Electroencephalogram signal real-time decoding and intention recognition method and equipment for attention glasses

Abstract

The invention discloses an electroencephalogram signal real-time decoding and intention recognition method and device for attention glasses. The method comprises the steps that a standardized signal sequence corresponding to an initial signal sequence collected by the attention glasses is obtained; obtaining a frequency domain component corresponding to the standardized signal sequence and an energy analysis result corresponding to the frequency domain component, constructing a double-domain feature sequence for association mapping, obtaining a feature breakpoint, establishing a time-frequency relation graph according to the feature breakpoint, determining a mapping interval according to the time-frequency relation graph, and enhancing the feature sequence according to the mapping interval to obtain an enhanced signal sequence. Acquiring an enhanced feature sequence; performing mapping conversion on the enhanced feature sequence to obtain initial components, determining main components in the initial components, and generating feature representation according to the main components; obtaining a feature interval and an interval feature corresponding to the feature interval, and generating an intention feature sequence; obtaining change nodes and time sequence features corresponding to the intention feature sequence, and generating an intention change sequence; and obtaining an intention recognition result corresponding to the intention change sequence.

Description

Technical Field

[0001] The present application relates to the field of brain-computer interface technology, and more particularly to a method and device for real-time decoding of electroencephalogram signals and intention recognition using attention glasses. Background Art

[0002] As a portable EEG acquisition and analysis device, the core function of attention glasses is to decode the user's EEG signals in real time and identify their cognitive intentions. The traditional EEG signal decoding methods have the following main problems: first, the signal acquisition and preprocessing stages use fixed sampling parameters and filtering strategies, which are difficult to cope with the dynamic changes in signal quality; second, the feature extraction process fails to effectively combine the complementary information of time domain and frequency domain features, resulting in incomplete feature expression; third, the feature mapping and enhancement stages lack an adaptive mechanism and cannot be dynamically adjusted according to feature importance; finally, the intention recognition process ignores the temporal correlation of feature sequences, affecting the accuracy of state transition judgment.

[0003] In practical applications, the factors that affect the decoding effect are mainly reflected in three aspects: first, interference at the signal level, including baseline drift, channel noise, etc., which affect the stability of feature extraction; second, changes at the feature level, users show diverse feature patterns and complex temporal dependencies under different cognitive tasks; third, differences at the intention level, users' attention state and intention expression are affected by factors such as emotions and fatigue, showing significant individual differences and time-varying characteristics. Existing technologies often use independent processing modules, lack of coordinated optimization between various links, and it is difficult to achieve a systematic solution to the above challenges.

[0004] In addition, the processing power of portable devices is limited, and existing methods generally use complex offline analysis algorithms, which are computationally intensive and have high latency, making it difficult to meet the needs of real-time interaction. At the same time, the algorithm is not adaptable enough and cannot dynamically adjust the processing strategy according to device resources.

[0005] Therefore, a method is urgently needed to solve at least one of the above problems. Summary of the invention

[0006] The embodiment of the present application provides a method and device for real-time decoding of EEG signals and recognition of intentions of attention glasses, aiming to solve the problem that the existing methods generally use complex offline analysis algorithms, which are computationally intensive and have high latency, and are difficult to meet the needs of real-time interaction. At the same time, the algorithm is not adaptable enough and cannot dynamically adjust the processing strategy according to the device resources.

[0007] In a first aspect, an embodiment of the present application provides a method for real-time decoding of EEG signals and recognition of intentions using attention glasses, comprising: Acquire a multi-channel EEG time domain signal collected by the attention glasses, perform multi-point sampling on the multi-channel EEG time domain signal to obtain original sampling data, and generate an initial signal sequence according to the original sampling data; Acquire a signal segment that needs to be corrected in the initial signal sequence, perform compensation adjustment on the signal segment, and generate a standardized signal sequence; acquire a time domain feature corresponding to the standardized signal sequence, perform frequency domain conversion on the time domain feature, acquire a frequency domain component and an energy analysis result corresponding to the frequency domain component, and construct a dual-domain feature sequence according to the energy analysis result; Performing association mapping on the dual-domain feature sequence to obtain feature breakpoints, establishing a time-frequency relationship diagram according to the feature breakpoints, determining a mapping interval according to the time-frequency relationship diagram, and enhancing the feature sequence according to the mapping interval to obtain an enhanced feature sequence; Performing mapping conversion on the enhanced feature sequence to obtain initial components, determining main components in the initial components, and generating feature representations according to the main components; segmenting the feature representations to obtain feature intervals and interval features corresponding to the feature intervals, and generating an intention feature sequence according to the interval features; Acquire the change nodes and the corresponding time series features corresponding to the intention feature sequence, and generate the intention change sequence according to a plurality of the time series features; Obtain a preliminary mark corresponding to the intention change sequence, and obtain an intention recognition result corresponding to the preliminary mark.

[0008] In a second aspect, the present application also provides a device for real-time decoding of EEG signals and recognition of intentions, comprising: A signal acquisition module, used to acquire multi-channel EEG time domain signals collected by the attention glasses, perform multi-point sampling on the multi-channel EEG time domain signals, acquire original sampling data, and generate an initial signal sequence according to the original sampling data; A compensation adjustment module is used to obtain a signal segment that needs to be corrected in the initial signal sequence, perform compensation adjustment on the signal segment, and generate a standardized signal sequence; obtain a time domain feature corresponding to the standardized signal sequence, perform frequency domain conversion on the time domain feature, obtain a frequency domain component and an energy analysis result corresponding to the frequency domain component, and construct a dual-domain feature sequence according to the energy analysis result; An association mapping module is used to perform association mapping on the dual-domain feature sequence, obtain feature breakpoints, establish a time-frequency relationship diagram according to the feature breakpoints, determine a mapping interval according to the time-frequency relationship diagram, and enhance the feature sequence according to the mapping interval to obtain an enhanced feature sequence; A mapping conversion module is used to perform mapping conversion on the enhanced feature sequence, obtain initial components, determine main components in the initial components, and generate feature representations according to the main components; segment the feature representations, obtain feature intervals and interval features corresponding to the feature intervals, and generate an intention feature sequence according to the interval features; A feature acquisition module, used to acquire the change nodes and corresponding time series features corresponding to the intention feature sequence, and generate an intention change sequence according to a plurality of the time series features; The result acquisition module is used to obtain the preliminary mark corresponding to the intention change sequence, and obtain the intention recognition result corresponding to the preliminary mark.

[0009] In a third aspect, the present application also provides a computer device, including 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 real-time decoding of EEG signals and the intention recognition method of the attention glasses as described in the first aspect are implemented.

[0010] 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 method for real-time decoding of EEG signals and intention recognition of the attention glasses as described in the first aspect.

[0011] Compared with the prior art, this application has at least the following beneficial effects: 1. The robustness of decoding is improved through multi-level signal processing strategies. Adaptive sampling and dynamic baseline correction are used in the signal acquisition stage to effectively suppress signal interference; in the feature extraction stage, time domain and frequency domain features are combined to enhance the integrity of feature expression; in the feature enhancement stage, adaptive weight adjustment is used to improve the feature discrimination ability. This multi-level processing method significantly improves the anti-interference ability of the system corresponding to the method (collectively referred to as the system in the manual) in complex environments, making the decoding results more reliable.

[0012] 2. Improved recognition accuracy through dynamic mapping of feature sequences and state optimization. The system performs multi-scale analysis on feature sequences to accurately capture key nodes of state changes; effectively eliminates the instability of state judgment through state merging and optimization; and uses a hierarchical intent recognition strategy to accurately distinguish different types of intent. This sequence processing-based method greatly improves the system's recognition accuracy of user intent changes.

[0013] 3. Efficient data processing is achieved through pipeline processing architecture. The system adopts segmented processing and parallel computing strategies, and uses adaptive computing strategies in each processing link to dynamically adjust the processing depth according to the importance of features; through feature screening and dimension optimization, data redundancy is reduced; using streaming processing architecture, efficient data transmission and processing are achieved. This optimized design improves the system processing efficiency, ensures the processing effect and meets the actual application needs.

