A user intention recognition method, device and terminal equipment

Abstract

The application provides a user intention recognition method and device and a terminal equipment, which are suitable for the technical field of data processing, and the method comprises: performing space-time slicing and projection processing on a plurality of to-be-processed electroencephalogram signals to generate a plurality of to-be-processed electroencephalogram signal sequence information; performing mask processing on the plurality of to-be-processed electroencephalogram signal sequence information according to a preset mask ratio and a preset dynamic mask processing rule to generate a plurality of electroencephalogram signal mask sequence information; and obtaining a target intention recognition model according to the plurality of electroencephalogram signal mask sequence information and an initial intention recognition model. The application significantly improves the robustness of the model to complex noises such as vehicle vibration and electromagnetic interference and the cross-user generalization ability, realizes precise recognition of the driver's motor imagery intention with low delay and high accuracy, and provides stable and efficient decision support for intelligent cabin non-sensing brain control and man-machine cooperation.

Description

Technical Field

[0001] This application belongs to the field of data processing technology, and in particular relates to user intent recognition methods, devices and terminal equipment. Background Technology

[0002] With the rapid development of intelligent connected vehicles and intelligent cockpit technology, cockpit human-computer interaction is upgrading from physical buttons and touch voice to multimodal intent perception. Brain-computer interface technology based on EEG signals has become a cutting-edge direction for realizing seamless cockpit control because it can directly interpret the driver's motor imagination intent.

[0003] The in-vehicle environment is a typical high-noise, non-stationary scenario. Drivers' EEG signals are easily affected by vehicle vibration, electromagnetic interference, and their own psychological state. Extracting high signal-to-noise ratio and highly separable features is crucial for the successful implementation of in-vehicle brain-computer interface (BCI) technology. Current methods for decoding motor imagery EEG signals can be broadly categorized into three types: traditional machine learning, conventional deep learning, and generative model-based methods. Traditional machine learning methods employ manually designed feature extraction algorithms, such as cosine patterns and their variants, combined with support vector machines for classification. Conventional deep learning methods utilize convolutional neural networks to extract spatial correlations between electrodes and recurrent neural networks to capture temporal dependencies, achieving end-to-end feature learning and classification. Generative model-based methods learn the latent distribution of data through signal reconstruction for EEG data enhancement and feature extraction. All of these methods directly process EEG features in Euclidean space and complete decoding and classification.

[0004] Traditional machine learning methods are sensitive to non-stationary signals, have poor generalization ability, and are prone to feature aliasing. Conventional deep learning methods are difficult to capture the global topological structure of the whole brain channels, lack adaptive recovery mechanisms for signal loss and strong interference, and their performance drops sharply after being disturbed. Generative model-based methods cannot capture the low-dimensional manifold geometry of EEG signals, have poor feature space decoupling effect, and features of different motion intentions are prone to aliasing. The models have insufficient generalization ability and are difficult to adapt to the complex dynamic noise environment of vehicles. Summary of the Invention

[0005] In view of this, embodiments of this application provide a user intent recognition method, apparatus, and terminal device, aiming to solve the problems existing in the prior art, such as difficulty in signal analysis, weak anti-interference ability, poor feature decoupling effect, insufficient cross-user generalization ability, and easy aliasing of intent feature space, making it difficult to accurately identify, in the complex environment of vehicle in-vehicle systems.

[0006] A first aspect of this application provides a user intent recognition method, including:

[0007] Acquire EEG signals from multiple users;

[0008] User intent recognition information is generated based on the multiple user EEG signals and the target intent recognition model.

[0009] One aspect of this application provides a step for generating a target intent recognition model, including:

[0010] Acquire multiple EEG signals to be processed;

[0011] Spatiotemporal slicing and projection processing are performed on the multiple EEG signals to be processed to generate multiple EEG signal sequence information to be processed.

[0012] Based on multiple preset masking ratios and preset dynamic masking processing rules, the multiple EEG signal sequence information to be processed is masked to generate multiple EEG signal masking sequence information.

[0013] Based on the multiple EEG signal mask sequence information and the initial intent recognition model, the target intent recognition model is obtained.

[0014] A second aspect of this application provides a user intent recognition device, including:

[0015] The user EEG signal acquisition module is used to acquire multiple user EEG signals.

[0016] The user intent recognition information generation module is used to generate user intent recognition information based on the multiple user EEG signals and the target intent recognition model.

[0017] A third aspect of this application provides a terminal device, the terminal device including a memory and a processor, the memory storing a computer program executable on the processor, and the processor executing the computer program to implement the steps of the user intent recognition method described in the first aspect above.

[0018] A fourth aspect of this application provides a computer-readable storage medium, comprising: storing a computer program, wherein when executed by a processor, the computer program implements the steps of the user intent recognition method described in the first aspect above.

[0019] Compared with the prior art, the beneficial effects of the embodiments of this application are: the application significantly improves the robustness of the model to complex noises such as vehicle vibration and electromagnetic interference and its cross-user generalization ability, and realizes accurate recognition of the driver's motion imagination intention with low latency and high accuracy, providing stable and efficient decision support for the seamless brain control and human-machine collaboration of the intelligent cockpit. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram illustrating the implementation process of the user intent recognition method provided in Embodiment 1 of this application;

[0022] Figure 2 This is a schematic diagram illustrating the implementation process of the target intent recognition model generation step provided in Embodiment 1 of this application;

[0023] Figure 3 This is a schematic diagram illustrating the implementation process of the user intent recognition method provided in Embodiment 2 of this application;

[0024] Figure 4 This is a schematic diagram illustrating the implementation process of the user intent recognition method provided in Embodiment 3 of this application;

[0025] Figure 5 This is a schematic diagram illustrating the implementation process of the user intent recognition method provided in Embodiment 4 of this application;

[0026] Figure 6 This is a schematic diagram illustrating the implementation process of the user intent recognition method provided in Embodiment 5 of this application;

[0027] Figure 7 This is a schematic diagram illustrating the implementation process of the user intent recognition method provided in Embodiment Six of this application;

[0028] Figure 8 This is a schematic diagram illustrating the implementation process of the user intent recognition method provided in Embodiment 7 of this application;

[0029] Figure 9 This is a schematic diagram of the user intent recognition device provided in the embodiments of this application;

[0030] Figure 10 This is a schematic diagram of the terminal device provided in the embodiments of this application. Detailed Implementation

[0031] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0032] To illustrate the technical solution described in this application, specific embodiments are provided below.

