Signal modulation recognition method and system based on domain incremental learning

Abstract

This invention relates to the field of signal modulation recognition technology, and in particular to a signal modulation recognition method and system based on domain incremental learning. The invention inputs current domain samples and historical domain samples into a residual convolutional neural network for multi-level feature extraction, and combines multi-level domain adversarial learning, cross-domain contrastive learning mechanisms, and a dual-prototype management system to calculate the total loss function. When minimizing the total loss function reaches the global minimum and the domain to be learned is empty, the residual convolutional neural network of the current domain, which was last updated, is used as the domain incremental learning model for signal modulation recognition. If it is not empty, the residual convolutional neural network of the current domain, which was last updated, is used as the residual convolutional neural network for the first multi-level feature extraction of the next domain for the next domain incremental learning. This invention, by combining a cross-domain contrastive learning mechanism and a dual-prototype management system, effectively alleviates the domain offset problem through dynamic adjustment of adversarial learning, further improving the accuracy of signal modulation recognition.

Description

Technical Field

[0001] This invention relates to the field of signal modulation recognition technology, and in particular to a signal modulation recognition method and system based on domain increment learning. Background Technology

[0002] In the context of the rapid development of wireless communication technology, communication signal modulation identification, as a key technology for radio and spectrum monitoring, has attracted much attention. With the rise of deep learning technology, signal modulation identification methods based on convolutional neural networks have been widely used due to their powerful feature learning capabilities. However, in real-world communication environments, the diversity and time-varying nature of channel conditions pose serious challenges to the stability and generalization ability of deep learning models. When models need to adapt to different channel environments (such as additive white Gaussian noise, Rayleigh fading, Doppler, etc.), domain drift and catastrophic forgetting problems often occur, severely affecting the practicality of the system.

[0003] In the field of domain adaptation, early research mainly focused on static domain adaptation methods. Ganin et al. proposed Domain-Adversarial Neural Networks (DANNs), which learn domain-invariant features through adversarial training. Long et al. further proposed Deep Adaptation Networks (DANs), which utilize multi-kernel maximum mean difference to align feature distributions across different domains. However, these methods are primarily designed for adaptation scenarios from a single source domain to a target domain and cannot effectively handle multiple domains occurring consecutively.

[0004] Research in the field of continuous learning focuses on addressing the catastrophic forgetting problem. Kirkpatrick et al. proposed the Elastic Weight Consolidation (EWC) method, which maintains knowledge from previous tasks by regularizing important parameters. Lopez-Paz et al. proposed Gradient Episodic Memory (GEM), which uses a memory buffer to constrain gradient updates. Although these methods alleviate the forgetting problem to some extent, insufficient feature representation still exists in domain incremental learning scenarios.

[0005] In recent years, contrastive learning has made significant progress in the field of representation learning. The SimCLR method proposed by Chen et al. learns high-quality feature representations by comparing positive and negative sample pairs. The MoCo method proposed by He et al. further introduces a momentum update mechanism to improve training stability. However, existing contrastive learning methods mainly focus on feature learning within the same domain, lacking consideration for cross-domain knowledge transfer.

[0006] In the field of signal processing, multi-level feature fusion has been proven to effectively improve model performance. Wang et al. proposed a multi-scale feature pyramid, and Zhang et al. studied the application of attention mechanisms in signal recognition. However, the effectiveness of these methods in domain incremental learning scenarios still needs to be verified.

[0007] Existing research reveals that traditional domain adaptation methods cannot effectively handle multiple consecutive domains, continuous learning methods lack modeling of differences in feature distributions between domains, and contrastive learning methods neglect the transfer of cross-domain knowledge. Therefore, combining incremental domain learning with signal modulation recognition to simultaneously address domain drift and catastrophic forgetting in incremental domain learning scenarios, and fully utilizing cross-domain knowledge for feature learning, thereby improving the accuracy of signal modulation recognition, has become a critical technical challenge that urgently needs to be addressed. Summary of the Invention

[0008] The purpose of this invention is to provide a signal modulation recognition method and system based on domain incremental learning. By applying domain incremental learning to signal modulation recognition and making full use of cross-domain knowledge for feature learning, the domain drift problem can be solved, and the accuracy of signal modulation recognition can be further improved.

[0009] To achieve the above objectives, the present invention provides a signal modulation recognition method based on domain increment learning, the method comprising:

[0010] Step S1: Select any one of the multiple domains as the current domain, and obtain the current domain sample and the historical domain sample; the current domain sample includes multiple samples corresponding to the current domain, and the historical domain sample includes multiple samples corresponding to the historical domain collected from the memory buffer, wherein the samples include modulation signals corresponding to different modulation types;

[0011] Step S2: Input the current domain samples and historical domain samples into the residual convolutional neural network to extract multi-level features, and obtain the multi-level features corresponding to the current domain samples and the multi-level features corresponding to the historical domain samples.

[0012] Step S3: Based on the multi-layer features corresponding to the current domain samples and the multi-layer features corresponding to the historical domain samples, and combining multi-level domain adversarial learning, cross-domain contrastive learning mechanisms and dual prototype management system, calculate the total loss function;

[0013] Step S4: Minimize the total loss function using the Adam optimizer and determine whether the minimized function has reached the global minimum. If the global minimum has not been reached, update the parameters of the residual convolutional neural network and return to "Step S2". If the global minimum has been reached, proceed to "Step S5".

[0014] Step S5: Determine if the learning domain is empty; if it is empty, use the residual convolutional neural network updated last time in the current domain as the domain incremental learning model so that the signal modulation recognition can be performed using the domain incremental learning model later; if it is not empty, use the residual convolutional neural network updated last time in the current domain as the residual convolutional neural network for the first multi-level feature extraction of the next domain, and arbitrarily select a domain from the learning domain as the current domain, while obtaining the current domain sample and the historical domain sample, and return to "Step S2".

[0015] Optionally, the step of calculating the total loss function based on the multi-layer features corresponding to the current domain samples and the multi-layer features corresponding to the historical domain samples, combined with multi-level domain adversarial learning, cross-domain contrastive learning mechanisms, and a dual-prototype management system, specifically includes:

[0016] Based on the multi-layer features corresponding to the current domain samples and the multi-layer features corresponding to the historical domain samples, multi-layer domain adversarial learning is used for training to obtain multi-layer domain adversarial loss values.

