Electroencephalogram signal processing method and device and brain electrode cap

Abstract

The invention discloses an electroencephalogram signal processing method and device and a brain electrode cap, relates to the technical field of brain-computer interfaces, and mainly aims to solve the problems of poor processing efficiency and inaccuracy of an existing electroencephalogram signal. The method comprises the following steps: when a force feedback sensor determines that a pressure value between an electrode plate and a detection part is matched with a preset pressure threshold value, collecting an electroencephalogram signal of the detection part through the electrode plate, the force feedback sensor is used for driving the motor and the pneumatic push rod in the detection body to adjust the contact angle and the contact strength between the electrode plate and the detection part when the pressure value is not matched with the preset pressure threshold value; feature extraction is carried out on residual electroencephalogram components of the electroencephalogram signals based on a deep residual shrinkage network model, and feature information and main components of the electroencephalogram signals are reconstructed; and performing uniform scale decomposition on the multi-channel electroencephalogram signals of the reconstructed electroencephalogram signals to obtain sub-band features, and performing view feature processing based on the sub-band features to obtain a multi-view feature set.

Description

Technical Field

[0001] The present invention relates to the technical field of brain-computer interface, and in particular to a method and device for processing electroencephalogram (EEG) signals. Background Art

[0002] The brain-computer interface is a signal exchange method that communicates with the outside world through brain nerve activity. The user's intention of the brain-computer interface is identified and judged by analyzing brain signals and identifying and analyzing them, so as to realize interactive processing of the external environment. Among them, since it is difficult to receive EEG signals, there are two existing ways to record EEG signals, namely implanted electrodes and scalp electrodes. Among them, since scalp electrode detection is non-invasive and the detection method is simple, scalp electrical detection is more widely used.

[0003] At present, the existing EEG equipment using scalp electrodes collects EEG signals by wearing an electrode cap on the patient's head, collecting EEG signals by contacting the electrode sheet with the patient's head, and performing signal analysis and processing based on the EEG signals, such as spectrum analysis. However, due to the fact that the head shapes of different patients are very different, the electrode sheet fixed on the electrode cap will not fit the patient's head tightly, resulting in a large difference between the collected EEG signals and the real brain reflection information. At the same time, simply performing spectrum analysis on the EEG signals cannot accurately obtain brain activity information. Therefore, a method for processing EEG signals is urgently needed to solve the above problems. Summary of the invention

[0004] In view of this, the present invention provides a method and device for processing EEG signals and an electrode cap, the main purpose of which is to solve the problems of poor efficiency and inaccuracy in existing EEG signal processing.

[0005] According to one aspect of the present invention, a method for processing an electroencephalogram signal is provided, comprising:

[0006] When the force feedback sensor determines that the pressure value between the electrode sheet and the detection part matches the preset pressure threshold, the electroencephalogram signal of the detection part is collected by the electrode sheet, and the pressure value is collected by the force feedback sensor in the detection body. The force feedback sensor is also used to drive the motor and the pneumatic push rod in the detection body to adjust the contact angle and contact strength between the electrode sheet and the detection part when the collected pressure value does not match the preset pressure threshold, until the pressure value matches the preset pressure threshold;

[0007] Extracting features of the residual EEG components of the EEG signal based on the deep residual shrinkage network model for which model training has been completed, and reconstructing the feature information after feature extraction with the main components of the EEG signal to obtain a reconstructed EEG signal;

[0008] Uniform scale decomposition is performed on the multi-channel EEG signal of the reconstructed EEG signal to obtain a decomposition result, and view feature processing is performed based on the decomposition result to obtain a multi-view feature set.

[0009] Furthermore, before extracting features of the residual EEG components of the EEG signal based on the deep residual shrinkage network model that has completed model training, the method further includes:

[0010] Performing singular spectrum analysis on the EEG signal to obtain multiple signal sub-components;

[0011] A main component is determined from the multiple signal subcomponents, and the main component is deleted from the EEG signal to obtain a residual EEG component.

[0012] Further, determining the main component from the plurality of signal subcomponents comprises:

[0013] Extracting relevant sub-components from the plurality of signal sub-components based on a preset autocorrelation coefficient threshold, and reconstructing the relevant sub-components to obtain the main component;

[0014] The signal subcomponents include frequency domain subcomponents, time domain subcomponents, space subcomponents, physiological and artifact shadow components, equipment artifact signal subcomponents, and nonlinear subcomponents.

[0015] Furthermore, the feature information after feature extraction is reconstructed with the main components of the EEG signal to obtain the reconstructed EEG signal, including:

[0016] Obtaining reconstruction weights, wherein the reconstruction weights include subject weights and feature weights;

[0017] Performing weighted sum calculation based on the reconstruction weight, the main component, and the feature information to obtain a reconstructed EEG signal;

[0018] The subject weight is determined based on the signal variance of the subject component and the signal variance of the feature information, and the sum of the feature weight and the subject weight is 1.

[0019] Furthermore, the main components include frequency main components, time domain main components, and space main components, and the weighted and calculated based on the reconstruction weights, the main components, and the feature information to obtain the reconstructed EEG signal includes:

[0020] Determine a first main body weight and a first characteristic weight of the main component in the frequency domain, and perform weighting and calculation based on the first main body weight, the first characteristic weight, the main component, and the characteristic information to obtain a first reconstructed EEG signal, wherein the first characteristic weight is a frequency domain dynamic weight function;

[0021] Determine a second subject weight and a second feature weight of the time domain subject component, and perform weighting and calculation based on the second subject weight, the second feature weight, the subject component, and the feature information to obtain a second reconstructed EEG signal, wherein the second subject weight is 1, and the second feature weight is a time-varying gain function;

[0022] Determine the third subject weight and the third feature weight of the spatial subject component, and perform weighting and calculation based on the third subject weight, the third feature weight, the subject component, and the feature information to obtain a third reconstructed EEG signal.

[0023] Furthermore, before extracting features of the residual EEG components of the EEG signal based on the deep residual shrinkage network model that has completed model training, the method further includes:

[0024] Constructing a deep residual shrinkage network, wherein the input of the deep residual shrinkage network is a multi-scale input, and a filter unit, an attention mechanism unit, and a recurrent unit are embedded in the deep residual shrinkage network model;

[0025] The deep residual shrinkage network model is trained based on the residual EEG feature training samples to obtain a deep residual shrinkage network model, wherein the dynamic threshold in the deep residual shrinkage network model is used to eliminate signal artifacts and noise.

[0026] Furthermore, before uniformly scaling the multi-channel EEG signal of the reconstructed EEG signal, the method further comprises:

[0027] Determine the mutual information between the reconstructed EEG signal and the standard EEG signal, and screen the multi-channel EEG signal of the reconstructed EEG signal based on the mutual information, wherein the mutual information is used to characterize the degree of correlation between the reconstructed EEG signal and the standard EEG signal.

[0028] Furthermore, after screening the multi-channel EEG signal of the reconstructed EEG signal based on the mutual information, the method further comprises:

[0029] The multi-channel EEG signal is filtered based on a spatial filter to perform uniform scale decomposition based on the filtered multi-channel EEG signal, and a weight matrix in the spatial filter is determined based on a deep learning network and a common spatial pattern.

[0030] Furthermore, the uniform-scale decomposition of the multi-channel EEG signal of the reconstructed EEG signal is performed to obtain a decomposition result, which includes:

[0031] Performing modal decomposition on the multi-channel EEG signal to obtain a modal decomposition result, and screening out a subband set from the modal decomposition result;

[0032] Calculating the energy of each subband in the subband set, and determining the optimal subband based on the dynamic correlation weight and the subband energy;

[0033] Based on the multi-layer convolutional model that has completed model training, feature extraction is performed on the optimal sub-band to obtain a decomposition result of the sub-band feature.

