A method and system for assessing attention levels based on brain-computer interfaces

By using dual probes to collect brain responses and employing the improved SPDNet model and HSMM algorithm, the accuracy and stability issues of attention level assessment in brain-computer interface technology have been resolved, enabling personalized attention level assessment.

CN123116013APending Publication Date: 2026-11-17NAOMI (HENAN) INTELLECTUAL PROPERTY MANAGEMENT CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611323084.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-28
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing brain-computer interface technologies struggle to effectively separate the common components in the brain responses induced by target probes and interference probes, leading to inaccurate attention level assessments and difficulty in adapting to differences in brain response structures among different subjects, thus affecting the stability and individualization of the assessments.

Method used

By employing dual probe brain response acquisition, individual attention anchor construction, improved SPDNet model and improved HSMM algorithm, selective attention differential evidence sequence is generated through shared manifold mapping, anchor constraint common response separation and attention differential projection. Then, the state persistence probability and ordered state decoding are performed by the improved HSMM algorithm to generate attention level temporal trajectory.

Benefits of technology

It improves the accuracy and reliability of attention level assessment, can adapt to differences in brain response structure among different subjects, reduces interference, and generates continuous and stable attention level temporal trajectories.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116013A_ABST
    Figure CN123116013A_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for assessing attention levels based on a brain-computer interface (BCI), belonging to the field of BCI technology. The method includes the following steps: receiving assessment task information and generating an attention level assessment constraint set; controlling the BCI to present target probes and interference probes, collecting and pairing brain responses; correcting brain responses and determining response confidence, constructing a dual brain response covariance sequence and an individual attention anchor set; inputting both into an improved SPDNet model, and generating a selective attention differential evidence sequence through shared manifold mapping, anchor constraint common response separation, and attention differential projection; using an improved HSMM algorithm to perform confidence gating, state persistence probability correction, and ordered state decoding to generate an attention level temporal trajectory; summarizing state persistence and determining the trend of change, and generating an attention level assessment result. This invention can reduce the interference of common responses and low-confidence brain responses, improving assessment accuracy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of brain-computer interface technology, and in particular to a method and system for assessing attention levels based on brain-computer interfaces. Background Technology

[0002] With the continuous development of brain-computer interface technology in cognitive training, human-computer interaction, and task status monitoring, the continuous assessment of subjects' attention levels using brain responses has become an important research direction in cognitive state analysis. Existing methods typically collect brain response signals during a task, extract temporal, frequency, spatial, or covariance features, and identify attentional states at different time points using classification or temporal models to obtain results on changes in attention levels.

[0003] However, the brain responses induced by the target probe and the interference probe contain many common components, and existing methods are unable to separate the differential responses directly related to selective attention; the brain response structures and attentional change ranges of different subjects are different, and it is difficult to form individualized assessments with a unified feature space and fixed judgment criteria.

[0004] Abnormalities in the acquisition channels, transient mutations, and a reduction in effective channels can also decrease the reliability of brain responses, which can easily be misjudged as a decline in attention levels. The relationship between the probability of fixed state persistence and the transition between ordinary states is also difficult to reflect the continuous changes in attention evidence, which can easily cause frequent jumps or cross-level transitions in attention states, affecting the stability of the temporal assessment of attention levels.

[0005] Therefore, how to provide a method and system for assessing attention levels based on brain-computer interfaces is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method and system for assessing attention levels based on brain-computer interfaces. This invention combines dual probe brain response acquisition, individual attention anchor construction, improved SPDNet model and improved HSMM algorithm to separate common responses and continuously decode attention states. It has the advantages of strong individual adaptability, high anti-interference ability and accurate and stable assessment results.

[0007] The brain-computer interface-based attention level assessment method according to embodiments of the present invention includes the following steps: Step 1: Receive attention level assessment task information, parse the brain-computer interface task and assessment requirements, and generate an attention level assessment constraint set; Step 2: Based on the attention level assessment constraint set, the brain-computer interface presents the target probe and interference probe, synchronously collects and pairs brain responses, and generates a dual probe brain response sequence. Step 3: Correct the brain response sequence of the dual probe according to the probe event location, complete the response credibility determination, and generate a set of credible dual brain response units; Step 4: Calculate the covariance matrix of the target probe brain response and the interference probe brain response in the trusted dual brain response unit set, respectively, perform positive definite processing, and generate the dual brain response covariance sequence. Step 5: Based on the focused attention calibration and attention shift calibration, select calibration units from the dual brain response covariance sequence and perform manifold aggregation to generate individual attention anchor set; Step 6: Input the dual brain response covariance sequence and individual attention anchor set into the improved SPDNet model, and generate a selective attention differential evidence sequence through shared manifold mapping, anchor-constrained common response separation and attention differential projection. Step 7: Input the selective attention differential evidence sequence into the improved HSMM algorithm, and generate the attention level time trajectory by passing the confidence gating, evidence continuity correction state persistence probability and ordered state decoding. Step 8: Continuously summarize the execution status and determine the trend of changes in the attention level time-series trajectory, and generate attention level assessment results.

[0008] Optionally, step one specifically includes: Receive attention level assessment task information and read probe presentation and calibration instructions in the brain-computer interaction task according to the task execution order; The target probes are determined based on the task content that the subjects need to continuously focus on, and the probes that have the same stimulus conditions but do not undertake the current task response are determined as interference probes. The target probes and interference probes are arranged alternately along the task time axis so that the paired probes are in the same task stage, and the corresponding brain response acquisition time period is configured for each pair of paired probes. According to the calibration instructions, set up a focused attention calibration stage and an attention transfer calibration stage, so that both calibration stages include the complete presentation process of the target probe and the interference probe; Based on the assessment requirements, the probe presentation order, brain response acquisition time period, and coverage relationship of the calibration phase are verified, and the verified task configuration is written into the attention level assessment constraint set.

[0009] Optionally, step two specifically includes: Read the attention level assessment constraint set, connect the probe presentation process and the brain response acquisition process to the same task timeline, and start continuous brain response acquisition before the first probe is presented; The target probe and interference probe are presented alternately along the probe presentation sequence, and the corresponding probe event position is written into the continuous brain response each time a probe is triggered. Based on the pairing arrangement of the attention level assessment constraint set records, the target probe event locations and interference probe event locations are bound as dual event groups; Each pair of events is associated with a continuous brain response along the task timeline, so that each group of target probes and interference probes has a corresponding brain response acquisition segment; The brain response acquisition segments are arranged according to the time sequence of the dual event groups to generate the dual probe brain response sequence.

[0010] Optionally, step three specifically includes: Read the brain response sequence of the dual probes, extract the brain response of the target probe and the brain response of the interference probe according to the probe event location of each dual event group, and unify the event starting point and sampling location of the two types of brain responses; Based on the average brain response of each acquisition channel before the probe event location, the target probe brain response and the interference probe brain response are corrected respectively. Examine the amplitude saturation and transient mutations of each acquisition channel in the calibrated brain response, isolate abnormal acquisition channels, and retain the common effective acquisition channels in each pair of events. The confidence level of the response is calculated based on the retention of the common effective acquisition channel and the correspondence of the sampling locations of the two types of brain responses. Assembly is stopped for the dual event group that lacks a common effective acquisition channel. The target probe brain response, interference probe brain response, and response confidence, formed based on a common effective acquisition channel, are bound into a trusted dual brain response unit. The trusted dual brain response unit set is generated by arranging the dual event groups in chronological order.

