Brain-computer interaction intention recognition method, device, equipment and medium based on AI glasses

Abstract

The present application discloses a method, device, equipment and medium for brain-computer interaction intention recognition based on AI glasses. The method includes: obtaining intention recognition data segments of a target user for constructing a discrete set of duration and a discrete set of frequency, obtaining a key time-frequency feature point position sequence and an overall feature point position sequence; obtaining the number of occurrences and confidence scores of the key time-frequency feature point position sequence in the overall feature point position sequence, and obtaining a local anomaly evaluation score corresponding to the key time-frequency feature point position sequence according to the number of occurrences and the confidence score; obtaining multiple data blocks, and constructing a frequency conversion set according to the multiple data blocks; obtaining data fluctuation information and a graph complexity score corresponding to the data block; determining a decoding threshold adjustment factor according to the intention complexity score and the anomaly evaluation score, standardizing the intention recognition data segments, and obtaining normalized decoded data to obtain the user intention instruction corresponding to the target user, so as to realize the interactive control function of the AI ​​glasses.

Description

Technical Field

[0001] The present application relates to the field of brain-computer interface technology, and in particular to a method, device, equipment and medium for identifying brain-computer interaction intention based on AI glasses. Background Art

[0002] Brain-computer interface technology is a technical means to connect the neural activities in an organism with external devices or systems. In recent years, with the development of artificial intelligence and wearable device technology, the combination of brain-computer interface and AI glasses has gradually become a new research hotspot. As a new type of intelligent wearable device, AI glasses provide users with a natural and intuitive human-computer interaction method by integrating EEG acquisition electrodes, processors and display units. Among them, accurately identifying the user's interaction intention is the key to realizing the intelligent control of AI glasses.

[0003] EEG signals are weak potential changes generated by the synchronized ion currents of neuronal groups below the cerebral cortex, which can be collected non-invasively by arranging electrode arrays on the surface of the scalp. In AI glasses, portable EEG signal collection can be achieved by integrating dry electrodes on the edge of the frame or lens. Due to its advantages such as high temporal resolution, portable equipment, and low cost, EEG signals are widely used in research fields such as brain-computer interfaces and cognitive neuroscience. However, in the application scenario of AI glasses, in addition to the characteristics of large noise, significant individual differences, and time-varying non-stationarity, EEG signals also face unique challenges such as glasses movement, blinking interference, and environmental electromagnetic interference.

[0004] In order to separate and decode useful intention information from raw EEG data, traditional methods usually require multiple preprocessing, such as filtering and denoising, manual annotation, feature extraction, etc., and then use machine learning and other algorithms to model and identify intention patterns. This processing flow has many limitations in mobile scenarios such as AI glasses: first, the complex preprocessing process is difficult to meet the needs of real-time interaction; second, there are problems such as parameter matching and optimization difficulties in the two links of preprocessing and decoding, which easily lead to information loss; in addition, traditional algorithms have high requirements for computing resources and are difficult to run efficiently on portable devices.

[0005] In addition, in the actual use of AI glasses, due to factors such as the changing user activity status and the possible changes in the wearing position of the glasses, the EEG data characteristics often show greater individual differences and time-varying properties. Therefore, when performing intent recognition, it is necessary to design a more intelligent and adaptive algorithm that can adjust the parameters in a personalized manner according to the specific characteristics of the data. Using a fixed one-size-fits-all solution will be difficult to adapt to the changing actual application scenarios, especially in dynamic environments such as walking and exercising, which puts higher requirements on the adaptability and robustness of the algorithm.

[0006] Therefore, a method is urgently needed to solve at least one of the above problems. Summary of the invention

[0007] In a first aspect, an embodiment of the present application provides a method for identifying brain-computer interaction intention based on AI glasses, comprising:

[0008] Obtaining EEG activity data of multiple brain regions of a target user, forming an EEG time series and dividing it into multiple intent recognition data segments; the target user wears AI glasses;

[0009] Construct a discrete time set and a discrete frequency set according to the intention recognition data fragments, and obtain a key time-frequency feature point position sequence and an overall feature point position sequence according to the intention recognition data fragments, the discrete time set and the discrete frequency set;

[0010] Obtaining the number of occurrences and the confidence score of the key time-frequency feature point position sequence in the overall feature point position sequence, and obtaining the local anomaly evaluation score corresponding to the key time-frequency feature point position sequence according to the number of occurrences and the confidence score;

[0011] Divide the intention recognition data fragment into blocks to obtain multiple data blocks, and construct a frequency conversion set according to the multiple data blocks;

[0012] Obtain data fluctuation information corresponding to the data block, and obtain an intention complexity score based on the frequency conversion set and each data fluctuation information based on a cluster analysis method; the fluctuation information at least includes a frequency change sequence and a total change amount;

[0013] Determine a decoding threshold adjustment factor according to the intention complexity score and the abnormality evaluation score, and standardize the intention recognition data segment according to the decoding threshold adjustment factor to obtain normalized decoding data;

[0014] A user intention instruction corresponding to the target user is obtained according to the normalized decoded data, and an interactive control function of the AI ​​glasses is implemented according to the user intention instruction.

[0015] In a second aspect, the present application also provides a brain-computer interaction intention recognition device, comprising:

[0016] A data acquisition module, used to acquire EEG activity data of multiple brain regions of a target user, form an EEG time series and divide it into multiple intent recognition data segments; the target user wears AI glasses;

[0017] A sequence generation module is used to construct a discrete time set and a discrete frequency set according to the intention recognition data fragment, and obtain a key time-frequency feature point position sequence and an overall feature point position sequence according to the intention recognition data fragment, the discrete time set and the discrete frequency set;

[0018] A score acquisition module, used to obtain the number of occurrences and confidence scores of the key time-frequency feature point position sequence in the overall feature point position sequence, and obtain the local anomaly evaluation score corresponding to the key time-frequency feature point position sequence according to the number of occurrences and the confidence score;

[0019] A set construction module, used for dividing the intention recognition data segment into blocks, obtaining multiple data blocks, and constructing a frequency conversion set according to the multiple data blocks;

[0020] A change acquisition module is used to obtain data fluctuation information corresponding to the data block, and obtain an intention complexity score based on the frequency conversion set and each data fluctuation information based on a cluster analysis method; the fluctuation information at least includes a frequency change sequence and a total change amount;

[0021] A decoding acquisition module, used to determine a decoding threshold adjustment factor according to the intention complexity score and the abnormality evaluation score, and to normalize the intention recognition data segment according to the decoding threshold adjustment factor to obtain normalized decoding data;

[0022] The interaction completion module is used to obtain the user intention instruction corresponding to the target user according to the normalized decoded data, and implement the interactive control function of the AI ​​glasses according to the user intention instruction.

[0023] In a third aspect, the present application also provides a computer device, comprising a processor and a memory, wherein the memory is used to store a computer program, and when the computer program is executed by the processor, the brain-computer interaction intention recognition method based on AI glasses as described in the first aspect is implemented.

[0024] In a fourth aspect, the present application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the brain-computer interaction intention recognition method based on AI glasses as described in the first aspect.

[0025] Compared with the prior art, this application has at least the following beneficial effects:

[0026] 1 Adaptive soft standardization This paper proposes an innovative soft standardization strategy, which uses fuzzy logic reasoning to fuse information such as intention complexity score and abnormality score to obtain the decoding threshold adjustment factor, and dynamically adjusts the amplitude distribution and frequency distribution of the original EEG data according to the factor. This personalized standardization effect of "teaching students in accordance with their aptitude" effectively reduces the influence of factors such as changes in the wearing position of AI glasses and changes in the use environment, and significantly improves the accuracy of intention recognition.

[0027] 2. Efficient intention decoding technology This invention draws on the harmonic demodulation technology in modern communication theory and regards the EEG signal as a composite signal composed of a baseband intention coding pattern and a carrier component. Through the harmonic demodulation algorithm, the intention pattern can be separated from the complex background noise efficiently and accurately, avoiding the parameter matching difficulties of the two-level algorithms of preprocessing and decoding in the traditional method. In particular, this method has been optimized and can realize real-time processing on portable devices such as AI glasses, and has strong engineering application value.

[0028] 3 Intelligent clustering classification The present invention integrates multiple algorithms such as wavelet transform, random forest, Gaussian mixture model, etc. to perform intelligent feature extraction and pattern recognition on intent patterns. Through algorithm optimization and lightweight design, it is suitable for running in the resource-constrained environment of AI glasses while maintaining a high recognition accuracy. This method not only has strong robustness and generalization capabilities, but also supports the expansion of the intent-operation mapping table to achieve richer interactive functions. It should be understood that the above general description and the detailed description below are only exemplary and explanatory, and cannot limit this application. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 A flowchart of a method for identifying brain-computer interaction intention based on AI glasses according to an embodiment of the present application;

[0030] Figure 2 This is a schematic diagram of the structure of a brain-computer interaction intention recognition device shown in an embodiment of the present application;

[0031] Figure 3 A schematic diagram of the structure of a computer device shown in an embodiment of the present application. DETAILED DESCRIPTION

[0032] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0033] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, wholes, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or combinations thereof.

[0034] It should also be understood that the term “and / or” used in the specification and appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0035] As used in the specification and appended claims of this application, the term "if" can be interpreted as "when" or "uponce" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "uponce it is determined" or "in response to determining" or "uponce [described condition or event] is detected" or "in response to detecting [described condition or event]", depending on the context.