[0034] Further, performing view feature processing based on the decomposition result to obtain a multi-view feature set includes:

[0035] The decomposition results are feature classified according to feature types, and a view set of the classified decomposition results is generated according to time, frequency, and time domain. The feature types include linear type, nonlinear type, and energy type.

[0036] According to another aspect of the present invention, there is provided a device for processing an electroencephalogram signal, comprising:

[0037] an acquisition module, configured to acquire an electroencephalogram signal of the detection part through the electrode sheet when the force feedback sensor determines that the pressure value between the electrode sheet and the detection part matches a preset pressure threshold, wherein the pressure value is acquired by the force feedback sensor in the detection body, and the force feedback sensor is further configured to drive a motor and a pneumatic push rod in the detection body to adjust a contact angle and a contact strength between the electrode sheet and the detection part when the acquired pressure value does not match the preset pressure threshold, until the pressure value matches the preset pressure threshold;

[0038] An extraction module, used for extracting features of the residual EEG components of the EEG signal based on the deep residual shrinkage network model that has completed model training, and reconstructing the feature information after feature extraction with the main components of the EEG signal to obtain a reconstructed EEG signal;

[0039] The processing module is used to perform uniform scale decomposition on the multi-channel EEG signal of the reconstructed EEG signal to obtain a decomposition result, and perform view feature processing based on the decomposition result to obtain a multi-view feature set.

[0040] Furthermore, the device also includes:

[0041] The determination module is used to perform singular spectrum analysis on the EEG signal to obtain multiple signal sub-components; determine the main component from the multiple signal sub-components, and delete the main component from the EEG signal to obtain a residual EEG component.

[0042] Furthermore, the determination module is specifically used to extract correlated sub-components from the multiple signal sub-components based on a preset autocorrelation coefficient threshold, and reconstruct the correlated sub-components to obtain the main components; wherein the signal sub-components include frequency domain sub-components, time domain sub-components, spatial sub-components, physiological and artifact components, equipment artifact signal sub-components, and nonlinear sub-components.

[0043] Furthermore, the extraction module is specifically used to obtain reconstruction weights, which include subject weights and feature weights; weights and calculations are performed based on the reconstruction weights, the subject components, and the feature information to obtain reconstructed EEG signals; wherein the subject weights are determined based on the signal variance of the subject components and the signal variance of the feature information, and the sum of the feature weights and the subject weights is 1.

[0044] Furthermore, the main components include frequency main components, time domain main components, and space main components. The extraction module is specifically used to determine the first main weight and the first characteristic weight of the frequency domain main component, and perform weights and calculations based on the first main weight, the first characteristic weight, the main component, and the characteristic information to obtain a first reconstructed EEG signal, and the first characteristic weight is a frequency domain dynamic weight function; determine the second main weight and the second characteristic weight of the time domain main component, and perform weights and calculations based on the second main weight, the second characteristic weight, the main component, and the characteristic information to obtain a second reconstructed EEG signal, and the second main weight is 1, and the second characteristic weight is a time-varying gain function; determine the third main weight and the third characteristic weight of the spatial main component, and perform weights and calculations based on the third main weight, the third characteristic weight, the main component, and the characteristic information to obtain a third reconstructed EEG signal.

[0045] Furthermore, the device also includes:

[0046] A training module is used to construct a deep residual shrinkage network, the input of the deep residual shrinkage network is a multi-scale input, and the deep residual shrinkage network model is embedded with a filtering unit, an attention mechanism unit, and a loop unit; the deep residual shrinkage network model is trained based on the residual EEG feature training samples to obtain a deep residual shrinkage network model, wherein the dynamic threshold in the deep residual shrinkage network model is used to eliminate signal artifacts and noise.

[0047] Furthermore, the determination module is also used to determine the mutual information between the reconstructed EEG signal and the standard EEG signal, and to screen the multi-channel EEG signal of the reconstructed EEG signal based on the mutual information, and the mutual information is used to characterize the degree of correlation between the reconstructed EEG signal and the standard EEG signal.

[0048] Furthermore, the processing module is also used to filter the multi-channel EEG signal based on a spatial filter to perform uniform scale decomposition based on the filtered multi-channel EEG signal, and the weight matrix in the spatial filter is determined based on a deep learning network and a common spatial pattern.

[0049] Furthermore, the processing module is specifically used to perform modal decomposition on the multi-channel EEG signal to obtain a modal decomposition result, and filter out a subband set from the modal decomposition result; calculate the energy of each subband in the subband set, and determine the optimal subband based on the dynamic correlation weight and the subband energy; extract features of the optimal subband based on the multi-layer convolution model that has completed model training, and obtain a decomposition result of the subband features.

[0050] Furthermore, the processing module is specifically used to perform feature classification on the decomposition results according to feature types, and generate a view set of the decomposition results after classification according to time, frequency, and time domain, and the feature types include linear type, nonlinear type, and energy type.

[0051] According to another aspect of the present invention, there is provided a brain electrode cap, comprising: a cap body, a detection body, a signal processor,

[0052] The cap body is provided with a plurality of cavities, the end of the detection body is placed in the cavity, the detection body includes a force feedback sensor, a motor and a pneumatic push rod, and the hemispherical front end of the detection body is provided with an electrode sheet outside.

[0053] The force feedback sensor exchanges data with the motor and the pneumatic push rod to collect the pressure value between the electrode sheet and the detection part, and sends a control instruction to the motor and the pneumatic push rod when the pressure value does not match the preset pressure threshold;

[0054] The motor and the pneumatic push rod are used to adjust the contact angle and contact strength between the electrode sheet and the detection part based on the control instruction;

[0055] The signal processor is used to collect the EEG signal of the detection part through the electrode sheet when the force feedback sensor determines that the pressure value between the electrode sheet and the detection part matches the preset pressure threshold; perform feature extraction on the residual EEG component of the EEG signal based on the deep residual shrinkage network model that has completed model training, and reconstruct the feature information after feature extraction with the main component of the EEG signal to obtain a reconstructed EEG signal; perform uniform scale decomposition on the multi-channel EEG signal of the reconstructed EEG signal to obtain a decomposition result, and perform view feature processing based on the decomposition result to obtain a multi-view feature set.

[0056] Furthermore, the force feedback sensor is placed inside the hemispherical front end of the detection body, and the detection surface of the force feedback sensor is connected to the electrode sheet, a contraction groove is provided on the inner wall of the cap body adjacent to the detection body, and a connecting component is provided on the side of the detection body adjacent to the contraction groove, one side of the connecting component extends into the contraction groove and slides in the contraction groove, and a compression spring is fixedly connected between the side of the connecting component facing away from the detection body and the inner wall of the contraction groove.

[0057] Furthermore, a driving shaft is fixedly connected to the output shaft of the motor, the opposite end of the driving shaft rotates and passes through the contraction groove, a sleeve rod is slidably mounted on the outer surface of the driving shaft, two limiting grooves are provided on the inner wall of the sleeve rod, and a limiting slider is fixedly connected to the outer surface of the driving shaft, and the opposite side of the limiting slider extends into the limiting groove and slides in the limiting groove.

[0058] Furthermore, a control cavity is opened in the cap body, the pneumatic push rod and the motor are fixedly installed in the control cavity, and one end of the telescopic rod of the pneumatic push rod slides through the contraction groove and contacts the connecting cap body.

[0059] Furthermore, a connecting column is fixedly connected to the side of the connecting cap body facing the detection body, a movable groove is opened on the surface of the side of the detection body connected to the cap body, a movable ball is arranged in the movable groove, the movable ball rotates in the movable groove, one side of the movable ball extends out of the movable groove and is fixedly connected to the connecting column, a plurality of control rods arranged in a circular array are provided in the connecting cap body, one end of the control rod slides through the connecting cap body, the end of the control rod located outside the connecting cap body is fixedly connected to a contact ball, the contact ball contacts the surface of the detection body, the end of the control rod located inside the connecting cap body is fixedly connected to an extrusion body, a support spring is fixedly connected between the extrusion body and the inner wall of the connecting cap body, and the support spring is sleeved on the control rod.

