A Deep Learning-Based EEG Acquisition and Analysis Method

Abstract

This invention discloses a deep learning-based method for EEG acquisition and analysis, comprising: Step 1: acquiring multi-channel EEG sequences, segmenting them according to time windows, and setting calibration time slots to form segments; Step 2: generating calibration sequences and writing them into the acquisition terminal register within the calibration time slots; Step 3: constructing an input labeled sequence and inputting it into an improved FNet model, outputting link transfer function parameters and contact state vectors; Step 4: performing inverse system reconstruction operations and applying stability constraints to form a reconstructed EEG sequence; Step 5: generating channel-time masks and performing suppression and interpolation to form a clean EEG sequence; Step 6: performing task analysis and processing to form task inference results and calculating a joint credibility index; Step 7: generating and issuing acquisition control command sets. This invention achieves online calibration, stability compensation, and closed-loop adaptive adjustment for EEG acquisition.

Description

Technical Field

[0001] This invention relates to the field of deep learning and physiological signal detection technology, and in particular to a deep learning-based method for electroencephalogram (EEG) acquisition and analysis. Background Technology

[0002] With the continuous expansion of applications such as brain-computer interfaces, neurorehabilitation assessment, consciousness monitoring, and intelligent human-computer interaction, the importance of EEG signal acquisition and analysis technology in medical diagnosis and intelligent computing is becoming increasingly prominent. Current technologies typically employ multi-channel EEG acquisition devices to obtain scalp electrode signals and use fixed filter banks, manual feature extraction methods, or general deep learning networks to perform rhythm analysis, state recognition, and task classification. However, in practical engineering deployment environments, the changing contact state between electrodes and the scalp over time, drift in acquisition link parameters, and frequent occurrences of electromagnetic interference and motion artifacts lead to amplitude and phase distortion of EEG signals, increased noise levels, and insufficient stability across time windows, severely affecting the reliability of subsequent analysis results. Existing methods often rely on offline calibration or simple gain adjustment strategies, lacking the ability to dynamically excite analog front-end parameters and identify links during acquisition, making it difficult to characterize the transient transmission characteristics of different channels within different frequency ranges. Simultaneously, traditional deep learning models typically treat the acquisition link as a fixed system, failing to introduce physically interpretable system parameters into the network output structure. This results in the compensation process relying on empirical filtering or static denoising modules, which cannot adapt to rapid changes in complex scenarios.

[0003] Existing EEG analysis methods often employ a serial processing approach with independent modules for artifact handling and quality assessment. They typically rely solely on amplitude thresholds, frequency band energy, or statistical characteristics for coarse-grained judgments, lacking a mechanism for collaborative modeling of electrode contact states and acquisition link characteristics. This makes them prone to residual noise under conditions of electromyographic interference, polarization drift, or local saturation. While existing studies have introduced deep neural networks to perform classification or regression tasks, most have not explicitly constrained system stability within the network, nor have they established a unified framework for closed-loop adjustment of acquisition end parameters. This makes it difficult for analysis results to drive hardware configuration optimization.

[0004] Therefore, how to provide a deep learning-based EEG acquisition and analysis method is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based EEG acquisition and analysis method. This invention introduces a calibration sequence based on controllable parameter modulation and an improved FNet model to collaboratively model the acquisition link characteristics and contact state, outputting transfer function parameters with stability domain constraints, and combining inverse system compensation and artifact suppression to achieve high-quality EEG recovery. Furthermore, based on a joint credibility index, a closed-loop adjustment mechanism for acquisition parameters is formed, enabling the acquisition process to have online calibration, adaptive compensation, and quality self-evaluation capabilities, thereby significantly improving the stability of EEG acquisition, the reliability of analysis, and the long-term robustness of the system.

[0006] A deep learning-based EEG acquisition and analysis method according to an embodiment of the present invention includes the following steps:

[0007] Step 1: Acquire multi-channel raw EEG sequences, divide them into windowed EEG matrices according to preset time windows, and set calibration time slots to form calibration time slot EEG segments;

[0008] Step 2: Obtain the set of controllable parameters to generate a calibration sequence, and write the calibration sequence into the acquisition terminal register within the calibration time slot;

[0009] Step 3: Encode and splice the calibration time slot EEG segments and calibration sequence, divide them into data blocks, construct input label sequences, and input them into an improved FNet model. The improved FNet model introduces a pole-zero stability domain parameter output mechanism to form the transfer function parameters and contact state vector of each channel acquisition link.

[0010] Step 4: Perform inverse system reconstruction operation on the window EEG matrix based on the acquisition link transfer function parameters and apply stability constraints to form a reconstructed EEG sequence;

[0011] Step 5: Based on the reconstructed EEG sequence and contact state vector, an artifact type identifier and channel-time mask are formed, and suppression and interpolation operations are performed on the reconstructed EEG sequence to form a clean EEG sequence;

[0012] Step 6: Perform task analysis processing on the clean EEG sequence to form task inference results, and calculate the joint reliability index based on the task inference results, contact state vector and acquisition link transfer function parameters;

[0013] Step 7: Based on the joint credibility index, contact state vector and acquisition link transfer function parameters, a set of acquisition control instructions is formed and sent to the acquisition terminal.

[0014] Optionally, step one specifically includes:

[0015] A multi-channel raw EEG sequence with a sampling rate of a preset value is acquired. The multi-channel raw EEG sequence includes the channel identifier of each acquisition channel and the amplitude data at the corresponding sampling time.

[0016] The multi-channel raw EEG sequence is subjected to sliding segmentation processing according to preset time window segmentation parameters. The preset time window segmentation parameters include time window length and window shift length. The window shift length is the number of sampling points between the starting sampling points of adjacent time windows. The sliding segmentation processing includes determining the starting sampling point and ending sampling point of each time window based on the time window length, and determining the interval between the starting sampling points of adjacent time windows based on the window shift length.

[0017] The amplitude data corresponding to each acquisition channel within each time window are arranged and spliced ​​according to the channel order to form a window EEG matrix. The rows of the window EEG matrix correspond to the acquisition channels, and the columns correspond to the sampling point numbers within the time window.

[0018] Within each time window, the start and end sampling points of the calibration time slot are determined based on the calibration time slot parameters, which include the calibration time slot length and the relative position of the calibration time slot within the time window.

[0019] Amplitude data from each acquisition channel within the calibration time slot range are extracted from the window EEG matrix to form a calibration time slot EEG segment.

[0020] Optionally, step two specifically includes:

[0021] The controllable parameter set is read based on the analog front-end circuit of the acquisition terminal and the digital control module. The controllable parameter set includes amplifier gain control quantity, reference electrode switching control quantity, input impedance switching control quantity, notch filter center frequency control quantity, and high-pass filter cutoff frequency control quantity.

