An individual cognitive baseline adaptive training method based on double time scale

Abstract

The application discloses an individual cognitive baseline adaptive training method based on a double time scale: data in the most optimal cognitive state is screened from multi-modal heterogeneous data of a target individual, and an optimal feature space is constructed as a cognitive baseline according to the data; the optimal feature space is taken as a Bayesian prior constraint, longitudinal time series data is collected, and a posterior probability is calculated online; the posterior probability is compared with a preset posterior probability threshold value; when the data characteristics of the longitudinal time series data are long-term stable characteristics, the cross-modal consistency of the longitudinal time series data is verified; if the longitudinal time series data is verified, the optimal feature space is updated, and adaptive calibration of the cognitive baseline is realized. By constructing an individual optimal feature space and combining a Bayesian online learning model, short-term fluctuation characteristics and long-term stable characteristics of the cognitive state of the target individual are effectively distinguished, individualized adaptation of the cognitive baseline is realized, and the problem that the cognitive baseline is fixed and not changed and individual adaptability is poor in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of individual cognitive assessment technology, specifically to an adaptive training method for individual cognitive baselines based on dual time scales. Background Technology

[0002] In various technology fields that are oriented towards personalization, the core premise for accurate judgment and scientific analysis of an individual's cognitive state is to establish a dynamic reference benchmark that fits the individual—that is, to clarify what the individual's cognitive performance is like under normal conditions. The establishment of this benchmark is inseparable from the continuous collection, tracking and analysis of the individual's long-term data.

[0003] It is important to clarify that a person's cognitive state is not static but rather dynamic: in the short term, it is affected by factors such as task load, emotional fluctuations, and physiological state, resulting in temporary fluctuations; in the long term, it undergoes slow and stable changes with age, changes in lifestyle, and the accumulation of experience. This dynamic characteristic of short-term fluctuations and long-term evolution means that there is no fixed standard for judging an individual's cognitive state, and it is impossible to use a uniform group benchmark to cover all individuals.

[0004] Existing technologies for monitoring and calibrating individual cognitive states mostly suffer from two major drawbacks: First, it ignores the key characteristic of dynamic changes in cognitive state and fails to establish a dynamic reference model that can track the long-term evolution of individual cognition and capture short-term fluctuations. It only uses fixed static benchmarks or group benchmarks, which cannot adapt to the dynamic changes in individual cognition. Second, there is a lack of systematic collection and analysis of long-term individual data, and no personalized reference benchmarks have been built to fit individuals. This makes it impossible to effectively distinguish between short-term fluctuations and long-term changes in individual cognition, and it is also difficult to adapt to the significant differences between different individuals, resulting in inaccurate calibration results and unreliable analysis conclusions.

[0005] Furthermore, existing methods do not fully consider the dynamic evolution of individual cognition. Even if some schemes attempt to collect individual data, they do not form a complete logic of long-term tracking, dynamic updating, and accurate differentiation. They cannot adjust the reference benchmark according to the long-term changes in individual cognition, resulting in the reference benchmark becoming disconnected from the individual's actual cognitive state after long-term use, further exacerbating the problem of analytical distortion.

[0006] Therefore, there is an urgent need for a technical solution that can track and collect individual data over a long period of time, adapt to dynamic changes in cognition, effectively distinguish between short-term fluctuations and long-term characteristics, and construct personalized dynamic reference benchmarks. This solution would address the analytical distortion caused by dynamic changes in individual cognition, large individual differences, and the lack of dynamic personal benchmarks, and provide reliable support for the accurate monitoring and scientific evaluation of individual cognitive states. Summary of the Invention

[0007] The purpose of this invention is to provide an adaptive training method for individual cognitive baselines based on dual time scales to solve the problems raised in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: an adaptive training method for individual cognitive baselines based on dual time scales, comprising the following steps: Step S1: Collect multimodal heterogeneous data of the target individual; Step S2: Select the optimal dataset from the multimodal heterogeneous data in which the target individual is in the optimal cognitive state, construct a feature distribution model representing the target individual in the optimal cognitive state based on the optimal dataset, obtain the optimal feature space, and use it as the initial cognitive baseline of the target individual; Step S3: Construct a dual-timescale identification model. The dual-timescale identification model uses the optimal feature space as a Bayesian prior constraint, collects longitudinal time-series data of the target individual and calculates the posterior probability online. By comparing the posterior probability with a preset posterior probability threshold, the data features of the longitudinal time-series data are identified. The data features include short-term fluctuation features and long-term stable features. Step S4: When the data features of the longitudinal time series data are identified as long-term stable features, the cross-modal consistency of the longitudinal time series data is verified, and the longitudinal time series data with cross-modal consistency is used as the new optimal dataset to reconstruct the optimal feature space, thereby realizing the adaptive calibration of the cognitive baseline and the dual time scale identification model.