[0011] Optionally, step four specifically includes: Read the set of trusted dual brain response units, and organize the target probe brain response and the interference probe brain response into a brain response matrix according to the arrangement order of the common effective acquisition channels; The centered brain response matrix is ​​obtained by subtracting the mean value of the corresponding channel from the sampled value of each acquisition channel in each brain response matrix. Calculate the product of the centered brain response matrix and its transpose, and scale it by one less effective sample size to form the target brain response covariance matrix and the interference brain response covariance matrix, respectively. Construct unit reference matrices of the same dimension based on the traces of the two covariance matrices, and determine the common contraction intensity by combining the deviations of the two covariance matrices relative to the corresponding unit reference matrices. According to the common contraction intensity, the target brain response covariance matrix and the interference brain response covariance matrix are respectively weighted and fused with the corresponding unit reference matrix; Perform eigenvalue decomposition on the weighted fusion covariance matrix, correct non-positive eigenvalues, and reconstruct a symmetric positive definite matrix according to the corresponding eigenvectors; The target brain response covariance matrix, interference brain response covariance matrix, and response confidence corresponding to the same reliable dual brain response unit are bound as dual brain response covariance units and arranged in chronological order to generate a dual brain response covariance sequence.

[0012] Optionally, step five specifically includes: Read the dual brain response covariance sequence and aggregate the corresponding dual brain response covariance units according to the concentration attention calibration stage and the attention transfer calibration stage; Within each calibration phase, the response confidence of each dual brain response covariance unit is read, the aggregation weight of the corresponding calibration unit is determined based on the response confidence, and the aggregation weight within the current calibration phase is normalized. The target brain response covariance matrix and the interference brain response covariance matrix are extracted from each calibration unit. Matrix logarithmic mapping is performed on the extracted covariance matrices to generate the target response logarithmic matrix and the interference response logarithmic matrix. The target response logarithmic matrix and the disturbance response logarithmic matrix within the same calibration stage are accumulated according to the aggregation weights to form the target response aggregation matrix and the disturbance response aggregation matrix. Perform matrix exponential mapping on the target response aggregation matrix and the interference response aggregation matrix to restore the symmetric positive definite structure and generate the target brain response anchor point and the interference brain response anchor point for the current calibration stage; By binding the target brain response anchor and the interference brain response anchor in the same calibration phase as the dual brain response anchor, and combining the dual brain response anchors of the two calibration phases, an individual attention anchor set is generated.

[0013] Optionally, the improved SPDNet model includes a shared manifold mapping layer, an anchor point tangent space mapping layer, a common response separation layer, and an attention evidence projection layer, wherein the common response separation layer embeds an anchor point constraint common response separation mechanism; The shared manifold mapping layer is configured with a target response branch, an interference response branch, and an anchor point mapping branch. The three branches share bilinear mapping weights and eigenvalue correction rules. The three branches process the target brain response covariance matrix, the interfering brain response covariance matrix, and the individual attention anchor set, respectively, to generate dual response manifold features and anchor manifold features; The anchor point tangent space mapping layer generates an individual manifold reference matrix based on the anchor point manifold features. The dual response manifold features and anchor point manifold features are uniformly transformed using the individual manifold reference matrix. The dual response tangent space features and anchor point tangent space features in the same tangent space are generated through matrix logarithmic mapping. The common response separation layer inputs the dual response tangent space features and anchor point tangent space features into the anchor point constraint common response separation mechanism to establish the target response benchmark and interference response benchmark with the anchor point tangent space features in the attention transfer calibration stage, and determines the tangent space offset of the current target response and interference response relative to the corresponding benchmark. Extract the overlapping portions with consistent directions and matching change ratios from the two types of tangent space offsets as common response components. Separate the common response components from the two types of tangent space offsets to generate target response difference components and disturbance response difference components. The target response difference component and the interference response difference component are directionally matched with the anchor point tangent space features corresponding to the focused attention calibration stage to generate target enhancement evidence and interference suppression evidence. The attention evidence projection layer adjusts the projection contributions of target enhancement evidence and interference suppression evidence based on response credibility; The target enhancement evidence, interference suppression evidence, and response credibility are bound into selective attention differential evidence units, which are then arranged in chronological order to generate a selective attention differential evidence sequence.

[0014] Optionally, step seven specifically includes: Read the selective attention differential evidence sequence, set ordered attention states according to attention level from low to high, and call the attention observation model corresponding to each attention state; Target enhancement evidence and interference suppression evidence are read from the selective attention differential evidence unit. The degree of matching between the two types of evidence and the evidence distribution of each attention state is calculated to generate the state observation probability. Read the response confidence carried by the selective attention differential evidence unit, adjust the differentiation amplitude of the state observation probability according to the response confidence, so that the low confidence brain response forms a neutral observation that tends to be evenly distributed among various attention states, and keep the probability of low attention state from increasing due to the decrease in response confidence. By combining consecutive selective attention differential evidence units in chronological order, candidate state segments with different durations are established according to each attention state. Compare the changing directions of target enhancement evidence and interference suppression evidence in adjacent selective attention differential evidence units within the candidate state segment. Based on the degree of continuity of the two types of evidence changing in the same direction and the corresponding response credibility, adjust the basic persistence probability and generate the adjusted state persistence probability corresponding to each candidate persistence length. Divide the probability of the modified state duration corresponding to the duration of each candidate under the same attention state by the sum of the probabilities of the modified state duration under the same attention state, so that the sum of the normalized probabilities of the modified state duration is one. Establish the transition relationship between adjacent attentional states according to the order of attentional levels, block direct transitions across intermediate attentional states, and generate ordered state transition relationships; The state scores of candidate state segments are recursively derived by combining the state observation probability, the corrected state persistence probability, and the ordered state transition relationship, and the optimal state path corresponding to each time position is retained. By backtracking the optimal state path, the temporal location and duration of each attentional state are determined, and a temporal trajectory of the attentional level is generated.

[0015] Optionally, step eight specifically includes: Read the attention level temporal trajectory, merge the positions of trajectories that maintain the same attention state continuously along the time sequence into attention state segments, and record the attention state and duration corresponding to each attention state segment; Each attentional state segment is aggregated according to the attentional state, the duration corresponding to the same attentional state is accumulated, and the proportion of the corresponding duration to the total duration of the attentional level time sequence trajectory is calculated to generate a summary result of the state duration. Based on the state levels corresponding to low attention, medium attention, and high attention, the attention level score is generated by multiplying each state level by the percentage of its corresponding duration and summing the results. The overall attention level is determined by comparing the attention level score with the attention level classification threshold. The task timeline is divided into a pre-assessment interval and a post-assessment interval. Attention level scores are calculated based on the duration of each attention state within the corresponding assessment interval. Calculate the difference between the attention level score of the later stage and the attention level score of the earlier stage, and generate the attention change trend based on the comparison between the difference and the attention change trend threshold; The results of continuous state summarization, attention level score, overall attention level, and attention change trend are combined to generate an attention level assessment result.

[0016] Optional, a brain-computer interface-based attention level assessment system includes the following modules: The task constraint configuration module is used to parse attention level assessment task information, configure probe presentation, brain response acquisition and calibration processes, and generate attention level assessment constraint set. The probe brain response acquisition module is used to control the brain-computer interface to present target probes and interference probes alternately, bind corresponding brain response acquisition segments, and generate dual probe brain response sequences. The brain response credibility processing module is used to correct the target probe brain response and the interference probe brain response according to the probe event location, screen the common effective acquisition channels and calculate the response credibility, and generate a set of credible dual brain response units. The positive definite covariance processing module is used to calculate the target brain response covariance matrix and the interference brain response covariance matrix, complete the positive definite processing and bind it with the response confidence, and generate the dual brain response covariance sequence. The individual attention anchor construction module is used to screen dual brain response covariance units according to the concentrated attention calibration stage and the attention transfer calibration stage, and perform manifold aggregation based on the response credibility to generate an individual attention anchor set; Attention differential evidence extraction module is used to call the improved SPDNet model to separate common responses and project attention differential evidence, bind target enhancement evidence, interference suppression evidence and response credibility to generate selective attention differential evidence sequence; Attention state temporal decoding module is used to call the improved HSMM algorithm to perform confidence gating, state persistence probability correction and ordered state decoding to generate attention level temporal trajectory; The attention assessment summary module is used to summarize the duration of each attention state, determine the trend of attention changes, and generate attention level assessment results.