[0022] Based on the spatial position of the acquisition channel in the electrode cap, the acquisition channel is divided into multiple channel groups, and different parameter combination excitation templates are assigned to each channel group. The parameter combination excitation template is formed by combining two or more types of parameters in the controllable parameter set according to different parameters on the time axis using mutually unrelated switching cycle or amplitude encoding rules.

[0023] For each parameter combination excitation template, a discrete state sequence is constructed. The discrete state sequence is arranged on the time axis according to a preset switching period, and uncorrelated change trajectories are generated based on the change magnitude constraints of different parameter dimensions.

[0024] Piecewise linear interpolation is performed on the discrete state sequence to form a continuously changing multi-parameter control trajectory, and quantization encoding is performed on the multi-parameter control trajectory to generate a calibration sequence containing the relationship between the changes of each control parameter over time.

[0025] The multidimensional calibration sequence is converted into the corresponding front-end circuit control word sequence according to the channel group mapping relationship and written into the acquisition end register during the calibration time slot.

[0026] Optionally, step three specifically includes:

[0027] The calibration time-slot EEG segments and calibration sequence encodings are processed by channel-dimensional splicing and time-dimensional block segmentation to form the input labeled sequence;

[0028] The improved FNet model includes a Fourier labeling mixture module, a feedforward mapping module, a pole-zero stability domain parameter output structure, and a contact state parameter output structure;

[0029] The Fourier label mixing module performs discrete Fourier transform on the input label sequence in the label dimension and feature dimension respectively, preserving the real and imaginary parts of the complex features to form a frequency domain mixed label;

[0030] The feedforward mapping module performs linear and nonlinear mapping processing on the frequency domain hybrid marker to form a channel feature vector.

[0031] The pole-zero stability domain parameter output structure generates pole radius parameters, pole angle parameters, zero radius parameters, and zero angle parameters for each acquisition channel based on the channel feature vector. The pole radius parameters are restricted to the range of greater than 0 and less than 1, the pole angle parameters are restricted to the range of negative π to positive π, the zero radius parameters are restricted to the range of greater than 0 and less than 1, and the zero angle parameters are restricted to the range of negative π to positive π.

[0032] A set of complex poles is constructed based on the pole radius parameters and pole angle parameters, and a set of real coefficient poles is generated according to the conjugate pairing rule;

[0033] A complex set of zeros is constructed based on the zero radius parameter and the zero angle parameter, and a set of real coefficient zeros is generated according to the conjugate pairing rule;

[0034] A denominator polynomial is generated based on the set of poles, a numerator polynomial is generated based on the set of zeros, and a transfer function for each acquisition link is constructed based on the numerator and denominator polynomials.

[0035] The contact state parameter output structure generates a multi-dimensional contact state parameter set based on the channel feature vector. The multi-dimensional contact state parameter set includes the channel contact impedance change, electrode polarization amplitude index, signal saturation ratio index, and baseline drift amplitude index. The multi-dimensional contact state parameter set is arranged according to the channel index to form a contact state vector.

[0036] Optionally, step four specifically includes:

[0037] Based on the transfer function parameters of each channel acquisition link, frequency domain transformation is performed on the window EEG matrix to obtain the frequency domain representation of each channel;

[0038] In the frequency domain representation, an inverse system operator is constructed for each frequency point. The inverse system operator is constructed by taking the conjugate of the transfer function parameters at the corresponding frequency point and superimposing a positive stable term formed by a preset constant and the square of the magnitude of the transfer function parameters in the denominator.

[0039] Based on the contact state vector and the mean and variance of the power spectrum of the EEG segment in the calibration time slot at each frequency point, the inverse system gain constraint threshold is calculated for each channel and each frequency point. The inverse system gain constraint threshold is obtained by normalizing the channel contact impedance change, electrode polarization amplitude index, signal saturation ratio index and baseline drift amplitude index respectively, and then weighting and summing them according to preset weight coefficients.

[0040] When the amplitude of the inverse system operator exceeds the corresponding inverse system gain constraint threshold, the amplitude of the inverse system operator is truncated to the inverse system gain constraint threshold, while keeping the phase component unchanged, thus forming a constrained inverse system operator;

[0041] The constrained inverse system operator is multiplied with the frequency domain representation point by point to form a compensated frequency domain signal, and an inverse frequency domain transformation is performed on the compensated frequency domain signal to form a reconstructed EEG sequence.

[0042] Optionally, step five specifically includes:

[0043] Based on the reconstructed EEG sequence, the root mean square value of amplitude, short-time energy, kurtosis coefficient and baseline drift are calculated for each acquisition channel in the time dimension, and the power spectral density corresponding to each preset frequency band is calculated in the frequency domain.

[0044] The root mean square value of amplitude, short-time energy, kurtosis coefficient, baseline drift and contact state vector are concatenated by channel index to form a set of artifact discrimination features. Based on the set of artifact discrimination features, a threshold judgment is performed on each time segment to form a corresponding artifact type identifier.

[0045] Based on the artifact type identifier, a binary or multi-valued channel-time mask is generated in the channel dimension and the time dimension to identify the effective and abnormal regions of each acquisition channel in each time segment.

[0046] Based on the channel-time mask, amplitude attenuation processing is performed on reconstructed EEG sequence samples in abnormal regions, time domain interpolation processing is performed on samples in continuous abnormal regions, and spatial neighborhood interpolation processing is performed on samples in cross-channel abnormal regions.

[0047] The samples from each channel, after amplitude attenuation and interpolation processing, are recombined to form a clean EEG sequence.

[0048] Optionally, step six specifically includes:

[0049] The clean EEG sequences were arranged into a time series matrix according to channel order, and corresponding analysis segments were extracted according to the preset time window index.

[0050] In each acquisition channel, a fast Fourier transform is performed on the analysis segment, the frequency domain amplitude is squared and averaged according to the frequency band boundaries to form the power spectral density corresponding to each preset frequency band.

[0051] In the time domain, a fixed-length sliding window is used to calculate the root mean square value of the analysis segment, and a fixed-dimensional set of phase space embedding vectors is formed by continuous sampling points. The matching ratio is statistically analyzed based on the similarity threshold to form a sample entropy index.

[0052] The power spectral density, root mean square value and sample entropy index corresponding to each channel are concatenated according to the channel index to form an analysis feature vector, and then input into the task inference module to perform linear mapping and nonlinear mapping operations to form the task inference result;

[0053] Based on the category probability distribution of the task reasoning results, the logarithm of each category probability is taken and multiplied by the corresponding probability, and then summed to form the information entropy value;

[0054] The mean and variance statistics of the contact state vector are calculated separately for each channel dimension. The reciprocal of the amplitude of the transfer function parameter of the acquisition link is calculated at each frequency point. The reciprocal of the amplitude is then averaged in the frequency dimension to form the frequency domain condition number index.

