Passive cross-individual intent recognition method based on domain-invariant features

Abstract

This invention discloses a passive cross-individual intention recognition method based on domain-invariant features, belonging to the field of EEG signal recognition technology. The invention first pre-trains the model using labeled source domain EEG information, and then re-adapts the model using unlabeled target domain EEG signals. The invention introduces a channel attention mechanism in the feature extractor module, concentrating the optimization space on a few feature intervals to obtain more valuable feature vector representations. During the domain adaptation training phase, based on the established feature decoupling module, the model focuses more on domain-invariant features related to relevant potential difference information in the EEG signal, while ignoring non-stationary characteristics strongly correlated with the individual subject, thus forming a strong correlation between features and the task. Simultaneously, considering the strong similarity between different intention activities in the intention recognition task, an adaptive feature regularization term is proposed to reduce the possibility of feature space confusion caused by similar intention activities.

Description

Technical Field

[0001] This invention belongs to the field of EEG signal recognition technology, specifically relating to a passive cross-individual intention recognition method based on domain-invariant features. Background Technology

[0002] Brain-computer interface (BCI) is a rapidly developing field that aims to establish direct communication between the human brain and external devices. BCI technology extracts information about a person's intentions or mental state by recording and analyzing brain signals. This information is then used to control devices such as computers, robotic devices, and prosthetics. The potential applications of BCI are vast, ranging from medical applications to games and entertainment. For people with disabilities, BCI technology has the potential to restore their independence and improve their quality of life. In the medical field, BCI has the potential to improve the diagnosis and treatment of neurological diseases and provide new insights into the workings of the human brain. Furthermore, research on brain-computer interfaces helps advance our understanding of the human brain and its functions. In summary, brain-computer interfaces have enormous potential to improve people's lives, advance our understanding of the brain, and provide new insights into the diagnosis and treatment of neurological diseases. Among these applications, intention recognition algorithms play a crucial role. BCI uses intention recognition algorithms to detect and interpret the user's intended actions and commands based on their brain activity. The accuracy and reliability of the intention recognition algorithm determine the effectiveness of the BCI system.

[0003] In recent years, the research and development of intent recognition algorithms in brain-computer interfaces (BCIs) has been an active research area. This research aims to improve the accuracy and reliability of BCI systems in detecting and interpreting users' expected actions or commands based on their brain activity. Since neurons in the human brain generate EEG signals during control, movement, and sensory functions, the intended activity can be obtained by analyzing the EEG signals generated by intentional activity. Deep learning, due to its automatic feature extraction and ability to handle high-dimensional data and nonlinear relationships, has made significant progress in data-driven deep models for intent recognition. However, due to individual differences and the non-stationarity of EEG signals, models built on existing individual data may not be applicable to new users, resulting in significant performance degradation. This poses a challenge to designing a universal intent recognition model for all users.

[0004] In brain-computer interface devices, due to privacy concerns regarding subject EEG data, access to source domain-adapted cross-individual intention recognition methods is typically unavailable. Domain adaptation allows for the formation of a decision boundary for the intention task, aligning the data domains of new and old subjects. This mitigates the problem of poor adaptability and decreased accuracy of existing models on new users due to user variability. A cross-individual intention recognition method based on passive domain adaptation, without using source domain data, enhances the model's generalization ability. Experiments on two publicly available datasets validate the superiority of this method.

[0005] Cross-individual intent recognition based on domain adaptation technology achieves accurate identification of the target individual's intent task by utilizing labeled data from known individuals and unlabeled data from the target individual. The key to domain adaptation technology's ability to solve cross-individual problems lies in its requirement for access to data from both the source and target domains to align their distributions. However, in certain scenarios, access to source domain data is unavailable; only models trained on the source domain are available. In such cases, traditional domain adaptation methods cannot achieve domain transfer. In fact, this limitation frequently exists in cross-individual intent scenarios. Due to the privacy concerns surrounding individual EEG signal data and the memory limitations of brain-computer interfaces, in many situations, it is impossible to obtain or store this data containing private information, leaving only a model trained on private data.

[0006] Source-Free Domain Adaptation (SFDA) technology has attracted widespread attention because it can fine-tune target domain data without requiring data from the source domain. Traditional unsupervised domain adaptation typically calculates the differences between two domains using adversarial or metric methods, and then optimizes this difference space to achieve knowledge transfer. The training process for SFDA usually involves two steps: first, pre-training an initial model using source domain data to obtain an initial model, and then fine-tuning this model using target domain data for knowledge transfer.