[0017] The beneficial effects of this invention are: The target probe and interference probe are presented alternately according to the same task stage, and the corresponding brain response acquisition segments are paired and processed to ensure that the two types of brain responses have consistent stimulation conditions and time references. By combining probe event location correction, screening of common effective acquisition channels and determination of response credibility, the interference of task stage differences, channel abnormalities and transient mutations on attention assessment data can be reduced, providing a reliable data basis for subsequent differential response analysis.

[0018] The dual brain response covariance sequence formed by positive definite processing preserves the correlation structure between brain response channels. The covariance matrices corresponding to the focused attention calibration and attention transfer calibration are aggregated by manifold to form an individual attention anchor set, which can adapt to the differences in brain response structure among different subjects. The improved SPDNet model uses the anchor-constrained common response separation mechanism to extract the target response difference component and the interference response difference component, reducing the influence of the co-induced components of the two types of probes and highlighting the evidence of target enhancement and interference inhibition.

[0019] After the selective attention differential evidence sequence is fed into the improved HSMM algorithm, the response confidence is used to adjust the state observation probability to avoid low confidence brain responses being directly interpreted as low attention states; the evidence continuity is used to correct the state persistence probability, and the ordered state transition relationship is used to limit cross-level jumps. This generates a continuous and stable attention level time-series trajectory, and the attention level assessment result is formed by continuously summarizing the states and judging the change trend, thereby improving the accuracy and reliability of attention level assessment. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1This is an overall flowchart of the attention level assessment method based on brain-computer interface proposed in this invention; Figure 2 This is a diagram showing the relationship between the brain response of the target probe and the brain response of the interference probe proposed in this invention. Figure 3 This is a structural diagram of the improved SPDNet model proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figures 1-3 A brain-computer interface-based method for assessing attention levels includes the following steps: Step 1: Receive attention level assessment task information, parse the brain-computer interface task and assessment requirements, and generate an attention level assessment constraint set; Step 2: Based on the attention level assessment constraint set, the brain-computer interface presents the target probe and interference probe, synchronously collects and pairs brain responses, and generates a dual probe brain response sequence. Step 3: Correct the brain response sequence of the dual probe according to the probe event location, complete the response credibility determination, and generate a set of credible dual brain response units; Step 4: Calculate the covariance matrix of the target probe brain response and the interference probe brain response in the trusted dual brain response unit set, respectively, perform positive definite processing, and generate the dual brain response covariance sequence. Step 5: Based on the focused attention calibration and attention shift calibration, select calibration units from the dual brain response covariance sequence and perform manifold aggregation to generate individual attention anchor set; Step 6: Input the dual brain response covariance sequence and individual attention anchor set into the improved SPDNet model, and generate a selective attention differential evidence sequence through shared manifold mapping, anchor-constrained common response separation and attention differential projection. Step 7: Input the selective attention differential evidence sequence into the improved HSMM algorithm, and generate the attention level time trajectory by passing the confidence gating, evidence continuity correction state persistence probability and ordered state decoding. Step 8: Continuously summarize the execution status and determine the trend of changes in the attention level time-series trajectory, and generate attention level assessment results.

[0023] In this embodiment, step one specifically includes: After receiving the attention level assessment task information, the brain-computer interaction task is parsed according to the task execution order. The probe presentation instructions and calibration instructions corresponding to each task stage are read, and the probe presentation instructions, calibration instructions and assessment requirements are bound to the corresponding task stage. If any task stage lacks necessary instructions, the generation of the attention level assessment constraint set is stopped. The target probe is determined based on the task content that the subject needs to continuously focus on in the current task stage. The stimulation conditions recorded by the candidate probe and the target probe in the brain-computer interaction task are compared one by one. The candidate probes with all the same stimulation conditions and do not bear the current task response are identified as interference probes. When no interference probe that meets the conditions is obtained, the configuration of the current task stage is stopped. The target probes and interference probes are arranged alternately along the task timeline. The target probes and interference probes in the same task stage are bound into paired probes according to the presentation order. A brain response acquisition period with the same start and end offset relative to the event position of each pair of paired probes is set. The calibration instructions divide the process into a focused attention calibration phase and an attention transfer calibration phase. Paired probes from the corresponding task phase are written into the corresponding calibration phase. The calibration phase is checked to see if both calibration phases contain the complete presentation process of the target probe and the interference probe. If either calibration phase lacks a complete presentation process, the calibration configuration is stopped. Verify whether the presentation position of each probe falls within the corresponding brain response acquisition time period, check the probe presentation order, pairing relationship and coverage relationship between the calibration stages, and write the verified task configuration into the attention level assessment constraint set according to the task time axis. By uniformly defining the presentation order of target probes and interference probes, the brain response acquisition period, and the individual attention calibration range through the attention level assessment constraint set, a task configuration basis is provided for the generation of dual probe brain response sequences.

[0024] In this embodiment, step two specifically involves: Read the attention level assessment constraint set, connect the probe presentation process and the brain response acquisition process to the same task timeline, start continuous brain response acquisition before the first probe presentation, and determine each sampling position according to the acquisition clock of the brain-computer interface; The target probe and interference probe are presented alternately in the probe presentation order recorded along the attention level evaluation constraint set. The sampling position corresponding to the effective time of the probe presentation command is determined as the probe event position. When the effective time is between adjacent sampling positions, the sampling position with the smaller time difference is selected. When the sampling position cannot be matched within a sampling interval before or after the effective time, the assembly of the corresponding probe is stopped. According to the pairing arrangement, the target probe event position and the interference probe event position in the same task stage are bound into a pair event group. When the pairing relationship is missing or the same probe event position corresponds to multiple pairing relationships, the assembly of the corresponding pair event group is stopped. Using the location of each probe event as the time reference, brain response acquisition segments are extracted from continuous brain responses according to the brain response acquisition period recorded in the attention level assessment constraint set. Assembly of corresponding pairwise event groups is stopped when the sampling locations before the event are insufficient or the sampling locations after the event are not fully obtained. Bind the two brain response acquisition segments corresponding to the same pair of events to their respective probe event locations, while maintaining the original acquisition channel order and sampling location records of each brain response acquisition segment; A stable sorting is performed according to the chronological order of the earlier probe event locations in each pair of event groups, and the sorted brain response acquisition segments and probe event locations are written into the pair probe brain response sequence. By using paired probe brain response sequences, the task stage, acquisition channel, and relative sampling position of the target probe and the interference probe are kept consistent, providing paired input for brain response correction and response reliability determination.