[0055] The reciprocal of the information entropy value, the mean and variance statistics, and the frequency domain condition number index are mapped to the 0-1 interval, and weighted and accumulated according to the preset weight coefficients to form a joint credibility index.

[0056] Optionally, step seven specifically includes:

[0057] The joint credibility index is compared with a preset trigger threshold. When the joint credibility index is lower than the trigger threshold, the calculation step of adjusting the acquisition parameters is initiated.

[0058] The priority adjustment level of the corresponding acquisition channel is formed by weighting and sorting the normalized values ​​of the contact impedance change, electrode polarization amplitude index, signal saturation ratio index and baseline drift amplitude index of each channel in the contact state vector.

[0059] Based on the difference between the amplitude of each frequency point in the acquisition link transfer function parameters and the preset reference spectrum, determine the adjustment direction of the amplifier gain, the center frequency of the notch filter, and the cutoff frequency of the high-pass filter.

[0060] Based on the trend of the joint credibility index change corresponding to the most recent preset number of time windows, the priority adjustment level and the adjustment direction, the adjustment range of each acquisition parameter is calculated. The acquisition parameters include amplifier gain, input impedance level, notch filter center frequency, high-pass filter cutoff frequency and reference electrode switching period.

[0061] The adjustment range is quantified and superimposed with the current collected parameter value to form an updated set of collected parameters.

[0062] The updated set of acquisition parameters is encoded into a set of acquisition control instructions and written into the acquisition terminal register according to the channel mapping relationship;

[0063] At the starting sampling point of the next time window, the acquisition control command set is activated and sent to the acquisition terminal, and the corresponding parameter update record is written to the log unit.

[0064] The beneficial effects of this invention are:

[0065] This invention addresses the challenges of online modeling of link parameter drift, quantitative characterization of electrode contact changes, and lack of physical constraints in compensation processes during EEG acquisition by introducing a calibration sequence based on controllable parameter combination modulation during the EEG acquisition stage and simultaneously outputting the acquisition link transfer function parameters with stability domain constraints and a multidimensional contact state vector using an improved FNet model. It constructs a joint modeling framework integrating system identification and deep learning. At the model output, a pole-zero stability domain parameterization structure is adopted to map the acquisition link into a transfer function form that satisfies stability constraints, providing an interpretable and controllable compensation basis for subsequent inverse system reconstruction. During the signal recovery stage, a frequency-adaptive gain constraint mechanism is constructed by combining the contact state vector and the frequency domain statistics of the calibration time slot, and frequency-point truncation is applied to the inverse system operator. The system employs phase-preserving operations to effectively suppress noise amplification and numerical instability. During artifact processing, it generates channel-time masks based on time-frequency statistical features and contact states, and repairs abnormal segments through amplitude attenuation, time interpolation, and spatial neighborhood interpolation, achieving refined suppression of EEG interference, electrode polarization, and saturation distortion. In the analysis and control phase, it further integrates task inference results, link condition number indicators, and contact state statistics to calculate a joint reliability index, thereby forming a closed-loop acquisition control instruction set for amplifier gain, filter parameters, and reference electrode switching. This enables online calibration, stability compensation, quality self-assessment, and adaptive adjustment of hardware parameters for EEG signals, significantly improving the reliability of EEG acquisition in continuous operation scenarios, the stability of analysis results, and the robustness of the system in long-term operation. Attached Figure Description

[0066] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0067] Figure 1 This is a schematic diagram of the overall process of a deep learning-based EEG acquisition and analysis method proposed in this invention;

[0068] Figure 2 This is a schematic diagram of the structure of the improved FNet model in the deep learning-based EEG acquisition and analysis method proposed in this invention;

[0069] Figure 3 This is a schematic diagram of the inverse system reconstruction and stability constraint processing flow in the deep learning-based EEG acquisition and analysis method proposed in this invention. Detailed Implementation

[0070] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0071] refer to Figure 1-3 A deep learning-based method for EEG acquisition and analysis includes the following steps:

[0072] Step 1: Acquire multi-channel raw EEG sequences, divide them into windowed EEG matrices according to preset time windows, and set calibration time slots to form calibration time slot EEG segments;

[0073] Step 2: Obtain the set of controllable parameters to generate a calibration sequence, and write the calibration sequence into the acquisition terminal register within the calibration time slot;

[0074] Step 3: Encode and splice the calibration time-slot EEG segments and calibration sequence, divide them into data blocks, construct input label sequences, input them into the improved FNet model, and introduce the pole-zero stability domain parameter output mechanism into the improved FNet model to form the transfer function parameters and contact state vector of each channel acquisition link;

[0075] Step 4: Perform inverse system reconstruction operation on the window EEG matrix based on the acquisition link transfer function parameters and apply stability constraints to form a reconstructed EEG sequence;

[0076] Step 5: Based on the reconstructed EEG sequence and contact state vector, artifact type identifiers and channel-time masks are generated, and suppression and interpolation operations are performed on the reconstructed EEG sequence to generate a clean EEG sequence;

[0077] Step 6: Perform task analysis on the clean EEG sequence to generate task inference results, and calculate the joint reliability index based on the task inference results, contact state vector and acquisition link transfer function parameters;

[0078] Step 7: Based on the joint credibility index, contact state vector and acquisition link transfer function parameters, form an acquisition control command set and send it to the acquisition terminal.

[0079] In this embodiment, step one specifically includes:

[0080] Acquire multi-channel raw EEG sequences with a preset sampling rate. The multi-channel raw EEG sequences include the channel identifier of each acquisition channel and the amplitude data at the corresponding sampling time.

[0081] The multi-channel raw EEG sequence is subjected to sliding segmentation processing based on preset time window segmentation parameters. The preset time window segmentation parameters include time window length and window shift length. The window shift length is the number of sampling points between the starting sampling points of adjacent time windows. The sliding segmentation processing includes determining the starting and ending sampling points of each time window based on the time window length, and determining the interval between the starting sampling points of adjacent time windows based on the window shift length.

[0082] The amplitude data corresponding to each acquisition channel within each time window are arranged and spliced ​​according to the channel order to form a window EEG matrix. The rows of the window EEG matrix correspond to the acquisition channels, and the columns correspond to the sampling point sequence numbers within the time window.

[0083] Within each time window, the start and end sampling points of the calibration time slot are determined based on the calibration time slot parameters, which include the calibration time slot length and the relative position of the calibration time slot within the time window.