[0014] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A flowchart of a method for real-time decoding of EEG signals and recognition of intentions using attention glasses according to an embodiment of the present application; Figure 2 This is a schematic diagram of the structure of a device for real-time decoding of EEG signals and recognition of intentions shown in an embodiment of the present application; Figure 3 A schematic diagram of the structure of a computer device shown in an embodiment of the present application. DETAILED DESCRIPTION

[0016] 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.

[0017] 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.

[0018] 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.

[0019] 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.

[0020] 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.

[0021] 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.

[0022] The technical solution of the embodiment of the present application is introduced below.

[0023] As a portable EEG acquisition and analysis device, the core function of attention glasses is to decode the user's EEG signals in real time and identify their cognitive intentions. The traditional EEG signal decoding methods have the following main problems: first, the signal acquisition and preprocessing stages use fixed sampling parameters and filtering strategies, which are difficult to cope with the dynamic changes in signal quality; second, the feature extraction process fails to effectively combine the complementary information of time domain and frequency domain features, resulting in incomplete feature expression; third, the feature mapping and enhancement stages lack an adaptive mechanism and cannot be dynamically adjusted according to feature importance; finally, the intention recognition process ignores the temporal correlation of feature sequences, affecting the accuracy of state transition judgment.

[0024] In practical applications, the factors that affect the decoding effect are mainly reflected in three aspects: first, interference at the signal level, including baseline drift, channel noise, etc., which affect the stability of feature extraction; second, changes at the feature level, users show diverse feature patterns and complex temporal dependencies under different cognitive tasks; third, differences at the intention level, users' attention state and intention expression are affected by factors such as emotions and fatigue, showing significant individual differences and time-varying characteristics. Existing technologies often use independent processing modules, lack of coordinated optimization between various links, and it is difficult to achieve a systematic solution to the above challenges.

[0025] In addition, the processing power of portable devices is limited. Existing methods generally use complex offline analysis algorithms, which are computationally intensive and have high latency, making it difficult to meet the needs of real-time interaction. At the same time, the algorithm is not adaptable enough and cannot dynamically adjust the processing strategy according to device resources. Therefore, it is necessary to develop an efficient decoding framework that can achieve reasonable utilization of computing resources while ensuring recognition accuracy through feature optimization and sequence processing.

[0026] Please refer to Figure 1 , Figure 1 The flowchart of a method for real-time decoding of EEG signals and recognition of intentions of an attention glasses provided in an embodiment of the present application is shown in FIG. The method for real-time decoding of EEG signals and recognition of intentions of an attention glasses provided 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 method for real-time decoding of EEG signals and recognition of intentions by the attention glasses of this embodiment includes steps S101 to S106, which are described in detail as follows: Step S101, obtaining a multi-channel EEG time domain signal collected by the attention glasses, performing multi-point sampling on the multi-channel EEG time domain signal, obtaining original sampling data, and generating an initial signal sequence according to the original sampling data.

[0027] Specifically, the realization process of acquiring multi-channel EEG time domain signals of attention glasses is first based on a multi-channel signal acquisition system, which includes a signal acquisition unit, a data conversion unit and a data transmission unit. The signal acquisition unit uses 4-8 dry electrodes to form an acquisition array. The electrodes are made of Ag / AgCl material with a diameter of 10mm. The electrode surface is specially treated to improve the conductivity. The electrode surface is coated with nano-scale conductive material with a material thickness of 0.5-1.0μm and a surface resistivity of less than 0.1Ω / cm². The electrode layout follows the international 10-20 system and is mainly distributed in cognitive activity-related areas such as the frontal lobe (Fp1, Fp2 position) and temporal lobe (T3, T4 position). The electrode spacing is maintained within the range of 3-5cm, and a reference electrode is placed at the mastoid position behind the left ear. Each electrode is signal conditioned by an independent preamplifier. The amplifier uses a low-noise operational amplifier chip with an adjustable gain range of 1000-10000 times, usually set to 5000 times. The input impedance of the preamplifier is greater than 100MΩ, the common mode rejection ratio is greater than 100dB, and the input reference noise is less than 1μV (pp). The data conversion unit uses a 24-bit high-precision ADC with a quantization accuracy of 0.1μV, a maximum sampling rate of 2000Hz, and a standard working sampling rate of 500Hz. The signal-to-noise ratio of the ADC is greater than 100dB, the integral nonlinearity error is less than 0.001%, and the differential nonlinearity error is less than 0.0005%. Before analog-to-digital conversion, the signal is processed by an analog bandpass filter, using a fourth-order Butterworth filter with a passband range of 0.1-100Hz, a stopband attenuation of greater than 60dB, and a passband ripple of less than 0.5dB. The data transmission unit uses a high-speed serial interface with a transmission rate of not less than 1Mbps. It uses a differential signal transmission method to support real-time data stream transmission and status monitoring. The transmission protocol includes a data packet header (synchronization word, packet sequence number, timestamp), a data payload (multi-channel sampling data), and a check field (CRC32 check code).

[0028] In some embodiments, generating the initial signal sequence according to the original sampling data includes: obtaining the sampling point distribution corresponding to the multi-point sampling; marking the valid sampling segments in the original sampling data according to the sampling point distribution; and generating the initial signal sequence according to the valid sampling segments.

[0029] After completing the basic configuration of the signal acquisition system, the acquired time domain signal is sampled at multiple points. This process samples multi-channel signals based on the principle of time division multiplexing, and each sampling cycle includes the data acquisition and conversion process of all channels. Taking a sampling rate of 500Hz as an example, the system completes full-channel sampling every 2ms, and the sampling time interval of each channel does not exceed 50μs. The system uses a high-precision crystal oscillator to provide a reference clock with a clock frequency of 20MHz and a frequency stability of better than ±20ppm. The sampling process adopts a synchronous sampling mechanism, and all channels complete data acquisition under the control of the same clock signal. The sampling timing deviation between channels is less than 1μs, ensuring the time consistency of multi-channel data. Each sampling point record contains a timestamp (64-bit accuracy, resolution of 1μs), a channel number (8 bits), and a voltage value (24 bits). The sampled data undergoes preliminary digital filtering processing, using a 100-order FIR filter. The filter coefficients are designed using the Hamming window function method. The passband cutoff frequency matches the analog filter, and the transition band width is 2Hz. The filtered data is organized into data frames in chronological order, and each frame contains 50ms of data, that is, 25 sampling points. The data frame structure includes a frame header (32 bytes, including synchronization words, frame number, and timestamp), a data area (variable according to the number of channels, each sampling point occupies 4 bytes) and a frame tail (4 bytes, including CRC32 checksum). The system uses a double buffer mechanism to manage data frames. The main buffer maintains a data length of 10 seconds, adopts a circular queue structure, and the buffer size is dynamically allocated according to the number of channels. The secondary buffer is used for data preprocessing, and its size is 1 / 5 of the main buffer. The marking of valid sampling points is based on signal amplitude and spectrum characteristics. The system calculates the voltage difference between adjacent sampling points (the threshold is set to 50μV) and the local frequency distribution (mainly focusing on the energy distribution of the 0.1-30Hz frequency band), and determines the validity of the sampling points in combination with the preset judgment rules.

[0030] After obtaining the valid sampling point mark, the system generates the initial signal sequence. This process first normalizes the amplitude of the valid sampling data, and uses the piecewise linear mapping method to unify the signal amplitudes of different channels to the [-1,1] interval. The normalization process is divided into three steps: first, the signal amplitude range of each channel is calculated to determine the slope and offset of the mapping function; then, the abnormal values ​​beyond the normal range are truncated, and the truncation threshold is set to the mean ± 3 times the standard deviation; finally, the linear mapping conversion is performed. The normalized data constitutes a standardized time domain sequence, and each data point contains a normalized voltage value, an original timestamp, and a channel identifier. For the data gaps between adjacent valid sampling segments, different interpolation methods are used according to the gap length: cubic spline interpolation is used for gaps less than 100ms, and the interpolation node interval is 1 / 4 of the original sampling interval; linear interpolation is used for gaps in the range of 100-300ms, and the number of interpolation points is equal to the original sampling rate; gaps exceeding 300ms are marked as signal breakpoints, and marker fields are added at the breakpoints. The interpolation process is performed channel by channel to ensure the continuity of the signal, and an interpolation marker is added in the interpolation interval. The final initial signal sequence is stored in a structured manner. The data structure includes: main data area (multi-channel time domain data, indexed by channel number), time information area (including original timestamp and interpolation point timestamp) and marker information area (valid segment marker, breakpoint marker, interpolation marker). The sequence is stored in segments of 1 second, with adjacent segments maintaining a 50% overlap rate. Each segment of data is attached with segment number and time range information. The signal sequence is stored in a compressed format, using a lossless compression algorithm, with a compression ratio of not less than 2:1, while ensuring that the decompression speed meets the real-time processing requirements.