[0009] Preferably, the multimodal heterogeneous data includes at least basic attribute data, physiological monitoring data, behavioral performance data, and task performance data, wherein: The basic attribute data includes age, gender, place of origin, and education level; The physiological monitoring data includes electroencephalogram (EEG), electrocardiogram (ECG), skin electrical activity, eye tracking, and surface electromyography (EMG). The behavioral data includes the number of clicks, switching frequency, and dwell time. The task performance data includes task completion time and accuracy rate.

[0010] Preferably, step S2 includes the following steps: Step S21: Determine that the target individual is in the optimal cognitive state during a steady-state period when there is no fatigue interference, the task load is adapted, and the cognitive efficiency is optimal. Select data samples from the multimodal heterogeneous data within the steady-state period and perform preprocessing to form the optimal dataset. Step S22: Use a Gaussian mixture model to perform density fitting on the optimal dataset. By superimposing multiple Gaussian components, the distribution law of each modality feature in the optimal cognitive state is characterized, and the parameters of each Gaussian component are determined. Step S23: Use the expectation-maximization algorithm to iteratively optimize and solve the parameters of the Gaussian mixture model, determine the core parameters of the Gaussian mixture model, and lock the global distribution law of each modality feature under the optimal cognitive state; Step S24: Based on the probability distribution of the optimized Gaussian mixture model, a steady-state probability threshold is preset and high-confidence feature regions are selected. The high-confidence feature regions are determined as the optimal feature space of the target individual. The optimal feature space is used to characterize the feature centers of each modality and the correlation between features in the optimal cognitive state.

[0011] Preferably, the density fitting of the optimal dataset using a Gaussian mixture model in step S22 is achieved through the following expression:

[0012] In the formula: The number of Gaussian components; For the first The mixing weights of the Gaussian components satisfy the following conditions: ; For the first The mean vector of Gaussian components represents the central value of the core features of each modality under the optimal cognitive state; For the first The covariance matrix of Gaussian components represents the degree of dispersion and correlation of each modality feature under the optimal cognitive state; The probability density function of the multidimensional Gaussian distribution is expressed as follows:

[0013] By superimposing multiple Gaussian components, the distribution patterns of modal features in the optimal cognitive state are characterized, and the weighting coefficients of each Gaussian component are determined. Core Feature Center and the degree of feature dispersion .

[0014] Preferably, the dual-timescale identification model includes a short-term scale branch and a long-term scale branch, wherein the short-term scale branch is used to capture the short-term fluctuation features and the long-term scale branch is used to capture the long-term stable features. Step S3 includes the following steps: Step S31: Continuously collect new multimodal heterogeneous data of the target individual under different times and different task loads, and input them into the Bayesian online learning model in real time as longitudinal time series data; Step S32: Use the optimal feature space of the target individual as the prior distribution of the parameters of the Bayesian online learning model, and update the prior distribution using the longitudinal time series data according to the Bayesian formula to obtain the posterior distribution; Step S33: Based on the posterior distribution, calculate the posterior probability of the longitudinal time series data relative to the optimal feature space. If the posterior probability is greater than a preset posterior probability threshold, then determine that the data feature of the longitudinal time series data is a short-term fluctuation feature; otherwise, determine that the data feature of the longitudinal time series data is a long-term stable feature.

[0015] Preferably, step S4 employs a cognitive baseline gating update model, which uses the correlation between modal features in the optimal feature space as a reference to verify the cross-modal consistency of the longitudinal time series data, and uses cross-modal consistency as a gating condition for updating the optimal feature space. The verification process is as follows: Obtain the core features corresponding to each modality in the current longitudinal time series data, and verify whether the correlation between the core features corresponding to each modality is consistent with the correlation between the modality features in the optimal feature space; If the changing trends and numerical correlations of the modal features conform to the correlation relationship of the optimal feature space, and the deviation is within a preset reasonable range, then it is determined to be cross-modal consistency, triggering the optimal feature space update; If the changing trend of a certain modal feature among the modal features contradicts that of other modal features, the numerical correlation exceeds a preset reasonable range, or the correlation with the optimal feature space is not matched, then it is determined to be cross-modal inconsistency.