[0036] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0037] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0038] The technical solution of the embodiment of the present application is introduced below.

[0039] Brain-computer interface technology is a technical means to connect the neural activities in an organism with external devices or methods. In recent years, with the development of artificial intelligence and wearable device technology, the combination of brain-computer interface and AI glasses has gradually become a new research hotspot. As a new type of intelligent wearable device, AI glasses provide users with a natural and intuitive human-computer interaction method by integrating EEG acquisition electrodes, processors and display units. Among them, accurately identifying the user's interaction intention is the key to realizing the intelligent control of AI glasses.

[0040] EEG signals are weak potential changes generated by the synchronized ion currents of neuronal groups below the cerebral cortex, which can be collected non-invasively by arranging electrode arrays on the surface of the scalp. In AI glasses, portable EEG signal collection can be achieved by integrating dry electrodes on the edge of the frame or lens. Due to its advantages such as high temporal resolution, portable equipment, and low cost, EEG signals are widely used in research fields such as brain-computer interfaces and cognitive neuroscience. However, in the application scenario of AI glasses, in addition to the characteristics of large noise, significant individual differences, and time-varying non-stationarity, EEG signals also face unique challenges such as glasses movement, blinking interference, and environmental electromagnetic interference.

[0041] In order to separate and decode useful intention information from raw EEG data, traditional methods usually require multiple preprocessing, such as filtering and denoising, manual annotation, feature extraction, etc., and then use machine learning and other algorithms to model and identify intention patterns. This processing flow has many limitations in mobile scenarios such as AI glasses: first, the complex preprocessing process is difficult to meet the needs of real-time interaction; second, there are problems such as parameter matching and optimization difficulties in the two links of preprocessing and decoding, which easily lead to information loss; in addition, traditional algorithms have high requirements for computing resources and are difficult to run efficiently on portable devices.

[0042] In addition, in the actual use of AI glasses, due to factors such as the changing user activity status and the possible changes in the wearing position of the glasses, the EEG data characteristics often show greater individual differences and time-varying properties. Therefore, when performing intent recognition, it is necessary to design a more intelligent and adaptive algorithm that can adjust the parameters in a personalized manner according to the specific characteristics of the data. Using a fixed one-size-fits-all solution will be difficult to adapt to the changing actual application scenarios, especially in dynamic environments such as walking and exercising, which puts higher requirements on the adaptability and robustness of the algorithm.

[0043] To solve the above problems, please refer to Figure 1 , Figure 1 The flowchart of a method for identifying brain-computer interaction intention based on AI glasses provided in an embodiment of the present application is shown in FIG. The method for identifying brain-computer interaction intention based on AI glasses in an embodiment of the present application can be applied to computer devices, including but not limited to smart phones, laptops, tablet computers, desktop computers, physical servers, cloud servers and other devices. Figure 1 As shown, the brain-computer interaction intention recognition method based on AI glasses of this embodiment includes steps S101 to S107, which are described in detail as follows:

[0044] Step S101, obtaining EEG activity data of multiple brain regions of the target user, forming an EEG time series and dividing it into multiple intention recognition data segments; the target user wears AI glasses.

[0045] Specifically, the dry electrode array integrated in the frame and lens edge of AI glasses simultaneously records the discharge activities of neuronal groups from different areas of the brain, and converts these activity signals into multi-channel EEG time series data. AI glasses use specially designed dry electrodes that can collect EEG signals through direct contact with the scalp. These electrodes are arranged at key positions where the glasses are in close contact with the scalp according to ergonomic principles, mainly covering the frontal lobe, temporal lobe and other brain areas that are closely related to intention recognition. Each electrode position actually measures the overall electrical activity of the neuronal group within a certain range below the area. The built-in bioelectric amplifier in the glasses amplifies the weak EEG signals and digitizes the analog signals through analog-to-digital conversion to obtain discrete time series data.

[0046] For example, using 8-12 channel AI glasses can synchronously collect a corresponding amount of EEG time series data. Specifically, the electrodes are mainly distributed in the prefrontal and temporal lobe areas, which are closely related to intention formation and decision-making processes. Each time series X(t) represents the overall discharge activity of a neuron group in a specific brain area at time t.

[0047] In order to obtain high-quality EEG data, the original time series needs to be preprocessed multiple times. In the AI ​​glasses scenario, the preprocessing needs to pay special attention to the following aspects: 1) Apply a bandpass filter to remove 50 / 60Hz power frequency interference, low-frequency drift caused by glasses movement, and environmental electromagnetic interference. Commonly used filters include FIR, IIR, etc.; 2) Use improved blind source separation algorithms such as independent component analysis (ICA) to separate and remove artifacts caused by blinking, eye movement, and facial muscle activity; 3) Design an adaptive baseline correction algorithm to compensate for signal drift caused by slight changes in the position of glasses in real time; 4) Use an adaptive filtering method assisted by inertial sensor data to remove interference caused by daily activities such as walking and turning the head. Only after the above targeted preprocessing can relatively "pure" EEG time series data be obtained.

[0048] Next, it is necessary to infer the user's interaction intention by identifying and decoding the neural electrical activity patterns related to specific intentions generated by the brain. Each intention recognition data segment represents an independent intention formation process, such as attention shifting, making a choice, etc. This intention activity often stimulates changes in the discharge pattern of neuronal groups in a specific brain area that last for a period of time, which is manifested as a sudden change in the statistical characteristics of the amplitude and frequency of the EEG time series. This solution proposes an innovative segmentation method, which combines wavelet transform with the Random Forest algorithm, uses wavelet energy entropy features to capture the random changes in EEG signals caused by intention activities, and then uses the trained random forest model to automatically segment the boundaries of the intention segments. Specifically:

[0049] First, for the input N-channel EEG time series X(t)=[X1(t),X2(t),...,XN(t)]^T, wavelet transform is performed on each channel. Wavelet transform can represent and analyze signals in both time domain and frequency domain, and is particularly suitable for analyzing non-stationary signals such as EEG signals. Specifically, Xk(t) is subjected to continuous wavelet transform: Wk(a,b) =∫Xk(t)ψ((tb) / a)dt (1) where ψ is the wavelet basis function, a is the scale factor, and b is the time translation factor. Wk(a,b) represents the wavelet coefficients of Xk(t) at time b and scale a, reflecting the amplitude distribution of the signal at different scales. The smaller the scale a, the higher the time resolution; the larger the scale a, the higher the frequency resolution.

[0050] In actual calculation, it is necessary to perform discrete wavelet transform, use a set of scales ak and time bj for discrete sampling, and obtain the wavelet coefficient matrix Wk[j,m]. Equation (2) gives the calculation formula. Generally, ak=2^m and bj=j are selected for binary sampling, m is the scale index, and j is the time displacement: Wk[j,m]=Σ[n]Xk[n]ψ[2^-m(nj)] (2).

[0051] The normalized energy Ek[m]=Σ[j]|Wk[j,m]|^2 is calculated for the wavelet coefficient Wk[*,m] of each scale m, then Ek[m] reflects the energy distribution of Xk(t) at the corresponding scale. Usually cognitive activities will change the energy distribution of different frequency components, so the Ek[m] sequence can better represent the changes in the state of cognitive activities.

[0052] Next, Ek[m] is projected to the logarithmic scale lj=log2(1 / aj), and the energy density function pk(lj)=Ek[m] / (ΣEk[m]) is obtained. pk(lj) reflects the normalized energy distribution at the logarithmic scale lj. According to the concept of entropy in information theory, the wavelet energy entropy index Hk is defined as: Hk=-Σ[j]pk(lj)logpk(lj) (3) Hk can reflect the randomness or uncertainty of the energy distribution of the EEG signal Xk(t) at different scales. Generally speaking, when in a resting state, the neuronal discharge activities in various regions of the brain are relatively "ordered", resulting in energy concentration in a specific scale component, and Hk is small; when performing a certain cognitive task, the neuronal activity becomes more "random", and the energy distribution at each scale is more dispersed, and Hk increases. Therefore, the value of Hk and its changes can well characterize the cognitive state transition of the brain.

[0053] For N-channel input data, the wavelet energy entropy Hk of each channel is calculated, and a feature vector H(t)=[H1(t),H2(t),...,HN(t)]^T of length N is constructed to characterize the overall random state of multi-channel EEG data at time t. A sudden change in H(t) means that the brain has undergone a cognitive state shift, which may be the beginning of a new cognitive activity or the termination of an old activity.

[0054] In order to automatically detect the moment of cognitive state transition and segment the unit data fragments, the supervised learning random forest (RF) model is used. RF consists of multiple decision trees, each tree corresponds to a self-sampling set of data. During the tree construction process, constraints are imposed on different components of the feature vector, and finally the prediction results of all trees are integrated to improve the classification performance. First, a training sample set containing a large number of labeled cognitive state boundaries is constructed. The feature vector of the sample is H'(t)=[H(t),dH(t) / dt]^T with a length of N+1, that is, its derivative component is added to H(t) to enhance the detection ability of mutation points. RF is used to model the training set, output the prediction results of each decision tree at each time point, and use the majority voting method to obtain the comprehensive prediction probability. When the probability of predicting "cognitive state transition" exceeds the set threshold, it is determined to be the start or end boundary of an intention recognition data fragment. By sliding the time window, the position sequence of all segmentation boundaries can be obtained. Finally, the original multi-channel time series is segmented into a series of intention recognition data fragments Sim(t) (i=1...M) according to the boundary position.