[0017] Based on the deep features corresponding to the current domain samples and the deep features corresponding to the historical domain samples, a cross-domain contrastive learning mechanism is used to calculate and obtain the contrastive learning loss value.

[0018] Determine the prototype comparison loss value based on the dual prototype management system;

[0019] The total loss function is constructed based on the multi-level domain adversarial loss value, the contrastive learning loss value, and the prototype contrastive loss value.

[0020] Optionally, the step of training using multi-level domain adversarial learning based on the multi-level features corresponding to the current domain sample and the multi-level features corresponding to the historical domain sample to obtain a multi-level domain adversarial loss value specifically includes:

[0021] Calculate the maximum mean difference between the deep features corresponding to the current domain sample and the deep features corresponding to the historical domain sample;

[0022] Calculate the domain similarity between the current domain sample and the corresponding historical domain sample based on the maximum mean difference value;

[0023] Calculate the dynamic adversarial strength coefficient based on domain similarity;

[0024] Domain discriminators are set on the third convolutional layer, the fifth convolutional layer, and the fully connected layer of the residual convolutional neural network, respectively;

[0025] The multi-layer features corresponding to the current domain sample and the multi-layer features corresponding to the historical domain sample are respectively input into the corresponding discriminator, and multi-layer domain adversarial learning is used for training to obtain the domain adversarial loss value corresponding to each layer.

[0026] The multi-layer domain adversarial loss value is calculated based on the dynamic adversarial strength coefficient and the corresponding domain adversarial loss value of each layer.

[0027] Optionally, the step of calculating the contrastive learning loss value using a cross-domain contrastive learning mechanism based on the deep features corresponding to the current domain sample and the deep features corresponding to the historical domain sample specifically includes:

[0028] The deep features corresponding to each sample in the current domain are concatenated with the deep features corresponding to each sample in the corresponding historical domain to obtain a concatenated feature matrix.

[0029] The concatenated feature matrix is ​​processed for similarity to obtain the sample similarity matrix between each sample;

[0030] A cross-domain contrastive learning mechanism is used to calculate the weighted positive sample matrix;

[0031] The contrastive learning loss value is determined based on the sample similarity matrix and the weighted positive sample matrix.

[0032] Optionally, determining the prototype comparison loss value based on the dual prototype management system specifically includes:

[0033] A dual-prototype management system is established based on the deep features and corresponding labels of the current domain samples, and the dual-prototype management system is used to determine the domain-invariant prototype and domain-specific prototype corresponding to each category.

[0034] Calculate the mean of deep features for all samples in each category of the current domain;

[0035] A momentum update strategy is adopted, which uses the mean of deep features of all samples in each category of the current domain to update the domain-invariant prototype and domain-specific prototype corresponding to each category respectively;

[0036] Based on the updated domain-invariant prototype and domain-specific prototype, a weighted similarity calculation is performed to obtain the comprehensive classification vector for each sample.

[0037] The prototype contrastive loss value is obtained by comparing the comprehensive classification vector corresponding to each sample with the true label.

[0038] Optionally, the similarity processing of the concatenated feature matrix to obtain the sample similarity matrix between each sample is specifically calculated using the following formula:

[0039]

[0040] in, Let S be the temperature hyperparameter and S be the sample similarity matrix. This is for splicing feature matrices.

[0041] Optionally, the cross-domain contrastive learning mechanism is used to calculate the weighted positive sample matrix, and the specific formula is as follows:

[0042]

[0043] in, Weights for samples of the same type and within the same domain. Weights for cross-domain samples of the same type, For the label mask matrix, For the domain mask matrix, For the weighted positive sample matrix, Multiply the elements at corresponding positions of the two matrices.

[0044] Optionally, the step of calculating the weighted similarity based on the domain-invariant prototype and domain-specific prototype corresponding to each category after the update to obtain the comprehensive classification vector corresponding to each sample is as follows:

[0045]

[0046] in, and These represent the domain-invariant prototype and the domain-specific prototype corresponding to the c-th category after the update, respectively. For the current domain ID, For domain-invariant prototype weights, Domain-specific prototype weights, For temperature parameters, Indicates sample The L2 normalized form of the deep feature vectors, This represents the comprehensive classification vector corresponding to sample n.

[0047] The present invention also provides a signal modulation recognition system based on domain incremental learning, the system comprising:

[0048] The acquisition module is used to select any one of multiple domains as the current domain, and acquire current domain samples and historical domain samples; the current domain samples include multiple samples corresponding to the current domain, and the historical domain samples include multiple samples corresponding to the historical domain collected from the memory buffer, wherein the samples include modulation signals corresponding to different modulation types;

[0049] The multi-level feature extraction module is used to input the current domain samples and historical domain samples into the residual convolutional neural network to perform multi-level feature extraction, and obtain the multi-level features corresponding to the current domain samples and the multi-level features corresponding to the historical domain samples.

[0050] The total loss function calculation module is used to calculate the total loss function based on the multi-layer features corresponding to the current domain samples and the multi-layer features corresponding to the historical domain samples, combined with multi-level domain adversarial learning, cross-domain contrastive learning mechanisms and dual prototype management system.

[0051] The first judgment module is used to minimize the total loss function through the Adam optimizer and determine whether the minimized function has reached the global minimum. If the global minimum has not been reached, the parameters of the residual convolutional neural network are updated and returned to the "multi-level feature extraction module". If the global minimum has been reached, the "second judgment module" is executed.

[0052] The second judgment module is used to determine whether the learning domain is empty. If it is empty, the residual convolutional neural network updated last time in the current domain is used as the domain incremental learning model so that the domain incremental learning model can be used for signal modulation recognition in the future. If it is not empty, the residual convolutional neural network updated last time in the current domain is used as the residual convolutional neural network for the first multi-level feature extraction of the next domain, and an arbitrary domain is selected from the learning domain as the current domain. At the same time, the current domain samples and historical domain samples are obtained, and the result is returned to the "multi-level feature extraction module".

[0053] Optionally, the total loss function calculation module includes:

[0054] The multi-level domain adversarial loss calculation unit is used to train the multi-level domain adversarial learning based on the multi-level features corresponding to the current domain sample and the multi-level features corresponding to the historical domain sample to obtain the multi-level domain adversarial loss value.

[0055] The contrastive learning loss calculation unit is used to calculate the contrastive learning loss value based on the deep features corresponding to the current domain sample and the deep features corresponding to the historical domain sample using a cross-domain contrastive learning mechanism.