[0016] This invention also claims protection for an electronic device, comprising: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the steps of the above method.

[0017] This invention also applies for protection of a computer-readable storage medium storing a computer program / instructions thereon, which, when executed by a processor, implements the steps of the above-described method.

[0018] This invention also applies for protection of a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the above-described method.

[0019] Compared with the prior art, the beneficial effects of this invention are as follows: 1. This invention collects multimodal heterogeneous data of target individuals to construct a personalized optimal feature space, and uses a Gaussian mixture model to fit the probability distribution, thereby achieving personalized adaptation of the cognitive baseline and solving the problems of fixed cognitive baseline and poor individual adaptability in the prior art. 2. This invention employs a dual-timescale identification model, combined with a Bayesian online learning model, to effectively distinguish between short-term fluctuations and long-term stable characteristics of the cognitive state of a target individual, avoiding interference from short-term temporary changes on the cognitive baseline calibration, thereby improving calibration accuracy. 3. This invention introduces a cross-modal consistency gating mechanism, which ensures the effectiveness of cognitive baseline updates through consistency verification of multimodal heterogeneous data, and at the same time realizes dynamic adjustment of dual time-scale identification models to adapt to long-term changes in the cognitive state of the target individual. Attached Figure Description

[0020] Figure 1 This is an overall flowchart of the adaptive training method of the present invention; Figure 2 This is a flowchart illustrating the adaptive training method of the present invention. Figure 3 This is an overall flowchart of step S2 of the adaptive training method of the present invention; Figure 4 This is the overall flowchart of step S3 of the adaptive training method of the present invention. Detailed Implementation

[0021] The first embodiment of this invention discloses an adaptive training method for individual cognitive baselines based on dual time scales; the overall process is described in [reference needed]. Figure 1 For detailed procedures, please refer to [link / reference]. Figure 2 The method includes the following steps: Step S1: Collect multimodal heterogeneous data of the target individual. This multimodal heterogeneous data includes at least basic attribute data, physiological monitoring data, behavioral performance data, and task performance data, wherein: Basic attribute data includes age, gender, place of origin, education level, etc. Physiological monitoring data include electroencephalogram (EEG), electrocardiogram (ECG), skin electrical activity, eye tracking, and surface electromyography (EMG). Behavioral performance data includes click count, switching frequency, and dwell time; Task performance data includes task completion time, accuracy rate, etc.

[0022] Step S2: Select the optimal dataset from the multimodal heterogeneous data in which the target individual is in the optimal cognitive state, construct a feature distribution model representing the target individual in the optimal cognitive state based on the optimal dataset, obtain the optimal feature space, and use it as the initial cognitive baseline of the target individual.

[0023] See Figure 3 Step S2 includes the following steps: Step S21: Determine the steady-state period when the target individual is in its optimal cognitive state, characterized by no fatigue interference, appropriate task load, and optimal cognitive efficiency. Select data samples from the multimodal heterogeneous data within this steady-state period and preprocess them to form the optimal dataset. .

[0024] The steady-state period can be set to 9:00-11:00 AM every day. Studies have shown that from 9:00-11:00 AM every day, the body's cortisol level is high, nerve excitability is sufficient, nighttime sleep fatigue completely subsides, there is no daytime drowsiness or mental exhaustion interference, and the refresh efficiency of short-term memory and working memory reaches its peak.

[0025] The preprocessing steps for the data samples include: using a sliding window combined with... The anomaly detection method removes outliers, sudden noise, and invalid missing data. Then, the Min-Max normalization method is used to standardize all valid data, unify the dimensions and numerical distribution range of each modality, eliminate the magnitude differences between different types of data, and obtain the normalized time series optimal dataset.

[0026] Optimal dataset middle, d is the fusion feature vector of a single sample, where i takes values ​​from 1 to n, and the dimension d is determined by the number of fusion features of basic attribute data, physiological test data, behavioral performance data, and task performance data.