[0055] The improved segmentation method can better meet the needs of intention recognition in AI glasses scenarios. This method has the following innovations: 1) The introduction of wavelet energy entropy features can effectively capture the random changes in EEG signals caused by intention activities, providing a new criterion for segmentation; 2) The use of a powerful integrated learning model RF reduces the dependence on a single threshold division rule, and has strong adaptability and anti-interference capabilities; 3) The feature vector contains time derivative information, which can enhance the detection effect of mutation points; 4) The segmentation is automatically completed by the machine learning model, which can achieve real-time processing and is suitable for running on portable devices such as AI glasses.

[0056] Step S102, constructing a discrete time set and a discrete frequency set according to the intention recognition data fragment, and obtaining a key time-frequency feature point position sequence and an overall feature point position sequence according to the intention recognition data fragment, the discrete time set and the discrete frequency set.

[0057] Specifically, let's first look at the process of constructing the duration set for intent recognition. The goal is to extract the duration Δt from each Sim(t) and combine the Δt of all segments to form a duration set D. The duration Δt is actually the difference between the start time t_start and the end time t_end of the unit segment, so the key is how to accurately determine t_start and t_end. A common method is to monitor the energy E_i,k(t) of Sim(t) on each channel k over time. When a significant mutation occurs, it can be regarded as a segment boundary point.

[0058] In some embodiments, constructing the frequency conversion set according to the plurality of data blocks includes: obtaining mean differences corresponding to adjacent data blocks; and constructing the frequency conversion set according to the absolute value corresponding to each mean difference.

[0059] Specifically, the normalized energy of Sim(t) on each channel k can be calculated as E_i,k(t)=X_i,k(t)^2 / ∑_t{X_i,k(t)^2}, where Xi_i,k(t) is the time series component of the i-th segment on the k-th channel. E_i,k(t) reflects the energy distribution of the channel at time t. At the beginning or end of cognitive activities, E_i,k(t) often mutates due to the change in the discharge pattern of the neuronal population. In order to improve the robustness of mutation detection, E_i,k(t) can be smoothed along the time direction to obtain S_i,k(t)=∑_τ{E_i,k(t+τ)g(τ)}, where g(τ) is a smooth kernel function, such as a Gaussian kernel or other window function.

[0060] Next, the derivative dS_i,k(t) / dt of S_i,k(t) is calculated. When the absolute value of the derivative exceeds the preset threshold θ, it can be determined as a segment boundary point. Assuming that the boundary time points detected on the kth channel are t_start and t_end, the duration of the i-th unit segment on the channel is Δt_i,k=t_end-t_start. Repeating the above process, the duration of all M segments on the channel {Δt_i,k} (i=1...M) can be obtained, which constitutes the intention recognition duration set of the channel, recorded as D_k.

[0061] For EEG data containing N channels, N unit duration sets D_k (k=1...N) can be obtained. Next, these N sets need to be fused to obtain the overall intention recognition duration set D. There are many fusion strategies to choose from, such as taking the union D=∪(D_k), calculating the median or weighted average. However, it should be noted that due to the distributed processing characteristics of the brain, the same cognitive activity will activate multiple brain areas, so there may be some differences in the Δt extracted on different channels. It is necessary to design a reasonable strategy to deal with this difference in order to obtain a more reliable duration estimate. For example, the median of N Δt_i,k can be taken as the final duration estimate Δt_i of the segment, or a weighted average can be used. In addition, for segments with abnormally long durations (such as more than 10 seconds) or too short durations (such as less than 200 milliseconds), they can be eliminated or treated separately in subsequent analysis.

[0062] After constructing the intention recognition duration set D, it is also necessary to analyze the statistical characteristics of each Sim(t) in the frequency domain, extract the energy distribution information of its different frequency components, and form the intention recognition frequency set F. The specific method is to first perform bandpass filtering decomposition on each Sim(t) on each channel k, and divide it into different frequency band components X_i,k,b(t) such as δ(0.5-4Hz), θ(4-8Hz), α(8-12Hz), β(12-30Hz), and γ(above 30Hz), where b is the subscript of the frequency band. Then calculate the energy of each frequency band component E_i,k,b=∑_t{X_i,k,b(t)^2}, and normalize it to obtain the energy ratio η_i,k,b=E_i,k,b / ∑_b{E_i,k,b}.

[0063] In this way, the normalized energy ratio η_i,k,b of each frequency band of the i-th segment on the k-th channel is obtained. The combination of these ratio values ​​constitutes the intention recognition frequency set of the channel, which is recorded as F_k={η_i,k,b}(i=1...M;b=δ,θ,α,β,γ). Similar to the duration set, it is necessary to fuse N F_k to obtain the overall intention recognition frequency set F. Here, we can simply take the arithmetic average of the energy ratios of each frequency band, that is, η_i,b=∑_k{η_i,k,b} / N, to obtain F={η_i,b}. Alternatively, different weights can be assigned according to the "sensitivity" of different brain regions to different frequency components to improve the recognition ability of frequency features for cognitive activity types.

[0064] Through the above analysis, two types of key statistical features were successfully extracted from the original multi-channel EEG data Sim(t): the intention recognition duration set D and the intention recognition frequency set F.

[0065] In some embodiments, constructing a discrete set of durations and a discrete set of frequencies based on the intention identification data segments includes: constructing an intention identification duration set and an intention identification frequency set based on the intention identification data segments; respectively obtaining data differences corresponding to the intention identification duration set and the intention identification frequency set; and obtaining the discrete set of durations and the discrete set of frequencies based on the data differences.

[0066] In the step, although the elements in D and F come from the same batch of Sim(t), due to the high individual differences in the brain and the influence of noise during data acquisition and preprocessing, they often present a certain discrete distribution rather than a single definite value. Therefore, in the step, it is necessary to construct a discrete set of duration and a discrete set of frequency based on the data differences in D and F to characterize and quantify this discrete characteristic.

[0067] The construction process of the discrete duration set includes: Through the above steps, the intention recognition duration set D={Δti}(i=1...M) containing M elements is obtained, where each Δti is the estimated duration value of the i-th unit segment. Ideally, the durations of cognitive activities of the same type should be equal or close, but there are often some differences in actual situations. The kernel density estimation method in probability theory can be used to estimate the probability density function p(Δt) of the duration Δt based on the sample data in D.

[0068] Exemplarily, the method of respectively obtaining the data differences corresponding to the intention identification duration set and the intention identification frequency set includes: determining a first probability density function corresponding to the intention identification duration set at the estimated duration based on a kernel density analysis method; determining the data differences corresponding to the intention identification duration set according to the first probability density function, so as to obtain the discrete duration set according to the data differences corresponding to the intention identification duration set; performing principal component analysis on the intention identification frequency set to obtain a low-dimensional projection corresponding to the intention identification frequency set; performing kernel density estimation on the low-dimensional projection to obtain a second probability density function; determining a second data difference corresponding to the intention identification frequency set according to the second probability density function, so as to obtain the discrete frequency set according to the data differences corresponding to the intention identification duration set.

[0069] Specifically, let φ(x) be the kernel function (such as Gaussian kernel) and h be the bandwidth parameter, then p(Δt) can be calculated by the following formula: p(Δt) = (1 / M)∑(i=1...M)φ((Δt-Δti) / h).

[0070] p(Δt) reflects the discrete distribution of Δt in D. When p(Δt) has multiple significant peaks, it means that the sample data in D has great differences, corresponding to different duration patterns. The duration discrete set T can be constructed based on the maximum point position {Δtk} (k=1...K) of p(Δt), where each element Δtk represents a typical duration pattern. A simple construction method is to first determine the local maximum point {Δtk} of p(Δt), and then divide the sample Δti into the category corresponding to the nearest Δtk based on some domain division criteria (such as based on the kurtosis and half-peak width of p(Δt)) with these maximum points as the center, so as to obtain K subsets {Di} (i=1...K), and these K subsets constitute the duration discrete set T={Di}.

[0071] Next, let's look at the construction process of the frequency discrete set. In the step, an intention recognition frequency set F={ηi}(i=1...M) containing M elements is obtained, where each ηi is a vector composed of the energy ratios of 5 frequency bands, such as δ, θ, α, β, and γ, representing the energy distribution characteristics of the i-th unit fragment on different frequency components. Due to individual differences and noise influences, each element ηi in F also has a certain degree of discreteness, and this discreteness needs to be modeled and quantified.

[0072] Similar to the discrete set of duration, we can first estimate the probability density function p(η) of the elements in F, and then construct a discrete set based on the peak position of p(η). However, it should be noted that since ηi is a 5-dimensional vector, the computational complexity of directly estimating its joint probability density function p(η) is relatively high. A common dimensionality reduction method is to perform principal component analysis (PCA) on F to obtain its low-dimensional projection η'i, and then perform kernel density estimation on η'i to calculate p(η'). When p(η') shows a significant multi-peak distribution, the samples can be divided into K subsets {Fk} consisting of the nearest peak according to the peak position {η'k} (k=1...K), thereby obtaining a discrete frequency set R={Fk}. Each subset Fk represents a typical frequency energy distribution pattern.

