Reference domain mixed drift alignment and conditional constraint characterization system for cross-domain eeg signals

Abstract

The application discloses a reference domain mixed drift alignment and conditional constraint representation system of cross-domain electroencephalogram signals, belongs to the technical field of electroencephalogram signal processing, and solves the problems that existing electroencephalogram preprocessing and potential source separation methods are difficult to process domain-specific mixed drift, inconsistent component order and symbol and shared potential representation failure under cross-subject or cross-acquisition period conditions. The system comprises an intra-domain standardization module, a reference domain mixed drift alignment module, a shared period conditional potential source representation module and a joint optimization module. Under the condition of first-order linear mixed drift enhancement, the system can improve the source recovery quality, cross-domain component order consistency and common observation space consistency; in real cross-subject and cross-acquisition period electroencephalogram experiments, the system can improve the representation separability on the training visible domain, and can be used for representation regularization and preprocessing alignment of cross-domain electroencephalogram signals.

Description

Technical Field

[0001] This invention belongs to the field of EEG signal processing technology, specifically relating to a reference domain hybrid drift alignment and conditional constraint characterization system for cross-domain EEG signals. Background Technology

[0002] EEG, with its non-invasive nature and easily portable acquisition devices, is frequently used in neuro-assisted medicine and human-computer interaction scenarios. In these applications, the model typically needs to process data from different subjects or different acquisition sessions from the same subject. However, due to differences in skull models, electrode placement, individual variations in neural rhythms, and physiological changes during acquisition, the same type of neural activity can exhibit different observational forms in different data domains (subject / session). These cross-subject and cross-acquisition-session differences not only manifest as changes in statistical distribution but may also represent changes in mixing mechanisms within the observation space, thus affecting the consistency of cross-domain EEG representations.

[0003] Current EEG preprocessing typically includes steps such as rereference, filtering, downsampling, handling of bad segments and channels, segmentation, baseline correction, and artifact removal. These steps primarily aim to reduce the impact of noise from electrooculography (EOG), electromyography (EMG), and power line interference, thereby improving signal quality within a single data domain. However, in scenarios involving multiple subjects or different acquisition time periods, inter-domain differences may manifest as inconsistencies in observation coordinates, component correspondences, and mixing methods. Relying solely on conventional preprocessing makes it difficult to guarantee direct comparison of the representations of the same potential neural activity across different domains, which can also affect the stability of subsequent classification, decoding, or latent source analysis.

[0004] Blind source separation methods are commonly used in EEG signal preprocessing to isolate latent source components and aid in artifact identification. Independent component analysis (ICA) is a typical method that estimates independent latent components from multichannel EEG observations through linear unmixing. This method can be used for artifact removal and component interpretation within a single data domain, but its source recovery results often suffer from permutation, sign, and scale uncertainties. In cross-subject or cross-acquisition time-slot applications, components obtained from independent unmixing in different domains may exhibit inconsistencies in position, sign, and scale, leading to cross-domain component mismatches and reducing the reliability of latent source alignment and shared representation construction.

[0005] To improve the expressive power of latent source separation, nonlinear latent source separation methods such as conditionally discriminative variational autoencoders have been used to model more complex observation-generated relationships. These methods introduce auxiliary conditions into the latent variable modeling process, offering certain advantages in nonlinear latent source recovery. However, in multi-domain EEG scenarios, these methods often implicitly contain shared or fixed observation mixing mechanisms. When domain-specific mixing drift exists in data across subjects or across acquisition time periods, the fixed mixing hypothesis may be inconsistent with the actual observation process, leading to component mismatch between latent sources in the target domain and those in the reference domain, thus affecting the stability of shared latent representations.

[0006] Therefore, cross-domain EEG signal processing still requires a technical solution oriented towards observation representation alignment: on the one hand, using a reference domain as a common coordinate benchmark to address the positional, sign, and scale ambiguities arising from domain-by-domain component estimation; on the other hand, explicitly correcting the observation space mixing drift, which can be approximated by a first-order linear transformation, and learning the source-related structures common to different domains under shared time-segment constraints. Through these processes, the consistency of EEG representations across subjects and across acquisition time periods can be improved, providing a more stable input foundation for subsequent EEG decoding, classification analysis, or neural signal representation learning. Summary of the Invention

[0007] To address the challenges of existing EEG preprocessing and latent source separation methods in handling domain-specific mixture drift, component position inconsistencies with symbols, and shared latent representation failures across subjects or acquisition time periods, this invention provides a reference domain mixture drift alignment and conditional constraint representation system for cross-domain EEG signals, specifically comprising: Intradomain normalization module, reference domain hybrid drift alignment module, shared time-segment conditional potential source characterization module, and joint optimization module; The intra-domain standardization module is used to scale-align the EEG observation matrices of each domain. The reference domain hybrid drift alignment module includes a reference domain component anchoring module and a reference-guided linear hybrid drift adapter. The reference domain component anchoring module uses the reference domain as a reference, establishes cross-domain component correspondences through component matching and sign correction, and provides a weak anchoring reference for subsequent observation space alignment. Based on the weak alignment prior, the reference-guided linear hybrid drift adapter learns a first-order linear alignment matrix constrained by the reference domain for each domain, and maps the standardized EEG observations of each domain to a common observation coordinate system to correct the linear drift of the observation space relative to the reference domain. The shared time-segment conditional latent source representation module is used to learn cross-domain shared source-related latent representations. The joint optimization module is used to train the system and output aligned observation representations and latent source-related representations.

[0008] Furthermore, the standardization module calculates the mean and standard deviation of the EEG observation data for each domain according to the channel dimension and / or time dimension, performs mean removal and variance normalization, and obtains a standardized EEG observation matrix. It also determines auxiliary conditions for module collaboration for each sample, including domain index conditions and shared time period conditions. The domain index conditions are directly obtained based on the data domain to which the sample belongs: the reference domain is denoted as... The target domains are denoted as follows: The shared time period conditions are obtained based on the experimental procedures common to each domain: if the data already contains experimental phases, stimulus periods, time window numbers, or task phase labels, then the corresponding labels are read directly; if no explicit labels are provided, then the observation sequences of each domain are arranged according to a uniform window length. Divided into The time period, and the first The time period number of each sample is used as... , by all The time period numbers of each sample constitute a shared time period condition sequence. .

[0009] Furthermore, the specific processing flow within the reference domain component anchoring module is as follows: S31, regarding the first Standardized EEG observation matrix for each data domain The fast independent component analysis algorithm is used to compute the estimated component representation within the domain. ; S32. Utilizing cross-domain shared time-series conditions right The component sequences are uniformly segmented, and the conditional variance characteristics of each component at different shared time periods are calculated. ; S33. Match the component variance features of the target domain with the corresponding features of the reference domain to obtain the matching cost matrix. The optimal permutation matrix relative to the reference domain is obtained by applying the Hungarian algorithm. ; S34. Based on the optimal permutation matrix Construct a symbolic diagonal matrix using the inner product of the components of the reference field. Thus, the weak anchoring component representation is obtained. .