[0084] Amplitude data from each acquisition channel within the calibration time slot range are extracted from the window EEG matrix to form calibration time slot EEG segments.

[0085] In this implementation, the sampling rate of the multi-channel raw EEG sequence is set to 250Hz, 500Hz, or 1000Hz, and the number of acquisition channels is set to 16, 32, or 64. In the preset time window segmentation parameters, the time window length is set to 2s-6s, the window shift length is set to 0.5s-2s, and the window shift length corresponds to the number of sampling points between the starting sampling points of adjacent time windows. The calibration time slot length is set to 0.2s-1s, and the relative position of the calibration time slot within the time window is set in the range of 0.1s-0.5s after the starting sampling point. The window EEG matrix is ​​formed by arranging the channels in order, with the number of rows equal to the number of acquisition channels and the number of columns equal to the number of sampling points corresponding to the time window length. The calibration time slot EEG segment is generated by index truncation and written into the buffer or storage unit along with the channel identifier of the corresponding time window, the starting sampling point, and the ending sampling point.

[0086] In this embodiment, step two specifically includes:

[0087] The controllable parameter set is read based on the analog front-end circuit and digital control module of the acquisition end. The controllable parameter set includes amplifier gain control quantity, reference electrode switching control quantity, input impedance switching control quantity, notch filter center frequency control quantity, and high-pass filter cutoff frequency control quantity.

[0088] Based on the spatial position of the acquisition channel in the electrode cap, the acquisition channel is divided into multiple channel groups, and different parameter combination excitation templates are assigned to each channel group. The parameter combination excitation template is formed by combining two or more types of parameters from the controllable parameter set according to different parameters on the time axis using mutually unrelated switching cycle or amplitude encoding rules.

[0089] For each parameter combination excitation template, a discrete state sequence is constructed. The discrete state sequence is arranged on the time axis according to a preset switching period, and independent change trajectories are generated based on the change magnitude constraints of different parameter dimensions.

[0090] Piecewise linear interpolation is performed on the discrete state sequence to form a continuously changing multi-parameter control trajectory. Quantization encoding is then performed on the multi-parameter control trajectory to generate a calibration sequence containing the relationship between each control parameter and time.

[0091] The multidimensional calibration sequence is converted into the corresponding front-end circuit control word sequence according to the channel group mapping relationship and written into the acquisition end register during the calibration time slot.

[0092] In this implementation, the channel groups are determined based on the electrode cap wiring structure and scalp region division rules, with each channel group corresponding to a continuously arranged physical lead position. The parameter combination excitation template is constructed by setting different switching periods, duty cycles, and amplitude change step sizes for different parameters within the same calibration time slot, ensuring that any parameter dimension remains uncorrelated in both the spectral and temporal domains. The switching period of the discrete state sequence is determined by an integer multiple of the sampling clock, and the interpolation process is completed at the boundaries of adjacent states by changing at a fixed slope. The quantization encoding uses a register bit width matching method to map the multi-parameter trajectory into a control word. The channel group mapping relationship is stored in a configuration table, and the configuration table is called and synchronously written to the front-end control interface when the calibration time slot trigger signal arrives.

[0093] In this embodiment, step three specifically includes:

[0094] The calibration time-slot EEG segments and calibration sequence encodings are processed by channel-dimensional splicing and time-dimensional block segmentation to form the input labeled sequence;

[0095] The improved FNet model includes a Fourier labeling mixture module, a feedforward mapping module, a pole-zero stability domain parameter output structure, and a contact state parameter output structure;

[0096] The Fourier labeling mixing module performs discrete Fourier transforms on the input labeling sequence in the label dimension and feature dimension respectively, preserving the real and imaginary parts of the complex features to form a frequency domain mixed label;

[0097] The feedforward mapping module performs linear and nonlinear mapping on the frequency domain hybrid markers to form channel feature vectors;

[0098] The pole-zero stability domain parameter output structure generates pole radius parameters, pole angle parameters, zero radius parameters, and zero angle parameters for each acquisition channel based on the channel feature vector. The pole radius parameters are restricted to the range of greater than 0 and less than 1, the pole angle parameters are restricted to the range of negative π to positive π, the zero radius parameters are restricted to the range of greater than 0 and less than 1, and the zero angle parameters are restricted to the range of negative π to positive π.

[0099] Construct a set of complex poles based on the pole radius and pole angle parameters, and generate a set of real coefficient poles according to the conjugate pairing rule;

[0100] The conjugate pairing rule includes: when the complex form corresponding to the pole angle parameter or the zero angle parameter contains an imaginary part, for each complex pole or complex zero, a conjugate complex pole or conjugate complex zero with opposite signs of the imaginary part and the same real part is generated simultaneously, and the conjugate complex pole or conjugate complex zero is written together with the corresponding original complex pole or original complex zero into the pole set or zero set; when the imaginary part of the complex form corresponding to the pole angle parameter or the zero angle parameter is zero, the corresponding pole or zero is directly written into the pole set or zero set.

[0101] Construct a set of complex zeros based on the zero radius parameter and the zero angle parameter, and generate a set of real coefficient zeros according to the conjugate pairing rule;

[0102] The denominator polynomial is generated based on the pole set, the numerator polynomial is generated based on the zero set, and the transfer function of each acquisition link is constructed based on the numerator and denominator polynomials.

[0103] The contact state parameter output structure generates a multi-dimensional contact state parameter set based on the channel feature vector. The multi-dimensional contact state parameter set includes the channel contact impedance change, electrode polarization amplitude index, signal saturation ratio index, and baseline drift amplitude index. The multi-dimensional contact state parameter set is arranged according to the channel index to form a contact state vector.

[0104] The improved FNet model adopts the label-level frequency domain mixing concept of the FNet model in its overall architecture. The input is a label sequence composed of calibration time-slot EEG segments and calibration sequence encoding. The model includes a Fourier label mixing module and a feedforward mapping module. The Fourier label mixing module performs discrete Fourier operations on the label sequence in both the label dimension and the feature dimension, preserving the real and imaginary parts of the complex features, so that each label obtains global correlation information across time position and across feature channels in a single operation. The feedforward mapping module consists of multi-layer linear mapping units and non-linear activation units, which are used to rearrange features and transform dimensions of the frequency domain mixed labels to form channel-level feature representations.

[0105] Unlike the original FNet model, the improved FNet model introduces a pole-zero stability domain parameter output structure and a contact state parameter output structure at the output end. The pole-zero stability domain parameter output structure takes the channel feature vector as input and generates pole radius parameters, pole angle parameters, zero radius parameters, and zero angle parameters. It then constructs corresponding pole and zero sets according to stability domain constraints and conjugate pairing rules, thereby forming the acquisition link transfer function parameters. The contact state parameter output structure takes the same channel feature vector as input and outputs multidimensional parameters including channel contact impedance change, electrode polarization amplitude index, signal saturation ratio index, and baseline drift amplitude index, arranged by channel index to form a contact state vector.