[0060] Furthermore, a rotating plate is provided in the connecting cap body, and the rotating plate rotates in the connecting cap body. A sleeve rod is fixedly connected to a surface of the rotating plate on one side facing away from the control rod, and the other end of the sleeve rod rotates and passes through the outer surface of the connecting cap body.

[0061] Furthermore, a pushing block is fixedly connected to a surface of one side of the rotating plate facing the extrusion body, and the extrusion body contacts and slides with the inclined surface of the pushing block.

[0062] According to another aspect of the present invention, a storage medium is provided, wherein at least one executable instruction is stored in the storage medium, and the executable instruction enables a processor to execute an operation corresponding to the above-mentioned method for processing electroencephalogram signals.

[0063] According to another aspect of the present invention, there is provided a terminal, comprising: a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other via the communication bus;

[0064] The memory is used to store at least one executable instruction, and the executable instruction enables the processor to execute an operation corresponding to the above-mentioned method for processing EEG signals.

[0065] By means of the above technical solution, the technical solution provided by the embodiment of the present invention has at least the following advantages:

[0066] The present invention provides a method and device for processing electroencephalogram (EEG) signals. Compared with the prior art, in an embodiment of the present invention, when a force feedback sensor determines that a pressure value between an electrode sheet and a detection part matches a preset pressure threshold, the electroencephalogram (EEG) signal of the detection part is collected by the electrode sheet, and the pressure value is collected by the force feedback sensor in the detection body. The force feedback sensor is also used to drive a motor and a pneumatic push rod in the detection body to adjust a contact angle and a contact strength between the electrode sheet and the detection part when the collected pressure value does not match the preset pressure threshold, until the pressure value matches the preset pressure threshold. Based on a deep residual shrinkage network model for which model training has been completed, feature extraction is performed on the residual electroencephalogram (EEG) components of the EEG signal. The feature information after feature extraction is reconstructed with the main components of the EEG signal to obtain a reconstructed EEG signal; the multi-channel EEG signal of the reconstructed EEG signal is uniformly scaled to obtain a decomposition result, and view feature processing is performed based on the decomposition result to obtain a multi-view feature set, so as to achieve the purpose of accurately collecting EEG signals while adjusting the pressure between the electrode and the head in real time, ensuring that the collected EEG signals can better reflect brain activities. At the same time, after feature extraction of the collected EEG signals, they are reconstructed and then multi-scale decomposition of the reconstructed EEG signals is performed, so as to meet the analysis requirements of brain activity characteristics in different dimensions, greatly improve the processing accuracy of EEG signals, and realize accurate recognition of brain activities.

[0067] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0069] Figure 1 A flow chart of a method for processing an electroencephalogram signal provided by an embodiment of the present invention is shown;

[0070] Figure 2 A block diagram of a brain electrode cap provided by an embodiment of the present invention is shown;

[0071] Figure 3 A schematic diagram of a cap structure provided by an embodiment of the present invention is shown;

[0072] Figure 4 A schematic diagram of the structure of a detection body provided by an embodiment of the present invention is shown;

[0073] Figure 5 A schematic diagram of the structure of a rotating plate provided by an embodiment of the present invention is shown;

[0074] Figure 6 A block diagram showing a composition of an electroencephalogram signal processing device provided by an embodiment of the present invention is shown;

[0075] Figure 7 A schematic diagram of the structure of a terminal provided by an embodiment of the present invention is shown;

[0076] Among them, 1-electrode cap; 2-detection body; 3-electrode sheet; 4-force feedback sensor; 5-contraction groove; 6-connecting shell; 7-compression spring; 8-control chamber; 9-pneumatic push rod; 10-motor; 11-connecting column; 12-movable groove; 13-movable ball; 14-control rod; 15-contact ball; 16-extrusion body; 17-support spring; 18-drive shaft; 19-rotating plate; 20-sleeve rod; 21-limiting slide groove; 22-limiting slider; 23-pushing block. DETAILED DESCRIPTION

[0077] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0078] The embodiment of the present invention provides a method for processing an electroencephalogram signal. Figure 1 As shown, the method includes:

[0079] 101. When the force feedback sensor determines that the pressure value between the electrode sheet and the detection part matches a preset pressure threshold, the electroencephalogram signal of the detection part is collected through the electrode sheet.

[0080] In an embodiment of the present invention, the user whose brain electrical signal is to be collected collects the brain electrical signal by wearing a brain electrode cap. At this time, the brain electrode cap includes a cap body, a detection body, and a signal processor as the current step 101-103. Among them, the force feedback sensor collects the pressure value between the electrode sheet and the detection part (such as a certain collection point on the head), and compares it with the preset pressure threshold. At this time, in order to make the brain electrical signal clearly and accurately reflect the brain activity, the preset pressure threshold can be configured based on different collection requirements or human body characteristics. The preset pressure threshold can be a numerical range to avoid making the pressure value equal to or greater than the preset pressure threshold (pressure is too high), or less than the preset pressure threshold (pressure is too low), and the collected brain electrical signal is used as the basis for brain electrical signal processing. Furthermore, when the force feedback sensor determines that the pressure value matches the preset pressure threshold, the current execution end of the signal processor collects the brain electrical signal through the electrode sheet. In addition, the force feedback sensor is also used to drive the motor and pneumatic push rod in the detection body to adjust the contact angle and contact strength between the electrode sheet and the detection part when the collected pressure value does not match the preset pressure threshold, until the pressure value matches the preset pressure threshold, and the motor and pneumatic push rod are no longer driven to adjust the contact angle and contact strength.

[0081] It should be noted that a plurality of cavities are provided on the cap body of the brain electrode cap, and the end of the detection body is placed in the cavity. The detection body includes a force feedback sensor, a motor and a pneumatic push rod. An electrode sheet is arranged on the outside of the hemispherical front end of the detection body. The force feedback sensor interacts with the motor and the pneumatic push rod for data collection, and is used to collect the pressure value between the electrode sheet and the detection part, and sends a control instruction to the motor and the pneumatic push rod when the pressure value does not match the preset pressure threshold. The motor and the pneumatic push rod are used to adjust the contact angle and contact strength between the electrode sheet and the detection part based on the control instruction until the collected pressure value matches the preset pressure threshold.

[0082] 102. Perform feature extraction on the residual EEG components of the EEG signal based on the deep residual shrinkage network model that has completed model training, and reconstruct the feature information after feature extraction with the main components of the EEG signal to obtain a reconstructed EEG signal.

[0083] In an embodiment of the present invention, in order to extract part of the EEG signal that is useful for signal analysis, the current execution end first determines the residual EEG component of the EEG signal. At this time, the residual EEG component is the remaining component after the main component is subtracted from the collected EEG signal. Then, the deep residual shrinkage network model is used to extract features of the residual EEG component to obtain feature information. The feature information at this time is used to characterize the high-frequency or abnormal signal features with signal analysis significance in the residual EEG signal, including but not limited to abnormal waveform features, artifact features, EEG high-frequency features, and detail texture features. The embodiment of the present invention does not make specific limitations. Then, this feature information is reconstructed with the main component of the EEG signal to obtain the reconstructed EEG signal as the object for real EEG signal analysis. Among them, the main component is the content of the EEG signal related to the main brain activity to reflect the core signal of the functional activity of the brain. At this time, the main component includes the frequency main component, the time domain main component, and the space main component. The embodiment of the present invention does not make specific limitations.

[0084] 103. Perform uniform scale decomposition on the multi-channel EEG signal of the reconstructed EEG signal to obtain a decomposition result, and perform view feature processing based on the decomposition result to obtain a multi-view feature set.