[0031] Step S102, obtain the signal segment that needs to be corrected in the initial signal sequence, compensate and adjust the signal segment, and generate a standardized signal sequence; obtain the time domain features corresponding to the standardized signal sequence, perform frequency domain conversion on the time domain features, obtain the frequency domain components and the energy analysis results corresponding to the frequency domain components, and construct a dual-domain feature sequence according to the energy analysis results.

[0032] Specifically, the process of performing baseline drift detection on the initial signal sequence obtained in step S101 first adopts a multi-scale analysis method. The system divides the initial signal sequence into segments of 2 seconds in length, and the overlap rate between adjacent segments is 50%. The local mean curve is calculated for each signal segment, using a sliding window of 100ms and a window sliding step of 20ms. The local mean curve reflects the slow-changing trend of the signal, which contains baseline drift information. The system performs wavelet decomposition on the local mean curve, using the db4 wavelet basis function, and the number of decomposition layers is 4. In each layer of decomposition results, low-frequency coefficients are extracted as baseline drift features. A threshold is set for the extracted feature sequence, and the threshold value is adaptively determined according to the statistical characteristics of the signal segment, generally taking 2 times the local standard deviation. When the feature value exceeds the threshold, the corresponding time point is marked as the starting point of the baseline drift. The system simultaneously calculates the slope and duration of the baseline drift and establishes a drift feature vector. The drift detection result is represented by a binary sequence, 1 indicates the presence of drift, and 0 indicates no drift. The detection process adopts a parallel processing architecture to support multi-channel simultaneous analysis. The system passes the detection results together with the original signal and the drift feature vector to the next processing stage.

[0033] In some embodiments, the method of obtaining a signal segment that needs to be corrected in the initial signal sequence, performing compensation adjustment on the signal segment, and generating a standardized signal sequence includes: performing baseline drift detection on the initial signal sequence to obtain the signal segment that needs to be corrected; performing the compensation adjustment on the signal segment that needs to be corrected; and generating a standardized signal sequence according to an adjustment result corresponding to the compensation adjustment.

[0034] Based on the detection results, the system marks the signal segments that need to be corrected. The marking process first obtains the binary sequence and drift feature vector of drift detection from the previous stage to determine the boundaries of each drift interval. It extends forward 100ms from the drift starting point and extends backward until the drift characteristics return to normal. For each marked interval, the main characteristic parameters of drift are calculated: drift amplitude (deviation relative to the local mean), drift rate (amount of change per unit time) and drift morphology (linear / nonlinear). The system classifies the drift interval according to these characteristic parameters into two categories: fast drift (duration less than 500ms) and slow drift (duration greater than 500ms). For fast drift, the system adds transition marks at the boundary of the drift interval for subsequent smoothing. For slow drift, the system further subdivides the drift interval into multiple subintervals, each with a length of 200ms and adjacent subintervals overlapping by 50ms. When marking the interval, the system also records the statistical characteristics of the interval, including mean, standard deviation, maximum and minimum values, etc. This information is passed to the next stage as a reference for compensation adjustment. The tag information is stored in a compressed format and includes timestamp, interval type, characteristic parameters, etc.

[0035] The system makes compensation adjustments to the marked signal segments that need to be corrected. According to the interval type and characteristic parameters transmitted in the previous stage, the system selects different compensation strategies: for fast drift, an adaptive filtering method is adopted, using a 50-order FIR filter, the filter coefficients are updated in real time through the LMS algorithm, and the step size parameters are dynamically adjusted according to the drift amplitude. For slow drift, a piecewise polynomial fitting method is adopted, and a third-order polynomial is used for baseline estimation for each sub-interval, and the polynomial coefficients are solved by the least squares method. A weighted average method is used at the interval boundary to achieve a smooth transition. The compensation process adopts an iterative method, and the residual drift is calculated after each iteration. The iteration is stopped when the residual is less than the preset threshold (usually 5% of the original drift amplitude). The compensation adjustment also takes into account the frequency characteristics of the signal to ensure that the compensation process does not affect the effective components of the signal. The system evaluates the quality of the compensated signal and calculates the stability index and signal-to-noise ratio change of the signal. The compensation results are saved in a temporary buffer, including the original signal, compensation value, corrected signal and various evaluation indicators, providing input for the next stage of standardization processing.

[0036] After the compensation adjustment is completed, the system generates a standardized signal sequence. The standardization process first reads the corrected signal data from the temporary buffer, including the compensated time domain signal and related evaluation indicators. The amplitude of the read corrected signal is normalized using a piecewise linear mapping method with a mapping interval of [-1,1]. The impact of the compensation adjustment is considered during the normalization process, and weighted mapping is used for areas with large compensation amplitudes. The signal sequence is segmented, with a segment length of 1 second and an overlap rate of 50% between adjacent segments. Each signal segment contains a complete data structure: signal data (corrected time domain data), time information (original timestamp and processing timestamp), feature markers (drift features, compensation parameters) and quality indicators (signal-to-noise ratio, stability score). The standardized sequence adopts a hierarchical storage structure, including a data layer (signal data), a feature layer (processing parameters and results) and a control layer (processing process information). The system performs an integrity check on the standardized sequence to ensure data continuity and feature consistency. After the check is passed, the standardized sequence is transmitted to the next processing link. The sequence transmission adopts a block transmission method, with each block of data being 2 seconds in size, and the transmission process includes a data verification mechanism.

[0037] In some embodiments, the method of acquiring the time domain features corresponding to the standardized signal sequence, performing frequency domain conversion on the time domain features, acquiring frequency domain components and energy analysis results corresponding to the frequency domain components, and constructing a dual-domain feature sequence according to the energy analysis results, includes: segmenting the standardized signal sequence to acquire multiple signal segments; acquiring the time domain features corresponding to the signal segments; performing frequency domain conversion on the time domain features to acquire the frequency domain components; acquiring energy analysis results corresponding to the frequency domain components; and constructing the dual-domain feature sequence according to the time domain features, frequency domain components, and energy analysis results.

[0038] Exemplarily, segmenting the standardized signal sequence to obtain multiple signal segments includes: segmenting the standardized signal according to a preset initial segmentation length to obtain multiple initial segments; the overlapping rates corresponding to adjacent initial segments are within a preset overlapping rate range; iteratively optimizing the initial segmentation length according to a preset segmentation range, and segmenting the standardized signal sequence according to the optimized initial segmentation length to obtain multiple signal segments.

[0039] An adaptive segmentation strategy is used to segment the standardized signal sequence. The system initially segments the signal sequence according to the basic length of 2 seconds, and maintains a 50% overlap rate between adjacent segments to ensure the continuity of signal features. The signal statistical characteristics are calculated for each basic segment, including the mean, variance, peak-to-peak value and zero-crossing rate, which reflect the baseline level, fluctuation intensity, amplitude range and frequency change of the signal respectively. An adaptive threshold is set based on the statistical characteristics for boundary judgment. When the feature difference between adjacent sub-segments exceeds the threshold, it is marked as a paragraph boundary. The segmentation process adopts an iterative optimization method. The initial segment length is 2 seconds. Each iteration can adjust the segment length to 1 / 2 or 2 times the original length according to the feature difference, and the adjustment range is limited to 1-4 seconds. The system introduces an adjustment limit mechanism to prevent excessive segmentation. The segmentation results include the start time, segment length and segment feature description of each segment, which are stored in a structured manner. After segmentation, the time domain features are extracted for each signal segment: amplitude features include mean, variance, skewness, kurtosis, maximum value, and minimum value; morphological features include zero-crossing rate, number of peaks, number of troughs, and waveform factor; statistical features include autocorrelation coefficient, mutual correlation coefficient, and entropy value. The feature extraction adopts the sliding window method with a window length of 200ms and a step size of 50ms to ensure the temporal continuity of the features.