[0056] The prototype comparison loss value calculation unit is used to determine the prototype comparison loss value based on the dual prototype management system.

[0057] The total loss function calculation unit is used to construct the total loss function based on the multi-level domain adversarial loss value, the contrastive learning loss value, and the prototype contrastive loss value.

[0058] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0059] This invention discloses a method for simultaneously performing domain adaptation at shallow, mid, and deep layers during domain incremental learning. Combined with a cross-domain contrastive learning mechanism and a dual-prototype management system, this enhances the cross-domain generalization performance of the residual convolutional neural network while maintaining its class discrimination ability. Dynamic adjustment of adversarial learning effectively mitigates the domain shift problem. Furthermore, the invention utilizes a memory buffer and knowledge distillation mechanism to effectively suppress catastrophic forgetting during domain incremental learning, achieving continuous learning capabilities in the domain incremental learning scenario. Dynamic adjustment of adversarial learning further mitigates the domain shift problem. Therefore, this invention applies domain incremental learning to signal modulation recognition, which can further and effectively improve the accuracy of signal modulation recognition. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is a flowchart of a signal modulation recognition method based on domain incremental learning according to an embodiment of the present invention. Detailed Implementation

[0062] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] The purpose of this invention is to provide a signal modulation recognition method and system based on domain incremental learning. By applying domain incremental learning to signal modulation recognition and making full use of cross-domain knowledge for feature learning, the domain drift problem can be solved, and the accuracy of signal modulation recognition can be further improved.

[0064] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0065] like Figure 1 As shown, this invention discloses a signal modulation recognition method based on domain increment learning, the method comprising:

[0066] Step S1: Select any one of the multiple domains as the current domain, and obtain the current domain sample and the historical domain sample. The current domain sample includes multiple samples corresponding to the current domain, and the historical domain sample includes multiple samples corresponding to the historical domain collected from the memory buffer. The samples include modulation signals corresponding to different modulation types.

[0067] Step S2: Input the current domain samples and historical domain samples into the residual convolutional neural network to extract multi-level features, and obtain the multi-level features corresponding to the current domain samples and the multi-level features corresponding to the historical domain samples.

[0068] Step S3: Based on the multi-layer features corresponding to the current domain samples and the multi-layer features corresponding to the historical domain samples, and combining multi-level domain adversarial learning, cross-domain contrastive learning mechanisms and dual prototype management system, calculate the total loss function.

[0069] Step S4: Minimize the total loss function using the Adam optimizer and determine whether the minimized function has reached the global minimum. If the global minimum has not been reached, update the parameters of the residual convolutional neural network and return to "Step S2". If the global minimum has been reached, proceed to "Step S5".

[0070] Step S5: Determine if the learning domain is empty; if it is empty, use the residual convolutional neural network updated last time in the current domain as the domain incremental learning model so that the signal modulation recognition can be performed using the domain incremental learning model later; if it is not empty, use the residual convolutional neural network updated last time in the current domain as the residual convolutional neural network for the first multi-level feature extraction of the next domain, and arbitrarily select a domain from the learning domain as the current domain, while obtaining the current domain sample and the historical domain sample, and return to "Step S2".

[0071] The following is a detailed discussion of each step:

[0072] Step S1: Select any one of the multiple domains as the current domain, and obtain the current domain sample and the historical domain sample.

[0073] In this embodiment, in the domain incremental learning scenario, the datasets of each channel environment are loaded sequentially. Each channel environment (such as AWGN, Rician, Rayleigh, Doppler, Suzuki) corresponds to a data distribution domain (referred to as a domain).

[0074] The current domain samples include multiple samples corresponding to the current domain, which is the domain currently being learned; the historical domain samples include multiple samples corresponding to the historical domain collected from the memory buffer, which is a domain that has been learned at least once; in this embodiment, the total capacity of the memory buffer is set to be at least greater than the capacity corresponding to the historical domain samples required for each training session. Assuming that the number of currently learned domains is D, the capacity allocated to each domain in the memory buffer is... This ensures that the number of samples in each domain is approximately equal. The batch size sampled from the buffer is the same as the batch size of the current domain. In this embodiment, the memory buffer is automatically updated by the system after the previous round of domain training, that is, a portion of samples and their corresponding old model prediction results (logits) are randomly sampled from the domains that have been learned and stored in the buffer for subsequent domain replay and knowledge distillation.

[0075] Each of the aforementioned samples includes modulation signals corresponding to different modulation types; the modulation signals corresponding to each modulation type exist in the form of IQ dual-channel complex signals; the modulation types mentioned in this invention include at least one of digital modulation types and analog modulation types; wherein, the digital modulation types include at least one of BPSK, QPSK, 8PSK, 16QAM, 64QAM, PAM4, GFSK and CPFSK, and the analog modulation types include at least one of B-FM, DSB-AM and SSB-AM.

[0076] Step S2: Input the current domain samples and historical domain samples into the residual convolutional neural network for multi-level feature extraction, obtaining the multi-level features corresponding to the current domain samples and the multi-level features corresponding to the historical domain samples. Multi-level features corresponding to the current domain samples. Where B is the batch size, also known as the total number of samples in the current domain; Including shallow features corresponding to each sample in the current domain , Including the mid-level features corresponding to each sample in the current domain , Including deep features corresponding to each sample in the current domain ,above Multi-layer features corresponding to historical domain samples , Where B is the batch size, also known as the total number of historical domain samples; Including shallow features corresponding to each sample in the historical domain , Including mid-level features corresponding to each sample in the historical domain , Including deep features corresponding to each sample in the historical domain ,above In this embodiment, the total number of samples in the current domain and the total number of samples in the historical domain are both B, meaning that the two batches are of equal size.

[0077] In this embodiment, the feature hierarchy set For any level Let the vector features of the p-th sample in the current domain and the q-th sample in the historical domain at the l-th layer be denoted as follows: , ,in Indicates the first The dimension of the layer feature vector.