[0007] Since passive domain adaptation can only fine-tune knowledge using unlabeled data from the target domain, current research approaches mainly fall into three categories: data augmentation, pseudo-label correction, and model structure optimization. Data augmentation refers to using generative models such as Generative Adversarial Networks (GANs) or diffusion models to obtain additional training samples, thereby achieving a more generalized feature space and a compact decision boundary. Existing SFDA methods focus on embedding prior distribution information into the source model to achieve a certain degree of knowledge transfer.

[0008] Currently, passive domain adaptation technology has received widespread attention and achieved encouraging results in the field of computer vision. However, in the privacy-conscious biomedical field, research on intention recognition applications based on physiological signals remains lacking. Although existing SFDA methods can address the knowledge transfer problem in the target domain to some extent, these methods fail to consider the following two issues inherent in intention recognition scenarios, leading to poor performance:

[0009] (1) Each individual's collected EEG signals have unique non-stationary characteristics (non-stationarity refers to changes in signal distribution parameters or distribution laws over time). In cross-individual intention recognition tasks, existing SFDA methods do not consider the impact of this non-stationary characteristic when performing knowledge transfer. Treating this characteristic as knowledge transfer causes unnecessary bias in the model's feature extraction, resulting in a significant decrease in the performance of these SFDA methods. In fact, transfer models should focus on the event-related potential (ERP) information in EEG signals caused by cognitive behavior, while ignoring the non-stationary characteristics caused by individual differences. Unsupervised domain adaptation has long utilized data from the source and target domains for domain-invariant feature (ERP) learning, but this requires simultaneous access to data from both the source and target domains, which is impossible in passive scenarios.

[0010] (2) In passive scenarios, due to the lack of source domain data, it is difficult to align the distributions of the two domains. Therefore, SFDA is usually defined as an unsupervised problem. Different intent activities in intent recognition tasks have strong similarities. Existing SFDA methods ignore the impact of this similarity on the feature space, resulting in feature confusion. Summary of the Invention

[0011] The purpose of this invention is to address the aforementioned problems by proposing a passive cross-individual intention recognition method based on domain-invariant features within a source-free intention recognition (SF-IR) framework, thereby reducing the impact of non-stationary characteristics of EEG signals.

[0012] The technical solution adopted in this invention is as follows:

[0013] A passive cross-individual intent recognition method based on domain-invariant features includes the following steps:

[0014] Step S1: Collect EEG signal data from several users in different intention tasks using an EEG signal acquisition device, and preprocess the collected EEG signal data from each user to obtain multiple data samples for each user and the intention activity label for each data sample. Use all data samples of the same user as the training set for the current user to obtain several training sets that constitute the source domain EEG signal dataset.

[0015] The data preprocessing involves: extracting M seconds of EEG signal data from the intention task as a data sample, and setting the true classification label for each dataset based on the user's actual intention activity task, where M is a preset value;

[0016] Step S2: Pre-train the intention recognition model based on the source domain EEG signal dataset;

[0017] The intent recognition pre-trained model includes: a feature extractor, a feature decoupling module, and a classifier;

[0018] Feature extractors are used to extract initial features from EEG signals;

[0019] The feature decoupling module is used to extract domain-invariant and domain-related features from EEG signals. The input of the feature decoupling module is the initial features of the EEG signal, and the feature decoupling module contains a domain discriminator for performing domain discrimination on the domain-invariant and domain-related features.

[0020] The classifier is used to output the category for intent recognition, and the input of the classifier is the domain-invariant features output by the feature decoupling module;

[0021] Step S3: The pre-trained intention recognition model is subjected to domain adaptation training based on the unlabeled target domain EEG signal dataset; the feature extractor, the network for outputting domain-invariant features in the feature decoupling module, and the classifier after domain adaptation training constitute an intention recognizer for the target object.

[0022] Furthermore, the feature extractor is a feature extractor based on the channel attention mechanism.

[0023] Furthermore, the backbone network of the feature extractor is a residual neural network, and a channel attention mechanism is introduced on the basis of the residual neural network.

[0024] Furthermore, the feature decoupling module includes two autoencoders with identical network structures and a domain discriminator. The autoencoder includes a convolutional layer and two parallel fully connected layers. The input feature map of the two fully connected layers is the output feature map of the convolutional layer. One fully connected layer is used to obtain the variance of its input feature map, and the other fully connected layer is used to obtain the mean of its input feature map. The extracted mean and variance are concatenated to obtain domain-invariant features or domain-related features. The input of the domain discriminator is the domain-invariant features and the domain-related features.