[0040] Based on the time domain feature sequence extracted in the previous stage, the system performs frequency domain conversion. First, the time domain feature sequence (including statistical features such as mean, variance, skewness, and kurtosis) is Fourier transformed to obtain the frequency distribution characteristics of the features. At the same time, the original signal segment is subjected to fast Fourier transform to obtain basic spectral features. Before the transformation, Hanning window processing is used to reduce spectral leakage and obtain the amplitude spectrum and phase spectrum of the signal. Then the signal segment is subjected to wavelet transform, and the db4 wavelet basis function is selected for 5-layer decomposition, and the decomposition depth covers the key frequency band. In each layer of decomposition, the system correlates the change trend of the time domain feature sequence with the distribution of the wavelet coefficients, and calculates and saves the approximate coefficients and detail coefficients. At the same time, the Hilbert transform is applied to obtain the instantaneous frequency and instantaneous amplitude characteristics of the signal, and establishes a corresponding relationship with the extracted time domain morphological features. The system organizes multiple transformation results into a frequency domain feature matrix, where the rows of the matrix correspond to time points and the columns correspond to different frequency domain features. All calculation processes are represented by high-precision numerical values, and the features obtained by each transformation method are timestamped to ensure synchronization with the time domain feature sequence.

[0041] After obtaining the frequency domain transformation results, the system performs energy analysis. Energy statistics are performed in standard EEG frequency bands, including the δ band, theta band, alpha band, beta band, and gamma band. The absolute energy and relative energy of each frequency band are calculated, and the energy of different frequency bands is made comparable through normalization. The energy ratios between frequency bands are analyzed, including θ / α, β / α, (θ+α) / β, etc. These ratios reflect the relative strength of different EEG rhythms. For the wavelet transform results, the energy distribution of each scale coefficient is calculated respectively, and an energy-scale diagram is constructed to show the distribution of energy at different decomposition scales. The system also calculates the time-varying characteristics of energy, including the short-term energy change rate and the energy fluctuation range. All energy features are organized into feature vectors as a complete description of the signal energy characteristics. The energy analysis results also have a time index to facilitate tracking the time process of energy changes.

[0042] Based on the analysis results of time domain features, frequency domain features and energy features, the system constructs a dual-domain feature sequence. First, each type of feature is standardized to make different types of features comparable. Feature fusion adopts a weighted combination method, and the weight coefficient is determined by analyzing the significance of the feature. For each signal segment, a feature descriptor is generated, which includes a time domain feature subset, a frequency domain feature subset and an energy feature subset. Adjacent feature descriptors establish a corresponding relationship through timestamps to ensure the continuity of the feature sequence. The system removes redundant features through feature analysis and reduces the dimension of the feature space. The final feature data is organized into a multi-layer structure: the basic feature matrix stores the time domain and frequency domain features of all channels, the frequency band energy matrix records the energy distribution of each frequency band, and the channel correlation matrix describes the feature relationship between channels. The feature sequence is stored in time segments, and each segment contains a complete feature description and time range information.

[0043] In this way, the system completes the feature extraction and representation process of the standardized signal sequence. The dual-domain feature sequence formed preserves the complete feature information of the signal in the time domain and frequency domain, including descriptions of multiple dimensions such as amplitude, morphology, frequency and energy. The feature data is stored in a hierarchical structure, and the correspondence with the original signal is guaranteed through time indexing, and a data basis is provided for subsequent feature association mapping processing. Each component of the feature sequence can be accessed independently, which facilitates the extraction and use of specific types of feature information in subsequent processing.

[0044] Step S103, performing association mapping on the dual-domain feature sequence, obtaining feature breakpoints, establishing a time-frequency relationship diagram according to the feature breakpoints, determining a mapping interval according to the time-frequency relationship diagram, and enhancing the feature sequence according to the mapping interval to obtain an enhanced feature sequence.

[0045] Specifically, the correlation mapping of the dual-domain feature sequence is first established to establish the feature correspondence. The system divides the dual-domain feature sequence into basic processing units according to the length of 2 seconds, and the adjacent units overlap by 1 second to ensure the continuity of the feature mapping. For each processing unit, the time domain feature matrix and the frequency domain feature matrix are extracted. The matrix rows represent the time sampling points, and the columns represent different feature types. The time domain features in the processing unit include waveform features (peak value, valley value, zero crossing point, waveform slope, amplitude mean), statistical features (mean, standard deviation, kurtosis, skewness, variance, interquartile range) and morphological features (waveform factor, pulse factor, margin factor, waveform steepness), and the frequency domain features include spectral features (dominant frequency, frequency center, bandwidth, frequency concentration, spectral kurtosis), energy features (band energy, energy ratio, power spectral density) and phase features (phase spectrum, group delay, phase coherence). The system establishes a feature mapping table with a multi-level index structure. The first-level index distinguishes time domain and frequency domain features, the second-level index is grouped by feature category, the third-level index points to specific feature attributes, and the bottom-level index links feature values ​​and timestamps. The strength of the association between features is calculated by the correlation coefficient, and the three-level association thresholds are set: strong association (above 0.8), medium association (0.5-0.8), weak association (below 0.5), and the association type, strength value and time duration are recorded.

[0046] Based on the results of feature association analysis, the system marks the breakpoints in the feature sequence. The marking process first uses the established feature mapping table and takes the association strength as an important basis for breakpoint judgment. Feature breakpoint detection uses an adaptive sliding window method. The basic window length is set to 500ms. The window length can be dynamically adjusted according to the signal change rate, ranging from 300-700ms, and the step length is 1 / 5 of the window length. In each sliding window, the rate of change of feature association (based on the calculated association coefficient), the fluctuation amplitude of the feature value, and the discrete degree of feature distribution are analyzed. When these indicators exceed their respective thresholds at the same time and last for more than 3 sampling points, the center point of the time period is marked as a candidate breakpoint. The system uses a hierarchical clustering method to process candidate breakpoints. The clustering feature vector contains the time position of the breakpoint, the feature change amplitude, and the similarity of the previous and next features. The final position of each breakpoint is determined by the centroid method, and the significance score is calculated.

[0047] According to the distribution of the marked feature breakpoints and the feature mapping results, the system establishes a time-frequency relationship diagram. The relationship diagram adopts a four-layer matrix structure, including the time layer, feature layer, association layer and control layer. The time layer records the timestamp and breakpoint position of the sampling point, the feature layer stores the values ​​and change trends of various features, the association layer integrates the feature mapping relationship and association strength, and the control layer manages the processing parameters and status tags. For each feature breakpoint, the system analyzes the feature change trend of the 50 sampling points before and after it, and fits the change curve in combination with the associated features. In the relationship diagram, the weight of the feature connection is set based on the association strength, the dominant feature is marked with solid lines of different thicknesses, and the secondary feature is marked with dotted lines. The system analyzes the feature propagation characteristics between breakpoints and establishes a propagation parameter model.

[0048] Using the established time-frequency relationship graph, the system determines the feature mapping interval. The mapping interval division first identifies the stable area in the relationship graph, and the judgment basis includes the coefficient of variation of the feature association strength, the stability index of the feature value, and the consistency measure of the feature distribution. The system sets the stability threshold based on the associated features. When these indicators meet the threshold requirements simultaneously in a continuous time, the time period is marked as a candidate interval. The system optimizes the boundaries of the candidate intervals, and the location of the boundary points is determined by feature gradient calculation. Each mapping interval generates a feature descriptor, which includes the time range, feature set, association map, and stability score.