[0073] In some embodiments, obtaining the key time-frequency feature point position sequence and the overall feature point position sequence based on the intention recognition data segment, the discrete set of duration and the discrete set of frequency includes: performing time-frequency analysis on the intention recognition data segment to obtain the time-frequency features corresponding to the intention recognition data segment; using the discrete set of duration and the discrete set of frequency as prior information, and obtaining a time-frequency heat map based on the prior information and the time-frequency features; and extracting the key time-frequency feature point position sequence and the overall feature point position sequence from the time-frequency heat map.

[0074] Different from traditional time-frequency analysis methods such as short-time Fourier transform, the main advantage of using wavelet transform here is that it has variable time window length and frequency bandwidth, which can better capture the time-varying characteristics of non-stationary signals. In the application scenario of AI glasses, due to the limited number of electrodes and their main distribution in the frontal and temporal lobe areas, special attention should be paid to the signal characteristics of these areas when performing time-frequency analysis. At the same time, considering the computing resource limitations of portable devices, the scale selection and calculation process of wavelet transform are optimized to achieve real-time processing requirements.

[0075] First, perform wavelet transform on each Sim(t) in the step on each channel k to obtain its wavelet coefficient matrix Wk[j,m]. The specific calculation formula is: Wk[j,m] = ∑[n]Xk[n]ψ[2^-m(nj)] (1) Where ψ[n] is the wavelet basis function, m is the scale index, j is the time displacement, and Wk[j,m] reflects the wavelet coefficient value of Xk(t) at time j and scale m. Generally speaking, small scale m corresponds to high-frequency components, and large scale m corresponds to low-frequency components. The time-frequency concentration characteristics of the wavelet transform can be optimized by selecting a suitable wavelet basis ψ[n]. Commonly used wavelet bases include Haar, Daubechies, Coiflet, etc.

[0076] After calculating Wk[j,m] of each Sim(t) on each channel k, we get the two-dimensional wavelet coefficient matrix of the unit fragment on the time-scale plane. However, it should be noted that since the scale ranges corresponding to different frequency components are different, it is necessary to further convert the scale m to the frequency f to obtain a unified time-frequency representation. The commonly used scale-to-frequency conversion formula is:

[0077] f = Fc / (a*Δ) (2).

[0078] Where Fc is the center frequency of the wavelet basis, a is the sampling interval, and Δ is the sampling time interval. By using formula (2), Wk[j,m] can be mapped to the time-frequency plane. In this way, each Sim(t) corresponds to a time-frequency coefficient matrix Ck[j,f] on each channel k, where each element Ck[j,f] reflects the signal energy at time j and frequency f.

[0079] Next, we need to normalize Ck[j,f] to get the time-frequency heatmap:

[0080] Hk[j,f] = Ck[j,f] / max(Ck[j,f]) (3).

[0081] The value range of Hk[j,f] is between [0,1]. The larger the value, the higher the energy at that time-frequency point. It can be plotted on a two-dimensional plane with different shades of color to intuitively present the energy distribution of the unit data segment in the entire time-frequency domain.

[0082] From the weighted time-frequency heat map Hk'[j,f], a simple method of setting a threshold η can be used to extract key time-frequency feature points, and all points (j,f) with Hk'[j,f] values ​​greater than η are recorded as feature points, forming the key time-frequency feature point position sequence Pk,i of the i-th intention recognition segment on the channel k. However, the setting of η directly affects the number and distribution of feature points in Pk,i, so the η value needs to be appropriately adjusted so that Pk,i meets the following two conditions:

[0083] 1. Pk,i contains a sufficient number of feature points, which can better characterize the time-frequency pattern of the intent recognition fragment; 2. The position distribution of the feature points in Pk,i is relatively concentrated and will not be too dispersed.

[0084] In order to meet the first condition, we can first count the functional relationship between the number of points Nk,i of Pk,i and η, that is, Nk,i=f(η). Generally speaking, as η gradually increases, Nk,i will first increase rapidly until it reaches a peak value, and then gradually decrease and approach 0. The Nk,i=f(η) curve can be fitted to obtain the Nk,i value corresponding to its maximum point η_max, which is recorded as N_max. Under normal circumstances, only when Nk,i>N_maxρ (ρ is a preset ratio threshold, such as 0.1~0.3), Pk,i is considered to contain a sufficient number of feature points. Therefore, the lower bound of η can be determined as the η value corresponding to the condition that Nk,i>N_maxρ.

[0085] In order to meet the second condition, it is necessary to evaluate the position distribution of feature points in Pk,i. A common approach is to calculate the distance d from all feature points in Pk,i to the geometric center (j_c, f_c) and construct a distance distribution histogram H(d). If H(d) shows an obvious unimodal distribution and the distance range where the peak is located is small (such as less than 1 / 3 of the time-frequency plane size), then it means that the position distribution of feature points in Pk,i is relatively concentrated. Based on this, a concentration score S can be defined, for example, using the peak height divided by the distance standard deviation as the calculation formula for S. The upper bound of η can be determined as the η value corresponding to when S>S_th (S_th is the preset concentration threshold).

[0086] After obtaining Pk,i on each channel k, it is necessary to further fuse these N sequences to obtain the overall feature point position sequence Pi. The most direct method is to take the union of each Pk,i, that is: Pi = ∪(k=1...N) Pk,i.

[0087] However, since different channels may respond differently to the same feature point, if we simply take the union, Pi will contain a large number of repeated points, which will bring certain difficulties to the subsequent feature processing. Therefore, we can consider assigning different weights and select points with higher comprehensive weights to form Pi.

[0088] A common weight calculation method is to first count the frequency C(j,f) of each feature point (j,f) in N Pk,i as its basic weight; then multiply it with the heat map value αk(j,f) of the point in each Hk'[j,f] to obtain the comprehensive weight:

[0089] W(j,f) = C(j,f) * Πk αk(j,f).

[0090] Where W(j,f) is the comprehensive weight at time point j and frequency f, C(j,f) represents the frequency of the feature point in N sequences, αk(j,f) is the heat map value of the kth channel at this time-frequency point, and Πk represents the multiplication operation of all channels k=1 to N. αk(j,f) can directly take the value of Hk'[j,f], or it can be some nonlinear transformation of the value, such as logarithmic or exponential transformation, to enhance the distinguishing ability of the heat map value. The design of this comprehensive weight takes into account both the degree of co-occurrence of feature points in multiple channels (through C(j,f)) and the intensity of each channel's response to the feature point (through the multiplication of αk(j,f)). Considering the real-time processing requirements of AI glasses, a simple linear transformation is used here to reduce the computational complexity. Next, given a weight threshold η_w, all points (j,f) with W(j,f)>η_w are recorded in the overall feature point position sequence Pi. Compared with simple union, this weighted fusion method can retain feature points with high cross-channel consistency and significant response, making Pi more compact and representative, and also suitable for running on resource-constrained portable devices such as AI glasses.

[0091] Step S103, obtaining the number of occurrences and confidence scores of the key time-frequency feature point position sequence in the overall feature point position sequence, and obtaining the local anomaly evaluation score corresponding to the key time-frequency feature point position sequence according to the number of occurrences and the confidence score.

[0092] Specifically, due to individual differences and noise, even for the same type of cognitive activities, the Pi extracted from different Sim(t) may have certain differences. Therefore, in this step, it is necessary to perform statistical analysis on these Pi, and evaluate and quantify the degree of abnormality based on the frequency and confidence of each feature point position in all Pi, so as to provide a reference for abnormal detection and activity classification.

[0093] First, count the number of occurrences C(j,f) of each feature point position (j,f) in all M Pi, which can be regarded as the frequency of the cognitive pattern corresponding to the position in the entire sample set. Generally speaking, if the value of C(j,f) is large, it means that (j,f) is a common and typical time-frequency feature, otherwise it is relatively rare. C(j,f) can be normalized to obtain the frequency p(j,f)=C(j,f) / M, and obviously Σp(j,f)=1. By calculating the Shannon entropy H=-Σp(j,f)logp(j,f) of the frequency p(j,f), the randomness of the entire feature position set can be evaluated.

[0094] In addition to frequency, the confidence score of each feature position (j, f) needs to be considered, that is, the significance of its response in all occurrences of Pi. A simple approach is to count the time-frequency heat map value α(j, f) corresponding to each occurrence, and take its mean or geometric mean as the confidence score s(j, f). The higher the confidence, the more significant the response of the feature position, and the more likely it is to correspond to the real cognitive activity pattern. The frequency p(j, f) and the confidence s(j, f) can be multiplied to obtain the comprehensive confidence weight:

[0095] w(j,f) = p(j,f) * s(j,f).

[0096] w(j,f) reflects the typicality of the feature position (j,f) in the entire sample set. The larger the value, the more likely it is a common and significant cognitive pattern, and vice versa.

[0097] With w(j,f), we can further calculate the local anomaly evaluation score of each feature position. The common practice is to first count the w(j,f) values ​​of all positions (j,f) in each Pi, calculate the mean μi and standard deviation σi, and then perform Z-Score normalization on the w(j,f) corresponding to each (j,f) to obtain its local anomaly score:

[0098] z(j,f) = (w(j,f) - μi) / σi.