[0085] In an embodiment of the present invention, in order to effectively analyze the EEG signal and thus display the functional activity in the user's brain, the current execution end first extracts the multi-channel EEG signal from the reconstructed EEG signal, and then performs uniform scale decomposition on the multi-channel EEG signal to obtain sub-band features. Among them, uniform scale decomposition is an adaptive, multi-scale, and intelligent decomposition method for multi-channel EEG signals, which is suitable for signal processing methods with high complexity. At this time, the characteristic changes of the EEG signal in different sub-bands or components after uniform scale decomposition can reflect the microscopic characteristics and potential significance of the signal brain activity, mainly including frequency characteristics, energy distribution, time domain fluctuations, nonlinear dynamics, and spatial distribution. The embodiment of the present invention does not make specific restrictions. In addition, in order to describe the characteristics of the signal in multiple dimensions, the current execution end performs view feature processing based on the decomposition result to obtain a multi-view feature set for comprehensively reflecting the details and laws in the EEG signal. Among them, the view feature processing includes view generation of features corresponding to linear type, nonlinear type, and energy type, thereby displaying multi-dimensional brain features.

[0086] In another embodiment of the present invention, for further definition and explanation, before the step of extracting features of the residual EEG components of the EEG signal based on the deep residual shrinkage network model that has completed model training, the method further includes:

[0087] Performing singular spectrum analysis on the EEG signal to obtain multiple signal sub-components;

[0088] A main component is determined from the multiple signal subcomponents, and the main component is deleted from the EEG signal to obtain a residual EEG component.

[0089] In order to improve the accuracy of EEG signals in reflecting brain activities, the current execution end first uses singular spectrum analysis to process the collected EEG signals before feature extraction to obtain multiple signal sub-components. At this time, the signal sub-components include frequency domain sub-components, time domain sub-components, space sub-components, physiological and pseudo shadow components, device artifact signal sub-components, and nonlinear sub-components. Specifically, singular spectrum analysis constructs a trajectory matrix based on the time series of EEG signals, and decomposes and reconstructs the trajectory matrix to obtain frequency domain sub-components, time domain sub-components, space sub-components, physiological and pseudo shadow components, device artifact signal sub-components, and nonlinear sub-components. Among them, the frequency domain subcomponents include: Delta waves (0.5-4Hz) related to deep sleep and brain damage, Theta waves (4-8Hz) related to light sleep, relaxation and attention fluctuations, Alpha waves (8-13Hz) reflecting the relaxation state when awake, Beta waves (13-30Hz) related to thinking activities, alertness and anxiety states, Gamma waves (>30Hz) related to advanced cognitive functions, memory processing and other activities, etc., the time domain subcomponents include: burst signals related to epilepsy or abnormal brain activity (such as spikes or instantaneous waveforms), background noise of non-electroencephalographic activity signals (low-amplitude baseline fluctuations), spatial distribution subcomponents include independent features such as visual cortex, motor cortex or frontal lobe activities, physiological and artifact shadow components include physiological signal artifacts such as electrocardiogram, electromyography or eye movement signals, such as device artifact signals introduced by poor device contact, power supply interference, etc., nonlinear subcomponents include chaotic EEG activities (such as epileptic discharges) or sudden strong non-periodic signals, which are not specifically limited in the embodiments of the present invention. Furthermore, a main component that mainly reflects the brain activity is determined from the multiple signal sub-components, and this part of the main component is deleted from the EEG signal to obtain a residual EEG component.

[0090] In another embodiment of the present invention, for further definition and illustration, the step of determining a main component from the plurality of signal subcomponents comprises:

[0091] The relevant sub-components are extracted from the plurality of signal sub-components based on a preset autocorrelation coefficient threshold, and the relevant sub-components are reconstructed to obtain the main component.

[0092] In order to improve the effectiveness of decomposition of EEG signals, thereby achieving the accuracy of reconstruction of EEG main components, so as to improve the accuracy of brain activity analysis, when determining the main components, the current execution end specifically first extracts relevant sub-components from multiple signal sub-components based on a pre-configured autocorrelation coefficient threshold. Among them, the autocorrelation coefficient is to calculate the correlation coefficient between two adjacent time series values, that is, to calculate the autocorrelation value between each two signal sub-components, and compare it with the preset autocorrelation coefficient threshold, and use the signal sub-component with the autocorrelation coefficient greater than the preset autocorrelation coefficient threshold as the relevant value component for reconstruction. Among them, since the signal sub-components include frequency domain sub-components, time domain sub-components, space sub-components, physiological and artifact components, device artifact signal sub-components, and nonlinear sub-components, when reconstructing after extracting the relevant sub-components, the reconstruction process of the frequency domain sub-components, time domain sub-components, space sub-components, and nonlinear sub-components includes decomposition and reconstruction, and the reconstruction process of the physiological and artifact components includes artifact removal and reconstruction.

[0093] In a specific implementation scenario, the reconstruction process of the frequency domain sub-components is as follows:

[0094] 1. Decomposition process: The EEG signal is x(t), which is the EEG signal x collected at time t. It can be decomposed by the frequency domain filter H ω (f) Decompose x(t) into signal value components in multiple frequency bands: x i (t) = F -1 {H ω (f)·F{x(t)}}, where F and F -1 are Fourier transform and inverse Fourier transform, respectively, ω (f) is a frequency selective filter, defined as: Each frequency value component x i (t) corresponds to Delta, Theta, Alpha, Beta or Gamma waves, f low ,f high are the lowest frequency and the highest frequency respectively, i=1,2...M, M is a positive integer.

[0095] 2. Reconstruction process: Reconstruct the complete frequency domain signal by weighted summation of sub-components in different frequency bands:

[0096] Among them, ω i is the weight of the frequency domain value component, which is dynamically adjusted according to specific tasks (such as sleep detection or cognitive status assessment), N bands is the number of frequency bands.

[0097] In a specific implementation scenario, the reconstruction process of the time domain sub-component is as follows:

[0098] 1. Decomposition process: EEG signals contain burst signals x spike (t) and background noise x noise (t), extract the burst signal through the sparse decomposition model: x(t) = x spike (t)+x noise (t); Burst signal extraction based on sparse representation optimization: Among them, ||·||1 is the sparse regularization term, which indicates the sparsity of the burst signal. is the reconstruction error constraint term.

[0099] 2. Reconstruction process: The burst signal and background noise are combined in a weighted manner during reconstruction: Weight ω spike and ω noise Dynamically adjust according to noise removal requirements.

[0100] In a specific implementation scenario, the reconstruction process of the spatial sub-components is as follows:

[0101] 1. Decomposition process: multi-channel EEG signal X∈R T×N (Number of time points T, number of channels N) can be decomposed into subcomponents related to different brain regions through spatial filtering: X space =W T X, where W∈R N×N is the spatial filter matrix, obtained by the common spatial pattern (CSP) algorithm, and the optimization goal is to maximize the difference between classes: S B and B W are the inter-class and intra-class covariance matrices of EEG signals, respectively.

[0102] 2. Reconstruction process: through the inverse filter matrix W -1 Reconstruct the spatial classification back into the original signal:

[0103]

[0104] In a specific implementation scenario, the reconstruction process of the nonlinear sub-component is as follows:

[0105] 1. Decomposition process: Use empirical mode decomposition (EMD) to decompose the signal into intrinsic mode classification (IMF): Among them, IMF k (t) is the modal classification of each order, r(t) is the residual term, where K is the selected modal classification set, k = 1, 2...K.

[0106] 2. Reconstruction process: Select the main modal classification for reconstruction:

[0107] In the specific implementation scenario, the reconstruction process of physiological and artifact shadow components is as follows:

[0108] 1. Decomposition process: The artifact signal A and the main component B of the EEG signal are separated by independent component analysis (ICA), which is expressed as X = M · [A, B], where M is a mixing matrix.