[0010] further, ; In the formula, Indicates the first The first data field Variance characteristics of each ICA component across different sharing time periods; Indicated in the data domain The first part, obtained by the fast independent principal component analysis algorithm, is... Each independent ICA component sequence; This indicates the calculation of conditional variance, i.e., the calculation of... Under the conditions The variance of the sequence.

[0011] Furthermore, diagonal matrix In the middle, the first Line number The elements in the column are : Thus, the weakly anchored component representation is obtained: .

[0012] Furthermore, the specific processing flow in the reference guided linear hybrid drift adapter is as follows: S61, regarding the first For each data field, its linear alignment matrix is ​​defined as: ,in, Represents a linear alignment matrix shared across domains. Indicates the first Each data field has a domain-specific offset relative to the shared transformation; S62, Calculate the first Alignment observation characterization of data domains : .

[0013] further, and The solution process is as follows: by For reference, Ridge regression estimation is performed on the linear mapping to the weakly anchored component space: ; In the formula, Indicates the first The initial linear alignment matrix for each data field; Let be the linear mapping matrix to be estimated; The number of components to be retained; Denotes the Frobenius norm; The ridge regularity coefficient is used to restrict... This increases the magnitude of the solution and improves the stability of closed-form solutions. , .

[0014] Furthermore, the operations performed in the shared time period conditional latent source characterization module include, in sequence, shared time period conditional embedding, conditional prior modeling, conditional posterior inference, latent variable reparameterization sampling, and aligned observation reconstruction; Conditional embedding of shared time periods: embedding shared time period tags Input the conditional embedding layer to obtain the corresponding shared time period conditional embedding representation; Conditional prior modeling: Based on shared time-segment conditional embedding, the prior distribution parameters of latent variables, including mean and standard deviation, are generated through a conditional prior network, and the conditional prior distribution is defined; Conditional posterior inference: Aligning observation samples with shared time period labels Input the conditional posterior network to obtain the posterior distribution parameters, and thus define the conditional posterior distribution; Latent variable reparameterization sampling: Based on the conditional posterior distribution, latent variables are obtained through reparameterization, and backpropagation-capable latent samples are constructed using random noise and posterior distribution parameters; for the ... The latent variables of all samples in the domain are arranged to obtain the latent source correlation representation. ; Alignment observation reconstruction: latent variables in The input is used to decode and reconstruct the network, which obtains the reconstructed distribution of the aligned observation samples and generates the reconstruction results. .

[0015] Furthermore, when the joint optimization module trains the system, the total loss function used is: ,in, The training loss function for the shared time-segment conditional latent source representation module is used to learn cross-domain shared latent representations under shared time-segment conditions. The anchoring consistency loss function of the shared time-period conditional latent source representation module is calculated by standardizing the shared latent representation output by the shared time-period conditional latent source representation module and the weakly anchored components output by the reference domain component anchoring module on a component-by-component basis. It is used to keep the two components consistent. , and These are respectively used to limit the deviation of the alignment matrix from the initialization direction, the domain-specific offset magnitude, and the alignment matrix degradation within the reference-guided linear hybrid drift adapter. , , and These are non-negative weighting coefficients, which control the relative strength of the above constraint terms in the overall objective.

[0016] Furthermore, the specific process of the characterization system in operation is as follows: Step 1: Acquire EEG observation data from multiple domains and determine a reference domain and at least one target domain; the domains are data domains corresponding to different subjects, different acquisition time periods of the same subject, or different batches of acquisition from different devices; Step 2: Perform intra-domain standardization on the EEG observation data of each domain using the intra-domain standardization module to obtain the standardized observation matrix, and determine the domain index conditions and shared time period conditions for each sample; Step 3: Use the domain index condition for front-end reference domain hybrid drift alignment, and use the shared time period condition for back-end shared time period condition latent source representation learning; Step 4: Estimate the intra-domain components of standardized observations in each domain using the reference domain component anchoring module, calculate the component statistical characteristics based on the shared time period conditions, and complete component position matching and sign correction relative to the reference domain. Step 5: Using the reference-guided linear hybrid drift adapter, initialize the linear alignment matrix based on the weak anchoring components output by the reference domain component anchoring module, and perform first-order linear hybrid drift alignment on each domain observation through shared linear basis and domain-specific drift terms; Step 6: Through the shared time-segment conditional latent source representation module, align the observation representation with the shared time-segment conditional input conditional prior network, conditional posterior network, and decoding reconstruction network to learn cross-domain shared latent source related representations; Step 7: Train the system using the joint optimization module and output aligned observation representations and potential source correlation representations.

[0017] The beneficial effects of the system are as follows: Using the reference domain as a unified coordinate benchmark, the uncertainty of position and sign generated by domain-by-domain component estimation is corrected, and combined with linear alignment of the observation space, the component mismatch caused by the failure of the fixed mixing hypothesis in multi-domain EEG scenarios is reduced.

[0018] By using reference domain anchoring and first-order linear alignment, the spatial mixing drift of observations across subjects and across acquisition periods is corrected, so that EEG observations from different domains are mapped to a more consistent common representation space.

[0019] A divide-and-conquer strategy is adopted, using domain index conditions and shared time period conditions. The domain index conditions are used for front-end domain-specific observation alignment, while the shared time period conditions are used for back-end shared latent representation learning, thereby reducing the interference of domain-specific information on the shared latent representation.

[0020] Experimental results show that, under first-order linear mixed drift enhancement conditions, the system described in this invention can improve source recovery quality, cross-domain component positional consistency, and common observation space consistency. In real cross-subject and cross-acquisition time period EEG experiments, the system described in this invention can improve the representational separability in the training visible domain and can be used for representation regularization and preprocessing alignment of cross-domain EEG signals. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the workflow of the reference domain hybrid drift alignment and conditional constraint characterization system for cross-domain EEG signals in this embodiment of the invention. Figure 2 This is a flowchart of the reference domain hybrid drift alignment module in an embodiment of the present invention, which includes a reference domain component anchoring module and a reference-guided linear hybrid drift adapter; Figure 3 This is a flowchart of the shared time-period conditional potential source characterization module in an embodiment of the present invention; Figure 4 This is a graph of the FULL MCC index under module ablation in an embodiment of the present invention; Figure 5 This is a graph showing the component positional consistency, sign consistency, and domain probe accuracy metrics in embodiments of the present invention. Figure 6 The figure shows the experimental results of how different linear drift intensities affect the source recovery quality, component sequence consistency, sign consistency, and domain probe accuracy of the standard condition identifiable variational autoencoder in this embodiment of the invention. Detailed Implementation

[0022] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0023] Example 1 This embodiment provides a reference-domain mixing-drift alignment and condition-constrained latent source representation system for cross-domain EEG signals. The system is a constructed Reference-Domain Mixing-Drift Alignment and Condition-Constrained Latent Source Representation Network, abbreviated as ReMiDA-CLSRNet. The network uses the reference domain as a common coordinate benchmark and combines positional and symbolic ambiguities in domain-by-domain component estimation with linear adaptation of the observation space to correct for mixing drifts that can be approximated as a first-order linear transformation. Simultaneously, by sharing time-segment condition-constrained latent representations, it learns the source correlation structures that coexist across domains.