[0078] In this embodiment, the residual convolutional neural network consists of six convolutional layers (conv1-conv6) and one fully connected layer (dense). The input to the residual convolutional neural network is an IQ dual-channel signal of [B, 2, 1024], where B is the batch size, 2 represents the IQ dual-channel signal, and 1024 is the signal length. Shallow features are obtained through the first to third convolutional layers, with a dimension of [B, 32, L']; mid-level features are obtained through the fourth to fifth convolutional layers, with a dimension of [B, 32, L'']; and deep features are obtained through the sixth convolutional layer and the fully connected layer, with a dimension of [B, 128]. Shallow features capture the local texture patterns of the signal, mid-level features capture structural features of medium abstraction, and deep features capture high-level semantic features.

[0079] Specifically, the residual convolutional neural network will be discussed using ResNet1D as an example: Selected samples from the current domain are input into ResNet1D. First, the input is expanded into a feature map of [B, 32, 1024] through the conv1 layer. Then, shallow features are obtained through the conv2 and conv3 layers. The dimensions are [B, 32, L'], where L' is the length of the sequence after pooling; forward propagation continues through conv4 and conv5 layers to obtain mid-level features. The dimensions are [B, 32, L'']; finally, after global average pooling through a conv6 layer and flattening, deep features are obtained through a dense fully connected layer. The three feature layers are [B, 128]. These three layers capture the local temporal texture, the intermediate-scale frequency domain structure, and the high-level semantic representation of the signal, respectively.

[0080] The historical domain samples are also input into the ResNet1D network to extract three layers of features output from layers conv1 to conv3, conv4 to conv5, and conv6 to dense. , and For convolutional layer features , Since its dimension is a three-dimensional tensor [B,C,L], it needs to be compressed into a two-dimensional feature vector [B,C] by the adaptive global average pooling operation AdaptiveAvgPool1d, which is convenient for subsequent domain discriminator processing.

[0081] The multi-level features obtained above will be used in subsequent steps for domain adversarial training, cross-domain contrastive learning, and prototype matching to achieve domain invariance and class discriminativeness of the features. The specific steps are as follows:

[0082] Step S3: Based on the multi-layer features corresponding to the current domain samples and the multi-layer features corresponding to the historical domain samples, and combining multi-layer domain adversarial learning, cross-domain contrastive learning mechanisms, and a dual-prototype management system, calculate the total loss function, specifically including:

[0083] Step S31: Based on the multi-layer features corresponding to the current domain samples and the multi-layer features corresponding to the historical domain samples, multi-layer domain adversarial learning is used for training to obtain the multi-layer domain adversarial loss value L. DA The specific steps include:

[0084] Step S311: Calculate the deep features corresponding to the current domain sample. Deep features corresponding to the historical domain samples The maximum mean difference between them is calculated using the following formula:

[0085]

[0086] in, This represents the expectation operation. This represents the kernel function operation, and MMD represents the maximum mean difference. This represents the deep feature corresponding to the p-th sample in the history domain. This represents the deep feature corresponding to the q-th sample in the history domain. This represents the deep feature corresponding to the p-th sample in the current domain. This represents the deep feature corresponding to the q-th sample in the current domain, where B is the batch size.

[0087] Step S312: Calculate the domain similarity between the current domain sample and the corresponding historical domain sample based on the maximum mean difference value. The specific formula is as follows:

[0088]

[0089] in, Representation domain similarity.

[0090] Step S313: Calculate the dynamic adversarial strength coefficient based on domain similarity. The specific formula is as follows:

[0091]

[0092] in, This represents the dynamic resistance intensity coefficient.

[0093] When faced with a new domain with significantly different distributions, this indicates... Smaller Automatically increasing the size to enhance adversarial training intensity forces the residual convolutional neural network to learn a stronger domain-invariant representation; conversely, when the new domain is similar to the historical domain, Automatic reduction is used to avoid excessive adversarial response that could lead to loss of discriminative power.

[0094] In this embodiment, the maximum mean discrepancy (MMD) is used to measure the distance between domain distributions. Domain similarity is used to measure whether the domain distributions are close; the higher the domain similarity, the closer the domain distributions are.

[0095] Step S314: Set up domain discriminators on the third convolutional layer, the fifth convolutional layer, and the fully connected layer of the residual convolutional neural network. , and .

[0096] In this embodiment, to achieve domain invariance of multi-level features, domain discriminators are set on the third convolutional layer conv3, the fifth convolutional layer conv5, and the fully connected layer dense of the residual convolutional neural network. , and Each discriminator employs a lightweight two-layer fully connected network structure to distinguish between current domain features and historical domain features. For each layer... Training the discriminator To distinguish domain labels, current domain samples are labeled as 1, and historical domain samples are labeled as 0. The discrimination ability is maximized through binary cross-entropy.

[0097] Step S315: Input the multi-layer features corresponding to the current domain samples and the multi-layer features corresponding to the historical domain samples into the corresponding discriminators, and train them using multi-layer domain adversarial learning to obtain the domain adversarial loss value corresponding to each layer. The specific formula is as follows:

[0098]

[0099] in, This represents the domain adversarial loss value corresponding to the l-th layer. and These represent the multi-layer features of the p-th sample in the current domain and the q-th sample in the historical domain at the l-th layer, respectively. Let E represent the discriminator corresponding to the l-th layer, and E() represent the expectation operation. .

[0100] Preferably, convolutional layers conv3 and conv5 correspond to shallow features, respectively. and mid-layer features Since their dimensions are respectively And [B,32,L''], so it still needs to be compressed into [B,32,L''] first through adaptive global average pooling. Then the data is input into the discriminator. For deep features corresponding to fully connected layers, the data is directly input into the discriminator.

[0101] In this embodiment, the residual convolutional neural network can be back-optimized to confuse the discriminator.

[0102] Step S316: Calculate the multi-layer domain adversarial loss value based on the dynamic adversarial strength coefficient and the corresponding domain adversarial loss value for each layer. The specific formula is as follows:

[0103]

[0104]

[0105] Among them, L DA For multi-level domain adversarial loss values, The learnable weights of the l-th layer, w l The weights obtained after Sigmoid activation normalization Activate the normalization function for the Sigmoid function. This represents the domain adversarial loss value corresponding to the l-th layer. This represents the dynamic resistance intensity coefficient.

[0106] This embodiment sets up a multi-level domain adversarial loss so that the parameters of each layer of the residual convolutional neural network can be optimized by gradient descent, so that the multi-level features generated can maintain class discriminativeness while possessing domain invariance, thereby alleviating the domain drift problem.