[0025] In this embodiment, step three specifically includes: Read the brain response sequence of the dual probes, extract the corresponding brain response acquisition segment according to the probe event position recorded in each dual event group, set the target probe event position and the interference probe event position as the relative time zero point respectively, and align the target probe brain response and the interference probe brain response according to the same relative sampling position. When the corresponding sampling position is missing, only the sampling position covered by both types of brain responses is retained. According to the brain response acquisition period recorded in the attention level assessment constraint set, the brain response before the event location of each probe is extracted as the baseline segment. The mean brain response of each acquisition channel in the baseline segment is calculated. The mean brain response of the corresponding channel is subtracted from each sampled value in the target probe brain response and the interference probe brain response. When the baseline segment lacks effective sampled values, the assembly of the corresponding pair of events is stopped. like Figure 2 As shown, the solid and dashed lines represent the brain responses of the target probe and the interference probe, respectively. The slanted area before the probe event location is the baseline segment, and the slanted area after the probe event location is used to represent the time range in which the differences between the two types of brain responses are concentrated. It is not used as a separate data extraction condition. Read the effective measurement upper and lower limits and instantaneous change limits recorded in the attention level assessment constraint set, check the sampled values ​​of each acquisition channel after correction, and determine the amplitude saturation when the sampled value reaches the effective measurement upper and lower limits. Determine the instantaneous change limit when the difference between adjacent sampled values ​​exceeds the instantaneous change limit. Identify the acquisition channels that show amplitude saturation or instantaneous change as abnormal acquisition channels. Abnormal acquisition channels in the target probe brain response and interference probe brain response are isolated. The intersection of the remaining acquisition channels of the two types of brain responses is taken as the common effective acquisition channel. The arrangement relationship of the common effective acquisition channels is maintained according to the original acquisition channel order. When the common effective acquisition channel is empty, the assembly of the corresponding pair of events is stopped. Calculate the ratio of the number of common effective acquisition channels to the original number of acquisition channels in the dual event group, and calculate the ratio of the number of sampling locations jointly covered by the two types of brain responses to the number of sampling locations specified during the brain response acquisition period. Multiply the two ratios to obtain the response confidence. The target probe brain response and interference probe brain response are reorganized according to the common effective acquisition channel and common coverage sampling location. The two types of brain responses are bound to the response confidence as a trusted dual brain response unit, and written into the trusted dual brain response unit set according to the time order of the dual event group. The set of trusted dual brain response units preserves paired brain responses that have completed event alignment, benchmark correction, and response credibility determination, providing a unified input for the calculation of the target brain response covariance matrix and the interference brain response covariance matrix.

[0026] In this embodiment, step four specifically includes: Read the set of trusted dual brain response units, and organize the target probe brain response and interference probe brain response in each trusted dual brain response unit into a brain response matrix with the number of channels multiplied by the number of sampling positions, according to the original arrangement order of the common effective acquisition channels. If the number of acquisition channels or the number of sampling positions in the two brain response matrices are inconsistent, stop processing the corresponding trusted dual brain response unit. The mean value of each acquisition channel in the brain response matrix is ​​calculated along the sampling location dimension. The mean value of the corresponding channel is subtracted from each sample value to form a centered brain response matrix. The calculation of the corresponding covariance matrix is ​​stopped when the number of effective samples is less than two. Multiply the centered brain response matrix by its transpose matrix, and divide the product matrix by the effective number of samples minus one to form the target brain response covariance matrix and the interference brain response covariance matrix respectively. Take half of the sum of each covariance matrix and its transpose matrix to eliminate the asymmetric bias caused by floating-point operations. The traces of the target brain response covariance matrix and the interference brain response covariance matrix are calculated separately. The traces of each matrix are divided by the number of common effective acquisition channels, and the results are multiplied by the same-dimensional identity matrix to form the corresponding identity reference matrix. The sampling column vectors of the two centered brain response matrices are extracted sequentially along the sampling position. The product of each sampling column vector and its transpose matrix is ​​calculated relative to the square of the Frobenius norm of the corresponding brain response covariance matrix. The calculation results corresponding to the two centered brain response matrices are accumulated and divided by the square of the effective sampling number to form the shrinkage bias. Calculate the sum of the squares of the Frobenius norm of the target brain response covariance matrix and the interference brain response covariance matrix relative to the corresponding unit reference matrix. When the current sum of squares is greater than zero, divide the contraction deviation by the current sum of squares and truncate the resulting ratio to the range of zero to one to obtain the common contraction intensity. When the current sum of squares is equal to zero, set the common contraction intensity to zero. According to the common contraction intensity, the target brain response covariance matrix and the interference brain response covariance matrix are weighted and fused with the corresponding unit reference matrix respectively. The weight of the brain response covariance matrix is ​​one minus the common contraction intensity, and the weight of the unit reference matrix is ​​the common contraction intensity. Eigenvalue decomposition is performed on the covariance matrix after weighted fusion. All positive eigenvalues ​​are read and the minimum value is determined. Eigenvalues ​​smaller than the minimum value are replaced with the minimum value. If no positive eigenvalue is obtained, the processing of the corresponding trusted dual brain response unit is stopped. The symmetric positive definite matrix is ​​reconstructed according to the updated eigenvalues ​​and corresponding eigenvectors, and the target brain response covariance matrix and the interference brain response covariance matrix are updated respectively. The target brain response covariance matrix, interference brain response covariance matrix and response confidence corresponding to the same trusted dual brain response unit are bound as dual brain response covariance units, and written into the dual brain response covariance sequence according to the time order of trusted dual brain response units. By uniformly preserving the channel correlation structure and response reliability of the target probe brain response and the interference probe brain response through the dual brain response covariance sequence, a symmetric positive definite input is provided for the construction of individual attention anchor set and the improvement of SPDNet model processing.

[0027] In this embodiment, step five specifically includes: Read the dual brain response covariance sequence, and according to the task time position corresponding to each dual brain response covariance unit, the dual brain response covariance units in the focused attention calibration stage and the attention transfer calibration stage are respectively grouped. The same dual brain response covariance unit is only grouped into one calibration stage. When no effective unit is obtained in any calibration stage, the generation of individual attention anchor set is stopped. The two calibration stages are processed separately. The acquisition channels corresponding to each dual brain response covariance unit in the current calibration stage are read. The acquisition channels commonly contained in each unit are taken. The corresponding rows and columns are extracted from the target brain response covariance matrix and the interference brain response covariance matrix according to the original acquisition channel order. The current calibration stage is stopped when no commonly contained acquisition channels are obtained. Eigenvalue decomposition is performed on the target brain response covariance matrix and the interference brain response covariance matrix after channel unification. The natural logarithm of each positive eigenvalue is taken while keeping the corresponding eigenvector unchanged. The target response logarithmic matrix and the interference response logarithmic matrix are reconstructed according to the eigenvector and the logarithmic eigenvalue. When a non-positive eigenvalue appears, the corresponding calibration unit is stopped from being processed. Read the response confidence of each calibration unit after completing the matrix logarithmic mapping, calculate the sum of all response confidence in the current calibration stage, divide each response confidence by the sum of the response confidence to obtain an aggregate weight with a total of one, and stop processing the current calibration stage when the sum of response confidence is zero. Multiply each target response log matrix in the current calibration phase by its corresponding aggregation weight and then sum them matrix by matrix to form the target response aggregation matrix. Process the interference response log matrix in the same way to form the interference response aggregation matrix. Eigenvalue decomposition is performed on the target response aggregation matrix and the interference response aggregation matrix respectively. Natural exponential operation is performed on each eigenvalue while keeping the corresponding eigenvector unchanged. The symmetric positive definite matrix is ​​reconstructed according to the eigenvector and exponential eigenvalue, and the target brain response anchor point and interference brain response anchor point of the current calibration stage are generated respectively. The target brain response anchor and the interference brain response anchor generated in the same calibration stage are bound as dual brain response anchors. The dual brain response anchors corresponding to the concentrated attention calibration stage and the attention transfer calibration stage are combined to generate an individual attention anchor set. By preserving the target response structure and interference response structure of the subjects during the concentrated attention calibration phase and the attention transfer calibration phase through individual attention anchor point sets, an individual manifold reference and a common response separation benchmark are established for improving the SPDNet model.