[0024] The system includes an intra-domain normalization module, a reference domain hybrid drift alignment module, a shared time-segment conditional latent source representation module, and a joint optimization module. The intra-domain normalization module is used to scale the EEG observation matrices of each domain. The reference domain hybrid drift alignment module includes a reference domain component anchoring module and a reference-guided linear hybrid drift adapter, which are used to establish cross-domain component correspondences and complete first-order linear hybrid drift alignment. The shared time-segment conditional latent source representation module is used to learn cross-domain shared source-related latent representations.

[0025] The reference domain hybrid drift alignment module includes a reference domain component anchoring module (ReCA) and a reference-guided linear hybrid drift adapter (ReLiMDA). The reference domain component anchoring module uses the reference domain as a benchmark, establishing cross-domain component correspondences through component matching and sign correction, and providing a weak anchoring reference for subsequent observation space alignment. Based on the weak alignment prior, the reference-guided linear hybrid drift adapter learns a first-order linear alignment matrix constrained by the reference domain for each domain, and maps the standardized EEG observations of each domain to a common observation coordinate system to correct for linear drift in the observation space relative to the reference domain.

[0026] The shared time-segment conditional latent source characterization module is a shared time-segment conditional discriminative variational autoencoder (S-iVAE), comprising five parts: shared time-segment conditional embedding, conditional prior modeling, conditional posterior inference, latent variable reparameterization sampling, and aligned observation reconstruction. This module receives the aligned observation characterization and cross-domain shared time-segment sequence conditions output from the reference domain hybrid drift alignment module. These conditions are used to characterize the cross-domain shared source-related nonlinear structure and learn the shared source-related latent representation. Domain index conditions do not enter this branch; their function is performed by the front-end reference domain hybrid drift alignment module.

[0027] When using the aforementioned cross-domain EEG signal reference domain hybrid drift alignment and conditional constraint representation system, the specific steps include: Step 1: Acquire EEG observation data from multiple domains and determine a reference domain and at least one target domain; the domains are data domains corresponding to different subjects, different acquisition time periods of the same subject, or different batches of acquisition from different devices.

[0028] Step 2: Perform intra-domain standardization on the EEG observation data of each domain using the intra-domain standardization module to obtain the standardized observation matrix, and determine the domain index conditions and shared time period conditions for each sample.

[0029] Step 3: Use the domain index condition for front-end reference domain hybrid drift alignment, and use the shared time period condition for back-end shared time period condition latent source representation learning.

[0030] Step 4: Estimate the intra-domain components of the standardized observations in each domain using the Reference Domain Component Anchoring (ReCA) module, calculate the component statistical characteristics based on the shared time period conditions, and complete component position matching and sign correction relative to the reference domain.

[0031] Step 5: Using the reference-guided linear hybrid drift adapter ReLiMDA, initialize the linear alignment matrix based on the weak anchoring components output by ReCA, and perform first-order linear hybrid drift alignment on each domain observation by sharing the linear basis and the domain-specific drift term.

[0032] Step 6: Through the shared time-segment conditional latent source representation module S-iVAE, the observation representation is aligned with the shared time-segment conditional input conditional prior network, conditional posterior network, and decoding reconstruction network to learn the cross-domain shared latent source related representation.

[0033] Step 7: Train the system using the joint optimization module and output aligned observation representations and potential source correlation representations.

[0034] Example 2 This embodiment further defines Embodiment 1 and provides a detailed description of the modules in the system.

[0035] I. Overall Architecture and Input Data Definition like Figure 1 As shown, the system of this invention consists of five parts: ① intra-domain normalization, used to scale the EEG observation matrices of each domain; ② reference domain anchoring, used to establish cross-domain component correspondence with the reference domain as a common coordinate benchmark; ③ domain-adaptive observation alignment, used to correct the first-order linear mixture drift caused by changes in Subject or Session; ④ shared time-segment conditional latent representation learning, used to model the cross-domain shared source correlation structure under shared time-segment condition constraints; ⑤ joint optimization constraints, used to simultaneously constrain the consistency of the front-end alignment matrix, weakly anchored components, and back-end latent representation. Among them, ①-④ correspond to the forward data flow in Figure 1, and ⑤ corresponds to the overall training objective in formula (3) in the following content.

[0036] II. Intra-domain Standardization Modules: The intradomain standardization module calculates the mean and standard deviation of the EEG observation data for each domain according to the channel dimension and / or time dimension, performs mean removal and variance normalization, and obtains the standardized EEG observation matrix.

[0037] After completing the intra-domain standardization of the observation matrices for each domain, auxiliary conditions for module collaboration need to be determined for each sample. These auxiliary conditions include domain index conditions and shared time period conditions. Domain index conditions are obtained directly based on the data domain to which the sample belongs: the reference domain is denoted as... The target domains are denoted as follows: This data can be encoded as a one-hot vector for front-end domain-specific alignment channel selection. Shared time periods are determined based on the experimental procedures common to each domain: if the data already contains experimental phases, stimulus periods, time window numbers, or task phase labels, the corresponding labels are directly read; if no explicit labels are provided, the observation sequences from each domain are arranged according to a uniform window length. Divided into The time period, and the first The time period number of each sample is used as... , by all The time period numbers of each sample constitute a shared time period condition sequence. .

[0038] This condition has the same definition across all domains, describes only the common temporal non-stationary structure, does not carry domain identity, and is used as a constraint on the backend latent source representation. Under the joint constraints of the domain index condition and the shared time period condition, the forward data flow of the system of this invention can be divided into two stages: front-end observation alignment and back-end latent representation learning. For the first... d Observation matrix of each domain First, we perform intra-domain standardization to obtain... Subsequently, ② reference domain anchoring and ③ domain adaptation observation alignment together constitute the front-end ReMiDA module, which performs domain indexing. d With the assistance of the reference domain, the aligned observations are completed and the aligned observation representations are output: (1) Based on this, ④ the shared time-segment conditional latent representation learning is completed by the backend S-iVAE. S-iVAE only receives... Conditions for sharing time series across domains It is used to characterize source-related nonlinear structures shared across domains and learn their shared source-related latent representations.

[0039] (2) In equation (2), Indicates the first The shared latent representation of the domain learned by S-iVAE; This represents the aligned observation representation output by the front-end ReMiDA module.

[0040] The above ①-④ describe the forward representation generation process of ReMiDA-CLSRNet. To ensure that reference domain anchoring, linear hybrid drift alignment, and shared latent representation learning are optimized collaboratively under the same objective, this invention further introduces ⑤ joint optimization constraint, which incorporates the S-iVAE representation learning objective, anchoring consistency constraint, and ReLiMDA alignment matrix regularization term into the overall objective: (3) In the formula, Used to learn cross-domain shared latent representations under shared time conditions; The anchoring consistency loss is calculated by standardizing the shared latent representation output by S-iVAE and the weakly anchored components output by the Reference-DomainComponent Anchoring Module (ReCA) on a component-by-component basis, and is used to keep the two components consistent. , and These are used to limit the alignment matrix from the initialization direction, the domain-specific offset magnitude, and the alignment matrix degradation, respectively. , , and These are non-negative weighting coefficients, which control the relative strength of the above constraint terms in the overall objective.