[0106] By adding a parallel mapping mechanism between the stable domain parameterized output and the contact state while keeping the FNet frequency domain label hybrid structure unchanged, the improved FNet model simultaneously forms the dynamic characteristic parameters of the acquisition link and the electrode contact state representation in a single network. This allows the subsequent inverse system reconstruction processing to be directly based on the transfer function expansion of the stable domain constraint, and shares a unified channel-level state description with the artifact processing and acquisition control steps. Thus, it achieves coordinated consistency in link compensation, quality assessment and closed-loop control parameter generation within the time window scale.

[0107] In this embodiment, step four specifically includes:

[0108] Based on the transfer function parameters of each channel acquisition link, frequency domain transformation is performed on the window EEG matrix to obtain the frequency domain representation of each channel;

[0109] In the frequency domain representation, an inverse system operator is constructed for each frequency point. The inverse system operator is constructed by taking the conjugate of the transfer function parameters at the corresponding frequency point and superimposing a positive stable term formed by a preset constant and the square of the magnitude of the transfer function parameters in the denominator.

[0110] Based on the contact state vector and the mean and variance of the power spectrum of the EEG segment in the calibration time slot at each frequency point, the inverse system gain constraint threshold is calculated for each channel and each frequency point. The inverse system gain constraint threshold is obtained by normalizing the channel contact impedance change, electrode polarization amplitude index, signal saturation ratio index and baseline drift amplitude index respectively, and then weighting and summing them according to the preset weight coefficients.

[0111] When the amplitude of the inverse system operator exceeds the corresponding inverse system gain constraint threshold, the amplitude of the inverse system operator is truncated to the inverse system gain constraint threshold, while the phase component remains unchanged, thus forming a constrained inverse system operator.

[0112] The constrained inverse system operator is multiplied with the frequency domain representation point by point to form a compensated frequency domain signal. The compensated frequency domain signal is then subjected to an inverse frequency domain transformation to form a reconstructed EEG sequence.

[0113] In this implementation, the frequency domain transformation employs Fast Fourier Transform (FFT), with the number of transformation points set to 256, 512, or 1024. The preset constant in the positive stability term is set to 0.001-0.05. In the calculation of the inverse system gain constraint threshold, the contact impedance change, electrode polarization amplitude, signal saturation ratio, and baseline drift amplitude of each channel are normalized to the 0-1 interval, and then weighted and summed according to weighting coefficients of 0.2-0.4, 0.2-0.3, 0.1-0.3, and 0.1-0.3. The mean and variance of the power spectrum at each frequency point are statistically analyzed using a sliding window within the calibration time slot length, with the window length set to 16-64 sampling points. The upper bound of the inverse system gain constraint threshold is set to 2-10 times the reciprocal of the amplitude at the corresponding frequency point, and the lower bound is set to 0.5-1.5 times the reciprocal of the amplitude at the corresponding frequency point, used to constrain the amplitude range of the inverse system operator.

[0114] In this embodiment, step five specifically includes:

[0115] Based on the reconstructed EEG sequence, the root mean square value of amplitude, short-time energy, kurtosis coefficient and baseline drift are calculated for each acquisition channel in the time dimension, and the power spectral density corresponding to each preset frequency band is calculated in the frequency domain.

[0116] The root mean square value of amplitude, short-time energy, kurtosis coefficient, baseline drift and contact state vector are concatenated by channel index to form a set of artifact discrimination features. Threshold judgment is performed on each time segment based on the set of artifact discrimination features to form the corresponding artifact type identifier.

[0117] Based on the artifact type identifier, a binary or multi-valued channel-time mask is generated in the channel dimension and the time dimension to identify the effective and abnormal regions of each acquisition channel in each time segment.

[0118] Based on the channel-time mask, amplitude attenuation processing is performed on reconstructed EEG sequence samples in abnormal regions, time domain interpolation processing is performed on samples in continuous abnormal regions, and spatial neighborhood interpolation processing is performed on samples in cross-channel abnormal regions.

[0119] The samples from each channel, after amplitude attenuation and interpolation processing, are recombined to form a clean EEG sequence.

[0120] In this implementation, a fixed-length time sliding window is used to extract continuous sampling points for each acquisition channel. The squared amplitude of each sampling point is averaged and the square root is taken to form the root mean square sequence of amplitude. The absolute value of the amplitude of each sampling point within the time sliding window is squared and accumulated, and then divided by the number of sampling points to form a normalized short-time energy sequence. The mean and standard deviation of the amplitude sequence within the time sliding window are calculated. The fourth power of the difference between each sampling point and the mean is averaged and divided by the fourth power of the standard deviation to form the kurtosis coefficient sequence. At the same time, low-pass filtering is performed on the amplitude sequence within the time sliding window to extract the slowly varying components, and the amplitude difference of the slowly varying components between the start and end positions of the time sliding window is calculated to form the baseline drift sequence. All indicators are arranged in chronological order and stored in correspondence with the channel index.

[0121] In this embodiment, step six specifically includes:

[0122] The clean EEG sequences were arranged into a time series matrix according to channel order, and corresponding analysis segments were extracted according to the preset time window index.

[0123] In each acquisition channel, a fast Fourier transform is performed on the analysis segment, the frequency domain amplitude is squared and averaged according to the frequency band boundaries to form the power spectral density corresponding to each preset frequency band.

[0124] In the time domain, a fixed-length sliding window is used to calculate the root mean square value of the analysis segment, and a fixed-dimensional set of phase space embedding vectors is constructed using continuous sampling points. The matching ratio is statistically analyzed based on the similarity threshold to form the sample entropy index.

[0125] The power spectral density, root mean square value and sample entropy index corresponding to each channel are concatenated according to the channel index to form an analysis feature vector, and then input into the task inference module to perform linear mapping and nonlinear mapping operations to form the task inference result;

[0126] Based on the category probability distribution of the task reasoning results, the logarithm of each category probability is taken and multiplied by the corresponding probability, and then summed to form the information entropy value;

[0127] The mean and variance statistics of the contact state vector are calculated separately for each channel dimension. The reciprocal of the amplitude of the transfer function parameter of the acquisition link is calculated at each frequency point. The reciprocal of the amplitude is then averaged in the frequency dimension to form the frequency domain condition number index.

[0128] The reciprocal of the information entropy value, the mean and variance statistics, and the frequency domain condition number index are mapped to the 0-1 interval, and weighted and accumulated according to the preset weight coefficients to form a joint credibility index.