[0033] Figure 1 A flowchart illustrating the implementation of the user intent recognition method provided in Embodiment 1 of this application is shown, and is described in detail below:

[0034] Step S101: Acquire EEG signals from multiple users.

[0035] In this embodiment, multiple user EEG signals refer to the multi-channel EEG signals collected on a high-fidelity driving simulator, following a standardized brain-computer interface experimental paradigm, for multiple subjects holding valid driver's licenses and having more than six months of driving experience. These signals are used to perform a five-category motor imagery task involving the driver's side window rising, the driver's side window falling, the passenger side window rising, the passenger side window falling, and a stationary state. A continuous dynamic motor imagery induction and data acquisition system can be constructed based on a flexible driving simulator. The acquisition process can begin with three minutes of resting-state baseline signal recording, followed by guiding the subjects to complete a four-second single first-person spatial motor imagery task according to on-screen visual prompts. Multiple rounds of data acquisition are completed according to a fixed time sequence and task intervals, thereby generating multiple user EEG signals.

[0036] Step S102: Generate user intent recognition information based on the multiple user EEG signals and the target intent recognition model.

[0037] In this embodiment, multiple user EEG signals may be preprocessed sequentially by performing fourth-order Butterworth bandpass filtering, 50 Hz notch filtering, whole-brain average reference, and electrooculography artifact removal. Then, the preprocessed user EEG signals are spatiotemporally sliced ​​and sequence embedded. The processed signals are then input into the target intent recognition model. Through a series of computational processes, including second-order statistical isomorphism and feature domain whitening, global spatiotemporal topological feature extraction of the visual transformer, latent manifold projection and distribution alignment, and low-rank adaptation for cockpit task fine-tuning, the target intent recognition model outputs user intent recognition information containing the driver's motor imagination category.

[0038] The user intent recognition method provided in this application significantly improves the robustness of the model to complex noises such as vehicle vibration and electromagnetic interference, as well as its cross-user generalization ability. It achieves low-latency and high-accuracy precise recognition of the driver's motor imagination intent, providing stable and efficient decision support for seamless brain control and human-machine collaboration in intelligent cockpits.

[0039] Figure 2 The following is a schematic diagram illustrating the implementation process of the target intent recognition model generation step provided in Embodiment 1 of this application, detailed below:

[0040] Step S1101: Acquire multiple EEG signals to be processed.

[0041] In this embodiment, the multiple EEG signals to be processed can refer to EEG signals obtained from a publicly available dataset, and can be multi-channel EEG signals that have undergone preprocessing such as fourth-order Butterworth bandpass filtering, 50 Hz notch filtering, whole-brain average reference, and removal of electrooculogram artifacts.

[0042] Step S1102: Spatiotemporal slicing and projection processing are performed on the multiple EEG signals to be processed to generate multiple EEG signal sequence information to be processed.

[0043] In this embodiment, multiple EEG signals to be processed can be divided into multiple non-overlapping spatiotemporal blocks according to a fixed size. Then, each spatiotemporal block is subjected to learnable linear projection processing and superimposed with learnable positional encoding. Finally, multiple EEG signal sequence information containing complete spatiotemporal information is generated.

[0044] Step S1103: According to multiple preset masking ratios and preset dynamic masking processing rules, the multiple EEG signal sequence information to be processed is masked to generate multiple EEG signal masking sequence information.

[0045] In this embodiment, multiple preset mask ratios can be preset manually, and preset dynamic mask processing rules can be preset manually. The mask ratios are adjusted in a non-linear growth manner from low to high, thereby performing multi-gradient occlusion replacement operations from random perturbation to regional failure on multiple EEG signal sequence information to be processed, and then generating multiple EEG signal mask sequence information.

[0046] Step S1104: Based on the multiple EEG signal mask sequence information and the initial intent recognition model, a target intent recognition model is obtained.

[0047] In this embodiment, the initial intent recognition model consists of four parts: an initial EEG signal encoding feature generation sub-model, an initial intent recognition inference distribution generation sub-model, an initial intent recognition reference distribution generation sub-model, and an initial intent recognition EEG signal reconstruction sub-model. Specifically, the initial EEG signal encoding feature generation sub-model can employ a global spatiotemporal topological feature extraction backbone network for EEG signals, built upon a visual transformer (ViT) and incorporating second-order statistical isomorphism and feature domain whitening, multi-head self-attention, and multilayer perceptron cascaded coding blocks. The initial intent recognition inference distribution generation sub-model can be a variational inference encoder that only receives masked incomplete features and lacks prior assistance from the original signal. The initial intent recognition reference distribution generation sub-model can be a reference encoder that can access the complete original EEG signal and is used to generate an ideal prior distribution. The initial intent recognition EEG signal reconstruction sub-model can be a lightweight reconstruction decoder that reconstructs spatiotemporal blocks of EEG signals based on reparameterized sampling and inverse projection.

[0048] In this embodiment, multiple EEG signal mask sequence information can be input into the initial intent recognition model, and then a self-supervised learning process can be completed, including second-order statistical isomorphism and feature domain whitening, global spatiotemporal topological feature extraction of the visual transformer, latent manifold projection and distribution alignment, signal reconstruction and network joint optimization. Then, the model is adapted by combining low-rank adaptation with cockpit task fine-tuning, thereby obtaining the target intent recognition model.

[0049] The user intent recognition method provided in this application improves the robustness of the target intent recognition model to interference from non-stationary EEG signals through a dynamic masking strategy. It effectively overcomes the defects of difficult signal analysis, weak noise resistance, and poor feature decoupling in complex vehicle environments, significantly improves the accuracy and stability of intent recognition, and effectively supports the efficient operation of contactless brain control and human-machine collaboration in intelligent cockpits.

[0050] Figure 3 The flowchart illustrating the implementation of the user intent recognition method provided in Embodiment 2 of this application is shown. The difference between this method and Embodiment 1 is that step S1102 specifically includes:

[0051] Step S201: Based on the preset spatiotemporal slice size information of the EEG signal, perform spatiotemporal slice processing on the multiple EEG signals to be processed to obtain multiple spatiotemporal block information of the EEG signals to be processed.