[0027] Step S22: Apply a Gaussian mixture model to the density of the optimal dataset, the expression of which is:

[0028] In the formula: The number of Gaussian components; For the first The mixing weights of the Gaussian components satisfy the following conditions: ; For the first The mean vector of Gaussian components represents the central value of the core features of each modality under the optimal cognitive state; For the first The covariance matrix of Gaussian components represents the degree of dispersion and correlation of modal features under optimal cognitive state.

[0029] The probability density function of the multidimensional Gaussian distribution is expressed as follows:

[0030] By superimposing multiple Gaussian components, the distribution patterns of modal features under optimal cognitive state are characterized, and the weighting coefficients of each Gaussian component are determined. Core Feature Center and the degree of feature dispersion .

[0031] Step S23: Use the expectation-maximization algorithm to optimize the parameters of the Gaussian mixture model. Perform iterative optimization to determine the final core parameters of the Gaussian mixture model and lock the global distribution law of each modality feature under the optimal cognitive state. Here, the global distribution law is different from the distribution law in step S22, and specifically refers to the distribution law under the Gaussian mixture model after the parameters have been iteratively optimized.

[0032] Step S24: Based on the optimized Gaussian mixture model probability distribution, preset the steady-state probability threshold. Filter out those that meet the requirements The high-confidence feature region is defined as the optimal feature space for the target individual. ,Right now .

[0033] The probability distribution of this optimal feature space is characterized by the fitting result of a Gaussian mixture model, namely:

[0034] Among them, the mean vector The central values ​​that characterize the modal features of each cognitive state, such as the optimal typical values ​​of EEG power, heart rate, and reaction time; covariance matrix Characterizing the correlation between modal features under optimal cognitive state, such as the sign and magnitude of covariance reflecting whether the features change in the same or opposite direction, and the strength of the correlation.

[0035] For example, under optimal cognitive conditions: EEG focus index and heart rate may be weakly positively correlated (positive covariance and small value); reaction time and accuracy may be negatively correlated (the faster the reaction, the slightly lower the accuracy, and the negative covariance).

[0036] Steady-state probability threshold The method for determining it is as follows: 1. Statistically analyze the distribution of membership probabilities of a large number of data samples of target individuals in their optimal cognitive state (such as the aforementioned optimal dataset) within the Gaussian mixture model. Combine this with the sample confidence interval in statistics to preliminarily determine an initial steady-state probability threshold. This initial threshold can ensure that the membership probability of the vast majority of data samples in the optimal cognitive state is greater than this threshold. 2. The threshold discrimination effect is verified by back-substitution test of labeled samples (samples clearly marked as "steady state" or "non-steady state") and comparison with the real state of time series. The initial optimization is completed based on the matching degree between the recognition accuracy (the proportion of sample membership probability ≥ the initial threshold) and the actual cognitive performance (such as task accuracy, reaction time, subjective fatigue score, etc.). 3. Combine long-term time-series monitoring data of the target individual (covering several months or years of daily multimodal heterogeneous data) to conduct long-term verification of the optimized threshold. Follow the slow evolution trend of individual cognitive ability (such as cognitive baseline shift due to age or long-term skill training) and make small adjustments to the threshold until the threshold can stably distinguish between the steady-state range of individual cognition and the short-term fluctuation deviation range, and the discrimination result is highly consistent with the actual state and the recognition effect tends to converge and stabilize. 4. Determine the final converged steady-state probability threshold from the above steps as a reasonable value for the target individual. This threshold is used to screen high-confidence feature regions from the Gaussian mixture model probability distribution, thus completing the optimal feature space. The quantitative modeling provides a unified and reliable prior constraint for the subsequent dual-timescale identification model.

[0037] Step S3: Construct a dual-timescale identification model. The dual-timescale identification model uses the optimal feature space as a Bayesian prior constraint, collects longitudinal time-series data of the target individual and calculates the posterior probability online. By comparing the posterior probability with a preset posterior probability threshold, the data features of the longitudinal time-series data are identified.

[0038] The data features include short-term fluctuation features and long-term stable features. The dual-timescale identification model includes a short-term scale branch and a long-term scale branch. The short-term scale branch is used to capture the short-term fluctuation features of longitudinal time series data, and the long-term scale branch is used to capture the long-term stable features of longitudinal time series data.