[0109] 2. Reconstruction process: Reconstruct by selecting the component of interest B and performing inverse matrix operations:

[0110]

[0111] In summary, combining the above reconstruction models of frequency domain, time domain, spatial distribution and nonlinear components, the weighted reconstruction of the signal is achieved through weighting: Each item can be dynamically weighted and adjusted to meet specific task requirements, and artifact removal operations are performed throughout the entire reconstruction process to ensure signal purity.

[0112] It should be noted that in a specific implementation scenario, the main components obtained by reconstructing the relevant sub-components are the frequency main components, the time domain main components, and the space main components. At this time, the frequency main components are the main components in the frequency domain, which are the main frequency components related to the characteristics of normal EEG waveforms, reflecting the basic activity state and main physiological functions of the brain, such as: Delta waves (0.5-4Hz) representing low-frequency states such as deep sleep and brain recovery, Theta waves (4-8Hz) representing attention fluctuations and mild relaxation states, Alpha waves (8-13Hz) representing relaxation waves in a wakeful resting state, Beta waves (13-30Hz) representing alertness, anxiety, and concentrated thinking, and Gamma waves (>30Hz) representing advanced cognitive and memory processing. Among these frequency components, the part related to the target task is preferably the main component. For example, if the goal is to analyze the attention level, Beta waves and Alpha waves may be the main components, and the embodiments of the present invention do not make specific limitations. The time domain main component is the main component in the time domain, which is manifested as a smooth, periodic or regular waveform in the time domain. These signal parts are related to specific brain activities (such as motor planning, visual response) or event-related potentials (ERP), which are not specifically limited in the embodiments of the present invention. The spatial main component is the main component in the sense of spatial distribution, which can be understood as the spatial features in the signal related to the activities of certain specific brain regions, for example, channel signals related to the motor cortex and visual cortex separated by spatial filtering (such as common spatial pattern, CSP) or independent component analysis (ICA), which are not specifically limited in the embodiments of the present invention.

[0113] In another embodiment of the present invention, for further definition and explanation, before the step of extracting features of the residual EEG components of the EEG signal based on the deep residual shrinkage network model that has completed model training, the method further includes:

[0114] Construct a deep residual shrinkage network;

[0115] The deep residual shrinkage network model is trained based on the residual EEG feature training samples to obtain a deep residual shrinkage network model.

[0116] In order to improve the accurate extraction of useful features in EEG signals, the current execution end pre-constructs a deep residual network, and uses residual EEG feature training samples to train the deep residual shrinkage network model to obtain a deep residual shrinkage network model. Among them, the input of the deep residual shrinkage network is a multi-scale input. The deep residual shrinkage network model is embedded with a filter unit, an attention mechanism unit, and a loop unit. The dynamic threshold in the deep residual shrinkage network model is used to eliminate signal artifacts and noise. Specifically, as a multi-scale input, time series features of different scales are simultaneously processed through multi-resolution convolution kernels (such as 3×3, 5×5, 7×7, etc.) to obtain a multi-scale residual signal, which is expressed as Among them, M represents the number of scales, k i is the convolution kernel size of different scales, Indicates feature concatenation. For the filtering unit, that is, a wavelet transform or filtering module is embedded in the network, the signal is decomposed into different frequency bands, processed separately and then merged, which can be expressed as N bands is the number of frequency bands, ω i is the weight for each frequency band. In addition, in order to solve the non-stationary and high noise characteristics of EEG signals, a dynamic shrinkage mechanism is used to adjust the local noise standard deviation. The adaptive shrinkage mechanism used is expressed as y i For network output, is the output of the network after contraction. The threshold τ(y) is determined by the statistical characteristics of the EEG signal. At this time, it can be automatically adjusted according to the local variance or energy of the signal. τ(y) = λ·σ(y), where λ is the scale factor and σ(y) is the standard deviation of the local noise. For the attention mechanism unit, in order to enhance the ability to extract signal detail features, the current execution end embeds the attention mechanism in the deep residual contraction network to weight key features: Q, K, V are query, key and value matrices respectively, d k is a dynamic threshold. At this point, the attention mechanism can locate key feature points in the EEG signal, such as sudden discharges, abnormal waveforms, etc. In addition, in order to use the temporal correlation of EEG signals for feature extraction, the current execution end embeds a recurrent unit in a deep residual contraction network (such as LSTM or Transformer) to model the dependency of the time series, expressed as: h t =σ(W x x t +W hh t-1 +b), h t is the hidden state at the current moment, x t is the input signal, W x is the input layer weight, W h is the cycle unit weight, and b is the bias term.

[0117] It should be noted that in order to carry out effective learning and extract features from the residual EEG components, at this time, the residual EEG feature training samples store high-frequency or abnormal features that need to be learned, including but not limited to abnormal waveforms (such as spike waves (Spike Wave) of epileptic patients, K complex waves and sleep spindles in the sleep stage), artifact features (such as high-frequency artifact features of electromyographic signals, periodic fluctuations caused by ECG signal interference), EEG high-frequency features (such as Gamma waves related to advanced cognitive and memory functions), non-linear features (such as chaotic characteristics of signals, non-periodic burst activities), signal samples corresponding to detail textures, etc., which are not specifically limited in the embodiments of the present invention.

[0118] In another embodiment of the present invention, for further definition and explanation, the step of reconstructing the feature information after feature extraction with the main components of the EEG signal to obtain the reconstructed EEG signal includes:

[0119] Obtain reconstruction weights;

[0120] The reconstruction weights are weighted and calculated based on the main components and the feature information to obtain a reconstructed EEG signal.

[0121] In order to reconstruct the EEG signal that can represent brain activity, taking into account the characteristics of time domain, frequency domain, spatial distribution, etc., to ensure that the reconstructed signal retains the main components and integrates the detailed features, while avoiding the reflux of artifacts or noise, the current execution end first obtains the reconstruction weights. At this time, the reconstruction weights include the main weights and feature weights, so as to perform weighting and calculation based on the reconstruction weights and the main components and feature information to obtain the reconstructed EEG signal. Specifically, x(t) is the computer signal, is the main component obtained by the main extraction module (such as singular spectrum analysis); The feature information extracted by the deep residual network, the reconstruction formula is: Among them, α is the main weight and β is the feature weight, which are used to balance the contribution of the main component and the feature information in the final signal. At this time, the main weight is determined based on the signal variance of the main component and the signal variance of the feature information, and the sum of the feature weight and the main weight is 1.

[0122] In another embodiment of the present invention, for further definition and explanation, the step of performing weighted sum calculation based on the reconstruction weight, the main component, and the feature information to obtain the reconstructed EEG signal includes:

[0123] Determine a first main body weight and a first characteristic weight of the main component in the frequency domain, and perform weighting and calculation based on the first main body weight, the first characteristic weight, the main component, and the characteristic information to obtain a first reconstructed EEG signal, wherein the first characteristic weight is a frequency domain dynamic weight function;

[0124] Determine a second subject weight and a second feature weight of the time domain subject component, and perform weighting and calculation based on the second subject weight, the second feature weight, the subject component, and the feature information to obtain a second reconstructed EEG signal, wherein the second subject weight is 1, and the second feature weight is a time-varying gain function;

[0125] Determine the third subject weight and the third feature weight of the spatial subject component, and perform weighting and calculation based on the third subject weight, the third feature weight, the subject component, and the feature information to obtain a third reconstructed EEG signal.

[0126] In order to improve the diversity and effectiveness of reconstructed EEG signals and meet the reconstruction of different main components, the current execution end can adjust the main weights and feature weights based on task requirements, such as denoising priority (for EEG signals with low signal-to-noise ratio, set α>β to rely more on the main components) and detail enhancement (corresponding to the scene that needs to identify abnormal or weak signals, such as epileptic seizure detection, set α<β to enhance feature details). Specifically, the weight adjustment method is β=1-α, where Var(·) represents the signal variance.