[0052] In this embodiment, the preset spatiotemporal slice size information of the EEG signal can be preset by humans. According to the preset spatiotemporal slice size information of the EEG signal, multiple EEG signals to be processed are divided into multiple non-overlapping regions along the time axis, thereby generating multiple spatiotemporal block information of EEG signals to be processed.

[0053] Step S202: Generate multiple EEG signal linear projection information based on the spatiotemporal block information of the multiple EEG signals to be processed and the preset EEG signal linear projection model.

[0054] In this embodiment, the preset EEG signal linear projection model can be pre-set by humans and can be a learnable linear projection layer. Multiple spatiotemporal block information of EEG signals to be processed can be flattened and sequentially input into the preset EEG signal linear projection model for mapping calculation, thereby generating multiple EEG signal linear projection information.

[0055] Step S203: Generate multiple EEG signal sequence information to be processed based on the multiple EEG signal linear projection information and the preset EEG signal position encoding information.

[0056] In this embodiment, the preset EEG signal location coding information can be manually preset or a learnable location coding. Multiple EEG signal linear projection information can be superimposed with the preset EEG signal location coding information to generate multiple EEG signal sequence information to be processed.

[0057] In this embodiment, the EEG acquisition device has a sampling rate of 250Hz and includes 32 electrode channels. The input EEG signal sample is defined as... The number of channels Time window length Since the Transformer cannot directly process two-dimensional maps, spatiotemporal slicing is required. The slice size can be set to [size missing]. (Along the timeline) Divided into Non-overlapping spacetime blocks, each block It contains local topological information for all channels within a short time. Then, a learnable linear projection layer maps each flattened slice to an initial embedding vector, and a learnable positional encoding is superimposed. The initial sequence embedding is obtained. :

[0058]

[0059] The user intent recognition method provided in this application accurately preserves the spatiotemporal structure information of the EEG signal to be processed, thereby improving the extraction accuracy and representation ability of the target intent recognition model for EEG signal features.

[0060] Figure 4 The flowchart illustrating the implementation of the user intent recognition method provided in Embodiment 3 of this application is shown. Its difference from Embodiment 1 described above lies in:

[0061] The preset mask ratios include a preset first mask ratio and a preset second mask ratio;

[0062] Wherein, the preset first mask ratio is less than the preset second mask ratio;

[0063] The plurality of EEG signal sequence information to be processed includes a plurality of first EEG signal sequence information to be processed and a plurality of second EEG signal sequence information to be processed;

[0064] The plurality of EEG signal mask sequence information includes a plurality of first EEG signal mask sequence information and a plurality of second EEG signal mask sequence information;

[0065] Step S1103 specifically includes:

[0066] Step S301: Generate first mask matrix information according to the preset first mask ratio and the preset dynamic mask processing rules.

[0067] In this embodiment, the preset mask ratio follows dynamic course scheduling, and the preset first mask ratio can be manually preset. The preset dynamic masking rules can also be manually preset, which can involve randomly occluding the values ​​in the matrix according to the manually preset mask ratio. Random occlusion calculation is performed according to the preset first mask ratio and the preset dynamic masking rules to generate the first mask matrix information.

[0068] In this embodiment, during the initial training phase, Setting it to a lower value primarily simulates slight signal perturbations, helping the model converge quickly; as the training rounds progress, As the nonlinearity increases to a high value, the simulation of large-scale signal loss under extreme and harsh conditions forces the model to learn deep topological dependencies. Based on current... Values, generate a binary mask matrix ,in This indicates that the spatiotemporal slice is damaged or missing. The sequence is embedded based on the mask matrix. With mask matrix Perform a dot product operation and remove the occluded part ( The region is replaced with a learnable mask token to obtain a mask sequence. The sequence will be fed into the initial intent recognition model, which may be the Visual Transformer (ViT) backbone network within the initial intent recognition model, used to extract features under conditions of incomplete information.

[0069] Step S302: Perform dot multiplication calculation based on the first mask matrix information and multiple first EEG signal sequence information to generate multiple first EEG signal mask sequence information.

[0070] In this embodiment, the first mask matrix information and multiple first EEG signal sequence information to be processed are multiplied by point, and then the occluded area is replaced with learnable mask words. Then, multiple first EEG signal mask sequence information are generated to simulate slight signal disturbances and local noise interference in the vehicle environment, which helps the initial intent recognition model to quickly and stably converge and learn basic contextual dependencies.

[0071] Step S303: Generate second mask matrix information according to the preset second mask ratio and the preset dynamic mask processing rules.

[0072] In this embodiment, the preset second mask ratio can be manually preset. The preset dynamic masking rule can also be manually preset, which can be to randomly occlude the values ​​in the matrix according to the manually preset mask ratio. Random occlusion calculation is performed according to the preset second mask ratio and the preset dynamic masking rule to generate the second mask matrix information.

[0073] Step S304: Perform dot multiplication calculation based on the second mask matrix information and multiple second EEG signal sequence information to generate multiple second EEG signal mask sequence information.

[0074] In this embodiment, the second mask matrix information is multiplied pointwise with multiple second EEG signal sequence information to be processed. The occluded region is then replaced with learnable mask terms, generating multiple second EEG signal mask sequence information. This simulates extreme conditions such as electrode detachment and large-area signal failure in a vehicle environment, forcing the initial intent recognition model to learn global deep topological features and improving robustness under complex noise. It is understood that the preset first mask ratio and preset second mask ratio are merely illustrative. In practical applications, the mask ratio can be manually set from 0 and linearly increased to 0.05 according to the training progress.

[0075] The user intent recognition method provided in this application enhances the adaptability of the target intent recognition model to interference of different intensities, thereby significantly improving the robustness and stability of the model in complex vehicle noise environments.

[0076] Figure 5 The flowchart illustrating the implementation of the user intent recognition method provided in Embodiment 4 of this application is shown. Its difference from Embodiment 1 described above lies in:

[0077] The initial intent recognition model includes an initial EEG signal coding feature generation sub-model, an initial intent recognition inference distribution generation sub-model, an initial intent recognition reference distribution generation sub-model, and an initial intent recognition EEG signal reconstruction sub-model.