[0039] See Figure 4 Step S3 includes the following steps: Step S31: Continuously collect longitudinal time-series data of the target individuals as input to the Bayesian online learning model. The longitudinal time-series data consists of new multimodal heterogeneous data under different times and different task loads.

[0040] Step S32: Using the optimal feature space Feature distribution as parameters of Bayesian online learning model The prior distribution, i.e. Given longitudinal time series data Calculate the posterior distribution using Bayes' theorem. posterior distribution (Model parameters) The distribution after Bayesian update satisfies Bayes' theorem:

[0041] in, It is the likelihood function, reflecting the longitudinal time series data. With model parameters The degree of matching.

[0042] Step S33: Based on the posterior distribution Calculate the matching degree (i.e., posterior probability) between the current longitudinal time series data and the optimal feature space, and then pass the posterior probability threshold. Distinguish between normal, predictable baseline drift and risky, anomalous deviations.

[0043] like If the data fluctuations are consistent with short-term temporary changes, then the data characteristics of the current longitudinal time series data are determined to be short-term fluctuation characteristics, classified as temporary interference, and the existing cognitive baseline is not changed. like If the data changes are stable and continuous, then the data characteristics of the current longitudinal time series data are determined to be long-term stable characteristics.

[0044] The posterior probability threshold The method for determining it is as follows: 1. Input a large number of confirmed steady-state data samples of the target individual (such as the optimal dataset used to construct the optimal feature space) into the Bayesian online learning model in batches, count the posterior probability distribution of all samples, and take the lower limit critical value of the probability distribution as the initial value of the threshold to ensure that the vast majority of normal steady-state samples fall within the threshold judgment range. 2. Retrieve long-term time-series monitoring data of the target individual (covering daily multimodal heterogeneous data over several months or years, such as longitudinal time-series data used in subsequent steps), distinguish between daily mild physiological baseline drift data and data with obvious abnormal deviations in cognitive state, and statistically determine the natural boundary interval of the posterior probability of these two types of data to lock in a reasonable range of threshold values. 3. Combine actual cognitive assessment results, task completion performance, and subjective state feedback for joint calibration, and repeatedly adjust the threshold size; avoid the threshold being too high and judging normal small drift as abnormal deviation, and also prevent the threshold being too low and failing to identify real cognitive risk deviation. Finally, through multi-time series verification, determine the optimal threshold Q with stable discrimination effect.

[0045] This invention uses Bayesian posterior probability quantification to determine the degree of matching between real-time multimodal heterogeneous data and the optimal feature space of the target individual. It relies on individual steady-state sample probability statistics, long-term data interval division, and actual evaluation effect calibration to reasonably determine the posterior probability threshold Q.

[0046] When the calculated posterior matching probability is greater than the threshold Q, it means that the data only fluctuates temporarily within the optimal feature interval, which is a predictable normal short-term cognitive fluctuation and slight baseline drift. When the posterior matching probability is consistently less than or equal to the threshold Q, it indicates that the data has deviated from the original optimal steady-state distribution for a long period of time. This signifies that the individual's cognition has undergone substantial and trend-like long-term trait changes and deviations from risky states. This judgment rule enables the accurate distinction between short-term instantaneous fluctuations and long-term inherent characteristic evolution.

[0047] Step S4: Construct a cognitive baseline gating update model. When the data features of the longitudinal time series data are identified as long-term stable features, the cognitive baseline gating update model uses the correlation relationship of each modal feature in the optimal feature space as a reference to verify the cross-modal consistency of the longitudinal time series data, and uses the longitudinal time series data with cross-modal consistency as the new optimal dataset to reconstruct the optimal feature space, thereby realizing the adaptive calibration of the cognitive baseline and the dual time scale identification model.

[0048] The cross-modal consistency of the cognitive baseline gating update model is used as the gating condition for updating the optimal feature space. The "gating condition" means that the admission logic for updating the optimal feature space is only allowed when the dual-timescale identification model determines that the data features of the current longitudinal time series data are long-term stable features and the cross-modal consistency verification of the longitudinal time series data passes.