[0127] In a specific implementation scenario, for the reconstruction of the main components in the frequency domain, the main components and feature information are decomposed into the frequency domain and then combined according to the frequency bands: and It is the representation of main components and characteristic information in the frequency domain, H 主体 (f) and H 细节 (f) is the frequency domain weight function, which can be adjusted dynamically: H 细节 (f) = 1-H 主体 (f) It can highlight the main components in the low-frequency part and enhance the detailed features in the high-frequency part.

[0128] In a specific implementation scenario, for the reconstruction of the main components in the time domain, the points are added in the time domain. The second characteristic weight g(t) is a time-varying gain function, which is used to dynamically adjust the contribution of detail features. The second main weight is 1. For example, when the intensity of the characteristic information is high (such as a spike wave or an abnormal signal), the gain g(t) increases to amplify the detail features, while in the normal band, g(t) decreases to highlight the main components.

[0129] In a specific implementation scenario, for the reconstruction of the spatial main component, if the signal is a multi-channel EEG signal X∈R T×N , the combination of main components and feature information needs to be processed separately in the channel dimension and time dimension: The multi-channel signals can be further decomposed by independent component analysis (ICA) and the combined signals can be mapped to the spatial distribution related to the brain regions.

[0130] It should be noted that, in the embodiment of the present invention, in order to achieve flexible adjustment of the weight value, according to different task requirements, the current execution end can combine the introduction of a dynamic weighting strategy to adjust the weight in real time according to the local characteristics of the signal (such as signal-to-noise ratio, peak strength, etc.): β(t)=1-α(t), SNR(·) represents the signal-to-noise ratio, and the dynamic weight can more flexibly adapt to the local characteristics of the signal. At the same time, the attention mechanism can be introduced in the combination process to automatically locate the part that contributes more to the overall signal, such as, The attention weight is determined by the temporal and spatial characteristics of the signal, ensuring that the key features are strengthened. Finally, the quality of the reconstructed signal can be gradually improved through multiple iterations of combination and optimization: 1. The first-level signal obtained by preliminary combination 2. Further optimization using deep learning models: f refine (·) is a deep learning algorithm, which is not specifically limited in the embodiments of the present invention.

[0131] In another embodiment of the present invention, for further definition and explanation, before the step of uniformly scaling the multi-channel EEG signal of the reconstructed EEG signal, the method further includes:

[0132] The mutual information between the reconstructed EEG signal and the standard EEG signal is determined, and the multi-channel EEG signal of the reconstructed EEG signal is screened based on the mutual information.

[0133] In order to accurately know the correlation or association degree between the EEG signal and the standard signal, the mutual information between the real-time collected EEG signal and the standard signal is used as a screening indicator to more comprehensively capture the importance of the channel. The current execution end first obtains the mutual information. At this time, the mutual information is used to characterize the degree of association between the reconstructed EEG signal and the standard EEG signal, which is defined as: Among them, X and Y represent EEG signals and standard signals respectively, P(x,y) is the joint probability distribution, which indicates the probability of two variables appearing at the same time, and P(x) and P(y) are marginal probability distributions. At this time, mutual information measures the amount of information shared between X and Y. High mutual information indicates that the two are highly dependent and highly correlated, while low mutual information indicates that the two are almost independent.

[0134] It should be noted that when screening EEG channels based on mutual information, the EEG signal X after acquisition or reconstruction is calculated. i The mutual information I(X i ,Y), we can filter out the channels that are most relevant to the target task (such as cognitive state, motor intention, etc.): I(X i ,Y)>τ, i=1,2...,N, where τ is the mutual information threshold, which can be adjusted according to the specific application scenario, and N is the total number of channels. At this time, the channels that meet the conditions are retained, and the remaining channels are removed. At the same time, since EEG signals are often mixed with artifacts such as electromyography, electrocardiography, and environmental noise, by calculating mutual information, channels that are irrelevant to or have low dependence on standard signals can be effectively identified and removed. For example, channels with low mutual information are reflected as artifacts or invalid data, and channels with high mutual information are target EEG signals.

[0135] In another embodiment of the present invention, for further definition and explanation, after the step of screening the multi-channel EEG signal of the reconstructed EEG signal based on the mutual information, the method further includes:

[0136] The multi-channel EEG signal is filtered based on a spatial filter to perform uniform scale decomposition based on the filtered multi-channel EEG signal.

[0137] In order to improve the signal processing accuracy, the current execution end filters the multi-channel EEG signals to optimize the feature extraction and classification capabilities of the multi-channel EEG signals by enhancing the target signal in the spatial domain and suppressing noise and artifacts. Specifically, the multi-channel EEG signals can be filtered by a spatial filter. At this time, the weight matrix in the spatial filter is determined based on the deep learning network and the common spatial pattern. Specifically, the weight matrix W of the spatial filter is first determined, and the training goal is determined to maximize the distinction between task-related information and noise, and then the model training of the deep learning network is performed. The training steps include: 1. Calculate the covariance matrices C1 and C2 of each task category: 2. Jointly optimize the weight matrix W: At this time, the common spatial pattern CSP used is to extract linear spatial features that distinguish different tasks, so as to achieve the goal of maximizing the covariance difference between task categories. At the same time, when using a deep learning network, the filter weights are automatically learned using a deep neural network to capture more complex nonlinear spatial features. The specific steps include: the input of the deep neural network is a multi-channel EEG signal X, the hidden layer includes a convolutional layer for capturing local features between channels and an attention mechanism for dynamic weight allocation, and the output layer is the optimized spatial weight W deep .

[0138] In specific implementation scenarios, the filtering objects and targets are different for different principal component signals. For example, the filtering object of a multi-channel EEG signal is a multi-channel signal after mutual information screening, and the filtering target is to enhance the contribution of the target channel signal to suppress the impact on task-irrelevant or low-correlation channel signals. The filtering object of a brain region-specific signal is a signal from a specific brain region (such as an occipital lobe channel signal), and the filtering target is to extract the spatial characteristics of the target brain region to suppress cross-interference with signals from other brain regions. The filtering object of the time-space joint signal is the signal formed after spatial filter processing, which may be further input into the timing analysis module. The filtering target is both to optimize the spatial features and for subsequent time domain analysis (such as event-related potential extraction) to provide high-quality signals.

[0139] In another embodiment of the present invention, for further definition and explanation, the step of uniformly scaling the multi-channel EEG signal of the reconstructed EEG signal to obtain the decomposition result includes:

[0140] Performing modal decomposition on the multi-channel EEG signal to obtain a modal decomposition result, and screening out a subband set from the modal decomposition result;

[0141] Calculating the energy of each subband in the subband set, and determining the optimal subband based on the dynamic correlation weight and the subband energy;

[0142] Based on the multi-layer convolutional model that has completed model training, feature extraction is performed on the optimal sub-band to obtain a decomposition result of the sub-band feature.