[0107] Step S32: Based on the deep features corresponding to the current domain samples and the deep features corresponding to the historical domain samples, a cross-domain contrastive learning mechanism is used to calculate and obtain the contrastive learning loss value. The specific steps include:

[0108] Step S321: Concatenate the deep features corresponding to each sample in the current domain with the deep features corresponding to each sample in the corresponding historical domain to obtain the concatenated feature matrix. The specific formula is as follows:

[0109]

[0110] in, To concatenate the feature matrix, This represents the deep feature corresponding to the p-th sample in the current domain. Let q be the deep feature corresponding to the q-th sample in the historical domain.

[0111] Step S322: Perform similarity processing on the concatenated feature matrix to obtain the sample similarity matrix S between each sample. The specific formula is as follows:

[0112]

[0113] in, Let S be the temperature hyperparameter and S be the sample similarity matrix. This is for splicing feature matrices.

[0114] Step S323: Calculate the weighted positive sample matrix using a cross-domain contrastive learning mechanism. The specific formula is as follows:

[0115]

[0116] in, Weights for samples of the same type and within the same domain. Weights for cross-domain samples of the same type, For the label mask matrix, Represents the label mask matrix The value corresponding to the element in the i-th row and j-th column represents the consistency between the two samples; a value of 1 indicates that the two samples belong to the same class, while a value of 0 indicates that they belong to different classes. For the domain mask matrix, Representation domain mask matrix The value corresponding to the element in the i-th row and j-th column indicates the consistency between the two sample domains; a value of 1 indicates that the two sample domains are the same, and a value of 0 indicates that they are different domains. For the weighted positive sample matrix, Multiply the elements at corresponding positions of the two matrices.

[0117] In this embodiment, the weighted positive sample matrix This is used to distinguish positive sample pairs in contrastive loss and assign differential weights, so that the model gives the highest priority when bringing similar samples within the same domain closer together, while also moderately bringing similar samples across different domains closer together to achieve knowledge transfer.

[0118] Step S324: Based on the sample similarity matrix and weighted positive sample matrix The specific formula for determining the contrastive learning loss value is as follows:

[0119]

[0120]

[0121] Where N is a positive integer, To compare the learning loss values, Let i be the contrastive learning loss value for the i-th row. It is the numerical stability constant. Weighted positive sample matrix The value corresponding to the element in the i-th row and j-th column represents the similarity weight coefficient between the two samples. Let N be the value corresponding to the element in the i-th row and j-th column of the sample similarity matrix S, representing the similarity between the two samples. In this embodiment, N is 2^B.

[0122] In this embodiment, the present invention enhances the cross-domain generalization ability of augmented features through a cross-domain contrastive learning mechanism. In addition, the contrastive learning loss maximizes the similarity of positive sample pairs while minimizing the similarity of negative sample pairs, thereby promoting the formation of a compact intra-class and separated inter-class distribution of features in space. The loss is optimized by gradient descent to achieve the aggregation of samples of the same class and the separation of samples of different classes. Furthermore, the positive transfer of knowledge between different domains is achieved by appropriately bringing cross-domain samples closer together.

[0123] Step S33: Determine the prototype comparison loss value L based on the dual prototype management system. Proto The specific steps include:

[0124] Step S331: Based on the deep features corresponding to the current domain samples and corresponding tags A dual-prototype management system is established to determine the domain-invariant prototype and domain-specific prototype corresponding to each category. The purpose of this invention is to enhance the stability of feature representation and intra-class aggregation.

[0125] Step S332: Calculate the mean of deep features for all samples of each category in the current domain. The specific formula is as follows:

[0126]

[0127] in, This indicates that all tags are categories. Sample Index , This represents the mean of deep features for all samples in the c-th category of the current domain. This represents the deep feature corresponding to the k-th sample in the current domain.

[0128] Step S333: Employ a momentum update strategy, using the mean of deep features of all samples in each category of the current domain to update the domain-invariant prototype and domain-specific prototype corresponding to each category, respectively. The specific formula is as follows:

[0129]

[0130]

[0131] in, and These represent the domain-invariant prototype and the domain-specific prototype corresponding to the c-th category, respectively. For the current domain ID, The momentum coefficient, This indicates an update to the rules. This represents the mean of deep features for all samples in the c-th category of the current domain.

[0132] In other words, the above will As the prototype of the current batch category, the momentum update strategy is used to update the prototype of the current batch category. Domain-invariant prototype Domain-specific prototypes After merging and updating, the updated version is stored in the prototype manager as a stable category center.

[0133] Step S334: Calculate the weighted similarity based on the updated domain-invariant prototype and domain-specific prototype for each category to obtain the comprehensive classification vector for each sample. The specific formula is as follows:

[0134]

[0135] in, and These represent the domain-invariant prototype and the domain-specific prototype corresponding to the c-th category after the update, respectively. For the current domain ID, For domain-invariant prototype weights, Domain-specific prototype weights, For temperature parameters, Indicates sample The L2 normalized form of the deep feature vectors, This represents the comprehensive classification vector corresponding to sample n. .

[0136] Step S335: Calculate the contrastive loss based on the comprehensive classification vector corresponding to each sample and the true label to obtain the prototype contrastive loss value L. Proto The specific calculation formula is as follows:

[0137]

[0138] Where B represents the batch size. Represents the cross-entropy loss function. This represents the true label corresponding to sample n. This represents the comprehensive classification vector corresponding to sample n.

[0139] Prototype contrast loss value L Proto By driving features toward the correct class prototype and away from the incorrect class prototype through gradient descent, a stable class representation center is maintained, effectively preventing drastic drift in the feature space when learning new domains.

[0140] Step S34: Based on the multi-level domain adversarial loss value L DA Contrast learning loss value L CDCL Loss value L compared with the prototype Proto Construct the total loss function L total The specific formula is as follows:

[0141]

[0142] in, and These are the classification losses for the current domain samples and the classification losses for the historical domain samples, respectively. For knowledge distillation loss value, , , This is the loss weight hyperparameter.

[0143] In this embodiment, both the historical domain classification loss and the knowledge distillation loss are multiplied by a coefficient of 0.5 to balance current domain learning and old domain preservation; in addition, the classification loss of current domain samples... Classification loss of historical domain samples All samples use the standard cross-entropy form; for historical domain samples taken from the memory buffer, a knowledge distillation loss is introduced to preserve old domain knowledge and prevent catastrophic forgetting: the specific formula for calculating the knowledge distillation loss value is as follows:

[0144]

[0145] in, Where B is the knowledge distillation loss value, and B is the batch size. For the current prediction The logits vector corresponding to each historical sample For historical predictions The logits vector corresponding to each historical sample Where K is the distillation temperature, and KL represents the KL divergence.