[0129] In this implementation, after the joint confidence index is mapped to the 0-1 range, it is compared with a preset threshold of 0.4-0.7 to trigger the generation of acquisition control commands. The acquisition control command set includes amplifier gain adjustment step size of 0.5dB-3dB, input impedance switching levels of 2-6, notch filter center frequency adjustment range of 0.5Hz-2Hz, high-pass filter cutoff frequency adjustment range of 0.1Hz-1Hz, and reference electrode switching cycle of 5s-30s. The control command issuance cycle is set to 1-3 time windows, and the parameter update amount is calculated based on the joint confidence change trend of the most recent 3-10 time windows. Each control command is written to the front-end control register through the digital control bus and takes effect synchronously at the beginning of the next time window sampling point, which is used to form a continuously updated acquisition parameter configuration record and stored in the log unit.

[0130] In this embodiment, step seven specifically includes:

[0131] The joint credibility index is compared with the preset trigger threshold. When the joint credibility index is lower than the trigger threshold, the calculation step of adjusting the collection parameters is initiated.

[0132] The priority adjustment level of the corresponding acquisition channel is formed by weighting and sorting the normalized values ​​of the contact impedance change, electrode polarization amplitude index, signal saturation ratio index and baseline drift amplitude index of each channel in the contact state vector.

[0133] Based on the difference between the amplitude of each frequency point in the acquisition link transfer function parameters and the preset reference spectrum, determine the adjustment direction of the amplifier gain, the center frequency of the notch filter, and the cutoff frequency of the high-pass filter.

[0134] Based on the trend of joint reliability index changes, priority adjustment level and adjustment direction corresponding to the most recent preset number of time windows, calculate the adjustment range of each acquisition parameter. The acquisition parameters include amplifier gain, input impedance level, notch filter center frequency, high-pass filter cutoff frequency and reference electrode switching period.

[0135] The adjustment range is quantified and superimposed with the current collected parameter value to form an updated set of collected parameters.

[0136] The updated set of acquisition parameters is encoded into a set of acquisition control instructions and written into the acquisition terminal register according to the channel mapping relationship;

[0137] At the starting sampling point of the next time window, the acquisition control command set is activated and sent to the acquisition terminal, and the corresponding parameter update record is written to the log unit.

[0138] In this implementation, the parameter adjustment calculation step is initiated when the joint confidence index falls below the trigger threshold of 0.4-0.7; the most recent preset number of time windows is set to 3-10. After normalization, each index in the contact state vector is weighted and sorted using weighting coefficients of 0.25-0.35, 0.20-0.30, 0.15-0.25, and 0.10-0.20 to form a priority adjustment level; the preset reference spectrum is obtained statistically from the reconstructed EEG sequences of the most recent 20-100 time windows, and the adjustment direction is updated when the amplitude deviation at each frequency point exceeds 2dB-6dB. The single amplifier gain adjustment amplitude is limited to 0.5dB-3dB, the input impedance level switching stage is set to 2-6 stages, the notch filter center frequency adjustment step size is set to 0.5Hz-2Hz, the high-pass cutoff frequency adjustment step size is set to 0.1Hz-1Hz, and the reference electrode switching cycle is set to 5s-30s; after quantization and encoding, the parameters are written to the acquisition terminal register at the starting sampling point of the next time window and recorded in the log unit.

[0139] Example 1: To verify the feasibility of this invention in practice, it was applied to a continuous EEG acquisition and cognitive task assessment scenario in a neurofunctional monitoring center of a tertiary hospital. This scenario primarily targets postoperative intensive care patients and rehabilitation subjects. During continuous monitoring for several to tens of hours, signal drift, amplitude abrupt changes, and EMG artifacts frequently occur due to scalp sweating, electrode loosening, posture changes, and electromagnetic interference from surrounding medical equipment. Existing systems typically use fixed gain and preset filter parameters for acquisition, supplemented by simple threshold-based artifact removal strategies. After long-term operation, signal quality deteriorates, misclassification rates increase, and frequent manual readjustment of electrodes or acquisition parameters is required, severely impacting clinical efficiency and analytical reliability. This example uses a 32-channel wet electrode EEG acquisition device. The front-end circuit at the acquisition end supports programmable amplifier gain, input impedance levels, notch filter center frequency, high-pass cutoff frequency, and reference electrode switching configurations. The deep learning-based EEG acquisition and analysis method described in this invention is deployed in the acquisition host as a real-time processing module.

[0140] In this embodiment, the system first automatically extracts a calibration time slot of preset length within each time window, and generates a multi-dimensional parameter combination calibration sequence based on the acquisition end register configuration table. This sequence is then written to the front-end circuit via a digital control bus, allowing different channel groups to experience uncorrelated parameter change trajectories within the same time window. The acquired calibration time slot EEG segments and corresponding parameter encoding sequences are synchronously fed into the computing unit and constructed as an input labeled sequence for inputting into the improved FNet model. The model performs frequency domain hybrid operations on the time and feature dimensions in the Fourier label mixing module, generating a set of pole-zero parameters that satisfy the stability domain constraints at the output end. Simultaneously, it outputs a contact state vector in parallel, characterizing changes in electrode contact impedance, electrode polarization, signal saturation ratio, and baseline drift amplitude. The system performs frequency domain inverse system compensation on the complete time window EEG signal based on the transfer function parameters, and combines the contact state with the spectral statistics of the calibration segment to form a frequency-by-frequency gain constraint, thereby avoiding excessive amplification in the noise band. The system then uses the time-frequency statistical features of the reconstructed signal and the contact state vector to construct a channel-time mask, and performs amplitude attenuation, time interpolation and spatial neighborhood interpolation on the electromyography burst segment, electrode polarization drift segment and local saturation segment respectively to form a clean EEG sequence for downstream analysis.

[0141] During continuous operation, the system further calculates the power spectral density, root mean square value, and sample entropy of the δ, θ, α, β, and γ frequency bands based on clean EEG sequences, and inputs these values ​​into the task inference module to complete the subject's working memory task and resting state classification. Combining the information entropy value formed by the classification probability distribution, contact state statistics, and the frequency domain condition number index of the acquisition link, the system generates a joint reliability index in real time. When this index remains below the threshold range, the acquisition control module automatically calculates the priority adjustment level of each channel based on the contact state ranking results and the trend of link parameter changes, and generates new amplifier gain, input impedance levels, and filter parameter control commands. These commands are written to the front-end control register at the beginning of the next time window, thereby achieving closed-loop adaptive adjustment of the acquisition process and significantly reducing the frequency of manual intervention.