[0025] The technical solution provided by this invention brings at least the following beneficial effects:

[0026] To address the impact of non-stationarity on feature shift, this invention proposes a deep neural network pre-trained model (intent recognition model) incorporating a decoupled generalization module (i.e., a feature decoupling module) to focus on EEG signal domain invariant features (ERPs). Simultaneously, it uses source domain data for pre-training and generalization, and introduces a convolutional channel attention mechanism to force the neural network to focus on specific channels that should be prioritized. To address the issue of feature confusion related to similar intent activities, this invention proposes an adaptive feature regularization term during the adaptation phase to reduce classification confusion. This regularization term encourages neighboring features in the feature space to have similar predictions by assigning the same label to similar features, while simultaneously reducing the negative impact of noisy neighbor samples through self-regularization loss, thereby achieving both efficient feature clustering and cluster assignment. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a schematic diagram of the network structure of the intent recognition model used in the embodiments of the present invention;

[0029] Figure 2 This is a schematic diagram of the channel attention principle used by the feature extractor in this embodiment of the invention;

[0030] Figure 3 This is a schematic diagram of the feature decoupling module Dis used in this embodiment of the invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0032] This invention discloses a passive cross-individual intent recognition method based on domain-invariant features. For the specific scenario of intent recognition, this invention proposes a novel passive framework for cross-individual intent recognition to address issues such as privacy protection and device memory limitations in brain-computer interfaces. Specifically, this invention can be used to solve the technical problems of feature shift and discrete features in passive frameworks.

[0033] In this embodiment of the invention, the first step is to collect electroencephalogram (EEG) signals, which include source domain EEG signals and target domain EEG signals. The source domain EEG signals are labeled datasets (with individual intention category labels), while the target domain EEG signals are unlabeled datasets.

[0034] One possible implementation involves acquiring the dataset as follows: EEG signal data from several users during different intention tasks is collected using an EEG signal acquisition device (e.g., an EEG cap). The collected EEG signal data from each user is preprocessed (M seconds of EEG signal data from the intention task are extracted as a data sample, and a true classification label is set for each dataset based on the user's actual intention activity, where M is a preset value). This yields multiple data samples for each user and the intention activity label for each data sample. All data samples from the same user are used as the training set for that user, resulting in the source domain EEG signal for each user. The acquisition of the target domain EEG signal is similar. To verify the recognition performance of the trained recognition model, the intention task of the target domain EEG signal can also be recorded. Furthermore, the training, testing, and validation datasets for the recognition model can also be obtained from publicly available datasets.

[0035] In this embodiment of the invention, the EEG signals used for source domain training are defined as the source domain, which contains a total of n s 1 labeled sample, training set The corresponding tag set is P(x s ,y s ) represents the source distribution, x s Represents source domain data, y s Indicates the source domain tag set, This represents the EEG signal of the i-th source domain individual. Simultaneously, a new subject is defined as the target domain containing n... t Unlabeled EEG samples P(x t ) represents the target domain distribution, x t Represents target domain data. This represents the EEG signal of the i-th target domain individual.

[0036] In this embodiment of the invention, the intent is to identify a pre-trained model such as Figure 1 As shown, it includes a feature decoupling module Dis, a classifier F, and a feature extractor g, wherein the feature decoupling module Dis includes a domain discriminator D. During the pre-training phase, it can only access the source domain n. sThe purpose of the pre-training phase is to train a feature extractor g and a classifier F for a specific model using labeled source domain data. During the domain adaptation phase, only the feature extractor g and classifier obtained in the pre-training phase are accessible. A regularization term is used to enable the pre-trained model to achieve high classification accuracy on the target domain data. Specifically, this embodiment of the invention introduces an adaptively adjusted feature clustering loss to control the formation of clear decision boundaries in the feature space, forcing ambiguous samples to be classified into the correct feature space.

[0037] Feature extractor g is used to extract initial features of the EEG signal. Inspired by the channel attention mechanism, it is introduced as a component into the feature extractor, forcing the neural network to automatically assign weights to different channels, allowing the feature extractor to automatically focus on channels that improve model performance. This embodiment introduces the channel attention mechanism based on the ResNetBlock module (residual neural network module) to improve the network's targeted learning ability. The channel attention module consists of three basic operations: F sq F ex and F scale ,refer to Figure 2 The three basic operations are described in detail below:

[0038] (1)F sq The operation uses global pooling to compress each channel of the feature map into a real number, providing a global view of the entire channel, F. sq The operation can be represented as follows:

[0039]

[0040] Among them, Z c The feature map is obtained after global pooling, u c The feature map of length L corresponds to each channel.