[0041] III. Reference Domain Component Anchoring Module (ReCA) The workflow diagram of the Reference Domain Component Anchoring Module (ReCA) is as follows: Figure 2 As shown, the reference domain component anchoring module performs rapid independent component analysis on the standardized EEG observation matrix of each domain, calculates the conditional variance of each component segmented according to the shared time period conditions, and matches the conditional variance profiles of the target domain and the reference domain.

[0042] The reference domain component anchoring module uses the Hungarian algorithm to obtain the permutation matrix, constructs the sign correction matrix according to the inner product sign of the matching components, obtains the weak anchoring component representation, and uses the weak anchoring component representation as the target to obtain the initial linear alignment matrix of each domain through ridge regular least squares estimation.

[0043] The source recovery results of Linear Independent Component Analysis (ICA) are only identifiable in the sense of permutation, sign, and scale equivalence classes. If different domains are unmixed independently, the resulting components may be inconsistent in position, sign orientation, and scale, and therefore cannot be directly used for cross-domain component comparison or shared representation learning. ReCA uses a reference domain as a benchmark to establish cross-domain component correspondences through component matching and sign correction, and provides a weak anchoring reference for subsequent observation space alignment.

[0044] set up This indicates the number of ICA components retained. Then, the normalized observations for each domain... The Fast Independent Component Analysis (FastICA) algorithm outputs an in-domain estimated component representation. This invention uses a full-rank setting, that is... .

[0045] set up For the first The observation matrix after domain standardization. First, apply FastICA to each domain to obtain the estimated component representation within the domain: (4) In the formula, Corresponding reference domain, subscript This is used to indicate that the component comes from the FastICA analysis results; to align the component order and sign across different domains, this invention utilizes cross-domain shared time series conditions. The component sequences are uniformly segmented, and the conditional variance characteristics of each component at different shared time periods are calculated. Specifically, let: (5) In the formula, Indicates the first Domain Variance characteristics of each ICA component across different sharing time periods; Indicated in the data domain Within, the first [unit / component] obtained by the Fast Independent Principal Component Analysis (FastICA) algorithm... Individual component sequences (ICA); Indicates to The variance of the subsequence is calculated along the sample dimension; This indicates the calculation of conditional variance, i.e., the calculation of... Under the conditions The variance of the sequence.

[0046] Then, the component variance features of the target domain are matched with the corresponding features of the reference domain to obtain the matching cost matrix. ,in Indicates the reference field number The component and the first Domain The variance characteristics of the components are different, and the Hungarian algorithm is applied: (6) The optimal permutation matrix relative to the reference domain is obtained based on equation (6). .in, Let represent the set of all feasible permutation matrices. The optimization aims to select the one with the minimum overall matching cost among all possible component matching relationships. Based on this, a sign correction matrix is ​​constructed according to the sign of the inner product between the corresponding components of the target domain and the corresponding components of the reference domain after matching. . Let be a diagonal matrix, and its first... Line number Column elements are denoted as Two of them All represent the aligned first... Component index.

[0047] ,in Represents the permutation matrix After adjusting the order, the first Domain The ICA component sequence is used to obtain the reference domain. The target domain components corresponding to each component; This represents the inner product between the target domain component sequence and the corresponding component sequence of the reference domain. This operation is used to determine whether the sign directions of the two are consistent. is a sign function used to convert the inner product result into a sign correction coefficient of +1 or -1.

[0048] Thus, the weakly anchored component representation is obtained: (7) In the formula, for Dimensional variance eigenvectors. Under the full-rank setting... Below, both the FastICA output and the weakly anchored component representation remain unchanged. Dimensions; permutation matrix diagonal matrix with sign All .

[0049] Weakly anchored components obtained from reference domain constraints This will be further used as anchoring consistency loss in the future. Weak priors, and shared representations Establish consistency constraints.

[0050] The aforementioned weakly anchored components do not constitute the final cross-domain alignment result; their role is to provide a consistent initial coordinate system in the reference domain for subsequent linear alignment. To transform this prior into an initial alignment matrix in the observation domain, this invention further employs ridge-regularized least-squares fitting in standardized observations. Representation of weakly anchored components Solve between them.

[0051] To obtain the first The initial alignment matrix of the domain, the present invention uses For reference, Ridge regression estimation is performed on the linear mapping to the weakly anchored component space: (8) In the formula, Indicates the first The initial linear alignment matrix of the domain; Let be the linear mapping matrix to be estimated; For observation dimensions; The number of components to be retained; Denotes the Frobenius norm; The ridge regularity coefficient is used to restrict... This increases the magnitude of the solution and improves the stability of closed-form solutions.

[0052] make , Substituting into formula (8) yields: (9) In the formula, This is the standardized observation matrix for the current domain. For the corresponding weakly anchored target, Indicates about The ridge regression objective function. about Taking the derivative and setting the gradient to zero, we get: (10) In the formula, Indicates matrix transpose; for An identity matrix of order 1; , .

[0053] Therefore, the first The closed-form solution to the initial alignment matrix of the domain is: (11) In the formula, This represents finding the inverse of a matrix. When... hour, It is easier to maintain invertibility, thereby improving the numerical stability of the initial alignment matrix solution.

[0054] IV. Reference Guided Linear Hybrid Drift Adapter ReLiMDA Refer to the workflow diagram of the ReLiMD linear hybrid drift adapter as follows: Figure 2 As shown, based on the weak alignment prior provided by ReCA, ReLiMDA performs learnable first-order linear reparameterization on the normalized observations of each domain to explicitly model the linear drift of the observation space relative to the reference domain. For the ... For a domain, its linear alignment matrix is ​​defined as: (12) in, Indicates the first A learnable linear alignment matrix for the domain, used for normalized observations. Perform linear reparameterization; Represents a linear alignment matrix shared across domains. Indicates the first Domain-specific offset terms relative to the shared transformation.

[0055] During the initialization phase, and Depend on structure: , (13) And further joint optimization is performed during training. Therefore, the first... d The alignment observation characterization of the domain is defined as: (14) Both the reference domain and the target domain are passed Mapped to the same common observation coordinate system, where This is used only to determine the reference coordinate baseline. To restrict the degrees of freedom of ReLiMDA, three regularization constraints are imposed on the alignment matrix during training: anchor-preservation constraint, domain offset magnitude constraint, and approximate orthogonality constraint. First, the anchor-preservation constraint is used to limit the learned alignment matrix from deviating from the initialization direction given by ReCA: (15) Secondly, drift regular expressions are used to limit the magnitude of domain-specific offsets, thereby ensuring that cross-domain differences are still interpreted as controlled first-order drifts, rather than completely independent domain-specific mappings: (16) Finally, approximate orthogonal regularization is used to suppress column correlation and condition number deterioration, improving the stability and interpretability of the alignment matrix: (17) Therefore, ReLiMDA only performs reference domain alignment for observation space mixing drifts that can be approximated by a first-order linear transformation; higher-order nonlinear domain shifts that are not absorbed by the linear term are not modeled in this module.