[0049] Based on the determined mapping interval, the system performs three-level feature enhancement processing based on the associated features. Feature-level enhancement processes a single feature, including median filtering to remove sudden noise, trend item extraction to enhance long-term changes, and outlier detection and correction. Association-level enhancement optimizes feature relationships based on association strength, including nonlinear mapping of association strength, selective enhancement of association patterns, and smoothing of time-varying associations. Sequence-level enhancement optimizes feature sequences globally to achieve temporal consistency optimization and spatial consistency enhancement of features. After three levels of feature enhancement, the enhanced feature sequence is finally obtained.

[0050] In some embodiments, enhancing the feature sequence according to the mapping interval to obtain an enhanced feature sequence includes: enhancing the sequence according to the mapping interval; dividing the enhanced feature sequence into paragraphs to obtain signal segments; calculating feature intensity values ​​corresponding to the signal segments; marking key feature segments according to the feature intensity values; and generating the enhanced feature sequence according to the key feature segments.

[0051] A multi-level segmentation strategy is used to segment the feature sequence after step enhancement. The system first segments the feature sequence according to the basic length of 1 second, and maintains a 30% overlap rate between adjacent segments to ensure the continuity of the features. The feature distribution indicators are calculated for each basic segment, including time domain feature distribution (mean, variance, skewness, kurtosis, waveform factor, pulse factor), frequency domain feature distribution (energy distribution, frequency center, bandwidth, spectrum peak distribution, spectrum entropy) and association feature distribution (feature correlation, association strength, synchronization, coupling degree). The calculation of distribution indicators adopts the sliding window method with a window length of 200ms and a step length of 50ms to ensure the capture of dynamic changes in features. The feature distribution in each window is obtained by the kernel density estimation method, the kernel function selects the Gaussian kernel, and the bandwidth is determined by cross-validation. The system uses these distribution indicators to segment the feature sequence through an improved hierarchical clustering method, and the cluster feature vector contains all distribution indicators and their first-order differences. The cluster distance adopts a weighted combination of the Euclidean distance and the Mahalanobis distance of the feature distribution, with a weight ratio of 7:3. The system performs post-processing on the clustering results, merging short paragraphs (less than 100ms) and segmenting long paragraphs (greater than 3 seconds). The resulting signal segments contain paragraph boundary timestamps, paragraph feature descriptions, paragraph type tags, and paragraph quality scores.

[0052] Based on the obtained signal fragments, the system calculates the characteristic strength index of each fragment. According to the segmentation results and characteristic distribution, the system designs three types of characteristic strength calculation schemes. The time domain strength index uses the time domain characteristic distribution to calculate the waveform amplitude strength (peak-to-peak value, root mean square value, waveform tilt, waveform steepness), morphological strength (curvature, sharpness, flatness, irregularity) and statistical strength (variance contribution rate, entropy value, complexity, prediction error). The frequency domain strength index is based on the frequency domain characteristic distribution, including band energy intensity (energy proportion of each frequency band and its time-varying characteristics), spectrum peak intensity (main frequency amplitude, number of spectrum peaks, spectrum peak spacing, spectrum peak shape) and modulation intensity (sideband energy ratio, modulation depth, modulation frequency, modulation index). The correlation strength index combines the correlation characteristic distribution to calculate the characteristic correlation strength, synchronization strength and coupling strength. The strength calculation process adopts the adaptive window method, the basic window length is 500ms, and the window length can be adjusted in the range of 300-700ms according to the signal change rate. For each strength index, the system calculates a set of statistical features and performs correlation analysis with the distribution index to ensure the accuracy of the strength calculation.

[0053] Using the calculated feature intensity values, the system performs feature tagging. The tagging process first combines all the intensity indicators of each fragment into a high-dimensional feature vector, which contains the calculation results of three types of indicators: time domain intensity, frequency domain intensity, and correlation intensity. The principal component analysis method with regularization constraints is used for dimensionality reduction to ensure that more than 90% of the information is retained after dimensionality reduction. In the reduced feature space, the improved DBSCAN density clustering algorithm is used to classify the fragments. The system calculates the significance score of the pattern based on three dimensions: intensity level (using the absolute value of intensity and its stability), feature stability (based on the calculated smoothness of feature changes), and pattern uniqueness (Mahalanobis distance with other patterns). The significance score is weighted summed, and the weight coefficient is determined by grid search optimization. The system marks the fragments with significance scores exceeding the threshold as key feature segments, and further identifies their dominant and secondary features.

[0054] Based on the feature labeling results, the system generates the final enhanced feature sequence. The generation process adopts a hierarchical enhancement strategy, and adopts differentiated processing schemes for different types of feature segments. For key feature segments: the dominant feature is selectively amplified through an adaptive nonlinear function, and the gain coefficient is dynamically adjusted between 1.2-1.5 according to the calculated feature strength; the secondary feature is relatively suppressed, and the suppression coefficient ranges from 0.6-0.8, and the suppression strength is inversely proportional to the significance of the dominant feature; the feature association is optimized and mapped through an S-type function to strengthen significant association patterns and weaken unstable associations. For non-key feature segments, basic feature retention processing is performed, and only necessary noise suppression and smoothing processing are performed to ensure the overall continuity of the sequence. The system establishes a feature temporal association model, including short-range associations and long-range associations. Feature consistency checks are performed simultaneously during the enhancement process to ensure that the enhanced features maintain physical meaning. The final enhanced feature sequence has clearer feature expression, more prominent key patterns, and more stable temporal associations.

[0055] Step S104, mapping and transforming the enhanced feature sequence, obtaining initial components, determining main components in the initial components, and generating feature representations based on the main components; segmenting the feature representations, obtaining feature intervals and interval features corresponding to the feature intervals, and generating an intention feature sequence based on the interval features.

[0056] Specifically, when mapping and transforming the enhanced feature sequence, the system adopts a multi-level transformation strategy. First, the enhanced feature sequence is time-aligned, using a sliding window method with a window length of 500ms and an overlap rate of 50%. In each window, the feature data is normalized and preprocessed, including amplitude normalization and variance normalization. Then, a nonlinear mapping method is used to transform the feature space, and the kernel function is used to map the features to a high-dimensional space. The kernel function selects the RBF kernel, and the kernel parameters are determined by cross-validation. The mapping process considers the temporal relationship of the features and introduces a time weight factor. The recent feature weight is large, and the long-term feature weight gradually decays. After the mapping transformation, the system obtains the initial feature components, each of which contains feature expressions in three aspects: time domain information, frequency domain information, and correlation information. The system sorts the initial components according to the degree of contribution, and establishes a mapping parameter index table for each component, recording its time attributes and corresponding weight coefficients.

[0057] In some embodiments, the mapping conversion of the enhanced feature sequence to obtain the initial component includes: time alignment of the enhanced feature sequence based on a sliding window method; normalization preprocessing of the aligned enhanced feature sequence; the normalization processing includes amplitude normalization and variance normalization; feature space conversion of the normalized enhanced feature sequence according to a nonlinear mapping method, mapping the enhanced feature sequence to a high-dimensional space, and obtaining the initial component.

[0058] Based on the generated initial feature components, the system performs feature screening. The screening process first uses the mapping parameter index table and the time weight of each component to calculate the significance index. The significance evaluation includes variance contribution rate (based on component ranking), information gain (calculated using feature expression) and stability score (weighted calculation based on time weight). The system performs a weighted combination of these three indicators, and the weight coefficient is dynamically adjusted according to the mapping effect. Then an adaptive threshold is set. The threshold value is determined according to the statistical characteristics of the score distribution, and the 75% quantile is generally taken. Components that exceed the threshold are retained, and components that are below the threshold are filtered out. At the same time, the system analyzes the correlation between components, and uses the established feature expression. For component pairs with a correlation coefficient exceeding 0.8, the component with a higher score is retained and the other component is deleted to avoid feature redundancy. The screening results contain the complete feature information of the retained components and the correlation analysis results.

[0059] Using the feature screening results, the system determines the main components. First, the filtered component set is clustered and analyzed using the spectral clustering algorithm. The similarity matrix is ​​constructed based on the calculated component scores. The optimal number of clusters is determined by eigenvalue analysis, and the inflection point position of the descending eigenvalue curve is generally selected. For each cluster, the system uses significance indicators to calculate the centrality index of the components within the class, including degree centrality, closeness centrality, and betweenness centrality. The component with the highest centrality index is selected as the representative component of the class. Then, combined with the correlation analysis results, all representative components are ranked in importance based on their projected area and coverage in the original feature space. Finally, the components ranked in the top 60% in importance are selected as the main components.