[0078] Step S1104 specifically includes:

[0079] Step S401: Generate multiple EEG signal mask sequence feature information based on the multiple EEG signal mask sequence information and the initial EEG signal coding feature generation sub-model.

[0080] In this embodiment, the initial EEG signal coding feature generation sub-model can be a visual transformer (ViT) backbone network composed of several cascaded coding blocks. Multiple EEG signal mask sequence information can be input into the initial EEG signal coding feature generation sub-model to complete the second-order statistical isomorphism and feature domain whitening, global spatiotemporal topological feature extraction of the visual transformer, and then generate multiple EEG signal mask sequence feature information.

[0081] In this embodiment, EEG signals inherently exhibit strong non-stationarity and inter-subject variability. To eliminate this cross-subject marginal distribution shift before feature embedding, this invention introduces an Euclidean alignment mechanism based on covariance whitening transform before the input sequence enters the transformer coding block. For a given subject... A pre-processed EEG test sequence The network first calculates the reference covariance matrix for the subject:

[0082]

[0083] Subsequently, the inverse square root of the reference matrix is ​​used to perform projection normalization on each experimental feature:

[0084]

[0085] After this spatial isomorphic transformation, the transformed EEG tests from different subjects or different collection batches... Forced to be projected into a consistent representational domain space with a unit covariance matrix (i.e. This process smooths out cross-subject brain physical differences and background baseline drift at the second-order statistical level, providing a highly consistent and bias-resistant standardized sequence input for the subsequent visual transformer. .

[0086] In this embodiment, EEG signals are essentially macroscopic potential mappings of neuronal activity in the cerebral cortex. Synchronous oscillations between different brain regions (such as the occipital and frontal lobes) often contain key information about motor intention. To reconstruct this cross-channel global topology in the feature space, this invention deploys a multi-head self-attention component at the core location of each coding block. Specifically, for the input sequence of layer 1... The network first performs projection onto the feature subspace. To simultaneously capture the properties of EEG signals in different dimensions, such as electrode synchronicity in the spatial domain and rhythmic fluctuations in the frequency domain, a multi-head design is introduced. The input sequence is first normalized to a standardized distribution through layers, and then passed through three learnable linear projection matrices. Mapped as a matrix of queries, keys, and values:

[0087]

[0088] After obtaining the above projections, the model performs global interactive computation. By calculating the dot product of the query matrix Q and the transpose of the key matrix K, the model quantifies the correlation strength between any two spatiotemporal slices in the sequence. To prevent the dot product result from becoming too large and causing Softmax gradient saturation, this invention introduces a scaling factor. Subsequently, the correlation is transformed into a probability distribution using the Softmax function, and the value matrix V is then weighted and aggregated. This process mathematically achieves information fusion that transcends the limitations of physical distance.

[0089]

[0090] Finally, to address the degradation problem in deep networks and preserve the original input information, residual connections are introduced. These connections add the aggregated features to the original input, yielding the output of the MSA sublayer. :

[0091]

[0092] In this embodiment, it is understood that although the MSA mechanism successfully captures the global context, the features are still in a linear combination state. To extract deeper semantic information, the output of the MSA sublayer... The input features are fed into a multilayer perceptron sublayer for nonlinear transformation. This process follows the design mechanism of "dimensionality increase-activation-dimensionality reduction". The input features are first normalized again by the sublayer, and then passed through the first linear layer. By increasing the dimensionality, EEG features can be fully unfolded in a higher-dimensional manifold space, thereby exposing subtle patterns hidden beneath the noise. Subsequently, a Gaussian error linear unit (GELU) is introduced as the activation function. Compared to the traditional ReLU, GELU has smoother gradient properties when processing highly random signals like EEG. Finally, a second linear layer... The features are restored to their original dimensions. The entire MLP sublayer operation, combined with residual connections, can be represented as:

[0093]

[0094] By stacking L layers of the above-mentioned coding blocks, the original sequence embedding is gradually transformed into a high-dimensional feature representation rich in high-order semantic information. :

[0095]

[0096] This feature H preserves the global spatiotemporal topology of the EEG signal acquired in the MSA stage, and completes denoising and semantic abstraction in the MLP stage.

[0097] Step S402: Based on the multiple EEG signal mask sequence feature information, multiple EEG signal sequence information to be processed, the initial intent recognition inference distribution generation sub-model, and the initial intent recognition reference distribution generation sub-model, generate multiple EEG signal mask sequence inference distribution information and multiple EEG signal sequence reference distribution information.

[0098] In this embodiment, the initial intent recognition inference distribution generation sub-model and the initial intent recognition reference distribution generation sub-model can be a pre-defined inference encoder and a pre-defined reference encoder. Multiple EEG signal mask sequence feature information can be input into the initial intent recognition inference distribution generation sub-model to generate multiple EEG signal mask sequence inference distribution information. Similarly, multiple EEG signal sequence information to be processed can be input into the initial intent recognition reference distribution generation sub-model to generate multiple EEG signal sequence reference distribution information.

[0099] In this embodiment, residual semantic features H representing the masked view are obtained after extraction via the visual transformer backbone network. Since these features are extracted based on incomplete observation data, they often suffer from semantic gaps and discontinuous distribution in high-dimensional Euclidean space. To recover the essential geometric structure of the EEG signal and address the feature aliasing problem in noisy environments, this step utilizes variational inference techniques to construct an asymmetric dual-flow manifold projection mechanism guided by prior conditions. Understandably, the original sequence, after processing by the visual transformer backbone network, extracts incomplete high-order semantic features H representing the masked view. Because EEG signals inherently possess extremely low signal-to-noise ratio (SNR), directly mapping these features to traditional high-dimensional Euclidean space easily leads to overfitting of the microscopic details of the EEG; in noisy in-vehicle environments, these features are often meaningless random noise, resulting in severe semantic fragmentation and inter-class aliasing in the feature space. To eliminate microscopic noise and accurately extract the driver's true intention, this application innovatively introduces a variational inference architecture, constructing an asymmetric dual-manifold projection mechanism guided by prior conditions. This mechanism transforms deterministic high-dimensional point mappings into low-dimensional, compact probability distribution representations. By leveraging this natural information bottleneck, it effectively filters out chaotic microscopic interference, thus preserving stable and discriminative macroscopic consciousness features completely and clearly in the probability manifold space, thoroughly restoring the essential geometric structure of the EEG signal. Specifically, the same set of residual semantic features H can be mapped to the probabilistic latent manifold space through both the reference encoder and the inference encoder. The core difference lies in whether the original EEG signal is used as a priori condition for auxiliary calibration.