[0028] In this embodiment, the improved SPDNet model includes a shared manifold mapping layer, an anchor-tangent space mapping layer, a common response separation layer, and an attention evidence projection layer. The common response separation layer embeds an anchor-constrained common response separation mechanism, specifically: The dual brain response covariance sequence and individual attention anchor set are read. In this implementation, a 32-channel brain-computer interface is used. After the attention calibration stage, attention transfer calibration stage and the current evaluation sequence are unified through the acquisition channels, 32 common acquisition channels are retained. This makes the target brain response covariance matrix, the interference brain response covariance matrix and the four brain response anchors all form a 32×32 symmetric positive definite matrix. The model processing is stopped when 32 common acquisition channels are not completely retained. The shared manifold mapping layer sets up a target response branch, an interference response branch, and an anchor point mapping branch. The three branches share two sets of bilinear mapping weights. The first set of weights maps a 32×32 matrix to a 24×24 manifold matrix, and the second set of weights maps a 24×24 manifold matrix to a 16×16 manifold matrix, so that the target response, interference response, and brain response anchor points are always in the same manifold coordinates. After each bilinear mapping, eigenvalue decomposition is performed on the manifold matrix, replacing non-positive eigenvalues ​​with the smallest positive eigenvalue in the current manifold matrix. Then, the symmetric positive definite matrix is ​​reconstructed according to the updated eigenvalues ​​and corresponding eigenvectors. When no positive eigenvalue is obtained, the processing of the corresponding dual brain response covariance unit is stopped, generating dual response manifold features and anchor point manifold features. The anchor point tangent space mapping layer performs matrix logarithmic mapping on the manifold matrices corresponding to the four brain response anchor points, sums the four logarithmic matrices with equal weights, divides by four, and then performs matrix exponential mapping to generate a 16×16 individual manifold reference matrix. Construct an inverse square root matrix based on the eigenvalues ​​and eigenvectors of the individual manifold reference matrix. Multiply the dual response manifold features and anchor point manifold features by the inverse square root matrix on the left and right, respectively. Then perform a matrix logarithmic mapping and extract elements in the order of the upper triangular positions. Keep the original values ​​of the diagonal elements and multiply the off-diagonal elements by the square root of two to generate 136-dimensional dual response tangent space features and anchor point tangent space features. The common response separation layer inputs the dual response tangent space features and anchor point tangent space features into the anchor point constraint common response separation mechanism. The target response benchmark and interference response benchmark are established with the target brain response anchor point tangent space features and interference brain response anchor point tangent space features corresponding to the attention transfer calibration stage. The 136-dimensional tangent space offset of the current target response and interference response relative to the corresponding benchmark is calculated respectively. The two types of tangent space offsets are compared dimension by dimension. When both components at the same position are zero, the corresponding common response component is set to zero. When the other two components are in the same direction, the change ratio matching coefficient is calculated by dividing the smaller absolute value by the larger absolute value. When the change ratio matching coefficient is not less than 0.8, the smaller absolute value and common direction are combined into a common response component, and the remaining positions are set to zero. Then, the common response component is subtracted from the two types of tangent space offsets respectively to generate the target response difference component and the interference response difference component. The spatial features of the target brain response anchor points corresponding to the attentional calibration stage and the attentional transfer calibration stage are subtracted to form the target enhancement reference direction. The interference suppression reference direction is formed in the same way. The cosine similarity between the target response difference component and the interference response difference component and the corresponding reference direction is calculated respectively. When the magnitude of any vector involved in the calculation is zero, the corresponding cosine similarity is set to zero. The remaining negative values ​​are set to zero and used to scale the corresponding response difference component to generate target enhancement evidence and interference suppression evidence. Note that the evidence projection layer uses an 8×136 target evidence projection matrix and interference evidence projection matrix respectively, mapping the target enhancement evidence and interference suppression evidence into two 8-dimensional evidence vectors, which are then multiplied by the response confidence. The two 8-dimensional evidence vectors and the response confidence are bound into selective attention differential evidence units, and then arranged in chronological order to generate a selective attention differential evidence sequence. The model training samples are derived from historical evaluation tasks that have completed focused attention calibration and attention shift calibration. Each training sample includes the dual brain response covariance unit and individual attention anchor set of the same subject. The component mean of the two 8-dimensional evidence vectors is used as the evidence strength. The ranking loss is constructed by the evidence strength of the focused attention calibration sample being higher than that of the attention shift calibration sample. The common response constraint loss is constructed by the absolute value mean of the remaining same-direction overlapping components in the difference components of the two types of responses after the common response separation is completed. Adam is used to update the bilinear mapping weights and two evidence projection matrices. The initial learning rate is 0.001, the maximum training time is 120 epochs, and the training is terminated when the verification loss does not decrease for 10 consecutive epochs. Orthogonalization is performed after each update of the bilinear mapping weights, and the model parameters are fixed after the termination condition is met. The improved SPDNet model retains the original SPDNet model's symmetric positive definite matrix input, bilinear manifold mapping, eigenvalue correction, and matrix logarithmic mapping structure. Simultaneously, it expands the original data pathway to meet the processing needs of dual brain responses. The manifold mapping pathway, originally designed for a single symmetric positive definite matrix, is adjusted to share bilinear mapping weights for the target response branch, interference response branch, and anchor point mapping branch, ensuring consistent manifold transformation coordinates across the three input types. After eigenvalue correction, an individual manifold reference matrix determined by the individual attention anchor point set is introduced to perform a unified transformation on the dual response manifold features and anchor point manifold features, allowing different response objects to enter the same tangent space. After the matrix logarithmic mapping forms the tangent space features, the common response separation layer invokes the anchor point-constrained common response separation mechanism to separate the common changing components in the target and interference responses. The model output replaces the original classification output with an attention evidence projection layer, converting the processing results into target enhancement evidence and interference suppression evidence adjusted by response credibility. This forms a dual brain response differential evidence extraction structure oriented towards individual attention benchmarks. By improving the SPDNet model to separate the common changes of the target response and the interference response in a unified manifold coordinate system, the difference information that conforms to the individual's focused attention direction is preserved, and the response confidence is written into the selective attention differential evidence sequence along with the attention evidence, providing continuous input for improving the HSMM algorithm to perform attention state temporal decoding.