[0099] Obviously, if z(j,f) is close to 0, it means that (j,f) is a typical feature position and there is no anomaly; if z(j,f) is larger (such as greater than 3) or smaller (such as less than -3), it means that (j,f) is a rare and abnormal position, which may be caused by noise or individual differences. You can traverse all (j,f) in Pi, calculate their z(j,f) values, and record the positions with larger |z| to form the abnormal position set Ai of Sim(t).

[0100] Step S104, dividing the intent recognition data segment into blocks, obtaining multiple data blocks, and constructing a frequency conversion set according to the multiple data blocks.

[0101] Specifically, first, for the component Xik(t) of the i-th intention recognition data segment Sim(t) on the k-th channel, it is equally divided into L sub-blocks of equal length, that is, Xik(t)={Xik,l(t)}(l=1...L), where the length of each Xik,l(t) is T=M / L, and M is the total length of Xik(t). Through this block division method, the entire Xik(t) is divided into L segments, each of which contains data of length T. Next, the absolute value of the mean difference between adjacent data blocks is calculated to construct the mean difference sequence Dik,l:

[0102] Dik,l ​​= |mean(Xik,l+1) - mean(Xik,l)| (l=1...L-1).

[0103] Where, mean(Xik,l) represents the mean of Xik,l(t). Obviously, if the mean difference between adjacent data blocks is large, it means that the signal frequency near the location may have changed; otherwise, it means that the frequency remains relatively stable. Therefore, Dik,l ​​can well reflect the time distribution characteristics of the frequency change in Xik(t).

[0104] However, it should be noted that due to the limited information of a single channel, it may be difficult to fully describe the frequency changes in the entire Sim(t) by directly using Dik,l. Therefore, it is necessary to fuse the Dik,l ​​from N channels to obtain a unified frequency conversion sequence Dl:

[0105] Dl = F(Dik,l) (k=1...N;l=1...L-1).

[0106] Where F(·) is some fusion function, such as a simple weighted average:

[0107] Dl = Σk (wk * Dik,l) / Σk wk.

[0108] Here wk is the weight of the kth channel, which can be assigned according to the importance of the brain area corresponding to the channel in cognitive activities. In addition to weighted averaging, other fusion methods can also be tried, such as applying nonlinear transformation to Dik,l ​​and then fusing it.

[0109] After the above processing, we get a frequency conversion sequence Dl with a length of L-1, which reflects the distribution of the frequency change of the entire Sim(t) in the time dimension. If the value of each element Dl in the Dl sequence is large, it means that a frequency conversion has occurred near this time point; otherwise, it means that the frequency at this position is relatively stable. Dl can be normalized to obtain the frequency conversion probability sequence Pl:

[0110] Pl = Dl / Σl Dl.

[0111] Then Pl reflects the probability of frequency conversion at each time position l of Sim(t). Next, we can construct a frequency conversion set G based on Pl. Specifically, we first set a probability threshold η, record all time positions l where Pl>η, and obtain a position set {li} (i=1...M), where M is the number of elements in the set. Then, taking each li as the center, we take δ time points on its left and right to form a subsequence Xi with a length of 2δ+1:

[0112] Xi = {Pl-δ, Pl-δ+1, ..., Pl, ..., Pl+δ}.

[0113] By combining all these Xi, we get the frequency transition set G = {Xi} (i = 1...M). Each Xi reflects a local change pattern of the time series Pl near the i-th frequency transition position li. This construction method based on the probability threshold η and the neighborhood length δ can well capture the time position of the frequency transition in Sim(t) while retaining the context information before and after the transition.

[0114] Step S105, obtain the data fluctuation information corresponding to the data block, and obtain the intention complexity score based on the frequency conversion set and each data fluctuation information based on the cluster analysis method; the fluctuation information at least includes the frequency change sequence and the total change amount.

[0115] Specifically, firstly, the data blocks divided by the steps are subjected to fluctuation analysis. For each data block Xik,l(t), a short-time Fourier transform (STFT) is performed to obtain its energy spectrum Eik,l(t,f) at different time and frequency. STFT divides the time series Xik,l(t) into overlapping small segments, and performs Fourier transform on each small segment, thereby obtaining local information in time and frequency at the same time. In general, by selecting a suitable window function w(t) and time step τ, Eik,l(t,f) can be calculated by the following formula: Eik,l(t,f) = |∫Xik,l(t')w(t'-t)e^(-j2πft')dt'|^2.

[0116] Where w(t) is a window function, such as a rectangular window, a Hamming window, a Gaussian window, etc. The window length determines the resolution of the STFT in the time and frequency domains. f is the frequency, and τ is the time step. Eik,l(t,f) reflects the energy distribution of Xik,l(t) at time t and frequency f.

[0117] With Eik,l(t,f), we can define a frequency change index Vik,l to measure the frequency fluctuation degree within the lth data block: Vik,l = ∑t,f |Eik,l(t,f) - Eik,l(t-τ,f)|.

[0118] That is, the changes of Eik,l(t,f) in the time dimension are accumulated and summed. Obviously, if the value of Vik,l is large, it means that the frequency inside the data block has changed greatly; otherwise, it means that the frequency is relatively stable. Vik,l can be normalized to obtain the normalized frequency change of the lth data block on the kth channel: Qik,l = Vik,l / ∑lVik,l.

[0119] Then, combined with the frequency conversion set G={Xi} of the step, Qik,l of each data block is weighted fused. First, according to the frequency conversion position {li} recorded in G, the weight coefficient wk of each channel k is calculated: wk = ∑i exp(-d(k,li) / σ) where d(k,li) is the distance from channel k to the conversion position li, and σ is the scale parameter. In this way, channels closer to the frequency conversion position will obtain larger weights. Qik,l from N channels is weighted fused: Ql = ∑k (wk * Qik,l) / ∑k wk.

[0120] The overall frequency change score Ql of the data block is obtained. The method analyzes these features in real time. When the local energy distribution deviation exceeds 25% or the network connection feature changes by more than 30%, the feature extraction parameters are adjusted in time to ensure the timeliness and reliability of the feature sequence. Next, the overall frequency change is calculated for the entire intent recognition data segment: Q = ∑lQl.

[0121] With {Ql} and Q, we can combine the frequency conversion set G obtained in the step to evaluate the complexity of the intent recognition data segment. The specific approach is to represent each data segment as a feature vector: Vi = [{Ql}i, Qi,Fi] where Fi is the feature extracted from the frequency conversion set G, including:

[0122] 1. Transform the autocorrelation coefficient of the sequence Xi to ri = corr(Xi, Xi).

[0123] 2. The mean μi and standard deviation σi of the intervals between adjacent transition points.

[0124] 3. Cumulative distribution function of conversion intensity Fi(x).

[0125] Then, cluster analysis is performed on all Vi and they are divided into K clusters, each cluster representing a level of complexity. In order to adapt to the resource constraints of AI glasses, the K-Means algorithm with lower computational complexity is selected here, and the operating efficiency is improved by optimizing the feature dimension and the number of iterations.

[0126] In cluster analysis, it is necessary to define a similarity measure between Vi, usually using Euclidean distance or other similar distance functions. However, it should be noted that since the numerical ranges of {Ql}i and Qi are quite different, they need to be normalized or standardized before calculating the distance. In addition, different weights can be assigned. For example, if the total frequency change Qi has a greater impact on the complexity of the intent, the weight of the Qi component can be increased.

[0127] After cluster analysis, all intent recognition data fragments are divided into K different clusters. According to the statistical values ​​of {Ql}i and Qi in each cluster, an intent complexity score can be assigned to the cluster. For example, for the kth cluster, the mean μk and standard deviation σk of all Qi in it are calculated, and the intent complexity score of the cluster can be defined as: Ck = μk + λ *σk.

[0128] Where λ is a preset weight parameter used to control the weight of the mean and standard deviation in the score. The larger the Ck, the more drastic the frequency change corresponding to the cluster and the higher the complexity of the intent. These K Ck values ​​can be normalized to the interval [0,1] to obtain the intent complexity score of each cluster. This score can serve as an important reference for the subsequent decoding process, helping the method to adopt corresponding decoding strategies for intents of different complexity.

[0129] Step S106, determining a decoding threshold adjustment factor according to the intent complexity score and the anomaly evaluation score, and normalizing the intent recognition data segment according to the decoding threshold adjustment factor to obtain normalized decoding data.

[0130] Specifically, the role of the decoding threshold adjustment factor is to adjust the parameter settings in the subsequent decoding process for the intention recognition data segments of different complexity and abnormality to improve the accuracy and robustness of decoding. Generally speaking, for segments with higher complexity or greater abnormality, the decoding threshold needs to be appropriately relaxed to avoid missing key intention patterns; while for relatively simple and normal segments, the decoding threshold can be appropriately increased to help reduce noise and improve the specificity of decoding.

[0131] In some embodiments, determining a decoding threshold adjustment factor according to the intention complexity score and the anomaly evaluation score includes: constructing a fuzzy logic reasoning system; the fuzzy logic reasoning system includes a fuzzification unit, an inference engine and a defuzzification unit; fuzzifying the intention complexity score and the anomaly evaluation score according to the fuzzification unit; inputting the fuzzified intention complexity score and the anomaly evaluation score into the inference engine and outputting a fuzzy reasoning result; inputting the fuzzy reasoning result into the defuzzification unit and outputting the decoding threshold adjustment factor.