[0056] V. Shared Time Period Conditional Potential Source Characterization Module (S-iVAE) As shown in Figure 3, S-iVAE is used to model cross-domain shared source correlation structures in multi-domain EEG observation representations aligned with ReMiDA. This module includes five parts: shared time-segment conditional embedding, conditional prior modeling, conditional posterior inference, latent variable reparameterization sampling, and aligned observation reconstruction.

[0057] As shown in Figure 3, S-iVAE is used to model cross-domain shared source correlation structures in ReMiDA-aligned multi-domain EEG observation representations. This module includes five parts: shared time-segment conditional embedding, conditional prior modeling, conditional posterior inference, latent variable reparameterization sampling, and aligned observation reconstruction. Let the first... The observation matrix after ReMiDA alignment of the domain is ,in Indicates the first Domain Each sample is aligned with an observation sample. S-iVAE introduces latent variables for each sample. and using shared time period tags As a condition variable, the domain index condition does not enter the S-iVAE branch; its function is already handled by the front-end ReMiDA module.

[0058] Specifically, the S-iVAE processing procedure includes the following five steps: First, embedding shared time period conditions.

[0059] Shared time period labels Input the conditional embedding layer to obtain the corresponding shared time period conditional embedding representation. This embedding only describes the temporal non-stationary structure that coexists across domains and does not carry domain identity information.

[0060] Second, conditional prior modeling.

[0061] Based on shared-time conditional embedding, prior distribution parameters of latent variables, including the mean, are generated through a conditional prior network. and standard deviation Thus, the conditional prior distribution is defined. This prior is used to constrain samples with the same shared time period label to follow a consistent distribution structure in the latent space.

[0062] Third, conditional posterior inference.

[0063] Align observation samples Shared time period tags Input a conditional posterior network to obtain the posterior distribution parameters. and Thus, the conditional posterior distribution is defined. In one implementation, the conditional posterior network uses a CNN-LSTM encoder to extract spatial redundancy and temporal dependence in the aligned observation representations, and outputs Gaussian distribution parameters through an MLP.

[0064] Fourth, latent variable reparameterization sampling.

[0065] Based on the conditional posterior distribution, latent variables are obtained using a reparameterization approach. That is, using random noise and posterior distribution parameters to construct potential samples that can be backpropagated. For the th The latent variables of all samples in the domain are arranged to obtain the latent source correlation representation. .

[0066] Fifth, align observation reconstruction.

[0067] latent variables The input is used to decode and reconstruct the network, which obtains the reconstructed distribution of the aligned observation samples and generates the reconstruction results. During training, S-iVAE optimizes the latent variables jointly through reconstruction and KL regularization terms, ensuring that the latent variables retain effective source-related information from aligned observations while also being subject to shared time-segment conditional prior constraints.

[0068] Let the first The first domain after ReMiDA alignment The observed samples are S-iVAE introduces latent variables for this sample. and using shared time period tags As a condition variable, it imposes conditional constraints on the distribution of latent variables. Through this constraint, samples with the same shared time-period conditions in different domains can maintain a relatively consistent temporal non-stationary structure in the latent space. Domain index conditions do not enter the S-iVAE branch; their corresponding inter-domain alignment is already completed by the front-end ReMiDA module.

[0069] This module is built upon a conditionally discriminative variational autoencoder. Considering the spatial redundancy and temporal dependence of EEG observations, the posterior network employs a CNN+LSTM encoding backbone, and outputs the Gaussian distribution parameters of the latent variables through a multilayer perceptron. Shared time-segment conditions are first mapped to conditional vectors via embedding layers, then input into the conditional prior network; the conditional prior network, composed of multilayer perceptrons, outputs the conditional mean and conditional variance of the latent variables. The decoder can be a combination of CNN and multilayer perceptrons, used to reconstruct the aligned EEG observation samples. The latent dimension can be related to the number of independent components retained. It can be consistent, or it can be set according to the observation dimensions and specific task requirements.

[0070] Therefore, the training objective of S-iVAE is written as: (18) In equation (18), Indicates the first d Domain t Aligned observation samples, This represents the latent variable corresponding to this sample. Indicates a shared time period label; This represents the posterior distribution given aligned observation samples and shared time-segment labels. This represents the prior distribution of shared time period conditions. This represents the decoded distribution of the aligned observation samples reconstructed from the latent variables. This indicates the expectation given the conditional posterior distribution. This represents the reconstructed log-likelihood. The KL divergence is used to measure the difference between the conditional posterior distribution and the conditional prior distribution of the shared time period.

[0071] The first item is a reconstruction item, and the second item is a KL regularization item. The weights of the KL terms are used to adjust the relative strength of the reconstruction constraints and the conditional prior constraints. To further ensure that the shared constraint representation maintains component identity consistency with the weak anchoring results of ReCA, this invention defines an anchoring consistency loss. .set up This represents the shared representation of the participants in the anchoring constraints. This indicates the weakly anchored component in the ReCA output. The number of ICA components to be retained. To eliminate scale uncertainty, both are standardized component-by-component. The row standardization operator is defined as: (19) in Representation matrix The OK, and These represent the mean and standard deviation, respectively. To prevent division by zero of constants. (Note: The last part is a typo and can be left as is.) , .

[0072] The anchoring consistency loss is defined as: (20) in, and They represent the first Domain j The loss consists of a standardized sequence of shared representation components and a sequence of weakly anchored components. The loss does not use real latent sources as monitoring signals; it only constrains the consistency of the shared representation and the weakly anchored components in terms of component identity. The minimum form is used to eliminate the effect of sign flipping, and the component order relationship is predetermined by ReCA.

[0073] In summary, S-iVAE receives the observation representation aligned with ReMiDA and learns cross-domain shared latent representations under shared time period conditions and anchor consistency constraints.

[0074] VI. Training Process of Joint Optimization Module In this invention, EEG observation data from multiple domains are first acquired, and a reference domain and at least one target domain are determined. Secondly, the EEG samples from each domain are divided into trials or time windows of uniform length, and shared time period conditions are generated based on the acquisition order, experimental stage, or fixed time window number.

[0075] Then, the EEG observation matrices of each domain are standardized within the domain, and ReCA is used to estimate the intra-domain components of the standardized observations. Cross-domain component position and symbol correspondence is established using shared time period conditions to obtain the weakly anchored component representation and the initial linear alignment matrix.

[0076] Furthermore, ReLiMDA is used to construct a cross-domain shared linear alignment matrix and a domain-specific drift term based on the initial linear alignment matrix. Through learnable first-order linear reparameterization, the reference domain and the target domain are mapped to the same common observation coordinate system to obtain the aligned observation characterization.