[0060] Based on the determined main components, the system generates feature representation. First, the basic framework of feature representation is constructed. The framework adopts tensor structure design, including time dimension (sequence of sampling points), feature dimension (sequence of main components) and correlation dimension (relationship between components). In the time dimension, the original sampling frequency is maintained, and each time point corresponds to a complete set of feature vectors. In the feature dimension, the main components are organized according to the determined importance order, and the components are orthogonal. In the correlation dimension, the component correlation graph is constructed using the correlation analysis results. Then, the main components are combined and optimized to design an improved autoencoder structure. The network parameters of the encoder and decoder are configured based on the component characteristics, and the L1 regularization term is introduced during the training process to promote the sparsity of feature representation. The encoded features are reconstructed, and the reconstruction process adopts an iterative optimization strategy, and the quality is evaluated using screening criteria. The reconstructed features are time-series smoothed, and the Kalman filter method is used. The filter parameters are adaptively adjusted according to the dynamic characteristics of the signal. The feature representation finally generated can effectively capture the time-varying characteristics of the original features, maintain the correlation structure between features, and reduce data redundancy.

[0061] When segmenting the feature representation, the system adopts an adaptive segmentation strategy. First, the feature representation is initially divided according to the basic time length of 1 second, and the overlap rate between adjacent segments is maintained at 40% to ensure the continuity of the feature and the complete capture of the change trend. In each basic segment, the statistical indicators of the feature are calculated, including the mean vector (reflecting the central tendency of the feature), the covariance matrix (describing the correlation between features), the autocorrelation coefficient (characterizing the temporal dependence of the feature), the entropy value (quantifying the uncertainty of the feature) and the complexity (measuring the degree of change of the feature). Based on these statistical indicators, the improved dynamic programming algorithm is used to determine the optimal segmentation point. The cost function of the algorithm is designed as: Cost = w1×Consistency + w2×Difference + w3×Length, where Consistency represents the consistency of the feature within the segment, Difference represents the difference of the feature between segments, Length is the segment length penalty term, and the weight coefficient [w1,w2,w3] is determined by grid search optimization. The algorithm is implemented in a sliding window manner, and the basic window length is 200ms, which can be adaptively adjusted in the range of 100-300ms according to the feature change rate. For each segment, the system records its time range (start time, end time, duration), feature statistics (mean, variance, kurtosis, skewness), and boundary features (boundary point feature value, boundary gradient). The system organizes the segmentation results into a segment feature table, which contains complete statistical indicators and boundary information for each segment.

[0062] Based on the segmented feature table, the system performs feature labeling and extraction for each interval. The labeling process first uses the calculated statistical indicators to construct the interval feature descriptor, which contains three types of information: statistical features, morphological features, and dynamic features. Statistical features inherit the calculated statistics such as mean, variance, skewness, and kurtosis; morphological features are based on boundary features, and calculate morphological indicators such as waveform factor (reflecting the complexity of the waveform), pulse factor (characterizing the impact of the waveform), and margin factor (describing the stability of the waveform); dynamic features use time series information, including dynamic parameters such as rate of change (instantaneous change speed of the feature), fluctuation range (variation amplitude of the feature), and trend direction (development trend of the feature). Feature extraction uses a sliding window method with a window length of 100ms and a step size of 20ms to ensure that the detailed changes of the feature are captured. The feature significance score is calculated for each interval, and the score calculation integrates statistical indicators, using a weighted combination method: Score = α×Amplitude + β×Stability + γ×Uniqueness, where the weight coefficients [α, β, γ] are determined by cross-validation. The system analyzes the characteristic patterns of key intervals, identifies the dominant and secondary features through hierarchical clustering methods, and constructs a hierarchical relationship diagram of the features.

[0063] Continuing to use the segmentation results, the system carries out multi-scale interval feature extraction. The extraction process captures the changing rules of features at different time scales based on the time range information. On the short time scale (200ms), the system extracts local features, including instantaneous feature values ​​(based on statistics), local change trends (using boundary gradients), and feature jump points (identified by boundary features). Short-term feature extraction uses a sliding window method with a window length of 100ms and a step length of 20ms. The first-order and second-order differences of the features are calculated for each window to reflect the change speed and acceleration of the features. On the medium time scale (500ms), the evolution pattern of the features is analyzed in combination with the autocorrelation coefficient, including the periodicity, stationarity, and correlation of the feature sequence. The system performs time-frequency analysis on each feature sequence, uses the wavelet transform method, selects Morlet wavelet as the basis function, sets the decomposition scale to 5 layers, and obtains the energy distribution of the features in different frequency bands. On the long time scale (more than 1s), global features are extracted based on complexity indicators, including the long-term trend, abnormal pattern, and repetitive pattern of the features. The system fuses feature information of different scales and adopts a weighted combination strategy. The weight coefficient is dynamically allocated according to the discriminative ability of features at each scale.

[0064] Using the feature information extracted from the first three segments, the system generates an intent feature sequence. The sequence generation process first establishes a feature conversion model based on the segmented structure. The model adopts a three-layer structure design: feature mapping layer, feature combination layer, and feature integration layer. The feature mapping layer uses an improved kernel function method to map the original feature space to the intent feature space. The kernel function selects the RBF kernel, and the kernel parameters are optimized by the feature significance score. The mapping process considers the temporal dependency of the feature and introduces the calculated time decay factor to give the recent features a larger weight. The feature combination layer uses the identified dominant features and secondary features for nonlinear combination. The combination methods include weighted summation, maximum pooling, and average pooling, which are adaptively selected according to the feature type. The weight coefficient is optimized by the least squares method, and the objective function includes the reconstruction error term and the regularization term. The feature integration layer arranges the combined features in chronological order, uses sequence alignment technology to process feature sequences of different lengths, and the alignment reference point is determined based on the results of multi-scale feature analysis. The system smoothes the generated sequence and uses an improved Kalman filter algorithm. The filter parameters are dynamically adjusted according to the temporal characteristics. Outlier detection and correction are performed simultaneously. Outlier determination is based on feature significance thresholds, and correction is performed using local interpolation methods. The final intention feature sequence achieves temporal continuity and consistency in feature expression while maintaining feature discriminability.

[0065] Step S105, obtaining the change nodes and corresponding time series features corresponding to the intention feature sequence, and generating the intention change sequence according to multiple time series features. In some embodiments, the step of obtaining the change nodes and the corresponding timing features corresponding to the intention feature sequence and generating the intention change sequence according to the multiple timing features includes: segmenting and marking the intention feature sequence to obtain the multiple change nodes; obtaining the timing features corresponding to each of the change nodes; merging states according to each of the timing features to obtain a merging result corresponding to the state merging; and obtaining the intention change sequence corresponding to the merging result.

[0066] When segmenting and marking the intention feature sequence, the system adopts a dynamic segmentation strategy of multi-feature fusion. First, the intention feature sequence is initially divided according to the basic length of 500ms, and a 30% overlap interval is set between adjacent segments to ensure the complete capture of the change nodes and the continuous expression of the change trend. For each basic segment, the system calculates the change index based on the feature sequence transmitted by the step, including: 1) amplitude change rate, obtained by calculating the local derivative of the feature sequence, reflecting the change speed of the feature intensity; 2) direction change rate, using the angle change of the feature vector to describe the trajectory change in the feature space; 3) morphological change rate, calculated based on the morphological features of the feature curve (such as curvature and torsion); 4) correlation change rate, characterized by the change speed of the cross-correlation function between features. The change rate calculation adopts the adaptive window method, and the window length is adjusted in the range of 100-300ms according to the severity of the feature change, and the window overlap rate is set to 50%. The system combines the feature mapping results of the step. When the feature changes more drastically, the window length is automatically shortened to improve the temporal resolution; when the feature changes slowly, the window length is increased to improve the statistical reliability. For the calculated change rate sequence, the system uses a multi-threshold detection method to identify the change nodes. The threshold value is dynamically set based on the statistical characteristics of the sequence, generally 2-3 times the local standard deviation. Each detected change node records its time position, change amplitude, change direction, and association strength to form a change node feature vector.

