition accuracy. In contrast, traditional domain adaptive methods such as WDDRL, MMD, DANN, and MDD incur significantly higher computational costs due to complex adversarial training or difference metric calculations, and their accuracy is lower in streaming data scenarios. Although the online baseline method KMPA is the fastest among all methods, its accuracy is lower than that of the method described in this invention, indicating that this embodiment achieves a better balance between recognition performance and time efficiency.

[0089] Based on the same inventive concept as the foregoing embodiments, this embodiment of the invention provides a computing device, such as... Figure 4 As shown, the device includes: a processor 310 and a memory 311 storing a computer program; wherein, Figure 4 The processor 310 shown in the diagram does not indicate that there is only one processor 310, but only indicates the positional relationship of the processor 310 relative to other devices. In practical applications, there can be one or more processors 310; similarly, Figure 4 The memory 311 illustrated herein has the same meaning, that is, it is only used to indicate the positional relationship of memory 311 relative to other devices. In practical applications, there can be one or more memories 311. When the processor 310 runs the computer program, the method applied to the above-mentioned device is implemented.

[0090] The device may also include at least one network interface 312. The various components of the device are coupled together via a bus system 313. It is understood that the bus system 313 is used to implement communication between these components. In addition to a data bus, the bus system 313 also includes a power bus, a control bus, and a status signal bus. However, for clarity, in... Figure 4 The general designated all buses as Bus System 313.

[0091] The memory 311 can be volatile memory or non-volatile memory, or both. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), ferromagnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc read-only memory (CD-ROM); magnetic surface memory can be disk storage or magnetic tape storage. Volatile memory can be random access memory (RAM), used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Synchronous Static Random Access Memory (SSRAM), Dynamic Random Access Memory (DRAM), Synchronous Dynamic Random Access Memory (SDRAM), Double Data Rate Synchronous Dynamic Random Access Memory (DDRSDRAM), Enhanced Synchronous Dynamic Random Access Memory (ESDRAM), SyncLink Dynamic Random Access Memory (SLDRAM), and Direct Rambus Random Access Memory (DRRAM).The memory 311 described in the embodiments of the present invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0092] The memory 311 in this embodiment of the invention is used to store various types of data to support the operation of the device. Examples of this data include: any computer programs used to operate on the device, such as operating systems and applications; contact data; phonebook data; messages; pictures; videos, etc. The operating system includes various system programs, such as the framework layer, core library layer, driver layer, etc., used to implement various basic services and handle hardware-based tasks. Applications can include various applications, such as media players, browsers, etc., used to implement various application services. Here, the program implementing the method of this embodiment of the invention can be included in the application.

[0093] Based on the same inventive concept as the foregoing embodiments, this embodiment also provides a computer-readable storage medium storing a computer program. The computer-readable storage medium can be a magnetic random access memory (FRAM), a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a flash memory, a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD-ROM), etc.; it can also be various devices including one or any combination of the above-mentioned memories, such as mobile phones, computers, tablet devices, personal digital assistants, etc. When the computer program stored in the computer-readable storage medium is run by a processor, it implements the above method. For the specific steps implemented when the computer program is executed by the processor, please refer to [link to relevant documentation]. Figure 1 The description of the illustrated embodiments will not be repeated here.

[0094] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0095] In this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, which includes not only the elements listed but also other elements not expressly listed.

[0096] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for online, fast, domain-adaptive radiation source signal identification, characterized in that, include: The radiation source signal identification model is pre-trained based on a labeled source domain sample set and a small number of unlabeled initial target domain sample sets. The target domain sample data that is dynamically flowing in is obtained, and the updated feature extractor and task classifier in the radiation source signal identification model are used for inference to obtain the high-dimensional feature vector, category prediction probability distribution and pseudo-label of each target domain sample. Based on the target domain sample data and its pseudo-labels, a first loss function is constructed to dynamically update the parameters of the feature extractor. The first loss function includes entropy loss and structure alignment loss. The hinge loss of the target domain sample data in the current batch is calculated based on the pseudo-label, and the parameters of the task classifier are updated online based on the kernel function when the hinge loss is greater than zero. The updated radiation source signal identification model outputs the category prediction results for the target domain sample data of the current batch.

2. The online rapid domain-adaptive radiation source signal identification method as described in claim 1, characterized in that, The radiation source signal identification model includes: a feature extractor, a task classifier, and a domain discriminator; The pre-training of the radiation source signal identification model includes: A gradient inversion layer is introduced between the feature extractor and the domain discriminator, and the model is trained by jointly minimizing the task classification loss, domain alignment loss, and adversarial loss. The domain alignment loss includes at least the use of multi-kernel maximum mean difference loss to align the feature distributions of the source and target domains.

3. The online rapid domain-adaptive radiation source signal identification method as described in claim 1 or 2, characterized in that, The feature extractor is a multi-scale feature alignment network, comprising at least two stacked modules. Each stacked module includes a multi-scale convolutional module, a dynamic adaptive attention module, a one-dimensional convolutional layer, group normalization, and a ReLU activation function.

4. The online rapid domain-adaptive radiation source signal identification method as described in claim 3, characterized in that, The multi-scale convolution module includes: multiple parallel convolution paths with different dilation rates and kernel sizes; The dynamic adaptive attention module is used to calculate channel attention weights based on input features and to perform channel weighting on the input features.

5. The online rapid domain-adaptive radiation source signal identification method as described in claim 1, characterized in that, The process of acquiring dynamically flowing target domain sample data and using the feature extractor and task classifier in the radiation source signal identification model for inference to obtain the high-dimensional feature vector, category prediction probability distribution, and pseudo-label for each target domain sample includes: Acquire dynamically flowing target domain sample data in batches; Input the target domain sample data of the current batch into the feature extractor to obtain the high-dimensional feature vector of each target domain sample in the target domain sample data; The high-dimensional feature vector is input into the task classifier to obtain the category prediction probability distribution of the target domain samples; Based on the predicted probability distribution of the categories and the maximum probability criterion, pseudo-labels are assigned to the target domain samples.

6. The online rapid domain-adaptive radiation source signal identification method as described in claim 1, characterized in that, The construction of the first loss function includes: Calculate the entropy loss based on the predicted probability distribution of the target domain samples in the current batch; Based on the pseudo-labels of the target domain samples in the current batch, calculate the temporary prototypes of each category of the target domain; and combine them with the historical target domain prototypes to obtain the updated target domain category prototypes, so as to construct the target domain category structure matrix. Calculate the difference between the target domain category structure matrix and the preset source domain category structure matrix to obtain the structure alignment loss; The entropy loss and the structure alignment loss are used to optimize the parameters of the feature extractor.

7. The online fast domain-adaptive radiation source signal identification method as described in claim 1, characterized in that, The step of calculating the hinge loss of the target domain sample data in the current batch based on the pseudo-label, and updating the parameters of the task classifier online based on the kernel function when the hinge loss is greater than zero, includes: When the hinge loss is greater than zero, the parameters of the task classifier are incrementally updated based on the kernel function; The incremental update of the parameters of the task classifier based on the kernel function includes: adding the current classifier function to a weighting term determined by the kernel function, the pseudo-label, and the hinge loss to correct the discriminant function weight vector.

8. The online fast domain-adaptive radiation source signal identification method as described in claim 1, characterized in that, The updated radiation source signal identification model outputs the category prediction results for the current batch of target domain sample data, followed by: The updated feature extractor and classifier parameters are saved for use in processing subsequent batches of target domain sample data.

Images