[0029] In this embodiment, step seven specifically includes: Read the selective attention differential evidence sequence, and concatenate the 8-dimensional target enhancement evidence and 8-dimensional interference suppression evidence in each selective attention differential evidence unit into a 16-dimensional attention observation vector in a fixed order, and read the response confidence at the same time. If any dimension of the evidence vector is missing, stop processing the corresponding selective attention differential evidence unit. According to the attention level, low attention state, medium attention state and high attention state are set up from low to high. The evidence distribution of low attention state is established by the sample of attention transfer calibration stage, and the evidence distribution of high attention state is established by the sample of attention concentration calibration stage. The 16-dimensional means corresponding to the two states are equally weighted and fused, and the corresponding 16×16 covariance matrices are equally weighted and fused to establish the evidence distribution of medium attention state. Based on the 16-dimensional mean and 16×16 covariance matrix corresponding to the three attention state evidence distributions, the multivariate Gaussian probability density of the current attention observation vector is calculated respectively. The three calculation results are normalized to the state observation probability. When the covariance matrix is ​​not invertible, eigenvalue decomposition is performed. Non-positive eigenvalues ​​are replaced with the smallest positive eigenvalue in the current covariance matrix and then recalculated. When no positive eigenvalue is obtained, the current selective attention differential evidence sequence is stopped from being processed. The confidence level is used to read the confidence level of the response. When the confidence level of the response is not lower than 0.6, the state observation probability is retained. When the confidence level of the response is lower than 0.6, the state observation probability is contracted to the equal probability of the three attention states according to the confidence level of the response. If the probability of the low attention state after contraction is higher than the probability of the low attention state before contraction, the original low attention state probability is maintained, and the excess probability is redistributed according to the original probability ratio of the medium attention state and the high attention state. Following the chronological order, with each time position as the candidate termination position, and combining consecutive selective attention differential evidence units forward according to the candidate duration length corresponding to each attention state, candidate state segments corresponding to different attention states and different duration lengths are established. The change direction of adjacent target enhancement evidence and adjacent interference suppression evidence in each candidate state segment is compared respectively. The number of positions with the same change direction in the two types of evidence is divided by the total number of corresponding comparison positions. The average of the two ratios is multiplied by the average confidence value of the response in the candidate state segment to generate the evidence continuity. No continuity correction is performed when the candidate state segment is less than two units. Based on the historical sequence of the completed calibration, the frequency of occurrence of three attention states with different number of units is statistically analyzed, normalized to form the basic persistence probability, and the 95th percentile of the historical persistence length of each attention state is determined as the corresponding maximum persistence length. Candidate state segments exceeding the corresponding maximum persistence length are not established. Multiply the base persistence probability corresponding to the persistence length of the candidate state segment by the evidence continuity, keep the base persistence probability of other persistence lengths in the same attention state unchanged, and then normalize all persistence probabilities of the current attention state to generate the corrected state persistence probability. The state transition probability is determined based on the number of occurrences of adjacent attention states in the historical calibration sequence. The duration of the same attention state is represented by the modified state persistence probability. Only the transition between low attention state and medium attention state and the transition between medium attention state and high attention state are retained. Cross-level transitions between low attention state and high attention state are blocked. The allowable transition probability corresponding to each attention state is normalized. The joint state observation probability, modified state duration probability, and state transition probability are dynamically recursively calculated. At each time point, the preceding state with the highest cumulative probability, the duration of the preceding state, and the duration of the current state are recorded. After recursively calculating to the end of the sequence, the process is reversed from the terminating state with the highest cumulative probability to determine the time position and duration range corresponding to each attention state and generate the attention level time sequence trajectory. The improved HSMM algorithm retains the hidden state setting, state observation probability calculation, explicit state persistence probability, state transition, and optimal path backtracking processing of the original HSMM algorithm. To address the issue of inconsistent credibility of selective attention differential evidence, after the state observation probability is generated, credibility gating reduces the state differentiation intensity of low-credibility observations, while restricting low-credibility observations from raising the low-attention state probability. Before the basic persistence probability enters dynamic recursion, evidence continuity correction is added, so that target enhancement evidence, interference suppression evidence, and response credibility jointly adjust the state persistence range. The original free state transition relationship is tightened into an adjacent state transition relationship set according to the attention level order, thus forming a time-series decoding process with controlled credibility, correctable persistence probability, and orderly state changes. By improving the HSMM algorithm, attentional state bias caused by low-reliability brain responses is suppressed, state persistence boundaries are corrected based on continuous changes in dual-type attentional evidence, and cross-level jumps that do not conform to the rules of attentional change are blocked, providing a temporal trajectory of attentional level for state persistence summarization and trend determination.

[0030] In this embodiment, step eight specifically includes: Read the attention level time sequence trajectory, extract the state type, start position, end position and duration of each attention state segment according to the task time position, and stop summarizing the corresponding attention state segment when the start position is missing, the end position is missing or the end position is earlier than the start position. The durations of low attention, medium attention and high attention are summed separately, and the sum of the durations of the three states is determined as the effective evaluation duration. The duration of each state is divided by the effective evaluation duration to obtain the percentage of low attention, medium attention and high attention. The generation of attention level evaluation results stops when the effective evaluation duration is zero. The attention levels are assigned as one, two, and three respectively, according to low attention, medium attention, and high attention. Each level is multiplied by the corresponding percentage of duration and then summed to obtain an attention level score ranging from one to three. A score below 1.5 indicates a low level of attention; a score between 1.5 and 2.5 indicates a medium level of attention; and a score above 2.5 indicates a high level of attention. The corresponding scores are then used to determine the overall level of attention. The effective evaluation time is divided into a pre-evaluation interval and a post-evaluation interval along the task timeline. The pre-evaluation level score and the post-evaluation level score are calculated based on the continuous proportion of the three types of attentional states in the corresponding evaluation intervals. Attentional state segments that cross the interval boundaries are counted into the corresponding evaluation intervals according to the actual coverage time on both sides. Subtracting the score of the first attention level from the score of the second attention level, the difference is determined to be an upward trend if the difference is not less than 0.2, a downward trend if the difference is not greater than -0.2, and a stable trend if the difference is greater than -0.2 and less than 0.2. If any assessment interval lacks an effective attention state, the trend is marked as undeterminable. The percentages of low, medium, and high attention states are combined to form a state persistence summary result. Then, the state persistence summary result, attention level score, overall attention level, and attention change trend are bound together according to the output order specified by the attention level assessment constraint set to generate the attention level assessment result. By using the results of attention level assessment, the state persistence information in the attention level time trajectory is converted into the overall attention level and its changing trend, providing a summary output for the evaluation of attention state in brain-computer interaction tasks.

[0031] In this embodiment, the attention level assessment system based on a brain-computer interface includes the following modules: The task constraint configuration module is used to parse attention level assessment task information, configure probe presentation, brain response acquisition and calibration processes, and generate attention level assessment constraint set. The probe brain response acquisition module is used to control the brain-computer interface to present target probes and interference probes alternately, bind corresponding brain response acquisition segments, and generate dual probe brain response sequences. The brain response credibility processing module is used to correct the target probe brain response and the interference probe brain response according to the probe event location, screen the common effective acquisition channels and calculate the response credibility, and generate a set of credible dual brain response units. The positive definite covariance processing module is used to calculate the target brain response covariance matrix and the interference brain response covariance matrix, complete the positive definite processing and bind it with the response confidence, and generate the dual brain response covariance sequence. The individual attention anchor construction module is used to screen dual brain response covariance units according to the concentrated attention calibration stage and the attention transfer calibration stage, and perform manifold aggregation based on the response credibility to generate an individual attention anchor set; Attention differential evidence extraction module is used to call the improved SPDNet model to separate common responses and project attention differential evidence, bind target enhancement evidence, interference suppression evidence and response credibility to generate selective attention differential evidence sequence; Attention state temporal decoding module is used to call the improved HSMM algorithm to perform confidence gating, state persistence probability correction and ordered state decoding to generate attention level temporal trajectory; The attention assessment summary module is used to summarize the duration of each attention state, determine the trend of attention changes, and generate attention level assessment results.

[0032] Example 1: To verify the feasibility of this invention in practice, it was applied to the assessment of attention levels in a continuous visual monitoring task. The task interface continuously presented the target area that the subject needed to track, while simultaneously inserting interference content unrelated to the current task response. The brain-computer interface used 32 acquisition channels to continuously acquire the brain responses corresponding to the target probe and the interference probe at a sampling frequency of 256Hz.

[0033] In the implementation process, firstly, an attention level assessment constraint set is generated based on the continuous visual monitoring task. Paired brain responses corresponding to target probes and interference probes are collected and processed to generate a dual brain response covariance sequence. Then, an individual attention anchor set is constructed based on the focused attention calibration stage and the attention transfer calibration stage. The dual brain response covariance sequence and the individual attention anchor set are input into the improved SPDNet model to generate a selective attention differential evidence sequence. Finally, the improved HSMM algorithm is used to generate the attention level temporal trajectory, and the attention level assessment result is generated based on the state persistence and changing trend.