[0142] To verify the beneficial effects of this invention, a comparative experiment was conducted with continuous monitoring for 72 hours. This embodiment cross-tested the traditional fixed-parameter data acquisition scheme with the method of this invention, with each scheme covering 20 subjects, for a total effective data acquisition time exceeding 2800 hours. Evaluation metrics included the improvement in signal-to-noise ratio, the proportion of effective time windows, task classification accuracy, the number of manual readjustments, and the system's stable runtime. The experimental results are shown in Table 1.

[0143] Table 1. Performance Comparison of the Invention Method and Traditional Schemes in Long-Term EEG Monitoring Scenarios.

[0144] Comparison indicators Traditional statistical thresholding method Method of the present invention Average signal-to-noise ratio (dB) 9.8±1.4 15.9±1.8 Effective time window ratio (%) 71.6 92.4 Task classification accuracy (%) 78.3 89.7 Number of manual adjustments / 24h 5.6 1.1 Duration of continuous stable operation (h) 18.2 46.5

[0145] As shown in Table 1, this invention improves the average signal-to-noise ratio by approximately 62% in long-term operation scenarios, increases the effective time window ratio by over 20 percentage points, and improves task classification accuracy by over 11%. Simultaneously, the number of manual readjustments is significantly reduced, and the continuous stable operation time is more than twice that of traditional methods. In scenarios where some electrodes are severely loose or sweat accumulates, the proportion of channels that become unusable for more than 15 minutes in traditional methods reaches 18%, while this proportion is reduced to below 3% using the method of this invention. This indicates that the proposed online identification and closed-loop adjustment mechanism for link parameters can maintain high acquisition stability in complex environments.

[0146] The results of the above implementation show that the present invention can effectively solve the problems of difficulty in modeling acquisition link drift, unquantifiable electrode contact changes, unstable inverse compensation, and lack of hardware closed-loop control in real clinical continuous monitoring scenarios. It enables the EEG acquisition system to have online calibration, dynamic compensation, self-evaluation and adaptive adjustment capabilities, thereby significantly improving the quality of EEG data, the reliability of analysis and the engineering practicality of long-term system operation.

[0147] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based method for EEG acquisition and analysis, characterized in that, Includes the following steps: Step 1: Acquire multi-channel raw EEG sequences, divide them into windowed EEG matrices according to preset time windows, and set calibration time slots to form calibration time slot EEG segments; Step 2: Obtain the set of controllable parameters to generate a calibration sequence, and write the calibration sequence into the acquisition terminal register within the calibration time slot; Step 3: Encode and splice the calibration time slot EEG segments and calibration sequence, divide them into data blocks, construct input label sequences, and input them into an improved FNet model. The improved FNet model introduces a pole-zero stability domain parameter output mechanism to form the transfer function parameters and contact state vector of each channel acquisition link. Step 4: Perform inverse system reconstruction operation on the window EEG matrix based on the acquisition link transfer function parameters and apply stability constraints to form a reconstructed EEG sequence; Step 5: Based on the reconstructed EEG sequence and contact state vector, an artifact type identifier and channel-time mask are formed, and suppression and interpolation operations are performed on the reconstructed EEG sequence to form a clean EEG sequence; Step 6: Perform task analysis processing on the clean EEG sequence to form task inference results, and calculate the joint reliability index based on the task inference results, contact state vector and acquisition link transfer function parameters; Step 7: Based on the joint credibility index, contact state vector and acquisition link transfer function parameters, a set of acquisition control instructions is formed and sent to the acquisition terminal.

2. The deep learning-based EEG acquisition and analysis method according to claim 1, characterized in that, Step one specifically involves: A multi-channel raw EEG sequence with a sampling rate of a preset value is acquired. The multi-channel raw EEG sequence includes the channel identifier of each acquisition channel and the amplitude data at the corresponding sampling time. The multi-channel raw EEG sequence is subjected to sliding segmentation processing according to preset time window segmentation parameters. The preset time window segmentation parameters include time window length and window shift length. The window shift length is the number of sampling points between the starting sampling points of adjacent time windows. The sliding segmentation processing includes determining the starting sampling point and ending sampling point of each time window based on the time window length, and determining the interval between the starting sampling points of adjacent time windows based on the window shift length. The amplitude data corresponding to each acquisition channel within each time window are arranged and spliced ​​according to the channel order to form a window EEG matrix. The rows of the window EEG matrix correspond to the acquisition channels, and the columns correspond to the sampling point numbers within the time window. Within each time window, the start and end sampling points of the calibration time slot are determined based on the calibration time slot parameters, which include the calibration time slot length and the relative position of the calibration time slot within the time window. Amplitude data from each acquisition channel within the calibration time slot range are extracted from the window EEG matrix to form a calibration time slot EEG segment.

3. The deep learning-based EEG acquisition and analysis method according to claim 1, characterized in that, Step two specifically involves: The controllable parameter set is read based on the analog front-end circuit of the acquisition terminal and the digital control module. The controllable parameter set includes amplifier gain control quantity, reference electrode switching control quantity, input impedance switching control quantity, notch filter center frequency control quantity, and high-pass filter cutoff frequency control quantity. Based on the spatial position of the acquisition channel in the electrode cap, the acquisition channel is divided into multiple channel groups, and different parameter combination excitation templates are assigned to each channel group. The parameter combination excitation template is formed by combining two or more types of parameters in the controllable parameter set according to different parameters on the time axis using mutually unrelated switching cycle or amplitude encoding rules. For each parameter combination excitation template, a discrete state sequence is constructed. The discrete state sequence is arranged on the time axis according to a preset switching period, and uncorrelated change trajectories are generated based on the change magnitude constraints of different parameter dimensions. Piecewise linear interpolation is performed on the discrete state sequence to form a continuously changing multi-parameter control trajectory, and quantization encoding is performed on the multi-parameter control trajectory to generate a calibration sequence containing the relationship between the changes of each control parameter over time. The multidimensional calibration sequence is converted into the corresponding front-end circuit control word sequence according to the channel group mapping relationship and written into the acquisition end register during the calibration time slot.