[0041] (2)F ex The operation reduces and restores the number of channels by using two 1-D convolutions (with a kernel size of 1).

[0042] Replacing linear layers with two convolutional layers maximizes the degree of nonlinear fitting with a minimum number of parameters. This allows for adaptive acquisition of importance weights for each channel during training, Fi. ex The operation can be represented as follows:

[0043] s = F ex (zW)=σ(g(zW))=σ(W2δ(W1z))

[0044] Where C×H×W represents the dimensionality information of the feature map (C represents the number of channels, H×W represents the feature map size), W1 and W2 represent the dimensionality reduction and dimensionality increase of the convolution, respectively, and z represents the result of F sq The operation generates an array of real numbers. (By F) sq The generated real number array with channel number C is restored after convolution. Then restore C. After obtaining the sigmoid activation function (σ), we obtain a weighted array of the original feature maps, where δ() represents the output of the activation function.

[0045] (3)F scale The operation multiplies the weighted array with the corresponding feature map channels to obtain the final result, making the network focus more on the key components in the multi-channel feature maps. scale The operation can be represented as follows:

[0046]

[0047] That is, the final feature map X c The feature map (uc) of the channel is multiplied by the weighted value (sc) of the corresponding channel.

[0048] This embodiment improves upon the ResNetBlock model by incorporating multiple convolutional channels composed of one-dimensional convolutions with kernel sizes of 1 and 3. This significantly reduces the number of training parameters during the convolution process while maintaining the expressiveness of the multiple convolutional channels. The fully connected classifier (Classifier C) uses a Global Average Pooling (GAP) layer to receive high-level representation features from the convolutions. By averaging the values ​​of each channel, the GAP layer reduces each feature vector to a single number, greatly reducing the number of parameters, enhancing training stability, and minimizing model overfitting.

[0049] Pre-training decoupling module: In order to force the network model to separate the domain-invariant features that are strongly related to the intent recognition task from the relatively confusing feature distribution of the source domain, a decoupling module is introduced into the pre-trained model. This module can greatly help the pre-trained model improve this content discrimination ability.

[0050] As one possible implementation, such as Figure 3As shown, the decoupling module of this embodiment is designed with reference to Variational Autoencoders (VAEs). It utilizes two latent space vectors to generate two feature spaces: one containing domain-irrelevant features and the other containing domain-specific features. Specifically, the decoupling module includes two autoencoders with identical network structures and a domain discriminator D. Each autoencoder includes a convolutional layer and two parallel fully connected layers. The input feature maps of the two weighted connected layers are the output feature maps of the convolutional layers. One fully connected layer is used to obtain the variance of its input feature map, and the other fully connected layer is used to obtain the mean of its input feature map. The extracted mean and variance are concatenated to obtain the domain-irrelevant or domain-specific features. In other words, one encoder outputs domain-irrelevant features, and the other outputs domain-specific features. The initial parameters of the two autoencoders are the same, but the network parameters change differently during training due to the presence of the domain discriminator D.

[0051] The overall optimization function approach is as follows:

[0052] (1) Ensure that the extracted domain-invariant features can confuse the domain discriminator (binary classification).

[0053] (2) Ensure that the extracted domain-related features can be accurately identified by the domain discriminator as to which domain they come from.

[0054] (3) Finally, domain-invariant features are used as input to the intention recognition task classifier.

[0055] Domain-invariant feature g 1 The corresponding Softmax score p 1 The corresponding domain discriminator outputs the probability score as follows: Domain-related features g 2 The corresponding Softmax score p 2 The corresponding domain discriminator outputs the probability score as follows: The overall optimization results are as follows:

[0056]

[0057] Among them, L cls () represents the optimization loss (classification loss) of classifier C. In this embodiment, cross-entropy loss is used. L represents the softmax score of the domain-invariant feature of sample i. dls Let D represent the optimization loss of the domain discriminator. In this embodiment, the optimization loss of the domain discriminator D is as follows:

[0058]

[0059] in, Let D() represent the mathematical expectation over the source domain, D() represent the output of the domain discriminator D, and KL() represent the standard KL (Kullback-Leibler) divergence formula. This represents the domain-related features of sample i.