[0049] The specific verification process is as follows: Obtain the core features corresponding to each modality in the current longitudinal time series data, and verify whether the correlation between the core features of each modality is consistent with the correlation between the features of each modality in the optimal feature space; If the changing trends and numerical correlations of each modal feature conform to the correlation relationship of the optimal feature space, and the deviation is within a preset reasonable range, then it is determined to be cross-modal consistency, triggering the update of the optimal feature space; If the changing trend of a certain modal feature contradicts that of other modal features, the numerical correlation exceeds the preset reasonable range, or the correlation with the optimal feature space does not match, it is judged as cross-modal inconsistency.

[0050] The preset reasonable range of the deviation refers to: First, the basic fluctuation range is obtained by statistically analyzing the multimodal feature linkage deviation distribution under the optimal cognitive state of the target individual. Then, a safety margin is added by combining the natural fluctuation range of the individual's long-term daily steady-state data, and the system is graded and adapted according to different task load scenarios. Subsequently, positive and negative samples were repeatedly back-substituted for verification to continuously optimize the interval boundary. Finally, a stable and reasonable deviation range that can accommodate normal physiological fluctuations and slight shifts in cognitive behavior, while effectively identifying modal correlation anomalies, was determined as the quantitative judgment boundary for whether cross-modal feature matching is possible.

[0051] The following uses quantitative data examples to illustrate the verification process, selecting four core features of three typical modalities: 1. Physiological modality: EEG characteristic value of attention E 2. Physiological modality: Resting heart rate H 3. Behavioral Modality: Task Response Speed ​​Value V 4. Performance Modality: Task Accuracy R The optimal correlation between modalities in the optimal feature space (optimal correlation): Higher focus leads to a slight and stable decrease in heart rate, faster response speed, and higher accuracy in answering questions. Optimal baseline values ​​for each mode: E_0=75, H_0=72 times / min, V_0=0.85, R_0=0.90; The preset reasonable deviation range is ±5%, and after adding the safety margin, the overall allowable fluctuation range is ±8%.

[0052] 1. Real-time data acquisition: E=71, H=69, V=0.81, R=0.87; The percentage fluctuations of each feature relative to the optimal baseline value are as follows: 5.33%, 4.17%, 4.71%, and 3.33%; All fluctuations are within a reasonable range of ±8%; Synchronized trends: Concentration decreased slightly, heart rate decreased slightly, response speed decreased slightly, and accuracy decreased slightly in sync, which is completely consistent with the optimal correlation.

[0053] Decision: Cross-modal consistency.

[0054] 2. Real-time data collection: E=62 (significantly decreased focus), H=70, V=0.92 (response speed actually increased), R=0.82; Trend logic conflict: Decreased cognitive focus should normally lead to slower response times, but the speed of this data set actually increased, which violates the optimal correlation relationship; Decision: Cross-modal inconsistency.

[0055] 3. Real-time data acquisition: E=76, H=85; The heart rate H fluctuated by 18.06% relative to the optimal baseline value, far exceeding the preset reasonable deviation range of ±8%; other characteristics fluctuated normally, but the deviation of a single mode exceeded the standard, and the modal correlation was unbalanced. Decision: Cross-modal inconsistency.

[0056] A second embodiment of the present invention also provides an electronic device, including at least one processor and a memory communicatively connected to the at least one processor. The memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the steps of the method described above.

[0057] The third embodiment of the present invention also provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0058] The fourth embodiment of the present invention also provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the method described above.