[0077] Subsequently, the aligned observation representations and shared time-period conditions are input into the S-iVAE. Through a conditional prior network, a conditional posterior inference network, latent variable reparameterization sampling, and decoding reconstruction process, the cross-domain shared latent source correlation representations are learned.

[0078] Finally, the joint optimization module trains the network according to the overall training objective shown in formula (3). After training, the observation representation is aligned with the output of the ReLiMDA guided linear hybrid drift adapter. The shared time-conditional potential source characterization module S-iVAE outputs a potential source correlation representation. The above output can be used as input representation for subsequent classifiers, decoding models, or neural signal analysis modules.

[0079] Example 3 This embodiment is a further explanation of Embodiments 1 and 2, and experimental verification and effect description are provided for the systems in Embodiments 1 and 2.

[0080] To verify the corrective effect of the reference domain component anchoring module and the reference-guided linear hybrid drift adapter on first-order linear hybrid drift in this invention, module ablation experiments were conducted on a linear drift-dominated synthesis benchmark. The experiments compared a standard condition identifiable variational autoencoder, a model B1 containing only the reference domain component anchoring module, and a complete model M1 containing both the reference domain component anchoring module and the reference-guided linear hybrid drift adapter. B1 was used to evaluate the effect of weak anchoring priors, and M1 was used to evaluate the complete alignment effect after linear adaptation.

[0081] like Figure 4 As shown, with drift intensity As γ increased from 0 to 0.40, the global average correlation coefficient (FULL MCC) of model B1, which only includes the reference domain component anchoring module, decreased from 76.0% to 47.0%, while the FULL MCC of the full model M1 decreased from 89.3% to 72.2%. At γ values ​​of 0, 0.05, 0.10, 0.20, and 0.40, the full model M1 showed significantly higher γ values ​​than B1, by 13.3, 16.2, 24.9, 20.7, and 25.2 percentage points, respectively. These results demonstrate that the primary gain of this invention stems from the explicit correction of the first-order linear hybrid drift in the observation space by the reference-guided linear hybrid drift adapter.

[0082] As shown in Figure 5, under low drift conditions, the component order consistency and sign consistency between B1 and M1 remain at a relatively high level. When the drift intensity increases, the complete model M1 can better maintain the component order consistency of the target domain relative to the reference domain. At γ=0.20, the component order consistency and sign consistency of M1 are 100% and 90%, respectively, which are significantly improved compared to B1's 54% and 72%. At γ=0.40, the component order consistency of M1 is still 100%, significantly higher than B1's 34%. The domain probe accuracy is close to the level of random three-domain classification under different drift intensities, indicating that the improvement of source recovery quality and component order consistency in this invention is not achieved by enhancing domain label separability.

[0083] The following experiments are only used to illustrate the technical effects of the present invention and are not intended to limit the dataset type, parameter settings or experimental classifier type of the present invention.

[0084] (I) Composite Benchmark and Evaluation Indicators To illustrate the impact of domain-specific linear drift on a fixed mixture hypothesis model and to verify the corrective effect of this invention on first-order linear mixture drift, a linear drift-dominated synthetic benchmark is constructed. All domains share the same latent source matrix and shared time-segment conditional sequences; inter-domain differences are primarily introduced by the domain-specific linear mixture matrix. Three domains are set up in the experiment: domain 0 is the reference domain, and the other two are the target domains; each domain includes 20 shared time periods, and each shared time period contains 500 samples, therefore each domain contains 10,000 samples; both the latent source dimension and the observation dimension are set to 5. The synthetic benchmark experiment is repeated under 5 random seeds.

[0085] The synthetic experiments used the Full Mean Correlation Coefficient (FULL MCC) to evaluate the recovery quality between the estimated latent representation and the true latent source. FULL MCC is an implementation of MCC (mean correlation coefficient) after optimal matching of aggregate components, and is not proposed as a new general metric. For the ... Each field is defined as: (twenty one) in, Indicates the true potential source, Indicates the first The estimated representation of each domain, This represents the optimal permutation that maximizes the mean absolute correlation coefficient. The FULL MCC in the experimental report represents the average value across the corresponding statistical units.

[0086] To further characterize the consistency of the constituent identities of the target domain relative to the reference domain, this invention uses two auxiliary indicators: permutation consistency and sign consistency. and : (twenty two) (twenty three) in, and Representing the reference domain and the first The optimal permutation of each target domain relative to the true potential source and Indicates the first match The sign orientation of each component. These two quantities are used to assist in analyzing whether the positional order and sign orientation of cross-domain components remain consistent.

[0087] In addition, to diagnose whether there is still any domain-specific information that can be read out in the estimated representation, probe accuracy is introduced. As an auxiliary diagnostic metric, specifically, after the main model training is complete, the estimated representation is frozen, and a lightweight linear classifier that does not participate in the main model optimization is additionally trained. This classifier takes the estimated latent representation as input and the sample domain labels as the prediction target, and reports the domain label classification accuracy on independent test splits. This metric measures the separability of domain label information in the estimated representation and is compared with the random classification level. Comparison, among which The number of fields. If If the value is close to random, it indicates that the domain label is difficult to read directly from the estimated representation; if it is significantly higher than random, it indicates that the representation still retains strong domain-specific information.

[0088] (II) Degeneracy verification of the fixed mixture model under linear drift To illustrate the effect of domain-specific linear drift on a fixed mixture hypothesis model, this invention progressively adjusts the drift intensity. We used a standard condition identifiable variational autoencoder as a control model to statistically analyze the source recovery quality, component sequence consistency, sign consistency, and domain probe accuracy under different drift intensities.

[0089] Table 1. Degradation results of the standard condition identifiable variational autoencoder under linear drift.

[0090] As shown in Figure 6 and Table 1, with the drift intensity As the value increased from 0 to 0.40, the full MCC of the standard conditionally identifiable variational autoencoder decreased from 76.0% to 47.6%, the component position consistency decreased from 100.0% to 28.0%, and the sign consistency decreased from 100.0% to 72.0%, while the domain probe accuracy remained generally between 32.9% and 37.0%. These results indicate that when there is a first-order linear mixing drift between different domains, the cross-domain shared latent representation of the fixed mixing hypothesis model exhibits significant degradation, mainly manifested in the inconsistency between the component position order in the target domain and the reference domain.

[0091] (III) Verification of the effect of reference domain anchoring and linear drift correction To verify the corrective effect of the reference domain component anchoring module and the reference-guided linear hybrid drift adapter on first-order linear hybrid drift in this invention, module ablation experiments were conducted on a linear drift-dominated synthesis benchmark. The experiments compared a standard condition identifiable variational autoencoder, a model B1 containing only the reference domain component anchoring module, and a complete model M1 containing both the reference domain component anchoring module and the reference-guided linear hybrid drift adapter. B1 was used to evaluate the effect of weak anchoring priors, and M1 was used to evaluate the complete alignment effect after linear adaptation.