[0060] During the domain adaptation training phase, this embodiment of the invention utilizes unsupervised clustering loss to re-cluster the feature space. Unlike directly trying to keep the labels of samples with similar features consistent, this embodiment categorizes samples into high-affinity and low-affinity groups based on defined sample structural affinity. This affinity is then used to weight the loss, ensuring that high-affinity samples maintain as consistent a label as possible, while keeping low-affinity samples as distinct a label as possible. In this way, the model can clearly partition the feature space as much as possible during the adaptation phase.

[0061] First, in order to find the semantically nearest neighbor, the nearest neighbor sample can be identified based on the Euclidean distance, cosine similarity, etc. between the samples.

[0062] A feature library for storing target features was constructed: Where, f(x) i The ) represents the features extracted by the feature extractor (i.e., the initial features of the EEG signal). The target domain dataset was represented, and then a score record module was created. This is used to store the corresponding softmax prediction scores (i.e., classification probability values), where c() represents the output of the classifier.

[0063] Then, a weight is assigned to each neighboring sample. If j and i are each other's nearest neighbors, then a weight of 2 is assigned. If the two are not nearest neighbors, then a small value (w) between 0 and 1 is assigned to control the label attributes of the two samples. The formula is defined as follows:

[0064]

[0065] Among them, V ij This represents the weight of the j-th nearest neighbor of the i-th sample. Let K represent the nearest neighbor set of sample i, and K represent the number of nearest neighbors of sample i. Let M represent the nearest neighbor set of sample j, and M represent the number of nearest neighbors of sample j.

[0066] For each sample x i In the feature library Find its k nearest neighbor samples and maximize the weighted dot product of their scores (sample scores are derived from the score table S). In this way, samples that are close to each other have high output similarity.

[0067] The specific formula is defined as follows:

[0068]

[0069] Where, n t Indicates the number of samples in the target domain. This represents the nearest neighbor set of sample i. p represents the predicted score of the k-th nearest neighbor sample (note that this term is a constant vector), and the superscript T indicates transpose. i V represents the predicted score of the i-th sample. ik This represents the weight of the k-th nearest neighbor of the i-th sample.

[0070] Meanwhile, optimizations were made to the above regularization terms to further reduce the nearest neighbor (N) value. k To mitigate the potential impact of samples in the dataset, a self-regulation term is introduced, forcing samples not to ignore the current prediction of their own features. The specific formula is as follows:

[0071]

[0072] Among them, S i p represents the predictions stored in the score database (note that this term is a constant vector). i This represents the predicted probability of the sample after applying Softmax. By maximizing the dot product of the output score and the predicted probability for each sample, it's easy to see that the dot product is maximized when the score is a one-hot vector, forcing the sample to prioritize the current prediction of its own features. Furthermore, inspired by the passive domain adaptation method SHOT, a diversification loss is introduced, forcing the model to output predictions for multiple categories. The specific formula is as follows:

[0073]

[0074] Where JS() represents the Jensen-Shannon divergence, and C represents the number of intent categories. Let q represent the classification prediction probability that sample i is of the c-th intention category. c This indicates a fixed distribution.

[0075] Therefore, the entire adaptation phase is obtained as follows:

[0076]

[0077] This invention addresses the issue of neglecting the influence of unique non-stationary characteristics of individual subjects on unbiased feature transfer in passive cross-individual intent recognition processing. It proposes a passive cross-individual intent recognition method based on domain-invariant features (SF-IR), which consists of two parts. First, the defined source model (intent recognition model) is pre-trained using labeled source domain EEG information. Then, the previously obtained pre-trained model is re-adapted using unlabeled target domain EEG signals.

[0078] As one possible implementation, in this embodiment of the invention, during the pre-training phase, the feature extractor g, feature decoupling module Dis, and classifier C of the intent recognition model are trained based on labeled source domain EEG signals, based on a preset maximum number of pre-training rounds. The training stops when the maximum number of pre-training rounds is reached. In the domain adaptation training phase, the extractor g, feature decoupling module Dis, and classifier C of the pre-trained intent recognition model are re-adapted based on a preset maximum number of domain adaptation training rounds and unlabeled target domain EEG signals. The training stops when the maximum number of domain adaptation training rounds is reached. The autoencoder used to output domain-invariant features in the trained feature extractor g and feature decoupling module Dis, along with the classifier C, constitute the intent recognizer for the target object.