[0100] The reference encoder aims to establish an ideal manifold distribution target. Although its input features are incomplete, the encoder is granted access to the raw EEG signal. First, the raw EEG signal is mapped to a global prior condition vector using a conditional embedding module. Subsequently, the prior condition vector is deeply fused with the residual semantic features. The reference encoder uses the fused features to infer the target prior distribution given the known physical truth value of the signal. Its prior mean... and prior log-variance The calculation is as follows:

[0101]

[0102] in, This represents the feature fusion operation. Due to the introduction of strong prior conditions... This distribution can effectively correct erroneous information in features and form high-confidence manifold structure anchor points.

[0103] The inference encoder simulates the reasoning process in real-world applications, namely blind source resolution when the truth value of the original signal is unavailable. The encoder directly receives residual semantic features and maps them to a variational posterior distribution using independent projection heads. This distribution represents the model's best estimate of the underlying motion intent based on only incomplete observations. Its posterior mean... and posterior log-variance The calculation is as follows:

[0104]

[0105] Step S403: Alignment processing is performed based on the distribution information inferred from the multiple EEG signal mask sequences and the reference distribution information of the multiple EEG signal sequences to generate multiple aligned EEG signal mask sequence distribution information.

[0106] In this embodiment, the distribution information inferred from multiple EEG signal mask sequences can be aligned with the reference distribution information of multiple EEG signal sequences using the Kolb-Leibler divergence method, thereby generating multiple aligned EEG signal mask sequence distribution information.

[0107] In this embodiment, the Kolb-Leibler (KL) divergence is introduced as a manifold alignment constraint. Mathematically, this constraint is defined as minimizing the difference between the student-generated posterior distribution Q and the teacher-constructed prior distribution P. The core objective is to train the student network to supplement missing information, enabling it to infer information similar to or related to the prior distribution even without prior knowledge. Consistent manifold structure during auxiliary operation:

[0108]

[0109] Here, J represents the feature dimension of the latent manifold space. Through this constraint, the model achieves a robust mapping from incomplete features to complete manifolds in the latent space.

[0110] Step S404: Based on the distribution information of the multiple aligned EEG signal mask sequences and the initial intention recognition EEG signal reconstruction sub-model, generate multiple EEG signal mask sequence reconstruction information.

[0111] In this embodiment, the distribution information of multiple aligned EEG signal mask sequences is input into the initial intention recognition EEG signal reconstruction sub-model. Then, reparameterized sampling and signal inverse mapping calculation are performed to generate multiple EEG signal mask sequence reconstruction information. It is understood that simply aligning the distribution in the probability space is insufficient to guarantee that the features learned by the target intention recognition model have sufficient physical semantics. To verify whether the latent manifold features truly preserve the spatiotemporal topology of the EEG signals and form an effective self-supervised learning loop, a signal reconstruction mechanism needs to be introduced. This mechanism calculates the physical error by inverse mapping the latent features back to the original signal space to drive the joint optimization of the entire network.

[0112] Step S405: Based on the reconstruction information of the multiple EEG signal mask sequences and the multiple EEG signals to be processed, calculate the EEG signal reconstruction loss information.

[0113] In this embodiment, point-by-point error calculation is performed on the reconstruction information of multiple EEG signal mask sequences and multiple EEG signals to be processed, thereby calculating the EEG signal reconstruction loss information.

[0114] In this embodiment, it can be understood that during the training phase, the core objective is to optimize the inference encoder to make it independently noise-resistant. Therefore, the variational posterior distribution generated by the inference encoder... Sampling is then performed. This process is no longer considered a non-differentiable random sampling operation, but is modeled as a deterministic transformation process: first, an auxiliary noise variable is sampled from a standard normal distribution. Then, it is linearly transformed to the target distribution space. Specifically, the posterior mean obtained in step three is used. and posterior standard deviation latent vector The formula for generating it is as follows:

[0115]

[0116] in, This represents element-wise multiplication. In this way, randomness is transferred to an auxiliary variable. This allows the subsequent reconstruction error gradient to smoothly pass through the sampling layer and propagate back to the parameters of the inference encoder. The generated latent vectors... At this point, the geometric structure of the motor intention in the EEG signal has been highly condensed, and redundant noise has been removed.

[0117] Obtaining latent vectors Subsequently, it needs to be restored to a sequence of EEG signal slices with the same dimension as the original input. This invention constructs a lightweight reconstruction decoder. The decoder performs the inverse transformation operation opposite to the encoder, first converting the low-dimensional latent vector... Mapping back to a high-dimensional feature space, the temporal information of the sequence is recovered using position encoding, and finally, the reconstructed EEG signal slice sequence is output. This process can be represented as:

[0118]

[0119] in, That is, the recovered A spacetime slice.

[0120] The calculation of EEG signal reconstruction loss information can be performed using the mean squared error (MSE) form of the reconstruction loss:

[0121]

[0122] in, This represents the set of slice indices that are obscured by a mask. The total number of occluded slices. Raw EEG signal slices, This corresponds to the reconstructed slice. Secondly, it combines the KL divergence manifold alignment loss defined in step three. Construct the final overall objective function:

[0123]

[0124] in, It is an adjustable hyperparameter used to balance the weights between physical reconstruction accuracy and manifold structure canonicity. During network backpropagation, The driving model learns the fine-grained spatiotemporal dependence of EEG signals to ensure the fidelity of the reconstructed signals; while This serves as a regularization term, constraining the latent distribution of the inference encoder to converge towards the prior distribution of the reference encoder, preventing overfitting or distribution collapse in the feature space. By minimizing... Through end-to-end training, the entire network gradually learns the robust ability to accurately analyze and represent the driver's movement intentions under dynamic masking interference.

[0125] Step S406: Based on the initial intent recognition model, the EEG signal reconstruction loss information, and the preset EEG signal reconstruction loss threshold, the target intent recognition model is obtained.

[0126] In this embodiment, the preset EEG signal reconstruction loss threshold can be preset manually. When the EEG signal reconstruction loss information is less than or equal to the preset EEG signal reconstruction loss threshold, the training of the initial intention recognition model is stopped, thereby obtaining the target intention recognition model.