[0146] The above formula can also be understood as: To utilize the current model in the first The logits vector predicted from historical samples. To utilize the logits vector predicted by the old model on the same historical samples, distillation temperature The KL divergence is used to soften the probability distribution, making knowledge transfer smoother, and it measures the difference between the two distributions. In other words, this invention uses knowledge distillation loss to constrain the output of the current model on historical samples to remain consistent with the old model, thereby preserving the learned domain knowledge.

[0147] Step S4: Minimize the total loss function L using the Adam optimizer total Then, determine whether the minimized function has reached the global minimum; if it has not reached the global minimum, update the parameters of the residual convolutional neural network and return to "step S2"; if it has reached the global minimum, execute "step S5".

[0148] Step S5: Determine if the learning domain is empty. If empty, use the residual convolutional neural network (RCN) updated last time in the current domain as the domain incremental learning model for subsequent signal modulation recognition. If not empty, use the residual RNN updated last time in the current domain as the residual RNN for the first multi-level feature extraction of the next domain, and arbitrarily select one domain from the learning domains as the current domain. Simultaneously, obtain the current domain samples and historical domain samples, and return to "Step S2". The learning domain is used to store multiple unlearned domains; the selection above can be understood as selection without replacement. The domain incremental learning main loop learns five domains sequentially in the order of AWGN → Rician → Rayleigh → Doppler → Suzuki.

[0149] This invention discloses a method that simultaneously performs domain adaptation in shallow, mid, and deep layers, combined with a cross-domain contrastive learning mechanism and a dual-prototype management system. This enables residual convolutional neural networks to enhance cross-domain generalization performance while maintaining class discrimination capabilities. The domain shift problem is effectively mitigated through dynamically adjusted adversarial learning. Furthermore, through a memory buffer and knowledge distillation mechanism, catastrophic forgetting is effectively suppressed, achieving continuous learning capabilities in domain incremental learning scenarios. The domain shift problem is effectively mitigated through dynamically adjusted adversarial learning.

[0150] The present invention also provides a signal modulation recognition system based on domain incremental learning, the system comprising:

[0151] The acquisition module is used to select any one of multiple domains as the current domain, and acquire current domain samples and historical domain samples. The current domain samples include multiple samples corresponding to the current domain, and the historical domain samples include multiple samples corresponding to the historical domain collected from the memory buffer. The samples include modulation signals corresponding to different modulation types.

[0152] The multi-level feature extraction module is used to input the current domain samples and historical domain samples into the residual convolutional neural network to perform multi-level feature extraction, thereby obtaining the multi-level features corresponding to the current domain samples and the multi-level features corresponding to the historical domain samples.

[0153] The total loss function calculation module is used to calculate the total loss function based on the multi-layer features corresponding to the current domain samples and the multi-layer features corresponding to the historical domain samples, combined with multi-level domain adversarial learning, cross-domain contrastive learning mechanisms and dual prototype management system.

[0154] The first judgment module is used to minimize the total loss function through the Adam optimizer and determine whether the minimized function has reached the global minimum. If the global minimum has not been reached, the parameters of the residual convolutional neural network are updated and returned to the "multi-level feature extraction module". If the global minimum has been reached, the "second judgment module" is executed.

[0155] The second judgment module is used to determine whether the learning domain is empty. If it is empty, the residual convolutional neural network updated last time in the current domain is used as the domain incremental learning model so that the domain incremental learning model can be used for signal modulation recognition in the future. If it is not empty, the residual convolutional neural network updated last time in the current domain is used as the residual convolutional neural network for the first multi-level feature extraction of the next domain, and an arbitrary domain is selected from the learning domain as the current domain. At the same time, the current domain samples and historical domain samples are obtained, and the result is returned to the "multi-level feature extraction module".

[0156] As an optional implementation, the total loss function calculation module of the present invention includes:

[0157] The multi-level domain adversarial loss calculation unit is used to train multi-level domain adversarial learning based on the multi-level features corresponding to the current domain sample and the multi-level features corresponding to the historical domain sample to obtain the multi-level domain adversarial loss value.

[0158] The contrastive learning loss calculation unit is used to calculate the contrastive learning loss value by using a cross-domain contrastive learning mechanism based on the deep features corresponding to the current domain samples and the deep features corresponding to the historical domain samples.

[0159] The prototype comparison loss value calculation unit is used to determine the prototype comparison loss value based on the dual prototype management system.

[0160] The total loss function calculation unit is used to construct the total loss function based on the multi-level domain adversarial loss value, the contrastive learning loss value, and the prototype contrastive loss value.

[0161] The steps and methods for each of the above calculation units are the same, and will not be repeated here. For details, please refer to the method steps.

[0162] Specific examples:

[0163] In this embodiment, the residual convolutional neural network is constructed based on a multi-channel communication dataset. The multi-channel communication dataset contains 11 modulation categories: 8 digital modulations (BPSK, QPSK, 8PSK, 16QAM, 64QAM, PAM4, GFSK, and CPFSK) and 3 analog modulations (B-FM, DSB-AM, and SSB-AM). The sampling rate is set to 200kHz, with 8 samples per symbol and 1024 samples per frame. For each modulation type, 1000 frames are generated under a specific channel environment, each sample being an IQ dual-channel complex signal. Channel modeling uses the standard channel models from the MATLAB Communication Systems Toolbox, including Rician fading channels (K-factor of 4, Doppler shift of 4Hz), Rayleigh fading channels (multipath delays of 0, 1.8, and 3.4 sampling intervals), AWGN channels, Doppler channels (maximum Doppler shift of 30Hz), and Suzuki channels (Rayleigh fading superimposed with log-normal shading). In the data preprocessing stage, all samples under each channel environment are tagged and encoded according to the modulation type, and divided into current domain samples and historical domain samples.

[0164] This invention uses the ResNet1D network as an example to illustrate the residual convolutional neural network. Lightweight domain discriminators are constructed sequentially at the conv3, conv5, and dense layers. For the conv3 and conv5 layers, since the input feature dimension is 32, the discriminator structure is Linear(32) → ReLU → Dropout(0.3) → Linear(1) → Sigmoid. For the dense layer, with an input dimension of 128, the structure is Linear(128) → ReLU → Dropout(0.3) → Linear(1) → Sigmoid. When training the three domain discriminators using the Adam optimizer, the learning rate is set to 0.001.