[0079] In summary, the passive cross-individual intent recognition method based on domain-invariant features provided in this invention has the following advantages:

[0080] (1) Introduce a channel attention mechanism in the feature extractor. This module can enhance the model’s targeted learning ability on the basis of ResNetBlock, concentrate the optimization space on a few feature intervals, and obtain more valuable feature vector representations.

[0081] (2) In the domain training phase of the model, this chapter designs a feature decoupling module to make the model pay more attention to the domain-invariant features related to ERP, while ignoring the non-stationary characteristics that are strongly correlated with the individual subjects, thereby forming a strong correlation between features and tasks.

[0082] (3) At the same time, since different intention activities in the intention recognition task have strong similarities, an adaptive feature regularization term (AFR) is proposed to reduce the possibility of feature space confusion caused by similar intention activities.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0084] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A passive cross-individual intent recognition method based on domain-invariant features, characterized in that, Includes the following steps: Step S1: Collect EEG signal data from several users in different intention tasks using an EEG signal acquisition device, and preprocess the collected EEG signal data from each user to obtain multiple data samples for each user and the intention activity label for each data sample. Use all data samples of the same user as the training set for the current user to obtain several training sets that constitute the source domain EEG signal dataset. The data preprocessing involves: extracting M seconds of EEG signal data from the intention task as a data sample, and setting the true classification label for each dataset based on the user's actual intention activity task, where M is a preset value; Step S2: Pre-train the intention recognition model based on the source domain EEG signal dataset; The intent recognition pre-trained model includes: a feature extractor, a feature decoupling module, and a classifier; Feature extractors are used to extract initial features from EEG signals; The feature decoupling module is used to extract domain-invariant and domain-related features from EEG signals. The input of the feature decoupling module is the initial features of the EEG signal, and the feature decoupling module contains a domain discriminator for performing domain discrimination on the domain-invariant and domain-related features. The classifier is used to output the category for intent recognition, and the input of the classifier is the domain-invariant features output by the feature decoupling module; Step S3: The pre-trained intention recognition model is subjected to domain adaptation training based on the unlabeled target domain EEG signal dataset; the feature extractor, the network for outputting domain-invariant features in the feature decoupling module, and the classifier after domain adaptation training constitute an intention recognizer for the target object. The feature decoupling module includes two autoencoders with identical network structures and a domain discriminator. The autoencoders consist of a convolutional layer and two parallel fully connected layers. The input feature maps of the two fully connected layers are the output feature maps of the convolutional layers. One fully connected layer is used to obtain the variance of its input feature map, and the other fully connected layer is used to obtain the mean of its input feature map. The extracted mean and variance are concatenated to obtain domain-invariant features or domain-related features. The input of the domain discriminator consists of domain-invariant features and domain-related features.

2. The method as described in claim 1, characterized in that, The feature extractor is a feature extractor based on the channel attention mechanism.

3. The method as described in claim 2, characterized in that, The backbone network of the feature extractor is a residual neural network, and a channel attention mechanism is introduced on the basis of the residual neural network.

4. The method according to any one of claims 1 to 3, characterized in that, The loss during pre-training is set as follows: Among them, L pre L represents the pre-training loss. cls () represents the classification loss of the classifier. L represents the classification probability value of the domain-invariant features of sample i in the classifier. dls The discriminant loss of the domain discriminator is expressed as: in, Let D() represent the mathematical expectation over the source domain, D() represent the output of the domain discriminator, and KL() represent the standard Kullback-Leibler divergence. This represents the domain-related features of sample i. The output probability score of the domain discriminator with respect to the domain-invariant features.

5. The method as described in claim 4, characterized in that, The classification loss is the cross-entropy loss.

6. The method according to any one of claims 1 to 3, characterized in that, The loss during domain adaptation training is set as follows: in, Indicates domain adaptation training loss, Indicates diversified losses, Indicates the loss due to the regularization term. The expressions for the regularization term loss are as follows: Where JS() represents the Jensen-Shannon divergence, and C represents the number of intent categories. This represents the classification prediction probability of sample i being the c-th intent category, with a fixed distribution. Where, n t Indicates the number of samples in the target domain. This represents the nearest neighbor set of sample i. This represents a constant vector consisting of the predicted scores of the k-th nearest neighbor sample, with the superscript T indicating transpose. i V represents the predicted score of the i-th sample. ik The weight of the k-th nearest neighbor of the i-th sample is given: if the two samples are each other’s nearest neighbors, the weight is set to 2; otherwise, the weight is a value between 0 and 1.

Images