[0034] To verify the processing effect of this invention, three sets of comparison methods were set up: the traditional single-probe classification method, the ordinary SPDNet combined with HSMM method, and the method of this invention. The test subjects included 30 anonymous subjects. Each subject completed 3 sets of continuous visual monitoring tasks, resulting in 90 sets of task records. After evaluation interval division and invalid record exclusion, 216 valid evaluation windows were obtained. Each method used the same probe sequence, the same brain response acquisition records, and the same attention phase annotation. The statistical results are shown in Table 1.

[0035] Table 1. Comparison of test results from different attention level assessment methods

[0036] As shown in Table 1, traditional single-probe classification methods fail to distinguish the common changes in the response of the target probe and the response of the interference probe, resulting in a false positive rate of 12.7% for low-confidence samples and 17 instances of repeated state switching. While the ordinary SPDNet combined with the HSMM method can extract brain response covariance features, it lacks individual attention anchor constraints and dynamic correction of persistent probabilities, and the average error of the state boundary remains at 2.5 evidence units. The state recognition accuracy of the method of this invention reaches 90.6%, the false positive rate of low-confidence samples is reduced to 3.1%, the number of repeated state switching is reduced to 4, and the consistency rate of change trend reaches 94.4%, indicating that a stable processing chain is formed between brain response differential extraction and temporal state decoding.

[0037] Among the above indicators, the state recognition accuracy reflects the overall recognition of low, medium, and high attention states; the macro-average F1 score reflects the balance of the recognition results of the three types of attention states; the state boundary average error reflects the distance between the predicted state switching position and the labeled state switching position; the low confidence sample misjudgment rate reflects the impact of abnormal acquisition channels and sampling omissions on state determination; the number of repeated state switching times reflects the stability of the attention level temporal trajectory; and the trend consistency rate reflects the degree of matching between the rising, stable, and falling trends and the labeled results.

[0038] The method of this invention achieves the above results mainly because the improved SPDNet model incorporates the target brain response covariance matrix, the interference brain response covariance matrix, and the individual attention anchor set into the same manifold processing process, and removes the common variation components in the two types of brain responses through the anchor constraint common response separation mechanism; the improved HSMM algorithm adjusts the state observation probability according to the response credibility, and uses the continuous changes of target enhancement evidence and interference suppression evidence to correct the state persistence probability, while restricting the cross-level jump between low attention state and high attention state, so that the attention state determination takes into account the differences in brain response, data credibility, and temporal continuity.

[0039] As can be seen from the above implementation results, the present invention can complete probe pairing, brain response credibility verification, individual attention anchor construction, selective attention differential evidence extraction, attention state temporal decoding and change trend determination in continuous visual monitoring scenarios. It is applicable to brain-computer interaction tasks that require continuous attention to target content and identification of attention decline process, and provides an engineering implementation basis for attention level assessment.

[0040] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A brain-computer interface-based method for assessing attention levels, characterized in that, Includes the following steps: Step 1: Receive attention level assessment task information, parse the brain-computer interface task and assessment requirements, and generate an attention level assessment constraint set; Step 2: Based on the attention level assessment constraint set, the brain-computer interface presents the target probe and interference probe, synchronously collects and pairs brain responses, and generates a dual probe brain response sequence. Step 3: Correct the brain response sequence of the dual probe according to the probe event location, complete the response credibility determination, and generate a set of credible dual brain response units; Step 4: Calculate the covariance matrix of the target probe brain response and the interference probe brain response in the trusted dual brain response unit set, respectively, perform positive definite processing, and generate the dual brain response covariance sequence. Step 5: Based on the focused attention calibration and attention shift calibration, select calibration units from the dual brain response covariance sequence and perform manifold aggregation to generate individual attention anchor set; Step 6: Input the dual brain response covariance sequence and individual attention anchor set into the improved SPDNet model, and generate a selective attention differential evidence sequence through shared manifold mapping, anchor-constrained common response separation and attention differential projection. Step 7: Input the selective attention differential evidence sequence into the improved HSMM algorithm, and generate the attention level time trajectory by passing the confidence gating, evidence continuity correction state persistence probability and ordered state decoding. Step 8: Continuously summarize the execution status and determine the trend of changes in the attention level time-series trajectory, and generate attention level assessment results.

2. The brain-computer interface-based attention level assessment method according to claim 1, characterized in that, Step one specifically involves: Receive attention level assessment task information and read probe presentation and calibration instructions in the brain-computer interaction task according to the task execution order; The target probes are determined based on the task content that the subjects need to continuously focus on, and the probes that have the same stimulus conditions but do not undertake the current task response are determined as interference probes. The target probes and interference probes are arranged alternately along the task time axis so that the paired probes are in the same task stage, and the corresponding brain response acquisition time period is configured for each pair of paired probes. According to the calibration instructions, set up a focused attention calibration stage and an attention transfer calibration stage, so that both calibration stages include the complete presentation process of the target probe and the interference probe; Based on the assessment requirements, the probe presentation order, brain response acquisition time period, and coverage relationship of the calibration phase are verified, and the verified task configuration is written into the attention level assessment constraint set.

3. The brain-computer interface-based attention level assessment method according to claim 1, characterized in that, Step two specifically involves: Read the attention level assessment constraint set, connect the probe presentation process and the brain response acquisition process to the same task timeline, and start continuous brain response acquisition before the first probe is presented; The target probe and interference probe are presented alternately along the probe presentation sequence, and the corresponding probe event position is written into the continuous brain response each time a probe is triggered. Based on the pairing arrangement of the attention level assessment constraint set records, the target probe event locations and interference probe event locations are bound as dual event groups; Each pair of events is associated with a continuous brain response along the task timeline, so that each group of target probes and interference probes has a corresponding brain response acquisition segment; The brain response acquisition segments are arranged according to the time sequence of the dual event groups to generate the dual probe brain response sequence.

4. The brain-computer interface-based attention level assessment method according to claim 1, characterized in that, Step three specifically involves: Read the brain response sequence of the dual probes, extract the brain response of the target probe and the brain response of the interference probe according to the probe event location of each dual event group, and unify the event starting point and sampling location of the two types of brain responses; Based on the average brain response of each acquisition channel before the probe event location, the target probe brain response and the interference probe brain response are corrected respectively. Examine the amplitude saturation and transient mutations of each acquisition channel in the calibrated brain response, isolate abnormal acquisition channels, and retain the common effective acquisition channels in each pair of events. The confidence level of the response is calculated based on the retention of the common effective acquisition channel and the corresponding sampling locations of the two types of brain responses. Assembly is stopped for dual event groups that lack a common effective acquisition channel. The target probe brain response, interference probe brain response, and response confidence, formed based on a common effective acquisition channel, are bound into a trusted dual brain response unit. The trusted dual brain response unit set is generated by arranging the dual event groups in chronological order.

5. The brain-computer interface-based attention level assessment method according to claim 1, characterized in that, Step four specifically involves: Read the set of trusted dual brain response units, and organize the target probe brain response and the interference probe brain response into a brain response matrix according to the arrangement order of the common effective acquisition channels; The centered brain response matrix is ​​obtained by subtracting the mean value of the corresponding channel from the sampled value of each acquisition channel in each brain response matrix. Calculate the product of the centered brain response matrix and its transpose, and scale it by one less effective sample size to form the target brain response covariance matrix and the interference brain response covariance matrix, respectively. Construct unit reference matrices of the same dimension based on the traces of the two covariance matrices, and determine the common contraction intensity by combining the deviations of the two covariance matrices relative to the corresponding unit reference matrices. According to the common contraction intensity, the target brain response covariance matrix and the interference brain response covariance matrix are respectively weighted and fused with the corresponding unit reference matrix; Perform eigenvalue decomposition on the weighted fusion covariance matrix, correct non-positive eigenvalues, and reconstruct a symmetric positive definite matrix according to the corresponding eigenvectors; The target brain response covariance matrix, interference brain response covariance matrix, and response confidence corresponding to the same reliable dual brain response unit are bound as dual brain response covariance units and arranged in chronological order to generate a dual brain response covariance sequence.