[0059] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An adaptive training method based on an individual cognitive baseline with dual time scales, characterized in that, Includes the following steps: Step S1: Collect multimodal heterogeneous data of the target individual; Step S2: Select the optimal dataset from the multimodal heterogeneous data in which the target individual is in the optimal cognitive state, construct a feature distribution model representing the target individual's optimal cognitive state based on the optimal dataset, obtain the optimal feature space, and use it as the initial cognitive baseline of the target individual; Step S2 includes the following steps: Step S21: Determine that the target individual is in the optimal cognitive state during a steady-state period when there is no fatigue interference, the task load is adapted, and the cognitive efficiency is optimal. Select data samples from the multimodal heterogeneous data within the steady-state period and perform preprocessing to form the optimal dataset. Step S22: Use a Gaussian mixture model to perform density fitting on the optimal dataset. By superimposing multiple Gaussian components, the distribution law of each modality feature in the optimal cognitive state is characterized, and the parameters of each Gaussian component are determined. Step S23: Use the expectation-maximization algorithm to iteratively optimize and solve the parameters of the Gaussian mixture model, determine the core parameters of the Gaussian mixture model, and lock the global distribution law of each modality feature under the optimal cognitive state; Step S24: Based on the probability distribution of the optimized Gaussian mixture model, a steady-state probability threshold is preset and high-confidence feature regions are selected. The high-confidence feature regions are determined as the optimal feature space of the target individual. The optimal feature space is used to characterize the feature centers of each modality and the correlation between features in the optimal cognitive state. Step S3: Construct a dual-timescale identification model. The dual-timescale identification model uses the optimal feature space as a Bayesian prior constraint, collects longitudinal time-series data of the target individual and calculates the posterior probability online. By comparing the posterior probability with a preset posterior probability threshold, the data features of the longitudinal time-series data are identified. The data features include short-term fluctuation features and long-term stable features. The dual-timescale identification model includes a short-term scale branch and a long-term scale branch. The short-term scale branch is used to capture the short-term fluctuation characteristics, and the long-term scale branch is used to capture the long-term stable characteristics. Step S3 includes the following steps: Step S31: Continuously collect new multimodal heterogeneous data of the target individual under different times and different task loads, and input them into the Bayesian online learning model in real time as longitudinal time series data; Step S32: Use the optimal feature space of the target individual as the prior distribution of the parameters of the Bayesian online learning model, and update the prior distribution using the longitudinal time series data according to the Bayesian formula to obtain the posterior distribution; Step S33: Based on the posterior distribution, calculate the posterior probability of the longitudinal time series data relative to the optimal feature space. If the posterior probability is greater than a preset posterior probability threshold, then determine that the data feature of the longitudinal time series data is a short-term fluctuation feature; otherwise, determine that the data feature of the longitudinal time series data is a long-term stable feature. Step S4: When the data features of the longitudinal time series data are identified as long-term stable features, the cross-modal consistency of the longitudinal time series data is verified, and the longitudinal time series data with cross-modal consistency is used as the new optimal dataset to reconstruct the optimal feature space, thereby realizing the adaptive calibration of the cognitive baseline and the dual-time-scale identification model; Step S4 adopts a cognitive baseline gating update model, which uses the correlation relationship of each modal feature in the optimal feature space as a reference to verify the cross-modal consistency of the longitudinal time series data, and uses cross-modal consistency as the gating condition for updating the optimal feature space; The verification process is as follows: Obtain the core features corresponding to each modality in the current longitudinal time series data, and verify whether the correlation between the core features corresponding to each modality is consistent with the correlation between the modality features in the optimal feature space; If the changing trends and numerical correlations of the modal features conform to the correlation relationship of the optimal feature space, and the deviation is within a preset reasonable range, then it is determined to be cross-modal consistency, triggering the optimal feature space update; If the changing trend of a certain modal feature among the modal features contradicts that of other modal features, the numerical correlation exceeds a preset reasonable range, or the correlation with the optimal feature space is not matched, then it is determined to be cross-modal inconsistency.

2. The adaptive training method according to claim 1, characterized in that, The multimodal heterogeneous data includes at least basic attribute data, physiological monitoring data, behavioral performance data, and task performance data, among which: The basic attribute data includes age, gender, place of origin, and education level; The physiological monitoring data includes electroencephalogram (EEG), electrocardiogram (ECG), skin electrical activity, eye tracking, and surface electromyography (EMG). The behavioral data includes the number of clicks, switching frequency, and dwell time. The task performance data includes task completion time and accuracy rate.

3. The adaptive training method according to claim 1, characterized in that, The density fitting of the optimal dataset using a Gaussian mixture model in step S22 is achieved through the following expression: In the formula: The number of Gaussian components; For the first The mixing weights of the Gaussian components satisfy the following conditions: ; For the first The mean vector of each Gaussian component represents the central value of the core feature of each modality under the optimal cognitive state; For the first The covariance matrix of Gaussian components represents the degree of dispersion and correlation of each modality feature under the optimal cognitive state; The probability density function of the multidimensional Gaussian distribution is expressed as follows: By superimposing multiple Gaussian components, the distribution patterns of modal features in the optimal cognitive state are characterized, and the weighting coefficients of each Gaussian component are determined. Core Feature Center and the degree of feature dispersion .

4. An electronic device, characterized in that, include: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the steps of the method according to any one of claims 1-3.

5. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1-3.

6. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1-3.

Images