[0067] Based on the marked change nodes, the system extracts time series features. The extraction process is divided into two levels: node layer and interval layer. The node layer features describe the characteristics of a single change node, making full use of the calculated change rate information, focusing on the following analysis: 1) node strength, calculated by integrating the amplitude change rate and the direction change rate; 2) node polarity, judging the positive and negative directions of the change based on the morphological change characteristics; 3) node sharpness, using local curvature information to measure the rapidity of the change; 4) node duration, determined according to the window analysis results. The system establishes a feature description model for each node, records the feature changes within 50ms before and after the node, and the model parameters are optimized based on the change rate sequence. The interval layer features focus on the change pattern between adjacent nodes, and the analysis includes: 1) change trend, judged by the combination of various calculated change rates; 2) change rate, integrating change information at multiple time scales; 3) change law, based on the periodic analysis results; 4) stability evaluation, judging by the associated change rate. The system captures the feature evolution process within the interval through the sliding window method, with a window length of 200ms and a step size of 50ms, to achieve continuous monitoring of feature changes.

[0068] Using the extracted time series features, the system performs state merging. The merging process first constructs a state description table, which records the content that integrates the change indicators and time series features: 1) basic features, including statistics of node layer features; 2) duration, based on node duration analysis; 3) influence range, determined by interval layer features. For each state, the system calculates two key indicators: 1) stability index, evaluated by combining the change rate sequence and interval features; 2) influence index, based on node strength and interval change trend analysis. State merging adopts an adaptive method to determine the merging conditions based on the similarity of state features and time series constraints. During the merging process, the system focuses on: 1) the temporal continuity of the state, which is guaranteed by the interval layer features; 2) the feature consistency of the state, which is verified by the multi-dimensional change rate; 3) the transition characteristics of the state, based on the node feature analysis. The system sets a dynamic merging threshold, and the threshold value comprehensively considers the change significance and feature stability, and varies in the range of 0.5-1.5 times the state spacing.

[0069] Based on the results of state merging, the system generates an intention change sequence. The generation process first establishes a state transition framework, which integrates the analysis results of the first three stages: 1) state description, including change characteristics, time series characteristics and merge characteristics; 2) state parameters, combined with node characteristics and state indicators; 3) state evolution rules, based on merge conditions and state transition constraints. The system generates a state descriptor for each time point, recording in detail: 1) the characteristic component of the current state, from the multidimensional change rate; 2) statistical indicators, calculated based on time series characteristics; 3) time series characteristics, inherited interval characteristics; 4) association characteristics, using state relationship analysis. The sequence generation process pays special attention to the continuity requirements of the state, and ensures smooth transitions between adjacent states by setting the minimum duration (based on node duration statistics) and the state inertia coefficient (using stability indicators). The system performs three-level optimization on the generated sequence: 1) local optimization, using change rate information to eliminate sudden noise; 2) regional optimization, based on interval characteristics to enhance the main change trend; 3) global optimization, using state relationships to repair possible state breakpoints. The final generated intention change sequence can accurately depict the dynamic change process of the user's cognitive state and realize the precise expression and tracking of the intention process.

[0070] Step S106, obtaining preliminary marks corresponding to the intention change sequence, and obtaining intention recognition results corresponding to the preliminary marks.

[0071] When matching the intention change sequence of the step, the system adopts a hierarchical matching strategy. First, the intention change sequence is segmented according to the basic length of 1 second, and the adjacent segments are kept with an overlap interval of 25% to ensure the continuity of the matching process. For each segment, the matching features are calculated based on the state descriptor of the step, including: 1) state features, which inherit the basic feature attributes in the state description of the step, including state value, state duration, state stability, state distribution characteristics and state boundary characteristics; 2) change features, which use the state transition information of the step, including change amplitude, change rate, change direction, change acceleration and change periodicity; 3) association features, which are based on the state association analysis of the step, including state transition probability, state similarity, state dependence, state coupling and state transferability. The feature matching adopts a multi-level comparison method. First, a rough match is performed based on the main state features transferred by the step to quickly screen the candidates; then a fine match is performed using the state transition information to compare the secondary features and change features of the state in detail. During the matching process, the system sets weight coefficients for different features, with the main feature weights ranging from 0.4 to 0.6, the secondary feature weights ranging from 0.2 to 0.4, and the auxiliary feature weights ranging from 0.1 to 0.2. The matching process obtains preliminary marking results, each of which contains a matching score, a matching feature list, and a reliability assessment. The system performs a preliminary screening of the matching results and removes marks with matching degrees below the threshold. The threshold value is dynamically set according to the statistical characteristics of the feature distribution.

[0072] In some embodiments, obtaining the intention recognition result corresponding to the preliminary mark includes: enhancing the features of the preliminary mark to obtain enhanced features; classifying and filtering the enhanced features; and generating the intention recognition result according to the filtering results corresponding to the classification and filtering.

[0073] Based on the generated preliminary tagging results, the system performs feature enhancement. The enhancement process first performs reliability analysis on each tag and calculates the tag confidence indicators based on the matching results, including: 1) feature matching, which inherits the matching score; 2) temporal consistency, which is evaluated based on the state transition feature; 3) context relevance, which is calculated using the associated features. Feature matching is obtained by calculating the cosine similarity of feature vectors, temporal consistency is determined based on the autocorrelation analysis of the sequence, and context relevance is calculated by the cross-correlation of the sliding window. For tags with low confidence (confidence < 0.6), the system starts the feature supplementation mechanism and derives supplementary features by analyzing the distribution of adjacent tag features. Feature supplementation considers the time continuity constraint, combines the state transition law, and uses interpolation and smoothing methods to ensure a smooth transition between the supplementary features and the original features. For tags with high confidence (confidence > 0.8), the system performs feature enhancement and performs differentiated enhancement according to the importance of the matching features, highlighting their dominant features and suppressing secondary features. The enhancement process adopts an adaptive weight strategy, and the weight coefficient is dynamically adjusted according to the confidence of the tag. At the same time, the system performs consistency checks on the enhanced features to ensure that they are consistent with the matching standards.

[0074] Using the enhanced features, the system performs classification screening. A three-layer screening mechanism is designed for the screening process: 1) Feature layer screening, based on the enhanced discrimination ability of a single feature, calculates the feature discrimination index, including feature value distribution, feature stability and feature uniqueness; 2) Pattern layer screening, focusing on the recognition effect of the generated feature combination, analyzing the integrity, typicality and representativeness of the feature pattern; 3) Sequence layer screening, evaluating the temporal coherence of the feature sequence, considering the smoothness, periodicity and trend of the sequence. The system calculates the information gain and feature importance score for each feature, and performs feature screening in combination with confidence assessment. Feature layer screening is based on the discrimination ability of a single feature, and calculates the feature discrimination index, including feature value distribution (mean, variance, skewness, kurtosis), feature stability (short-term stability, long-term stability), and feature uniqueness (the degree of distinction from other features). Information gain and feature importance scores are calculated for each feature, and features with scores below the threshold are filtered out. Pattern-level screening focuses on the recognition effect of feature combinations, analyzing the integrity of feature patterns (the completeness of the elements that make up the pattern), typicality (the degree of matching with the standard pattern), and representativeness (the ability to express the current intention). The system performs cluster analysis on feature patterns to identify the main pattern categories and variant pattern categories. Sequence-level screening evaluates the temporal coherence of feature sequences, taking into account the smoothness (the degree of change between adjacent points), periodicity (the regularity of repeated patterns), and trend (the direction of long-term change).