[0132] In order to achieve the above goals, we first need to construct a fuzzy logic reasoning method. The method includes three core parts: fuzzification, inference engine and defuzzification. First, the intention complexity score Ck and the anomaly score z(j,f) are fuzzified. For the complexity score Ck, three fuzzy subsets are defined: low complexity (Low), medium complexity (Medium) and high complexity (High), and their membership functions are as follows: μLow(Ck) = exp(-(Ck-α)^2 / (2σ^2)) (Ck < α) μMedium(Ck) = 1 - |Ck - α| / β(α - β / 2 <= Ck <= α + β / 2) μHigh(Ck) = exp(-(Ck-1)^2 / (2σ^2)) (Ck > 1 - σ) Where α, β, and σ are preset parameters, which control the center position and range of each fuzzy subset respectively. For the anomaly score z(j,f), three fuzzy subsets are designed: normal, moderately abnormal, and severely abnormal, and the membership function adopts a similar Gaussian form. In the AI ​​glasses application, in order to reduce the amount of calculation, a simple piecewise linear form is selected for the membership function.

[0133] Next is the inference engine. Based on the membership distribution of the input quantity, a targeted fuzzy rule base is designed, mainly considering the following rule combinations:

[0134] When complexity is low and abnormality is normal, set a higher decoding threshold to improve specificity Rule 1: If Complexity is Low and Abnormality is Normal Then the adjustment factor is reduced.

[0135] When complexity is medium and anomaly is normal, keep the default threshold Rule 2: If complexity is Medium and anomaly is Normal Then the adjustment factor is unchanged.

[0136] When complexity is high or abnormality is obvious, lower the decoding threshold to increase sensitivity Rule 3: If complexity is High or abnormality is severely abnormal Then the adjustment factor is significantly increased.

[0137] For each rule, first calculate the truth value of its antecedent part, that is, the combination of the input membership. The combination method usually adopts the minimum (AND) or maximum (OR) operation: τi = min(μA(x), μB(y)) (for AND rules) τi = max(μA(x), μB(y)) (for OR rules) Where τi is the activation of the i-th rule, μA(x) and μB(y) are the membership values ​​of the fuzzy subsets A and B corresponding to the input variables x and y respectively. The method calculates the activation of each rule in real time. For rules with higher activation, their output values ​​have a greater impact on the final result.

[0138] Finally, defuzzification is performed to convert the results of fuzzy reasoning into a certain adjustment factor value. Considering the computing power limitations of the AI ​​glasses processor, the centroid method with less computational effort is chosen here: y = (∫y * μ(y)dy) / (∫μ(y)dy) where y is the value range of the output quantity, and μ(y) is the membership function of the output fuzzy subset. Through the above calculations, the final decoding threshold adjustment factor y is obtained. This adjustment factor will be used to dynamically adjust the decoding parameters in the subsequent decoding process, so that the method can adaptively optimize the decoding performance according to the complexity and abnormality of the input data. For example, when processing high-complexity intent fragments, the method will appropriately lower the decoding threshold to avoid missing key intent patterns; when processing low-complexity normal fragments, the threshold will be increased to enhance noise resistance.

[0139] In some embodiments, the standardizing the intention recognition data segment according to the decoding threshold adjustment factor includes: adaptively normalizing the amplitude of the intention recognition data segment according to the decoding threshold adjustment factor, and softly adjusting the amplitude distribution corresponding to the amplitude; adaptively normalizing the frequency distribution of the intention recognition data segment according to the decoding threshold adjustment factor; generating the normalized decoded data based on the adaptively standardized amplitude and frequency, and completing the standardization of the intention recognition data segment.

[0140] By adopting a novel soft normalization method, the amplitude and frequency distribution of Sim(t) are adaptively adjusted using y, and finally the normalized Xnorm(t) is obtained. The innovation of this method is that it does not simply apply linear or nonlinear transformation to the intention recognition data fragment, but dynamically adjusts its normalization strategy according to the value of y, and adopts different processing procedures for fragments of different complexity and abnormality, thus achieving the personalized normalization effect of "teaching students in accordance with their aptitude". Considering the real-time processing requirements of AI glasses, this solution optimizes the calculation process.

[0141] First, let's look at the steps of amplitude normalization. For the component Xik(t) of Sim(t) on each channel k, its amplitude distribution is often affected by individual differences and environmental noise, and there will be a certain degree of drift or distortion. The amplitude of Xik(t) needs to be adjusted to a standard reference range to enhance the distinguishability between neural coding patterns. The usual practice is to estimate the mean μik and standard deviation σik of Xik(t), and then perform the following Z-Score normalization transformation: Zik(t) = (Xik(t) - μik) / σik (1) In this way, the mean of Zik(t) is 0, the standard deviation is 1, and the amplitude distribution is more concentrated and standardized. However, the above standard Z-Score normalization assumes that the amplitude of Xik(t) follows a normal distribution. For non-Gaussian distributions with skewness or kurtosis, its effect may be affected to a certain extent.

[0142] Therefore, in this scheme, an adaptive amplitude normalization (AAN) method is proposed, which uses the decoding threshold adjustment factor y to softly adjust the amplitude distribution of Xik(t). Specifically, a controllable, smooth nonlinear transformation function f(y,x) is defined, with the original amplitude x of Xik(t) as the independent variable and y as the control parameter. f(y,x) needs to meet the following conditions: 1) When y=0, f(0,x)=x, that is, no transformation is performed 2) When y>0, f(y,x) has a smaller compression degree for large values ​​x and a larger compression degree for small values ​​x, that is, it tends to Gaussianization 3) When y<0, f(y,x) has a larger stretch degree for large values ​​x and a smaller stretch degree for small values ​​x, that is, it tends to be sparse-tailed distribution.

[0143] Based on the above conditions, the following transformation function can be selected: f(y,x) = sign(x) * |x|^(1+y) / (1+|x|^y) (2) When y=0, f(0,x)=sign(x)*|x| / 2=x; when y>0, the function has a smaller compression degree for large values ​​of x and a larger compression degree for small values ​​of x; when y<0, the function has a larger stretch degree for large values ​​of x and a smaller stretch degree for small values ​​of x. Through this soft compression or soft stretching, the amplitude distribution shape of Xik(t) can be adjusted to achieve adaptive amplitude standardization.

[0144] Specifically, substituting Xik(t) into the transformation function (2) yields its amplitude-normalized version: Aik(t) = f(y, Xik(t)) (3) The amplitude distribution of Aik(t) will be more concentrated or dispersed than that of Xik(t), and the specific effect is determined by the sign of y. Next, Aik(t) can be subjected to a simple Z-Score normalization: Zik(t) = (Aik(t) - μik) / σik (4) where μik and σik are the mean and standard deviation of Aik(t), respectively. In this way, Zik(t) obtains a standardized amplitude distribution with a mean of 0 and a standard deviation of 1, while also retaining and amplifying the characteristics of the neural coding pattern in Aik(t), achieving the goal of amplitude normalization.

[0145] In addition to amplitude normalization, the frequency distribution of Sim(t) also needs to be normalized. This is because different Sim(t) often have different dominant frequency components, and neural coding is often concentrated in certain specific frequency bands. Therefore, it is necessary to concentrate the frequency energy of all fragments on these frequency bands to enhance the separability of the coding mode. Each Zik(t) can be bandpass filtered to retain only the key frequency band components. For example, it is known that the brain's β (12-30Hz) and γ (above 30Hz) frequency bands are highly correlated with cognitive activities. Then a bandpass filter can be designed to filter Zik(t) to obtain: Yik(t) = BPF(Zik(t), fl, fh) (5) Where BPF is a bandpass filter, fl and fh are the low-end and high-end frequencies respectively, and the components in the range of [fl,fh] can be passbanded, and the components of the remaining frequencies are blocked. In practical applications, the settings of fl and fh can refer to statistical data and prior knowledge, or they can be dynamically adjusted according to the specific type of cognitive activity.

[0146] In addition, it is also necessary to impose certain restrictions on the frequency distribution according to the decoding threshold adjustment factor y. For example, when y>0, it means that the corresponding Sim(t) is more complex or abnormal, and the passband frequency band [fl,fh] needs to be expanded to avoid losing key coding features. Specifically, an adaptively adjusted passband interval [fl',fh'] can be defined: fl' = fl * (1 - a * y)fh' = fh * (1 + b * y) (6) where a and b are preset positive weight factors used to control the degree of frequency band expansion. Next, Yik(t) is re-bandpass filtered: Yik'(t) = BPF(Yik(t), fl', fh') (7) In this way, the frequency distribution Yik'(t) after adaptive normalization is obtained. When y>0, [fl',fh'] is greater than [fl,fh], covering a wider frequency band; when y<0, [fl',fh'] is less than [fl,fh], only the information of the dominant frequency band is retained and some irrelevant components are filtered out.