[0165] The Adam optimizer was used to minimize the total loss function described above, with a learning rate of 0.001, a batch size of 128, and 20 epochs per domain. The learning rate scheduler employed a StepLR strategy, decaying the learning rate to 0.8 times its original value every 10 epochs.

[0166] The main loop for incremental domain learning learns five domains sequentially in the order of AWGN → Rician → Rayleigh → Doppler → Suzuki. If the last domain is not reached, the learning of the next domain begins, and the above process is repeated.

[0167] Application scenarios:

[0168] This invention applies the aforementioned signal modulation recognition method to a mobile spectrum monitoring system deployed in a mobile monitoring vehicle. When performing monitoring tasks in different geographical locations, this system encounters drastically different channel environments: densely populated urban areas suffer from severe multipath effects, open suburban areas are close to AWGN channels, and mountainous areas experience significant shadow fading. After incrementally learning four channel environments, this invention controls the decrease in signal modulation recognition accuracy for historical channel environments to within 5%, achieving an average cross-channel signal modulation recognition accuracy of over 87%, effectively addressing the continuous recognition capability requirement of spectrum monitoring systems in dynamic channel environments.

[0169] Traditional modulation recognition models, after being trained in one environment, experience a sharp performance drop when deployed to a new environment. However, the domain incremental learning model proposed in this application can incrementally learn the recognition capability for new channel environments online without re-collecting historical environmental data or interrupting existing monitoring tasks. This effectively alleviates the domain drift problem caused by the diversity of channel environments and significantly improves the accuracy of subsequent signal modulation recognition. Furthermore, compared to re-inputting all samples into the model for each new domain data instance, the scheme disclosed in this application greatly reduces training time and saves computational resources.

[0170] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0171] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A signal modulation recognition method based on domain increment learning, characterized in that, The method includes: Step S1: Select any one of the multiple domains as the current domain, and obtain the current domain sample and the historical domain sample; the current domain sample includes multiple samples corresponding to the current domain, and the historical domain sample includes multiple samples corresponding to the historical domain collected from the memory buffer, wherein the samples include modulation signals corresponding to different modulation types; Step S2: Input the current domain samples and historical domain samples into the residual convolutional neural network to extract multi-level features, and obtain the multi-level features corresponding to the current domain samples and the multi-level features corresponding to the historical domain samples. Step S3: Based on the multi-layer features corresponding to the current domain samples and the multi-layer features corresponding to the historical domain samples, and combining multi-level domain adversarial learning, cross-domain contrastive learning mechanisms and dual prototype management system, calculate the total loss function; Step S4: Minimize the total loss function using the Adam optimizer and determine whether the minimized function has reached the global minimum. If the global minimum has not been reached, update the parameters of the residual convolutional neural network and return to "Step S2". If the global minimum has been reached, proceed to "Step S5". Step S5: Determine if the learning domain is empty; if it is empty, use the residual convolutional neural network updated last time in the current domain as the domain incremental learning model so that the domain incremental learning model can be used for signal modulation recognition in the future; if it is not empty, use the residual convolutional neural network updated last time in the current domain as the residual convolutional neural network for the first multi-level feature extraction of the next domain, and arbitrarily select a domain from the learning domain as the current domain, and obtain the current domain sample and the historical domain sample, and return to "Step S2"; The step involves calculating the total loss function based on the multi-layer features corresponding to the current domain samples and the multi-layer features corresponding to the historical domain samples, combined with multi-level domain adversarial learning, cross-domain contrastive learning mechanisms, and a dual-prototype management system. Specifically, this includes: Based on the multi-layer features corresponding to the current domain sample and the multi-layer features corresponding to the historical domain sample, multi-layer domain adversarial learning is used for training to obtain multi-layer domain adversarial loss values. Based on the deep features corresponding to the current domain samples and the deep features corresponding to the historical domain samples, a cross-domain contrastive learning mechanism is used to calculate and obtain the contrastive learning loss value. Determine the prototype comparison loss value based on the dual prototype management system; Construct a total loss function based on the multi-level domain adversarial loss value, the contrastive learning loss value, and the prototype contrastive loss value; The step involves training using multi-level domain adversarial learning based on the multi-level features corresponding to the current domain samples and the multi-level features corresponding to the historical domain samples, to obtain a multi-level domain adversarial loss value. Specifically, this includes: Calculate the maximum mean difference between the deep features corresponding to the current domain sample and the deep features corresponding to the historical domain sample; Calculate the domain similarity between the current domain sample and the corresponding historical domain sample based on the maximum mean difference value; Calculate the dynamic adversarial strength coefficient based on domain similarity; Domain discriminators are set on the third convolutional layer, the fifth convolutional layer, and the fully connected layer of the residual convolutional neural network, respectively; The multi-layer features corresponding to the current domain sample and the multi-layer features corresponding to the historical domain sample are respectively input into the corresponding discriminator, and multi-layer domain adversarial learning is used for training to obtain the domain adversarial loss value corresponding to each layer. Calculate the multi-level domain adversarial loss value based on the dynamic adversarial intensity coefficient and the corresponding domain adversarial loss value of each layer; The step involves calculating the contrastive learning loss value using a cross-domain contrastive learning mechanism based on the deep features corresponding to the current domain sample and the deep features corresponding to the historical domain sample. This specifically includes: The deep features corresponding to each sample in the current domain are concatenated with the deep features corresponding to each sample in the corresponding historical domain to obtain a concatenated feature matrix. The concatenated feature matrix is ​​processed for similarity to obtain the sample similarity matrix between each sample; A cross-domain contrastive learning mechanism is used to calculate the weighted positive sample matrix; The contrastive learning loss value is determined based on the sample similarity matrix and the weighted positive sample matrix; The determination of the prototype comparison loss value based on the dual prototype management system specifically includes: A dual-prototype management system is established based on the deep features and corresponding labels of the current domain samples, and the dual-prototype management system is used to determine the domain-invariant prototype and domain-specific prototype corresponding to each category. Calculate the mean of deep features for all samples in each category of the current domain; A momentum update strategy is adopted, which uses the mean of deep features of all samples in each category of the current domain to update the domain-invariant prototype and domain-specific prototype corresponding to each category respectively; Based on the updated domain-invariant prototype and domain-specific prototype, a weighted similarity calculation is performed to obtain the comprehensive classification vector for each sample. The prototype contrastive loss value is obtained by comparing the comprehensive classification vector corresponding to each sample with the true label.