4. The deep learning-based EEG acquisition and analysis method according to claim 1, characterized in that, Step three specifically involves: The calibration time-slot EEG segments and calibration sequence encodings are processed by channel-dimensional splicing and time-dimensional block segmentation to form the input labeled sequence; The improved FNet model includes a Fourier labeling mixture module, a feedforward mapping module, a pole-zero stability domain parameter output structure, and a contact state parameter output structure; The Fourier label mixing module performs discrete Fourier transform on the input label sequence in the label dimension and feature dimension respectively, preserving the real and imaginary parts of the complex features to form a frequency domain mixed label; The feedforward mapping module performs linear and nonlinear mapping processing on the frequency domain hybrid marker to form a channel feature vector. The pole-zero stability domain parameter output structure generates pole radius parameters, pole angle parameters, zero radius parameters, and zero angle parameters for each acquisition channel based on the channel feature vector. The pole radius parameters are restricted to the range of greater than 0 and less than 1, the pole angle parameters are restricted to the range of negative π to positive π, the zero radius parameters are restricted to the range of greater than 0 and less than 1, and the zero angle parameters are restricted to the range of negative π to positive π. A set of complex poles is constructed based on the pole radius parameters and pole angle parameters, and a set of real coefficient poles is generated according to the conjugate pairing rule; A complex set of zeros is constructed based on the zero radius parameter and the zero angle parameter, and a set of real coefficient zeros is generated according to the conjugate pairing rule; A denominator polynomial is generated based on the set of poles, a numerator polynomial is generated based on the set of zeros, and a transfer function for each acquisition link is constructed based on the numerator and denominator polynomials. The contact state parameter output structure generates a multi-dimensional contact state parameter set based on the channel feature vector. The multi-dimensional contact state parameter set includes the channel contact impedance change, electrode polarization amplitude index, signal saturation ratio index, and baseline drift amplitude index. The multi-dimensional contact state parameter set is arranged according to the channel index to form a contact state vector.

5. The deep learning-based EEG acquisition and analysis method according to claim 1, characterized in that, Step four specifically involves: Based on the transfer function parameters of each channel acquisition link, frequency domain transformation is performed on the window EEG matrix to obtain the frequency domain representation of each channel; In the frequency domain representation, an inverse system operator is constructed for each frequency point. The inverse system operator is constructed by taking the conjugate of the transfer function parameters at the corresponding frequency point and superimposing a positive stable term formed by a preset constant and the square of the magnitude of the transfer function parameters in the denominator. Based on the contact state vector and the mean and variance of the power spectrum of the EEG segment in the calibration time slot at each frequency point, the inverse system gain constraint threshold is calculated for each channel and each frequency point. The inverse system gain constraint threshold is obtained by normalizing the channel contact impedance change, electrode polarization amplitude index, signal saturation ratio index and baseline drift amplitude index respectively, and then weighting and summing them according to preset weight coefficients. When the amplitude of the inverse system operator exceeds the corresponding inverse system gain constraint threshold, the amplitude of the inverse system operator is truncated to the inverse system gain constraint threshold, while keeping the phase component unchanged, thus forming a constrained inverse system operator; The constrained inverse system operator is multiplied with the frequency domain representation point by point to form a compensated frequency domain signal, and an inverse frequency domain transformation is performed on the compensated frequency domain signal to form a reconstructed EEG sequence.

6. The deep learning-based EEG acquisition and analysis method according to claim 1, characterized in that, Step five specifically involves: Based on the reconstructed EEG sequence, the root mean square value of amplitude, short-time energy, kurtosis coefficient and baseline drift are calculated for each acquisition channel in the time dimension, and the power spectral density corresponding to each preset frequency band is calculated in the frequency domain. The root mean square value of amplitude, short-time energy, kurtosis coefficient, baseline drift and contact state vector are concatenated by channel index to form a set of artifact discrimination features. Based on the set of artifact discrimination features, a threshold judgment is performed on each time segment to form a corresponding artifact type identifier. Based on the artifact type identifier, a binary or multi-valued channel-time mask is generated in the channel dimension and the time dimension to identify the effective and abnormal regions of each acquisition channel in each time segment. Based on the channel-time mask, amplitude attenuation processing is performed on reconstructed EEG sequence samples in abnormal regions, time domain interpolation processing is performed on samples in continuous abnormal regions, and spatial neighborhood interpolation processing is performed on samples in cross-channel abnormal regions. The samples from each channel, after amplitude attenuation and interpolation processing, are recombined to form a clean EEG sequence.

7. The deep learning-based EEG acquisition and analysis method according to claim 1, characterized in that, Step six specifically involves: The clean EEG sequences were arranged into a time series matrix according to channel order, and corresponding analysis segments were extracted according to the preset time window index. In each acquisition channel, a fast Fourier transform is performed on the analysis segment, the frequency domain amplitude is squared and averaged according to the frequency band boundaries to form the power spectral density corresponding to each preset frequency band. In the time domain, a fixed-length sliding window is used to calculate the root mean square value of the analysis segment, and a fixed-dimensional set of phase space embedding vectors is formed by continuous sampling points. The matching ratio is statistically analyzed based on the similarity threshold to form a sample entropy index. The power spectral density, root mean square value and sample entropy index corresponding to each channel are concatenated according to the channel index to form an analysis feature vector, and then input into the task inference module to perform linear mapping and nonlinear mapping operations to form the task inference result; Based on the category probability distribution of the task reasoning results, the logarithm of each category probability is taken and multiplied by the corresponding probability, and then summed to form the information entropy value; The mean and variance statistics of the contact state vector are calculated separately for each channel dimension. The reciprocal of the amplitude of the transfer function parameter of the acquisition link is calculated at each frequency point. The reciprocal of the amplitude is then averaged in the frequency dimension to form the frequency domain condition number index. The reciprocal of the information entropy value, the mean and variance statistics, and the frequency domain condition number index are mapped to the 0-1 interval, and weighted and accumulated according to the preset weight coefficients to form a joint credibility index.

8. The deep learning-based EEG acquisition and analysis method according to claim 1, characterized in that, Step seven specifically involves: The joint credibility index is compared with a preset trigger threshold. When the joint credibility index is lower than the trigger threshold, the calculation step of adjusting the acquisition parameters is initiated. The priority adjustment level of the corresponding acquisition channel is formed by weighting and sorting the normalized values ​​of the contact impedance change, electrode polarization amplitude index, signal saturation ratio index and baseline drift amplitude index of each channel in the contact state vector. Based on the difference between the amplitude of each frequency point in the acquisition link transfer function parameters and the preset reference spectrum, determine the adjustment direction of the amplifier gain, the center frequency of the notch filter, and the cutoff frequency of the high-pass filter. Based on the trend of the joint credibility index change corresponding to the most recent preset number of time windows, the priority adjustment level and the adjustment direction, the adjustment range of each acquisition parameter is calculated. The acquisition parameters include amplifier gain, input impedance level, notch filter center frequency, high-pass filter cutoff frequency and reference electrode switching period. The adjustment range is quantified and superimposed with the current collected parameter value to form an updated set of collected parameters. The updated set of acquisition parameters is encoded into a set of acquisition control instructions and written into the acquisition terminal register according to the channel mapping relationship; At the starting sampling point of the next time window, the acquisition control command set is activated and sent to the acquisition terminal, and the corresponding parameter update record is written to the log unit.

Images