[0127] The user intent recognition method provided in this application standardizes the latent manifold space structure of the target intent recognition model, effectively suppresses feature aliasing and model overfitting, and improves the feature decoupling ability and recognition stability of the target intent recognition model in a noisy vehicle environment.

[0128] Figure 6 The flowchart illustrating the implementation of the user intent recognition method provided in Embodiment 5 of this application is shown. The difference between this method and Embodiment 4 above is that step S402 specifically includes:

[0129] Step S501: Generate multiple EEG signal mask sequence inference distribution information based on the feature information of the multiple EEG signal mask sequences and the initial intention recognition inference distribution generation sub-model.

[0130] In this embodiment, multiple EEG signal mask sequence feature information are individually input into the initial intent recognition inference distribution generation sub-model, then the latent manifold projection calculation is completed, and then multiple EEG signal mask sequence inference distribution information is generated.

[0131] Step S502: Generate multiple EEG signal sequence reference distribution information based on the multiple EEG signal sequence information to be processed and the initial intention recognition reference distribution generation sub-model.

[0132] In this embodiment, multiple EEG signal sequence information to be processed are individually input into the initial intention recognition reference distribution generation sub-model, and then the latent manifold projection calculation is completed, and then multiple EEG signal sequence reference distribution information are generated.

[0133] The user intent recognition method provided in this application embodiment enables the inferred distribution and the reference distribution to independently learn the corresponding signal features, thereby improving the accuracy of distribution alignment and the stability of model training, and thus strengthening the global feature inference capability of the target intent recognition model under incomplete signals.

[0134] Figure 7 The flowchart illustrating the implementation of the user intent recognition method provided in Embodiment Six of this application is shown. The difference between this method and Embodiment One is that step S102 specifically includes:

[0135] Step S601: Perform spatial filtering processing on the multiple user EEG signals to obtain multiple filtered user EEG signals.

[0136] In this embodiment, spatial filtering processing of second-order statistical isomorphism and feature domain whitening can be sequentially performed on multiple user EEG signals to eliminate the edge distribution offset across user EEG signals, thereby generating multiple filtered user EEG signals.

[0137] Step S602: Based on the multiple filtered user EEG signals, the target intent recognition model is fine-tuned to generate a fine-tuned target intent recognition model.

[0138] In this embodiment, a low-rank adapted cockpit task fine-tuning method can be adopted, in which multiple filtered user EEG signals are input into the target intent recognition model, the backbone network parameters of the model are frozen and the bypass adaptation parameters are trained, and then the fine-tuned target intent recognition model is generated.

[0139] In this embodiment, the fine-tuning process adopts the principle of low-rank adaptation for cockpit task fine-tuning. First, all parameters of the backbone network of the pre-trained target intent recognition model are frozen. Only the low-rank adaptation bypass matrix is ​​connected in parallel to the multi-head self-attention layer of the target intent recognition model. The lightweight bypass parameters are trained using a small amount of labeled EEG data in the cockpit scenario. This avoids the full parameter fine-tuning from destroying the noise-resistant manifold representation and global spatiotemporal topological features learned in the pre-training stage. At the same time, it reduces the computing power consumption of the vehicle edge device. The high-dimensional semantic features are mapped to the probability distribution of motion imagination categories through global average pooling and lightweight classification heads. The bypass parameters are updated with cross-entropy loss supervision, allowing the model to quickly adapt to the personalized EEG signal distribution of the smart cockpit. While retaining strong noise resistance, it improves the accuracy and cross-user generalization of vehicle intent recognition.

[0140] Step S603: Generate user intent recognition information based on the multiple user EEG signals and the fine-tuned target intent recognition model.

[0141] In this embodiment, multiple user EEG signals can be input into a finely tuned target intent recognition model, which then completes the calculation process of spatiotemporal slicing, global spatiotemporal topological feature extraction, latent manifold projection and distribution alignment, and finally generates user intent recognition information containing the driver's motor imagination category.

[0142] In this embodiment, a LoRA-based fine-tuning strategy can be employed. The target intent recognition model can be a pre-trained initial intent recognition inference distribution generation sub-model used as the backbone network for downstream classification tasks. Let the pre-trained weight matrix of the multi-head self-attention layer (MSA) in this backbone network be... During the fine-tuning phase, it can be strictly frozen. That is, its parameters are not updated during backpropagation. To endow the model with the ability to discriminate specific categories of motion images, in A low-rank update branch is connected in parallel to the side. This branch contains two low-rank matrices, including the dimension reduction matrix. and the increasing dimension matrix , where rank Much smaller than the feature dimension ,Right now For any input features The output of this layer It is composed of the frozen main path output and the trainable side path output:

[0143]

[0144] During the initialization phase, the matrix Initialize using a Gaussian distribution, matrix Initialize to a zero matrix. This ensures that at the start of fine-tuning, The model's behavior is completely consistent with the pre-trained state, thus smoothly inheriting the robust manifold representation capability. The LoRA-adapted inference encoder outputs high-dimensional semantic features optimized for the specific task. To map this to specific control commands, a lightweight task classification head is attached to the end of the encoder. First, the feature sequence... Perform global average pooling to obtain the global intent vector. Subsequently, a fully connected layer and a softmax activation function are used to calculate the probability distribution of each motion imagery category. :

[0145]

[0146] in, and These are the trainable parameters of the classification head. The class with the highest output probability is the driver's intention as determined by the system. The model is then trained in a supervised manner using a small amount of labeled calibration data. The classification loss function is defined as cross-entropy loss.

[0147]

[0148] in One-hot encoding for the real label, The model predicts the first Class probability. During backpropagation, the gradient only flows to the classification head parameters. and LoRA bypass parameters The massive backbone network parameters It remains unchanged. Through this LoRA-based fine-tuning strategy, not only is the amount of parameter updates reduced by orders of magnitude, enabling rapid deployment and personalized calibration on automotive edge devices, but the noise-resistant manifold structure learned during the pre-training stage is also effectively preserved. This allows the final classifier to still make accurate decisions using robust feature boundaries when faced with electrode detachment or signal interference.