[0092] like Figure 4 As shown, with drift intensity As the value increased from 0 to 0.40, the full MCC of model B1, which only contains the reference domain component anchoring module, decreased from 76.0% to 47.0%, while the full MCC of the complete model M1 decreased from 89.3% to 72.2%. When the coefficients are 0, 0.05, 0.10, 0.20, and 0.40, the complete model M1 is 13.3, 16.2, 24.9, 20.7, and 25.2 percentage points higher than B1, respectively. This result indicates that the main gain of this invention comes from the explicit correction of the first-order linear hybrid drift of the observation space by the reference-guided linear hybrid drift adapter.

[0093] like Figure 5 As shown, under low drift conditions, the consistency of component order and sign between B1 and M1 remains at a relatively high level; when the drift intensity increases, the intact model M1 can better maintain the consistency of component order between the target domain and the reference domain. At a value of 0.20, the component ordinal consistency and sign consistency of M1 are 100% and 90%, respectively, a significant improvement compared to B1's 54% and 72%. At a value of 0.40, the component order consistency of M1 remains 100%, significantly higher than B1's 34%. The domain probe accuracy, under different drift intensities, is generally close to the random level of three-domain classification, indicating that the improvement in source recovery quality and component order consistency achieved by this invention is not achieved by enhancing domain label separability.

[0094] (iv) Setup of a real EEG experiment To verify the representation alignment effect of this invention on real EEG data, experiments were conducted using the publicly available InnerSpeech2021 EEG dataset. This dataset is designed for implicit speech brain-computer interface tasks, includes 10 participants, 3 acquisition time periods, and 4 classification labels, and provides high-density EEG recordings. This invention uses its standard preprocessed version as the original preprocessed observation input and compares the impact of different input representations on the downstream EEGNet classifier under cross-participant and cross-acquisition time period protocols.

[0095] Real-world EEG experiments employed a reference-target domain separation data partitioning approach. For each cross-domain task, reference domain samples were stratified by category into a reference domain training set and a reference domain validation set, with a ratio of 8:2; target domain samples were stratified by category into an unlabeled fitting set and a reserved test set, with a ratio of 5:5. In cross-subject experiments, sub-01 was fixed as the reference subject, and other subjects were used sequentially as the target domain within the same acquisition time period; in cross-acquisition time period experiments, one acquisition time period was fixed as the reference domain within the same subject, and the remaining acquisition time periods were used sequentially as the target domain. The alignment model learned a common observation coordinate system using only the reference domain training set and the unlabeled target domain fitting set; the classifier used the reference domain training set for supervised training and selected the model based on the reference domain validation set; the reserved target domain test set was not visible during alignment training, classifier training, and model selection stages, and was only used for the final test.

[0096] This invention uses Accuracy and Macro-F1 as evaluation metrics for real-world EEG four-class classification. Val-acc and Val-F1 on the validation set are used to assess the separability of representations in the visible training domain; Test-acc and Test-F1 on the test set retaining the target domain are used to assess transfer performance in the unseen target domain. In real-world EEG experiments, a fixed random seed of 42 is used, and the maximum number of training epochs for both the linear alignment model and the EEGNet classification model is set to 120 epochs. An early stopping strategy based on validation set metrics is employed.

[0097] The experiment compares four types of input representations: raw-preprocessed input (r1), observation representations aligned by ReMiDA (aligned, a1), linear latent representations of S-iVAE outputs (linear-latent, l1), and joint inputs of raw-preprocessed inputs and linear latent representations (raw-preprocess + linear-latent, rL).

[0098] (v) Results of cross-subject real EEG experiments Under the cross-subject protocol, the observation representations aligned by reference domain hybrid drift show superior performance on the training visible validation set, with the average validation accuracy increasing from 45.83% of the original preprocessed observations to 50.00%, and the average validation F1 score increasing from 43.49% to 47.87%. These results demonstrate that, in cross-subject scenarios, this invention, through reference domain constraints and linear hybrid drift correction, can enhance the separability of representations on the training visible domain.

[0099] On the target domain test set, the joint input of the original preprocessed observations and latent representations achieved the highest average test accuracy of 25.67%, an improvement of 1.19 percentage points compared to 24.48% for the original preprocessed observations; the corresponding test F1 score improved from 23.58% to 24.39%, an improvement of 0.81 percentage points. These results demonstrate that the aligned representations and latent representations learned in this invention can effectively supplement cross-subject EEG decoding.

[0100] (vi) Results of real EEG experiments across different data collection periods Under the cross-acquisition-period protocol, the joint input of the original preprocessed observations and the latent representation achieves superior results on the training visible validation set. The average validation accuracy increases from 35.00% for the original preprocessed observations to 40.62%, and the average validation F1 score increases from 33.54% to 39.10%. This result demonstrates that, in cross-acquisition-period scenarios, jointly using the original preprocessed observations with the latent source-related representations learned in this invention can enhance class separability on the training visible domain.

[0101] On the target domain test set, the latent representation achieved the highest average test accuracy of 25.89%, an improvement of 0.89 percentage points compared to the original preprocessed observation of 25.00%; the test F1 score was also higher than the original preprocessed observation of 23.98%. These results indicate that cross-domain differences remain complex in real-world EEG scenarios involving multiple acquisition time periods. This invention can serve as a front-end for representation alignment and preprocessing alignment, providing a foundation for further stability improvements through subsequent integration with target domain calibration or downstream decoding models.

[0102] This invention improves source recovery quality, cross-domain component positional consistency, and common observation space consistency in cross-domain scenarios dominated by first-order linear hybrid drift. It also enhances representation alignment and class separability across the training visible domain in scenarios spanning multiple subjects and acquisition time periods. These results collectively demonstrate that this invention is suitable as a reference domain hybrid drift alignment and conditionally constrained latent source representation method in cross-domain EEG signal processing links.

Claims

1. A reference domain hybrid drift alignment and conditional constraint representation system for cross-domain EEG signals, characterized in that, The system includes: an intra-domain normalization module, a reference domain hybrid drift alignment module, a shared time-segment conditional potential source characterization module, and a joint optimization module; The intra-domain standardization module is used to scale-align the EEG observation matrices of each domain. The reference domain hybrid drift alignment module includes a reference domain component anchoring module and a reference-guided linear hybrid drift adapter. The reference domain component anchoring module uses the reference domain as a reference, establishes cross-domain component correspondences through component matching and sign correction, and provides a weak anchoring reference for subsequent observation space alignment. Based on the weak alignment prior, the reference-guided linear hybrid drift adapter learns a first-order linear alignment matrix constrained by the reference domain for each domain, and maps the standardized EEG observations of each domain to a common observation coordinate system to correct the linear drift of the observation space relative to the reference domain. The shared time-segment conditional latent source representation module is used to learn cross-domain shared source-related latent representations. The joint optimization module is used to train the system and output aligned observation representations and latent source-related representations.