6. The brain-computer interface-based attention level assessment method according to claim 1, characterized in that, Step five specifically involves: Read the dual brain response covariance sequence and aggregate the corresponding dual brain response covariance units according to the concentration attention calibration stage and the attention transfer calibration stage; Within each calibration phase, the response confidence of each dual brain response covariance unit is read, the aggregation weight of the corresponding calibration unit is determined based on the response confidence, and the aggregation weight within the current calibration phase is normalized. The target brain response covariance matrix and the interference brain response covariance matrix are extracted from each calibration unit. Matrix logarithmic mapping is performed on the extracted covariance matrices to generate the target response logarithmic matrix and the interference response logarithmic matrix. The target response logarithmic matrix and the disturbance response logarithmic matrix within the same calibration stage are accumulated according to the aggregation weights to form the target response aggregation matrix and the disturbance response aggregation matrix. Perform matrix exponential mapping on the target response aggregation matrix and the interference response aggregation matrix to restore the symmetric positive definite structure and generate the target brain response anchor point and the interference brain response anchor point for the current calibration stage; By binding the target brain response anchor and the interference brain response anchor in the same calibration phase as the dual brain response anchor, and combining the dual brain response anchors of the two calibration phases, an individual attention anchor set is generated.

7. The brain-computer interface-based attention level assessment method according to claim 1, characterized in that, The improved SPDNet model includes a shared manifold mapping layer, an anchor point tangent space mapping layer, a common response separation layer, and an attention evidence projection layer. The common response separation layer embeds an anchor point constraint common response separation mechanism. The shared manifold mapping layer is configured with a target response branch, an interference response branch, and an anchor point mapping branch. The three branches share bilinear mapping weights and eigenvalue correction rules. The three branches process the target brain response covariance matrix, the interfering brain response covariance matrix, and the individual attention anchor set, respectively, to generate dual response manifold features and anchor manifold features; The anchor point tangent space mapping layer generates an individual manifold reference matrix based on the anchor point manifold features. The dual response manifold features and anchor point manifold features are uniformly transformed using the individual manifold reference matrix. The dual response tangent space features and anchor point tangent space features in the same tangent space are generated through matrix logarithmic mapping. The common response separation layer inputs the dual response tangent space features and anchor point tangent space features into the anchor point constraint common response separation mechanism to establish the target response benchmark and interference response benchmark with the anchor point tangent space features in the attention transfer calibration stage, and determines the tangent space offset of the current target response and interference response relative to the corresponding benchmark. Extract the overlapping portions with consistent directions and matching change ratios from the two types of tangent space offsets as common response components. Separate the common response components from the two types of tangent space offsets to generate target response difference components and disturbance response difference components. The target response difference component and the interference response difference component are directionally matched with the anchor point tangent space features corresponding to the focused attention calibration stage to generate target enhancement evidence and interference suppression evidence. The attention evidence projection layer adjusts the projection contributions of target enhancement evidence and interference suppression evidence based on response credibility; The target enhancement evidence, interference suppression evidence, and response credibility are bound into selective attention differential evidence units, which are then arranged in chronological order to generate a selective attention differential evidence sequence.

8. The brain-computer interface-based attention level assessment method according to claim 1, characterized in that, Step seven specifically involves: Read the selective attention differential evidence sequence, set ordered attention states according to attention level from low to high, and call the attention observation model corresponding to each attention state; Target enhancement evidence and interference suppression evidence are read from the selective attention differential evidence unit. The degree of matching between the two types of evidence and the evidence distribution of each attention state is calculated to generate the state observation probability. Read the response confidence carried by the selective attention differential evidence unit, adjust the differentiation amplitude of the state observation probability according to the response confidence, so that the low confidence brain response forms a neutral observation that tends to be evenly distributed among various attention states, and keep the probability of low attention state from increasing due to the decrease in response confidence. By combining consecutive selective attention differential evidence units in chronological order, candidate state segments with different durations are established according to each attention state. Compare the changing directions of target enhancement evidence and interference suppression evidence in adjacent selective attention differential evidence units within the candidate state segment. Based on the degree of continuity of the two types of evidence changing in the same direction and the corresponding response credibility, adjust the basic persistence probability and generate the adjusted state persistence probability corresponding to each candidate persistence length. Divide the probability of the modified state duration corresponding to the duration of each candidate under the same attention state by the sum of the probabilities of the modified state duration under the same attention state, so that the sum of the normalized probabilities of the modified state duration is one. Establish the transition relationship between adjacent attentional states according to the order of attentional levels, block direct transitions across intermediate attentional states, and generate ordered state transition relationships; The state scores of candidate state segments are recursively derived by combining the state observation probability, the corrected state persistence probability, and the ordered state transition relationship, and the optimal state path corresponding to each time position is retained. By backtracking the optimal state path, the temporal location and duration of each attentional state are determined, and a temporal trajectory of the attentional level is generated.

9. The brain-computer interface-based attention level assessment method according to claim 1, characterized in that, Step eight specifically involves: Read the attention level temporal trajectory, merge the positions of trajectories that maintain the same attention state continuously along the time sequence into attention state segments, and record the attention state and duration corresponding to each attention state segment; Each attentional state segment is aggregated according to the attentional state, the duration corresponding to the same attentional state is accumulated, and the proportion of the corresponding duration to the total duration of the attentional level time sequence trajectory is calculated to generate a summary result of the state duration. Based on the state levels corresponding to low attention, medium attention, and high attention, the attention level score is generated by multiplying each state level by the percentage of its corresponding duration and summing the results. The overall attention level is determined by comparing the attention level score with the attention level classification threshold. The task timeline is divided into a pre-assessment interval and a post-assessment interval. Attention level scores are calculated based on the duration of each attention state within the corresponding assessment interval. Calculate the difference between the attention level score of the later stage and the attention level score of the earlier stage, and generate the attention change trend based on the comparison between the difference and the attention change trend threshold; The results of continuous state summarization, attention level score, overall attention level, and attention change trend are combined to generate an attention level assessment result.

10. A brain-computer interface-based attention level assessment system, performing the brain-computer interface-based attention level assessment method according to any one of claims 1 to 9, characterized in that, Includes the following modules: The task constraint configuration module is used to parse attention level assessment task information, configure probe presentation, brain response acquisition and calibration processes, and generate attention level assessment constraint set. The probe brain response acquisition module is used to control the brain-computer interface to present target probes and interference probes alternately, bind corresponding brain response acquisition segments, and generate dual probe brain response sequences. The brain response credibility processing module is used to correct the target probe brain response and the interference probe brain response according to the probe event location, screen the common effective acquisition channels and calculate the response credibility, and generate a set of credible dual brain response units. The positive definite covariance processing module is used to calculate the target brain response covariance matrix and the interference brain response covariance matrix, complete the positive definite processing and bind it with the response confidence, and generate the dual brain response covariance sequence. The individual attention anchor construction module is used to screen dual brain response covariance units according to the concentrated attention calibration stage and the attention transfer calibration stage, and perform manifold aggregation based on the response credibility to generate an individual attention anchor set; Attention differential evidence extraction module is used to call the improved SPDNet model to separate common responses and project attention differential evidence, bind target enhancement evidence, interference suppression evidence and response credibility to generate selective attention differential evidence sequence; Attention state temporal decoding module is used to call the improved HSMM algorithm to perform confidence gating, state persistence probability correction and ordered state decoding to generate attention level temporal trajectory; The attention assessment summary module is used to summarize the duration of each attention state, determine the trend of attention changes, and generate attention level assessment results.