[0149] The user intent recognition method provided in this application improves the cross-user generalization ability of the target intent recognition model, making the target intent recognition model more adaptable to the personalized EEG signal characteristics in real vehicle cockpit scenarios, thereby outputting more stable and accurate user intent recognition information.

[0150] Figure 8 The flowchart illustrating the implementation of the user intent recognition method provided in Embodiment Seven of this application is shown. The difference between this method and Embodiment Six is ​​that step S602 specifically includes:

[0151] Step S701: Generate multiple user EEG signal encoding feature information based on the multiple filtered user EEG signals and the target EEG signal encoding feature generation sub-model.

[0152] In this embodiment, multiple filtered user EEG signals can be input into the target EEG signal encoding feature generation sub-model, thereby completing global spatiotemporal topological feature extraction and nonlinear decoupling abstraction, and then generating multiple user EEG signal encoding feature information.

[0153] Step S702: Generate a sub-model based on the multiple user EEG signal encoding feature information and target intent recognition information to generate user EEG signal intent recognition information.

[0154] In this embodiment, multiple user EEG signal encoded feature information can be input into the target intent recognition information generation sub-model, and then global average pooling and probability classification calculation can be performed to generate user EEG signal intent recognition information.

[0155] Step S703: Based on the user's EEG signal intention recognition information and the preset user intention recognition label information, the target intention recognition model is fine-tuned to generate a fine-tuned target intention recognition model.

[0156] In this embodiment, the preset user intent recognition label information can be preset by humans. The classification loss can be calculated based on the user's EEG signal intent recognition information and the preset user intent recognition label information. Then, the model bypass parameters are updated based on the classification loss and the backbone parameters are frozen, thereby generating a fine-tuned target intent recognition model.

[0157] The user intent recognition method provided in this application enhances the ability of the target intent recognition model to specifically identify motion imagination tasks in the vehicle cockpit, effectively improving the accuracy and noise resistance stability of intent recognition.

[0158] Corresponding to the method in the above embodiments, Figure 9 A structural block diagram of a user intent recognition device provided in an embodiment of this application is shown. For ease of explanation, only the parts related to the embodiment of this application are shown. Figure 9 The example user intent recognition device can be the execution subject of the user intent recognition method provided in the aforementioned embodiment 1.

[0159] Reference Figure 9 The user intent recognition device includes:

[0160] User EEG signal acquisition module 810 is used to acquire multiple user EEG signals;

[0161] The user intent recognition information generation module 820 is used to generate user intent recognition information based on the multiple user EEG signals and the target intent recognition model.

[0162] The process by which each module in the user intent recognition device provided in this application implements its respective function can be specifically referred to the foregoing. Figure 1 The description of Embodiment 1 shown will not be repeated here.

[0163] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0164] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0165] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0166] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0167] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance. It should also be understood that although the terms "first," "second," etc., are used in the text to describe various elements in some embodiments of this application, these elements should not be limited by these terms. These terms are merely used to distinguish one element from another. For example, a first table may be named a second table, and similarly, a second table may be named a first table, without departing from the scope of the various described embodiments. Both the first table and the second table are tables, but they are not the same table.

[0168] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0169] The user intent recognition method provided in this application embodiment can be applied to terminal devices such as mobile phones, tablets, wearable devices, in-vehicle devices, augmented reality / virtual reality devices, laptops, super mobile personal computers, netbooks, and personal digital assistants. This application embodiment does not impose any restrictions on the specific type of terminal device.

[0170] For example, the terminal device may be a station in a WLAN, a cellular phone, a cordless phone, a session initiation protocol phone, a wireless local loop station, a personal digital processing device, a handheld device with wireless communication capabilities, a computing device or other processing device connected to a wireless modem, an in-vehicle device, a vehicle networking terminal, a computer, a laptop computer, a handheld communication device, a handheld computing device, a satellite wireless device, a wireless modem card, a set-top box, a user premises equipment, and / or other devices for communication over a wireless system, as well as next-generation communication systems, such as mobile terminals in 5G networks or mobile terminals in future evolved public terrestrial mobile networks, etc.

[0171] Figure 10 This is a schematic diagram of the structure of a terminal device provided in an embodiment of this application. For example... Figure 10 As shown, the terminal device 9 of this embodiment includes: at least one processor 90 ( Figure 10 (Only one is shown in the image) a memory 91, which stores a computer program 92 that can run on the processor 90. When the processor 90 executes the computer program 92, it implements the steps in the various user intent recognition method embodiments described above, for example... Figure 1 Steps S101 to S102 are shown. Alternatively, when the processor 90 executes the computer program 92, it implements the functions of each module / unit in the above-described device embodiments, for example... Figure 9 The functions of modules 810 to 820 are shown.

[0172] The terminal device 9 can be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor 90 and a memory 91. Those skilled in the art will understand that... Figure 10 This is merely an example of terminal device 9 and does not constitute a limitation on terminal device 9. It may include more or fewer components than shown, or combine certain components, or different components. For example, the terminal device may also include input transmission devices, network access devices, buses, etc.

[0173] The processor 90 may be a central processing unit, or it may be other general-purpose processors, digital signal processors, application-specific integrated circuits, off-the-shelf programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0174] In some embodiments, the memory 91 may be an internal storage unit of the terminal device 9, such as a hard disk or memory of the terminal device 9. The memory 91 may also be an external storage device of the terminal device 9, such as a plug-in hard disk, smart memory card, secure digital card, flash memory card, etc., equipped on the terminal device 9. Furthermore, the memory 91 may include both internal and external storage units of the terminal device 9. The memory 91 is used to store operating systems, applications, bootloaders, data, and other programs, such as the program code of computer programs. The memory 91 can also be used to temporarily store data that has been sent or will be sent.

[0175] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0176] This application also provides a terminal device, which includes at least one memory, at least one processor, and a computer program stored in the at least one memory and executable on the at least one processor. When the processor executes the computer program, it causes the terminal device to implement the steps in any of the above method embodiments.

[0177] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps described in the various method embodiments above.

[0178] This application provides a computer program product that, when run on a terminal device, enables the terminal device to implement the steps described in the various method embodiments above.

[0179] If the integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory, a random access memory, an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0180] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0181] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0182] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0183] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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. Such 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 this application, and should all be included within the protection scope of this application.