[0075] Based on the screening results, the system outputs the final intent recognition results. The result generation process first aggregates the features of high-confidence tags, and combines tags with similar time (time interval <200ms) and similar features (similarity >0.8) into intent units. Each intent unit inherits the enhanced features, including: 1) core feature set, which is based on the key features determined by screening; 2) auxiliary feature set, which contains the retained secondary features; 3) time attribute set, which integrates the timing information. The core features include three categories: state features, change features, and association features. The feature weights are determined based on the screening scores. Auxiliary features provide contextual information of the intent, including environmental features, historical features, and predictive features. The time attribute describes the continuous characteristics of the intent, and records the start time, duration, and change time of the intent in combination with the temporal consistency analysis. The system organizes the intent units in time sequence, and establishes the association relationship between the intent units according to the sequence screening results, including causal association (pre- and post-causality), parallel association (simultaneous occurrence), and hierarchical association (inclusion relationship). By analyzing the temporal pattern of the intent unit, the system can accurately capture the evolution process and transformation law of the intent. The final output sequence of intent recognition results not only includes the intent judgment at each time point, but also includes complete feature support information and reliability evaluation.

[0076] In order to implement the real-time EEG signal decoding and intention recognition method of the attention glasses corresponding to the above method embodiment, to achieve the corresponding functions and technical effects. Figure 2 , Figure 2 The structural block diagram of a device 200 for real-time decoding of EEG signals and recognition of intentions provided in an embodiment of the present application is shown. For ease of description, only the parts related to the present embodiment are shown. The device 200 for real-time decoding of EEG signals and recognition of intentions provided in an embodiment of the present application includes: The signal acquisition module 201 is used to acquire the multi-channel EEG time domain signal collected by the attention glasses, perform multi-point sampling on the multi-channel EEG time domain signal, acquire original sampling data, and generate an initial signal sequence according to the original sampling data; The compensation adjustment module 202 is used to obtain the signal segment that needs to be corrected in the initial signal sequence, perform compensation adjustment on the signal segment, and generate a standardized signal sequence; obtain the time domain features corresponding to the standardized signal sequence, perform frequency domain conversion on the time domain features, obtain the frequency domain components and the energy analysis results corresponding to the frequency domain components, and construct a dual-domain feature sequence according to the energy analysis results; The association mapping module 203 is used to perform association mapping on the dual-domain feature sequence, obtain feature breakpoints, establish a time-frequency relationship diagram according to the feature breakpoints, determine a mapping interval according to the time-frequency relationship diagram, and enhance the feature sequence according to the mapping interval to obtain an enhanced feature sequence; A mapping conversion module 204 is used to perform mapping conversion on the enhanced feature sequence, obtain initial components, determine main components in the initial components, and generate feature representations according to the main components; segment the feature representations, obtain feature intervals and interval features corresponding to the feature intervals, and generate an intention feature sequence according to the interval features; A feature acquisition module 205 is used to acquire the change nodes and the corresponding time series features corresponding to the intention feature sequence, and generate an intention change sequence according to a plurality of the time series features; The result acquisition module 206 is used to acquire the preliminary mark corresponding to the intention change sequence, and acquire the intention recognition result corresponding to the preliminary mark.

[0077] The above-mentioned EEG signal real-time decoding and intention recognition device 200 can implement the EEG signal real-time decoding and intention recognition method of the attention 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.

[0078] Figure 3 This is a schematic diagram of the structure of a computer device provided in one embodiment of the present application. Figure 3As shown, the computer device 3 of this embodiment includes: at least one processor 30 ( Figure 3 Only 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.

[0079] 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.

[0080] 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.

[0081] 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.

[0082] 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.

[0083] 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.

[0084] 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.

[0085] 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.

[0086] 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 real-time decoding of EEG signals and recognition of intentions using attention glasses, characterized in that: include: Acquire a multi-channel EEG time domain signal collected by the attention glasses, perform multi-point sampling on the multi-channel EEG time domain signal to obtain original sampling data, and generate an initial signal sequence according to the original sampling data; Acquire a signal segment that needs to be corrected in the initial signal sequence, perform compensation adjustment on the signal segment, and generate a standardized signal sequence; acquire a time domain feature corresponding to the standardized signal sequence, perform frequency domain conversion on the time domain feature, acquire a frequency domain component and an energy analysis result corresponding to the frequency domain component, and construct a dual-domain feature sequence according to the energy analysis result; Performing association mapping on the dual-domain feature sequence to obtain feature breakpoints, establishing a time-frequency relationship diagram according to the feature breakpoints, determining a mapping interval according to the time-frequency relationship diagram, and enhancing the feature sequence according to the mapping interval to obtain an enhanced feature sequence; Performing mapping conversion on the enhanced feature sequence to obtain initial components, determining main components in the initial components, and generating feature representations according to the main components; segmenting the feature representations to obtain feature intervals and interval features corresponding to the feature intervals, and generating an intention feature sequence according to the interval features; Acquire the change nodes and the corresponding time series features corresponding to the intention feature sequence, and generate the intention change sequence according to a plurality of the time series features; Obtain a preliminary mark corresponding to the intention change sequence, and obtain an intention recognition result corresponding to the preliminary mark.

2. The method according to claim 1, characterized in that The generating an initial signal sequence according to the original sampling data comprises: Obtaining the sampling point distribution corresponding to the multi-point sampling; Marking valid sampling segments in the original sampling data according to the sampling point distribution; The initial signal sequence is generated according to the effective sampling segment.

3. The method according to claim 1, characterized in that The obtaining of the signal segment to be corrected in the initial signal sequence, performing compensation adjustment on the signal segment, and generating a standardized signal sequence includes: Performing baseline drift detection on the initial signal sequence to obtain a signal segment that needs to be corrected; Performing the compensation adjustment on the signal segment to be corrected; A standardized signal sequence is generated according to an adjustment result corresponding to the compensation adjustment.

4. The method according to claim 1, characterized in that The step of enhancing the feature sequence according to the mapping interval to obtain an enhanced feature sequence includes: enhancing the sequence according to the mapping interval; Divide the enhanced feature sequence into sections to obtain signal segments; Calculating a characteristic intensity value corresponding to the signal segment; Marking key feature segments according to the feature strength values; The enhanced feature sequence is generated according to the key feature segment.

5. The method according to claim 1, characterized in that The acquiring the change nodes and the corresponding time series features corresponding to the intention feature sequence, and generating the intention change sequence according to the multiple time series features, includes: Marking the intention feature sequence in segments to obtain a plurality of change nodes; Obtaining the time series features corresponding to each of the change nodes; Perform state merging according to each of the time series features, and obtain a merging result corresponding to the state merging; Obtain the intent change sequence corresponding to the merge result.

6. The method according to claim 1, characterized in that The obtaining of the intention recognition result corresponding to the preliminary mark includes: Performing feature enhancement on the preliminary mark to obtain enhanced features; Classifying and screening the enhanced features; The intention recognition result is generated according to the screening result corresponding to the classification screening.

7. The method according to claim 1, characterized in that The step of acquiring the time domain features corresponding to the standardized signal sequence, performing frequency domain conversion on the time domain features, acquiring frequency domain components and energy analysis results corresponding to the frequency domain components, and constructing a dual-domain feature sequence according to the energy analysis results, includes: Segmenting the standardized signal sequence to obtain multiple signal segments; Acquire the time domain feature corresponding to the signal segment; Performing frequency domain conversion on the time domain features to obtain the frequency domain components; Obtaining energy analysis results corresponding to the frequency domain components; The dual-domain feature sequence is constructed according to the time domain features, frequency domain components and energy analysis results.

8. The method according to claim 7, characterized in that The step of segmenting the standardized signal sequence to obtain a plurality of signal segments comprises: The standardized signal is segmented according to a preset initial segment length to obtain a plurality of initial segments; the overlapping rates corresponding to adjacent initial segments are within a preset overlapping rate range; The initial segment length is iteratively optimized according to a preset segment range, and the standardized signal sequence is segmented according to the optimized initial segment length to obtain a plurality of signal segments.

9. The method according to claim 1, characterized in that: The mapping conversion of the enhanced feature sequence to obtain the initial component includes: Performing time alignment on the enhanced feature sequence based on a sliding window method; Performing normalization preprocessing on the aligned enhanced feature sequence; the normalization processing includes amplitude normalization and variance normalization; The normalized enhanced feature sequence is subjected to feature space transformation according to a nonlinear mapping method, the enhanced feature sequence is mapped to a high-dimensional space, and the initial component is obtained.

10. 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 9 when executing the computer program.

Images