2. The signal modulation recognition method based on domain increment learning according to claim 1, characterized in that, The similarity processing of the concatenated feature matrix yields a sample similarity matrix between each sample, using the following formula: ; in, Let S be the temperature hyperparameter and S be the sample similarity matrix. This is for splicing feature matrices.

3. The signal modulation recognition method based on domain increment learning according to claim 1, characterized in that, The cross-domain contrastive learning mechanism is used to calculate the weighted positive sample matrix, and the specific formula is as follows: ; in, Weights for samples of the same type and within the same domain. Weights for cross-domain samples of the same type, For the label mask matrix, For the domain mask matrix, For the weighted positive sample matrix, Multiply the elements at corresponding positions of the two matrices.

4. The signal modulation recognition method based on domain increment learning according to claim 1, characterized in that, The weighted similarity calculation is performed based on the updated domain-invariant prototype and domain-specific prototype corresponding to each category to obtain the comprehensive classification vector for each sample. The specific formula is as follows: ; in, and These represent the domain-invariant prototype and the domain-specific prototype corresponding to the c-th category after the update, respectively. For the current domain ID, For domain-invariant prototype weights, Domain-specific prototype weights, For temperature parameters, Indicates sample The L2 normalized form of the deep feature vectors, This represents the comprehensive classification vector corresponding to sample n.

5. A signal modulation recognition system based on domain increment learning, characterized in that, The system includes: The acquisition module is used to select any one of multiple domains as the current domain, and acquire current domain samples and historical domain samples; the current domain samples include multiple samples corresponding to the current domain, and the historical domain samples include multiple samples corresponding to the historical domain collected from the memory buffer, wherein the samples include modulation signals corresponding to different modulation types; The multi-level feature extraction module is used to input the current domain samples and historical domain samples into the residual convolutional neural network to perform multi-level feature extraction, and obtain the multi-level features corresponding to the current domain samples and the multi-level features corresponding to the historical domain samples. The total loss function calculation module is used to calculate the total loss function based on the multi-layer features corresponding to the current domain samples and the multi-layer features corresponding to the historical domain samples, combined with multi-level domain adversarial learning, cross-domain contrastive learning mechanisms and dual prototype management system. The first judgment module is used to minimize the total loss function through the Adam optimizer and determine whether the minimized function has reached the global minimum. If the global minimum has not been reached, the parameters of the residual convolutional neural network are updated and returned to the "multi-level feature extraction module". If the global minimum has been reached, the "second judgment module" is executed. The second judgment module is used to determine whether the learning domain is empty. If it is empty, the residual convolutional neural network updated last time in the current domain is used as the domain incremental learning model so that the domain incremental learning model can be used for signal modulation recognition in the future. If it is not empty, the residual convolutional neural network updated last time in the current domain is used as the residual convolutional neural network for the first multi-level feature extraction of the next domain, and an arbitrary domain is selected from the learning domain as the current domain. At the same time, the current domain samples and historical domain samples are obtained, and the results are returned to the "multi-level feature extraction module". The total loss function calculation module includes: The multi-level domain adversarial loss calculation unit is used to train the multi-level domain adversarial learning based on the multi-level features corresponding to the current domain sample and the multi-level features corresponding to the historical domain sample to obtain the multi-level domain adversarial loss value. The contrastive learning loss calculation unit is used to calculate the contrastive learning loss value based on the deep features corresponding to the current domain sample and the deep features corresponding to the historical domain sample using a cross-domain contrastive learning mechanism. The prototype comparison loss value calculation unit is used to determine the prototype comparison loss value based on the dual prototype management system. The total loss function calculation unit is used to construct the total loss function based on the multi-level domain adversarial loss value, the contrastive learning loss value, and the prototype contrastive loss value. The step involves training using multi-level domain adversarial learning based on the multi-level features corresponding to the current domain samples and the multi-level features corresponding to the historical domain samples, to obtain a multi-level domain adversarial loss value. Specifically, this includes: Calculate the maximum mean difference between the deep features corresponding to the current domain sample and the deep features corresponding to the historical domain sample; Calculate the domain similarity between the current domain sample and the corresponding historical domain sample based on the maximum mean difference value; Calculate the dynamic adversarial strength coefficient based on domain similarity; Domain discriminators are set on the third convolutional layer, the fifth convolutional layer, and the fully connected layer of the residual convolutional neural network, respectively; The multi-layer features corresponding to the current domain sample and the multi-layer features corresponding to the historical domain sample are respectively input into the corresponding discriminator, and multi-layer domain adversarial learning is used for training to obtain the domain adversarial loss value corresponding to each layer. Calculate the multi-level domain adversarial loss value based on the dynamic adversarial intensity coefficient and the corresponding domain adversarial loss value of each layer; The step involves calculating the contrastive learning loss value using a cross-domain contrastive learning mechanism based on the deep features corresponding to the current domain sample and the deep features corresponding to the historical domain sample. This specifically includes: The deep features corresponding to each sample in the current domain are concatenated with the deep features corresponding to each sample in the corresponding historical domain to obtain a concatenated feature matrix. The concatenated feature matrix is ​​processed for similarity to obtain the sample similarity matrix between each sample; A cross-domain contrastive learning mechanism is used to calculate the weighted positive sample matrix; The contrastive learning loss value is determined based on the sample similarity matrix and the weighted positive sample matrix; The determination of the prototype comparison loss value based on the dual prototype management system specifically includes: A dual-prototype management system is established based on the deep features and corresponding labels of the current domain samples, and the dual-prototype management system is used to determine the domain-invariant prototype and domain-specific prototype corresponding to each category. Calculate the mean of deep features for all samples in each category of the current domain; A momentum update strategy is adopted, which uses the mean of deep features of all samples in each category of the current domain to update the domain-invariant prototype and domain-specific prototype corresponding to each category respectively; Based on the updated domain-invariant prototype and domain-specific prototype, a weighted similarity calculation is performed to obtain the comprehensive classification vector for each sample. The prototype contrastive loss value is obtained by comparing the comprehensive classification vector corresponding to each sample with the true label.

Images