[0147] By recombining Yik'(t) on all channels k, the final normalized decoding data Xnorm(t) is obtained: Xnorm(t) = [Y1'(t), Y2'(t), ..., YN'(t)]^T (8) where N is the number of electrode channels on the AI ​​glasses. Since Xnorm(t) combines the effects of amplitude normalization and frequency normalization, it not only has a standardized amplitude distribution, but also the frequency energy is concentrated in the key intention-related frequency band, creating favorable conditions for the subsequent decoding process. For example, in attention monitoring, when it is detected that the user is concentrating, the method will appropriately strengthen the energy of the β band; in the motor imagery task, special attention will be paid to the changes in the μ rhythm. Through this y-based adaptive normalization mechanism, the method can flexibly adjust the processing strategy according to the specific characteristics of the data, improving the adaptability and reliability in actual scenarios.

[0148] Step S107, obtaining the user intention instruction corresponding to the target user according to the normalized decoded data, and implementing the interactive control function of the AI ​​glasses according to the user intention instruction.

[0149] Specifically, harmonic demodulation is a commonly used communication signal demodulation method, which uses the product relationship between the carrier signal and the baseband signal to separate the baseband signal from the carrier. In the field of brain-computer interface, the neural coding pattern can be regarded as a brain activity pattern similar to the baseband signal, and the EEG data also contains some periodic carrier components, such as α, β and other harmonic frequencies. Therefore, we can learn from the idea of ​​harmonic demodulation to separate the coding pattern from Xnorm(t) and then decode the control instructions.

[0150] Specifically, the harmonic demodulation algorithm includes the following key steps:

[0151] 1. Carrier extraction: First, the carrier signal needs to be extracted from Xnorm(t). Generally, frequency components such as α and β with significant spectrum peaks are selected. Xnorm(t) can be subjected to a short-time Fourier transform (STFT) to obtain its time spectrum Snorm(t,f). For each time t, find the frequency fmax(t) where the significant peak is located in Snorm(t,f) as the carrier frequency at that moment. A simple peak detection method is to first calculate the average power of the spectrum at each frequency f:

[0152] Pf = mean(|Snorm(t,f)|^2) (1).

[0153] Then smoothing is performed in the frequency domain to obtain the power spectrum density Pdf:

[0154] Pdf = smooth(Pf) (2).

[0155] Where smooth(·) is a smooth kernel function, such as a Gaussian kernel or other window function. Next, find the position corresponding to the significant peak fmax(t) in the PDF, which must satisfy:

[0156] Pdf(fmax(t)) >= η * max(Pdf) (3).

[0157] Where η is the set peak detection threshold, which is usually set to 0.6~0.9. Obviously, for the time t when there is a significant spectrum peak, fmax(t) is the carrier frequency component at that moment. For the time period when the spectrum peak is not obvious, fmax(t) can be estimated by interpolation or tracking the time continuity of the peak.

[0158] After obtaining fmax(t), we can construct the carrier signal xc(t):

[0159] xc(t) = A(t) * cos(2πfmax(t)t + φ(t)) (4).

[0160] Where A(t) and φ(t) are the envelope and phase of the carrier signal respectively. The component of Xnorm(t) at fmax(t) can be estimated by methods such as Hilbert transform.

[0161] 2. Harmonic mixing: Once we have the carrier signal xc(t), we can mix it with Xnorm(t) to get the baseband signal xb(t):

[0162] xb(t) = xc(t) * Xnorm(t) (5).

[0163] Obviously, xb(t) contains the baseband information corresponding to the frequency component of fmax(t), that is, the neural coding pattern. Since the brain's information coding is often concentrated in specific frequency bands such as α and β, xb(t) contains the coding pattern that needs to be extracted.

[0164] 3. Low-pass filtering: In order to separate the coding mode from xb(t), xb(t) needs to be low-pass filtered to remove the unnecessary high-frequency components. Commonly used low-pass filters include FIR, IIR, etc. Here we choose a more efficient Butterworth low-pass filter:

[0165] y(t) = LPF(xb(t), fc) (6).

[0166] Where fc is the low-pass cutoff frequency, which is generally set to the upper limit of the α or β frequency band, such as 12Hz or 30Hz. LPF is the calculation formula of the Butterworth low-pass filter:

[0167] LPF(x(t)) = Σ(n=0...N) bn * x(tn) (7).

[0168] Where bn is the filter coefficient and N is the filter order. Through low-pass filtering, a pure coding pattern y(t) is obtained.

[0169] 4. Clustering and classification: After obtaining the coding pattern y(t), it needs to be classified and identified to decode the corresponding control instructions. Here, a clustering-based method is used to model the coding pattern using a pre-labeled training data set.

[0170] Specifically, we first construct a training set {yi(t)} (i=1...M) containing M different control commands, and extract the statistical features of each yi(t) in the time domain and frequency domain, such as mean, standard deviation, energy, frequency domain entropy, etc., to form a feature vector fi. Then, we perform unsupervised clustering on all fi to obtain K pattern clusters {Ck} (k=1...K). The Gaussian mixture model (GMM) can be selected as the clustering model. Since GMM can naturally perform soft classification on data, it is very suitable for processing noisy EEG data.

[0171] Specifically, GMM regards each data point fi as generated by a mixture of K Gaussian distributions, and its probability density function is:

[0172] p(fi) = Σ(k=1...K) πk * N(fi|μk,Σk) (8).

[0173] Where πk is the mixed weight of the kth Gaussian distribution, μk and Σk are its mean vector and covariance matrix respectively, and N(·|μk,Σk) is the probability density function of the Gaussian distribution with μk as mean and Σk as covariance. The expectation maximization (EM) algorithm can be used to perform unsupervised estimation on the parameters πk, μk, and Σk in the above formula to obtain the final GMM model.

[0174] With the GMM model, we can input the feature vector f corresponding to the new coding mode y(t) into it and calculate its posterior probability on each cluster Ck:

[0175] p(k|f) = (πk * N(f|μk,Σk)) / p(f) (9).

[0176] By dividing f into the cluster with the highest probability, we can get the corresponding control command type. According to this method, we successfully establish a mapping relationship between the neural coding mode and the control command.

[0177] 5. The last step of command execution is to execute the corresponding AI glasses control operation according to the decoded interaction command. Specifically, an intent-operation mapping table can be constructed to correspond each intent to a specific interaction function. For example, the "imagine 'left move'" intent corresponds to the "slide the displayed content to the previous page" operation; the "imagine 'right move'" intent corresponds to the "slide the displayed content to the next page" operation; the "imagine 'click'" intent corresponds to the "select the currently highlighted item" operation; the "imagine 'return'" intent corresponds to the "return to the previous interface" operation, and so on.

[0178] Once a specific interaction intention is decoded, the corresponding function instruction can be found according to the intention-operation mapping table, and the corresponding operation can be performed through the control module of the AI ​​glasses. In this way, users can achieve natural interactive control of the AI ​​glasses interface only through thoughts, which greatly improves the user experience and is particularly suitable for interactive needs in people with limited mobility or special scenarios.

[0179] In order to execute the brain-computer interaction intention recognition method based on AI glasses corresponding to the above method embodiment, so as to achieve the corresponding functions and technical effects. Figure 2 , Figure 2 The structural block diagram of a brain-computer interaction intention recognition device 200 provided in an embodiment of the present application is shown. For the convenience of explanation, only the parts related to the present embodiment are shown. The brain-computer interaction intention recognition device 200 provided in an embodiment of the present application includes:

[0180] The data acquisition module 201 is used to acquire the EEG activity data of multiple brain regions of the target user, form an EEG time series and divide it into multiple intention recognition data segments; the target user wears AI glasses;

[0181] A sequence generation module 202 is used to construct a discrete time set and a discrete frequency set according to the intention recognition data fragment, and obtain a key time-frequency feature point position sequence and an overall feature point position sequence according to the intention recognition data fragment, the discrete time set and the discrete frequency set;

[0182] The score acquisition module 203 is used to acquire the number of occurrences and the confidence score of the key time-frequency feature point position sequence in the overall feature point position sequence, and acquire the local anomaly evaluation score corresponding to the key time-frequency feature point position sequence according to the number of occurrences and the confidence score;

[0183] A set construction module 204 is used to divide the intention recognition data segment into blocks, obtain multiple data blocks, and construct a frequency conversion set according to the multiple data blocks;

[0184] The change acquisition module 205 is used to obtain the data fluctuation information corresponding to the data block, and obtain the intention complexity score according to the frequency conversion set and each data fluctuation information based on the cluster analysis method; the fluctuation information at least includes the frequency change sequence and the total change amount;

[0185] A decoding acquisition module 206, configured to determine a decoding threshold adjustment factor according to the intention complexity score and the abnormality evaluation score, and to normalize the intention recognition data segment according to the decoding threshold adjustment factor to obtain normalized decoding data;

[0186] The interaction completion module 207 is used to obtain the user intention instruction corresponding to the target user according to the normalized decoded data, and implement the interactive control function of the AI ​​glasses according to the user intention instruction.

[0187] The above-mentioned brain-computer interaction intention recognition device 200 can implement the brain-computer interaction intention recognition method based on AI glasses of the above-mentioned method embodiment. The optional items in the above-mentioned method embodiment are also applicable to this embodiment and will not be described in detail here. The rest of the contents of the embodiment of this application can refer to the contents of the above-mentioned method embodiment, and will not be repeated in this embodiment.

[0188] Figure 3 This is a schematic diagram of the structure of a computer device provided in one embodiment of the present application. Figure 3 As shown, the computer device 3 of this embodiment includes: at least one processor 30 ( Figure 3Only one is shown in the figure), a memory 31 and a computer program 32 stored in the memory 31 and executable on the at least one processor 30, wherein the processor 30 implements the steps of any of the above method embodiments when executing the computer program 32.