2. The reference domain hybrid drift alignment and conditional constraint representation system for cross-domain EEG signals according to claim 1, characterized in that, The intradomain standardization module calculates the mean and standard deviation of the EEG observation data for each domain according to the channel dimension and / or time dimension, performs mean removal and variance normalization, and obtains the standardized EEG observation matrix. Furthermore, auxiliary conditions for module collaboration are determined for each sample. These auxiliary conditions include domain index conditions and shared time period conditions. The domain index conditions are obtained directly based on the data domain to which the sample belongs: the reference domain is denoted as... The target domains are denoted as follows: The shared time period conditions are obtained based on the experimental procedures common to each domain: if the data already contains experimental phases, stimulus periods, time window numbers, or task phase labels, then the corresponding labels are read directly; if no explicit labels are provided, then the observation sequences of each domain are arranged according to a uniform window length. Divided into The time period, and the first The time period number of each sample is used as... , by all The time period numbers of each sample constitute a shared time period condition sequence. .

3. The reference domain hybrid drift alignment and conditional constraint representation system for cross-domain EEG signals according to claim 2, characterized in that, The specific processing flow within the reference domain component anchoring module is as follows: S31, regarding the first Standardized EEG observation matrix for each data domain The fast independent component analysis algorithm is used to compute the estimated component representation within the domain. ; S32. Utilizing cross-domain shared time-series conditions right The component sequences are uniformly segmented, and the conditional variance characteristics of each component at different shared time periods are calculated. ; S33. Match the component variance features of the target domain with the corresponding features of the reference domain to obtain the matching cost matrix. The optimal permutation matrix relative to the reference domain is obtained by applying the Hungarian algorithm. ; S34. Based on the optimal permutation matrix Construct a symbolic diagonal matrix using the inner product sign of the components of the reference field. Thus, the weak anchoring component representation is obtained. .

4. The reference domain hybrid drift alignment and conditional constraint representation system for cross-domain EEG signals according to claim 3, characterized in that, ； In the formula, Indicates the first The first data field Variance characteristics of each ICA component across different sharing time periods; Indicated in the data domain The first part, obtained by the fast independent principal component analysis algorithm, is... Each independent ICA component sequence; This indicates the calculation of conditional variance, i.e., the calculation of... Under the conditions The variance of the sequence.

5. The reference domain hybrid drift alignment and conditional constraint representation system for cross-domain EEG signals according to claim 4, characterized in that, diagonal matrix In the middle, the first Line 1 The elements in the column are : Thus, the weakly anchored component representation is obtained: .

6. The reference domain hybrid drift alignment and conditional constraint representation system for cross-domain EEG signals according to claim 5, characterized in that, The specific processing flow in the reference-guided linear hybrid drift adapter is as follows: S61, regarding the first For each data field, its linear alignment matrix is ​​defined as: ,in, Represents a linear alignment matrix shared across domains. Indicates the first Each data field has a domain-specific offset relative to the shared transformation; S62, Calculate the first Alignment observation characterization of each data domain : .

7. The reference domain hybrid drift alignment and conditional constraint representation system for cross-domain EEG signals according to claim 6, characterized in that, and The solution process is as follows: by For reference, Ridge regression estimation is performed on the linear mapping to the weakly anchored component space: ; In the formula, Indicates the first The initial linear alignment matrix for each data field; Let be the linear mapping matrix to be estimated; The number of components to be retained; Denotes the Frobenius norm; The ridge regularity coefficient is used to restrict... This increases the magnitude of the solution and improves the stability of closed-form solutions. ， 。 8. The reference domain hybrid drift alignment and conditional constraint representation system for cross-domain EEG signals according to claim 7, characterized in that, The operations performed in the shared time period conditional latent source characterization module include, in sequence, shared time period conditional embedding, conditional prior modeling, conditional posterior inference, latent variable reparameterization sampling, and aligned observation reconstruction; Conditional embedding of shared time periods: embedding shared time period tags Input the conditional embedding layer to obtain the corresponding shared time period conditional embedding representation; Conditional prior modeling: Based on shared time-segment conditional embedding, the prior distribution parameters of latent variables, including mean and standard deviation, are generated through a conditional prior network, and the conditional prior distribution is defined; Conditional posterior inference: Aligning observation samples with shared time period labels Input the conditional posterior network to obtain the posterior distribution parameters, and thus define the conditional posterior distribution; Latent variable reparameterization sampling: Based on the conditional posterior distribution, latent variables are obtained through reparameterization, and backpropagation-capable latent samples are constructed using random noise and posterior distribution parameters; for the ... The latent variables of all samples in the domain are arranged to obtain the latent source correlation representation. ; Alignment observation reconstruction: latent variables in The input is used to decode and reconstruct the network, which obtains the reconstructed distribution of the aligned observation samples and generates the reconstruction results. .

9. The reference domain hybrid drift alignment and conditional constraint representation system for cross-domain EEG signals according to claim 8, characterized in that, When the joint optimization module trains the system, the total loss function used is: ,in, The training loss function for the shared time-segment conditional latent source representation module is used to learn cross-domain shared latent representations under shared time-segment conditions. The anchoring consistency loss function of the shared time-period conditional latent source representation module is calculated by standardizing the shared latent representation output by the shared time-period conditional latent source representation module and the weakly anchored components output by the reference domain component anchoring module on a component-by-component basis. It is used to keep the component identities of the two consistent. , and These are respectively used to limit the deviation of the alignment matrix from the initialization direction, the domain-specific offset magnitude, and the alignment matrix degradation within the reference-guided linear hybrid drift adapter. , , and These are non-negative weighting coefficients, which control the relative strength of the above constraint terms in the overall objective.

10. The reference domain hybrid drift alignment and conditional constraint representation system for cross-domain EEG signals according to claim 9, characterized in that, The specific process of the characterization system in operation is as follows: Step 1: Acquire EEG observation data from multiple domains and determine a reference domain and at least one target domain; the domains are data domains corresponding to different subjects, different acquisition time periods of the same subject, or different batches of acquisition from different devices; Step 2: Perform intra-domain standardization on the EEG observation data of each domain using the intra-domain standardization module to obtain the standardized observation matrix, and determine the domain index conditions and shared time period conditions for each sample; Step 3: Use the domain index condition for front-end reference domain hybrid drift alignment, and use the shared time period condition for back-end shared time period condition latent source representation learning; Step 4: Estimate the intra-domain components of standardized observations in each domain using the reference domain component anchoring module, calculate the component statistical characteristics based on the shared time period conditions, and complete component position matching and sign correction relative to the reference domain. Step 5: Using the reference-guided linear hybrid drift adapter, initialize the linear alignment matrix based on the weak anchoring components output by the reference domain component anchoring module, and perform first-order linear hybrid drift alignment on each domain observation through shared linear basis and domain-specific drift terms; Step 6: Through the shared time-segment conditional latent source representation module, align the observation representation with the shared time-segment conditional input conditional prior network, conditional posterior network, and decoding reconstruction network to learn cross-domain shared latent source related representations; Step 7: Train the system using the joint optimization module and output aligned observation representations and potential source correlation representations.

Images