[0143] In order to improve the comprehensive capture of time domain, frequency domain and other features in EEG signals, the modal decomposition method is used to process multi-channel EEG signals to filter out sub-band sets from the modal decomposition results. When performing modal decomposition on multi-channel EEG signals, specifically, the pre-processed multi-channel EEG signals x filtered (t) Perform wavelet decomposition and empirical mode decomposition simultaneously to form a multi-modal decomposition result. During wavelet decomposition, a compactly supported wavelet (such as Daubechies wavelet) that is sensitive to the EEG signal is determined, and the signal is decomposed into J layers to obtain a subband D as the detail component.j (t) and the approximate component A J (t): j=1,2...J. At this time, the frequency band range corresponding to each layer component is: Among them, f s is the sampling frequency. At the same time, the EMD first adaptively decomposes the EEG signal into a set of intrinsic mode components (IMFs) and residuals: At this time, the frequency range estimation estimates the instantaneous frequency f of each IMF through Hilbert transform IMFk After decomposition, the subband D of wavelet decomposition j (t) IFM decomposed with IMF k (t) Merge to form a collection of modal decomposition results: F = {D1, ..., D J ,IFM1,...,IFM k}. Then, the subband set f is selected from the obtained modal decomposition results. i Specifically, a dynamic optimization mechanism is introduced. First, the energy of each subband in the subband set is calculated, and the optimal subband is determined based on the dynamic correlation weight and the subband energy. Specifically, for each subband f i Calculating Energy High-energy subbands usually contain important features of the target signal. At the same time, the task relevance weight is introduced, and the relevance of each subband to the task is calculated in combination with the task goal (such as motion intention decoding, epilepsy feature detection, etc.), which is used as the dynamic relevance weight Task is the task relevance weight, which is used to calculate the task relevance of each subband and help select the optimal subband. At this time, the contribution of different subbands can be weighted according to specific tasks (such as motion intention decoding or epilepsy detection, etc.). The dynamic optimization mechanism based on energy and task relevance weights enhances system adaptability and ensures the efficiency and accuracy of subband selection. When selecting the optimal subband, N optimal subbands are dynamically selected according to real-time task requirements and computing resource constraints: F selected =Top-N{ω i ·E i}, Top is the optimal subband set selected by the dynamic optimization mechanism. After the optimal subband is obtained, in order to improve the decomposition accuracy, realize real-time reconstruction and noise suppression, the optimal subband is extracted based on the multi-layer convolution model that has completed model training, and the decomposition result of the subband feature is obtained. At this time, the optimal subband F selected Input into the deep neural network DNN, at this time, the hidden layer of the deep neural network DNN is used to capture the local spatiotemporal features between sub-bands to output the reconstructed optimized EEG signal as the decomposition result of the sub-band features, that is,

[0144] In another embodiment of the present invention, for further definition and explanation, the step of performing view feature processing based on the decomposition result to obtain a multi-view feature set includes:

[0145] The decomposition results are feature classified according to feature types, and a view set of the classified decomposition results is generated according to time, frequency, and time domain.

[0146] In order to show the different properties of different EEG signals, thereby describing the characteristics of EEG signals, so as to fully reflect the details and laws of brain activities in EEG signals, the current execution end performs view feature processing on the decomposition results to obtain a multi-view feature set. Specifically, the decomposition results are classified according to the feature type. Among them, the feature type includes linear type, nonlinear type, and energy type. The linear type refers to the classification of linear features, that is, extracting statistical indicators based on the linear model of the signal to reflect the mean, variance, covariance and other properties of the signal, including time domain linear features (such as mean, variance, autocorrelation function, root mean square value), frequency domain linear features (such as power spectrum density), covariance matrix, etc., which are not specifically limited in the embodiment of the present invention. The nonlinear type refers to the classification of nonlinear features, that is, based on the complexity, chaos and dynamic change mode of the signal, which is suitable for processing the non-stationary and nonlinear characteristics of EEG signals, including entropy-related features (such as sample entropy, approximate entropy, etc.), fractal dimension, chaotic features, etc., which are not specifically limited in the embodiment of the present invention. Energy type refers to the classification of energy distribution in different time periods or frequency ranges of the signal reflected by energy characteristics, which is used to characterize signal strength and spectral characteristics, including total energy, sub-band energy, instantaneous energy, energy distribution diagram, etc., which are not specifically limited in the embodiments of the present invention. Furthermore, a view set of the decomposition results after classification is generated according to time, frequency, and time domain. At this time, the view set comprehensively describes the image content of the EEG signal through time, frequency, and space dimensions, and visualizes the characteristics of the EEG signal as a set of multimodal view features. Among them, the content of the view feature may include a time view, a frequency view, and a space view, that is, the time view is to extract time domain features (such as mean, variance, amplitude) to reflect the global trend and time series characteristics of the signal; the frequency view is to extract frequency domain features (such as power spectrum density, frequency band energy) to describe the distribution characteristics of the signal at different frequencies; the spatial view is to extract spatial features of multi-channel signals (such as covariance matrix, common spatial pattern features) to reflect the functional distribution of different brain regions; the nonlinear view is to extract chaos and complexity features (such as sample entropy, fractal dimension) to describe the non-stationary dynamics of the signal; the joint view is to integrate multiple views through feature fusion technology (such as deep learning, dimensionality reduction method) to form a unified feature expression, which is not specifically limited in the embodiment of the present invention.

[0147] The embodiment of the present invention provides a method for processing EEG signals, which achieves the purpose of accurately collecting EEG signals while adjusting the pressure between the electrode sheet and the head in real time, ensuring that the collected EEG signals can better reflect brain activities. At the same time, the collected EEG signals are feature extracted and then reconstructed, and then the reconstructed EEG signals are decomposed at multiple scales, thereby meeting the analysis requirements of brain activity characteristics in different dimensions, greatly improving the processing accuracy of EEG signals, and realizing accurate identification of brain activities.

[0148] Furthermore, as a response to the above Figure 1 To implement the method shown in FIG. 1 , an embodiment of the present invention provides a brain electrode cap, such as Figure 2 As shown, it includes: a cap body 1, a detection body 2, a signal processor 24,

[0149] The cap body 1 is provided with a plurality of cavities, the end of the detection body 2 is placed in the cavity, the detection body 2 comprises a force feedback sensor 4, a motor 10 and a pneumatic push rod 9, and an electrode sheet 3 is arranged outside the hemispherical front end of the detection body 2.

[0150] The force feedback sensor 4 performs data exchange with the motor 10 and the pneumatic push rod 9 to collect the pressure value between the electrode sheet 3 and the detection part, and sends a control instruction to the motor 10 and the pneumatic push rod 9 when the pressure value does not match the preset pressure threshold;

[0151] The motor 10 and the pneumatic push rod 9 are used to adjust the contact angle and contact strength between the electrode sheet 3 and the detection part based on the control instruction;

[0152] The signal processor 24 is used to collect the EEG signal of the detection part through the electrode sheet 3 when the force feedback sensor 4 determines that the pressure value between the electrode sheet 3 and the detection part matches the preset pressure threshold; perform feature extraction on the residual EEG component of the EEG signal based on the deep residual shrinkage network model that has completed model training, and reconstruct the feature information after feature extraction with the main component of the EEG signal to obtain a reconstructed EEG signal; perform uniform scale decomposition on the multi-channel EEG signal of the reconstructed EEG signal to obtain a decomposition result, and perform view feature processing based on the decomposition result to obtain a multi-view feature set.

[0153] In the embodiment of the present invention, Figure 3As shown in the structural diagram of the cap body 1, the main function of the pneumatic push rod 9 is to adjust the contact strength between the electrode sheet 3 of the EEG signal and the head of the detection part. At this time, the dynamic adjustment of the contact pressure is realized through linear motion, so as to ensure that the pressure value detected by the force feedback sensor 4 is in the normal range, and the signal acquisition quality and patient comfort are improved. The force feedback sensor 4 detects abnormal pressure (too large or too small), and the feedback trigger module controls the pneumatic push rod 9 to extend and retract. At this time, the extension and retraction of the pneumatic push rod 9 include extension and retraction, that is, when extending, the pressure between the electrode sheet 3 and the head of the detection part is increased, so that the electrode sheet 3 is more closely in contact with the head, or when retracting, the pressure between the electrode sheet 3 and the head of the detection part is reduced to avoid discomfort caused by excessive pressure. Furthermore, when the force feedback sensor 4 detects that the pressure returns to the normal range, the operation of the pneumatic push rod 9 is stopped. In addition, the movement of the pneumatic push rod 9 is linear push and pull, which is used to adjust the contact pressure. At this time, it does not involve the angle adjustment or the driving of the rotating parts. The motor 10 adjusts the contact angle between the electrode sheet 3 and the head through rotational motion, so as to achieve the purpose of accurately collecting EEG signals while adjusting the pressure between the electrode sheet 3 and the head in real time, ensuring that the collected EEG signals can better reflect brain activity. In addition, the signal processor 24 can be a terminal processor, which is arranged on the outside of the cap body 1 of the EEG cap, or a remote processor, which is arranged in a remote workbench, and the embodiment of the present invention does not make specific limitations.