[0189] The computer device 3 may be a computing device such as a smart phone, a tablet computer, a desktop computer, a cloud server, etc. The computer device may include but is not limited to a processor 30 and a memory 31. Those skilled in the art will understand that Figure 3 It is only an example of computer device 3 and does not constitute a limitation on computer device 3. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components, for example, it may also include input and output devices, network access devices, etc.

[0190] The processor 30 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0191] In some embodiments, the memory 31 may be an internal storage unit of the computer device 3, such as a hard disk or memory of the computer device 3. In other embodiments, the memory 31 may also be an external storage device of the computer device 3, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the computer device 3. Further, the memory 31 may also include both an internal storage unit and an external storage device of the computer device 3. The memory 31 is used to store an operating system, an application program, a boot loader (BootLoader), data, and other programs, such as the program code of the computer program. The memory 31 may also be used to temporarily store data that has been output or is to be output.

[0192] In addition, an embodiment of the present application further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in any of the above method embodiments are implemented.

[0193] An embodiment of the present application provides a computer program product. When the computer program product is run on a computer device, the computer device implements the steps in the above-mentioned method embodiments when executing the computer device.

[0194] In several embodiments provided in the present application, it is understood that each box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, a program segment or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from that marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved.

[0195] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, and other media that can store program codes.

[0196] The specific embodiments described above further describe the purpose, technical solutions and beneficial effects of the present application in detail. It should be understood that the above description is only a specific embodiment of the present application and is not intended to limit the scope of protection of the present application. It is particularly pointed out that for those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of protection of the present application.

Claims

1. A method for brain-computer interaction intention recognition based on AI glasses, characterized in that: include: Obtain EEG activity data from multiple brain regions of the target user, form an EEG time series and divide it into multiple intent recognition data segments; The target user wears AI glasses; Construct a discrete time set and a discrete frequency set according to the intention recognition data fragments, and obtain a key time-frequency feature point position sequence and an overall feature point position sequence according to the intention recognition data fragments, the discrete time set and the discrete frequency set; Obtain the number of occurrences and confidence scores of the key time-frequency feature point position sequence in the overall feature point position sequence, and obtain the local anomaly evaluation score corresponding to the key time-frequency feature point position sequence according to the number of occurrences and the confidence score; obtain the comprehensive confidence weight through the number of occurrences and the confidence score, normalize each comprehensive confidence weight, and obtain the local anomaly evaluation score; The intention recognition data fragment is divided into blocks to obtain multiple data blocks, and a frequency conversion set is constructed based on the multiple data blocks; a frequency conversion sequence is obtained by weighted average fusion of the mean difference sequences of multiple channels; the frequency conversion sequence is normalized to obtain a frequency conversion probability sequence, a probability threshold is set, and all time positions where the frequency conversion probability sequence is greater than the probability threshold are recorded to form a position set; a subsequence is constructed with each time position as the center; and each subsequence set is formed into the frequency conversion set; Obtain data fluctuation information corresponding to the data block, and obtain an intention complexity score based on the frequency conversion set and each data fluctuation information based on a cluster analysis method; The fluctuation information includes at least a frequency change sequence and a total change amount; the total change amount is the overall frequency change amount of the intended identification data segment; Determine a decoding threshold adjustment factor according to the intention complexity score and the local anomaly evaluation score, and standardize the intention recognition data segment according to the decoding threshold adjustment factor to obtain normalized decoding data; A user intention instruction corresponding to the target user is obtained according to the normalized decoded data, and an interactive control function of the AI ​​glasses is implemented according to the user intention instruction.

2. The method according to claim 1, characterized in that: The step of constructing a discrete duration set and a discrete frequency set according to the intention recognition data segments includes: Constructing an intention recognition duration set and an intention recognition frequency set according to the intention recognition data fragment; Obtain data differences corresponding to the intention recognition duration set and the intention recognition frequency set respectively; The discrete time set and the discrete frequency set are obtained according to the data difference.

3. The method according to claim 2, characterized in that The obtaining of data differences corresponding to the intention recognition duration set and the intention recognition frequency set respectively includes: Determine, based on a kernel density analysis method, a first probability density function corresponding to the intention recognition duration set under the estimated duration; Determine the data difference corresponding to the intention recognition duration set according to the first probability density function, so as to obtain the discrete duration set according to the data difference corresponding to the intention recognition duration set; Performing principal component analysis on the intention recognition frequency set to obtain a low-dimensional projection corresponding to the intention recognition frequency set; Performing kernel density estimation on the low-dimensional projection to obtain a second probability density function; The second data difference corresponding to the intention recognition frequency set is determined according to the second probability density function, so as to obtain the frequency discrete set according to the data difference corresponding to the intention recognition duration set.

4. The method according to claim 1, characterized in that: The step of obtaining a key time-frequency feature point position sequence and an overall feature point position sequence according to the intention recognition data segment, the discrete time length set and the discrete frequency set includes: Performing time-frequency analysis on the intention recognition data segment to obtain time-frequency features corresponding to the intention recognition data segment; Taking the discrete set of durations and the discrete set of frequencies as prior information, and obtaining a time-frequency heat map according to the prior information and the time-frequency features; The key time-frequency feature point position sequence and the overall feature point position sequence are extracted from the time-frequency heat map.

5. The method according to claim 1, characterized in that The step of determining a decoding threshold adjustment factor according to the intention complexity score and the local anomaly evaluation score includes: Constructing a fuzzy logic reasoning system; the fuzzy logic reasoning system includes a fuzzification unit, a reasoning engine and a defuzzification unit; Performing fuzzification processing on the intention complexity score and the local anomaly evaluation score according to the fuzzification unit; Inputting the fuzzified intention complexity score and local anomaly evaluation score into the reasoning engine, and outputting the fuzzy reasoning result; The fuzzy inference result is input into the defuzzification unit, and the decoding threshold adjustment factor is output.

6. The method according to claim 1, characterized in that The step of normalizing the intention recognition data segment according to the decoding threshold adjustment factor includes: Adaptively normalizing the amplitude of the intention recognition data segment according to the decoding threshold adjustment factor, and softly adjusting the amplitude distribution corresponding to the amplitude; the soft adjustment is based on a controllable, smooth nonlinear transformation function, in which the original amplitude corresponding to the amplitude is used as an independent variable and the decoding threshold adjustment factor is used as a control parameter to compress or stretch the original amplitude; Adaptively normalizing the frequency distribution of the intent recognition data segment according to a decoding threshold adjustment factor; The normalized decoding data is generated according to the adaptively standardized amplitude and frequency, thereby completing the standardization of the intent recognition data segment.

7. The method according to claim 1, characterized in that The step of constructing a frequency conversion set according to the plurality of data blocks comprises: Get the mean difference corresponding to adjacent data blocks; The frequency conversion set is constructed according to the absolute value corresponding to each of the mean differences.

8. A brain-computer interaction intention recognition device, characterized in that: include: The data acquisition module is used to acquire the EEG activity data of multiple brain regions of the target user, form an EEG time series and divide it into multiple intent recognition data segments; The target user wears AI glasses; A sequence generation module is used to construct a discrete time set and a discrete frequency set according to the intention recognition data fragment, and obtain a key time-frequency feature point position sequence and an overall feature point position sequence according to the intention recognition data fragment, the discrete time set and the discrete frequency set; A score acquisition module is used to obtain the number of occurrences and confidence scores of the key time-frequency feature point position sequence in the overall feature point position sequence, and obtain the local anomaly evaluation score corresponding to the key time-frequency feature point position sequence according to the number of occurrences and the confidence score; obtain the comprehensive confidence weight through the number of occurrences and the confidence score, normalize each comprehensive confidence weight, and obtain the local anomaly evaluation score; A set construction module is used to divide the intention recognition data fragment into blocks, obtain multiple data blocks, and construct a frequency conversion set based on the multiple data blocks; obtain a frequency conversion sequence by weighted average fusion of the mean difference sequences of multiple channels; normalize the frequency conversion sequence to obtain a frequency conversion probability sequence, set a probability threshold, and record all time positions where the frequency conversion probability sequence is greater than the probability threshold to form a position set; construct a subsequence with each time position as the center; and form each subsequence set into the frequency conversion set; A change acquisition module, used to acquire data fluctuation information corresponding to the data block, and acquire an intention complexity score according to the frequency conversion set and each data fluctuation information based on a cluster analysis method; The fluctuation information includes at least a frequency change sequence and a total change amount; the total change amount is the overall frequency change amount of the intended identification data segment; A decoding acquisition module, used to determine a decoding threshold adjustment factor according to the intention complexity score and the local anomaly evaluation score, and to normalize the intention recognition data segment according to the decoding threshold adjustment factor to obtain normalized decoding data; The interaction completion module is used to obtain the user intention instruction corresponding to the target user according to the normalized decoded data, and implement the interactive control function of the AI ​​glasses according to the user intention instruction.

9. A computer device, characterized in that: The method comprises a memory and a processor; the memory is used to store a computer program; the processor is used to execute the computer program and implement the method according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a processor, causes the processor to implement the method according to any one of claims 1 to 7.

Images