Claims

1. A method for recognizing user intent, characterized in that, include: Acquire EEG signals from multiple users; Based on the multiple user EEG signals and the target intent recognition model, user intent recognition information is generated; The target intent recognition model is obtained through the following steps: Acquire multiple EEG signals to be processed; Spatiotemporal slicing and projection processing are performed on the multiple EEG signals to be processed to generate multiple EEG signal sequence information to be processed. Based on multiple preset masking ratios and preset dynamic masking processing rules, the multiple EEG signal sequence information to be processed is masked to generate multiple EEG signal masking sequence information; wherein, the masking ratio increases with the increase of the training rounds of the initial intent recognition model; Based on the multiple EEG signal mask sequence information and the initial intent recognition model, the target intent recognition model is obtained; The initial intent recognition model includes an initial EEG signal coding feature generation sub-model, an initial intent recognition inference distribution generation sub-model, an initial intent recognition reference distribution generation sub-model, and an initial intent recognition EEG signal reconstruction sub-model. The step of obtaining the target intent recognition model based on the multiple EEG signal mask sequence information and the initial intent recognition model specifically includes: Based on the multiple EEG signal mask sequence information and the initial EEG signal coding feature generation sub-model, multiple EEG signal mask sequence feature information are generated; Based on the multiple EEG signal mask sequence feature information, multiple EEG signal sequence information to be processed, the initial intent recognition inference distribution generation sub-model, and the initial intent recognition reference distribution generation sub-model, multiple EEG signal mask sequence inference distribution information and multiple EEG signal sequence reference distribution information are generated. Based on the distribution information inferred from the multiple EEG signal mask sequences and the reference distribution information of the multiple EEG signal sequences, alignment processing is performed to generate multiple aligned EEG signal mask sequence distribution information; Based on the distribution information of the multiple aligned EEG signal mask sequences and the initial intention recognition EEG signal reconstruction sub-model, multiple EEG signal mask sequence reconstruction information are generated. Based on the reconstruction information of the multiple EEG signal mask sequences and the multiple EEG signals to be processed, the EEG signal reconstruction loss information is calculated. Based on the initial intent recognition model, the EEG signal reconstruction loss information, and the preset EEG signal reconstruction loss threshold, the target intent recognition model is obtained.

2. The user intent recognition method as described in claim 1, characterized in that, The step of performing spatiotemporal slicing and projection processing on the plurality of EEG signals to be processed to generate a plurality of EEG signal sequence information specifically includes: Based on the preset spatiotemporal slice size information of the EEG signal, the multiple EEG signals to be processed are spatiotemporal sliced ​​to obtain multiple spatiotemporal block information of the EEG signals to be processed. Based on the spatiotemporal block information of the multiple EEG signals to be processed and the preset EEG signal linear projection model, multiple EEG signal linear projection information is generated. Based on the linear projection information of the multiple EEG signals and the preset EEG signal location encoding information, multiple EEG signal sequence information to be processed is generated.

3. The user intent recognition method as described in claim 1, characterized in that, The preset mask ratios include a preset first mask ratio and a preset second mask ratio; Wherein, the preset first mask ratio is less than the preset second mask ratio; The plurality of EEG signal sequence information to be processed includes a plurality of first EEG signal sequence information to be processed and a plurality of second EEG signal sequence information to be processed; The plurality of EEG signal mask sequence information includes a plurality of first EEG signal mask sequence information and a plurality of second EEG signal mask sequence information; The step of masking the multiple EEG signal sequence information to be processed according to multiple preset masking ratios and preset dynamic masking processing rules to generate multiple EEG signal mask sequence information specifically includes: The first mask matrix information is generated based on the preset first mask ratio and the preset dynamic mask processing rules; Based on the first mask matrix information and multiple first EEG signal sequence information to be processed, a dot product calculation is performed to generate multiple first EEG signal mask sequence information; The second mask matrix information is generated based on the preset second mask ratio and the preset dynamic mask processing rules; Based on the second mask matrix information and multiple second EEG signal sequence information to be processed, a dot product calculation is performed to generate multiple second EEG signal mask sequence information.

4. The user intent recognition method as described in claim 1, characterized in that, The step of generating multiple EEG signal mask sequence inference distribution information and multiple EEG signal sequence reference distribution information based on the multiple EEG signal mask sequence feature information, multiple EEG signal sequence information to be processed, the initial intent recognition inference distribution generation sub-model, and the initial intent recognition reference distribution generation sub-model specifically includes: Based on the feature information of the multiple EEG signal mask sequences and the initial intention recognition inference distribution generation sub-model, multiple EEG signal mask sequence inference distribution information is generated; Based on the multiple EEG signal sequence information to be processed and the initial intention recognition reference distribution generation sub-model, multiple EEG signal sequence reference distribution information are generated.

5. The user intent recognition method as described in claim 1, characterized in that, The step of generating user intent recognition information based on the multiple user EEG signals and the target intent recognition model specifically includes: Spatial filtering is performed on the multiple user EEG signals to obtain multiple filtered user EEG signals; Based on the multiple filtered user EEG signals, the target intent recognition model is fine-tuned to generate a fine-tuned target intent recognition model. User intent recognition information is generated based on the multiple user EEG signals and the fine-tuned target intent recognition model.

6. The user intent recognition method as described in claim 5, characterized in that, The target intent recognition model includes a target EEG signal encoding feature generation sub-model and a target intent recognition information generation sub-model. The step of fine-tuning the target intent recognition model based on the multiple filtered user EEG signals to generate a fine-tuned target intent recognition model specifically includes: Based on the multiple filtered user EEG signals and the target EEG signal coding feature generation sub-model, multiple user EEG signal coding feature information is generated. Based on the encoded feature information of the multiple user EEG signals and the target intent recognition information, a sub-model is generated to generate user EEG signal intent recognition information; Based on the user's EEG signal intention recognition information and the preset user intention recognition label information, the target intention recognition model is fine-tuned to generate a fine-tuned target intention recognition model.

7. A user intent recognition device, characterized in that, For performing the user intent recognition method as described in claim 1, comprising: The user EEG signal acquisition module is used to acquire multiple user EEG signals. The user intent recognition information generation module is used to generate user intent recognition information based on the multiple user EEG signals and the target intent recognition model.

8. A terminal device, characterized in that, The terminal device includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 6.

Images