[0154] Furthermore, in order to enable the detection body 2 to effectively move up and down when embedded in the cap body 1, so as to ensure that the detection body 2 can effectively adjust the angle and force inside the cap body 1, the force feedback sensor 4 is placed inside the hemispherical front end of the detection body 2, and the detection surface of the force feedback sensor 4 is connected to the electrode sheet 3, and a contraction groove 5 is provided on the inner wall of the cap body 1 on the side adjacent to the detection body 2, and a connecting component is provided on the side of the detection body 2 adjacent to the contraction groove 5, one side of the connecting component extends into the contraction groove 5 and slides in the contraction groove 5, and a compression spring 7 is fixedly connected between the side of the connecting component facing away from the detection body 2 and the inner wall of the contraction groove 5.

[0155] Furthermore, if Figure 4The detection body 2 shown, in order to ensure that the driving shaft 18 can effectively drive the motor 10 to adjust the contact angle between the electrode sheet 3 and the detection part, so as to coordinate the pressure and angle between the electrode sheet 3 and the detection part, ensure the effectiveness of the adjustment process, and ensure the safety of the detection part, the output shaft of the motor 10 is fixedly connected with the driving shaft 18, and the opposite end of the driving shaft 18 rotates and penetrates into the shrinkage groove 5, and the sleeve rod 20 is slidably mounted on the outer surface of the driving shaft 18, and two limiting grooves 21 are provided on the inner wall of the sleeve rod 20, and the outer surface of the driving shaft 18 is fixedly connected with a limiting slider 22, and the opposite side of the limiting slider 22 extends into the limiting groove 21 and slides in the limiting groove 21.

[0156] Furthermore, in order to enable the pneumatic push rod 9 and the motor 10 to effectively adjust the contact angle and contact pressure in the cap body 1, a control cavity 8 is opened in the cap body 1, and the pneumatic push rod 9 and the motor 10 are fixedly installed in the control cavity 8. One end of the telescopic rod of the pneumatic push rod 9 slides through the contraction groove 5 and contacts the connecting cap body 61.

[0157] Furthermore, in order to ensure that the detection body 2 embedded in the cap body 1 is stable and can be flexibly adjusted when adjusting the contact angle and contact pressure between the electrode sheet 3 and the detection part, a connecting column 11 is fixedly connected to the side of the connecting cap body 61 facing the detection body 2, and a movable groove 12 is opened on the surface of the side of the detection body 2 connected to the cap body 1, and a movable ball 13 is arranged in the movable groove 12. The movable ball 13 rotates in the movable groove 12, and one side of the movable ball 13 extends out of the movable groove 12 and is fixedly connected to the connecting column 11. A plurality of control rods 14 arranged in a circular array are provided in the connecting cap body 61, one end of the control rod 14 slides through the connecting cap body 61, the end of the control rod 14 located outside the connecting cap body 61 is fixedly connected to a contact ball 15, and the contact ball 15 contacts the surface of the detection body 2, and the end of the control rod 14 located inside the connecting cap body 61 is fixedly connected to an extrusion body 16, and a support spring 17 is fixedly connected between the extrusion body 16 and the inner wall of the connecting cap body 61, and the support spring 17 is sleeved on the control rod 14.

[0158] Furthermore, in order to ensure that the connecting cap body 61 can be effectively supported during the process of adjusting the angle and pressure, a rotating plate 19 is provided inside the connecting cap body 61, and the rotating plate 19 rotates inside the connecting cap body 61. A sleeve rod 20 is fixedly connected to the surface of the rotating plate 19 on the side facing away from the control rod 14, and the other end of the sleeve rod 20 rotates to pass through the outer surface of the connecting cap body 61.

[0159] Furthermore, if Figure 5As shown, in order to increase the rotational stability of the rotating panel, a push block 23 is fixedly connected to the surface of the rotating plate 19 facing the extrusion body 16 , and the extrusion body 16 contacts and slides with the inclined surface of the push block 23 .

[0160] For the above embodiment, the telescopic rod of the pneumatic push rod 9 pushes the connecting cap body 61, and one side of the connecting cap body 61 extends into the contraction groove 5. The telescopic movement of the pneumatic push rod 9 causes the connecting cap body 61 to be displaced in the contraction groove 5, and this displacement changes the contact pressure between the detection body 2 and the patient's head. At the same time, the output shaft of the motor 10 is fixedly connected to the drive shaft 18, and the drive shaft 18 drives the rotating plate 19 to operate, that is, the drive shaft 18 is an extension component of the output shaft of the motor 10, which is directly driven by the motor 10 to achieve rotation transmission. At the same time, when the drive shaft 18 rotates, it will drive the rotating plate 19 to rotate. The rotating plate 19 further controls the angle adjustment of the detection body 2, that is, the rotating plate 19 squeezes the extrusion body 16 through the push block 23, acting on the control rod 14, and the movement of the control rod 14 causes the detection body 2 to tilt, thereby adjusting the contact angle of the electrode sheet 3.

[0161] It should be noted that the pneumatic push rod 9 is a linear motion, which is used to adjust the contact pressure between the electrode sheet 3 and the head, and acts on the connecting cap body 61. The driving shaft 18 is a rotational motion, which is used to adjust the contact angle between the electrode sheet 3 and the head, and acts on the rotating plate 19 and the detection body 2. In the embodiment of the present invention, the two are coordinated through the force feedback sensor 4 to ensure that the contact angle and pressure are in the optimal state.

[0162] An embodiment of the present invention provides a brain electrode cap, which can achieve the purpose of accurately collecting brain electrical signals while adjusting the pressure between the electrode sheet and the head in real time, so as to ensure that the collected brain electrical signals can better reflect brain activities.

[0163] Furthermore, as a response to the above Figure 1 The embodiment of the present invention provides a device for processing EEG signals, such as Figure 6 As shown, the device comprises:

[0164] The acquisition module 31 is used to acquire the EEG signal of the detection part through the electrode sheet when the force feedback sensor determines that the pressure value between the electrode sheet and the detection part matches the preset pressure threshold, the pressure value is acquired by the force feedback sensor in the detection body, and the force feedback sensor is also used to drive the motor and the pneumatic push rod in the detection body to adjust the contact angle and contact strength between the electrode sheet and the detection part when the acquired pressure value does not match the preset pressure threshold, until the pressure value matches the preset pressure threshold;

[0165] An extraction module 32 is used to extract features of the residual EEG components of the EEG signal based on the deep residual shrinkage network model that has completed model training, and reconstruct the feature information after feature extraction with the main components of the EEG signal to obtain a reconstructed EEG signal;

[0166] The processing module 33 is used to perform uniform scale decomposition on the multi-channel EEG signal of the reconstructed EEG signal to obtain a decomposition result, and perform view feature processing based on the decomposition result to obtain a multi-view feature set.

[0167] Furthermore, the device also includes:

[0168] The determination module is used to perform singular spectrum analysis on the EEG signal to obtain multiple signal sub-components; determine the main component from the multiple signal sub-components, and delete the main component from the EEG signal to obtain a residual EEG component.

[0169] Furthermore, the determination module is specifically used to extract correlated sub-components from the multiple signal sub-components based on a preset autocorrelation coefficient threshold, and reconstruct the correlated sub-components to obtain the main components; wherein the signal sub-components include frequency domain sub-components, time domain sub-components, spatial sub-components, physiological and artifact components, equipment artifact signal sub-components, and nonlinear sub-components.

[0170] Furthermore, the device also includes:

[0171] A training module is used to construct a deep residual shrinkage network, the input of the deep residual shrinkage network is a multi-scale input, and the deep residual shrinkage network model is embedded with a filtering unit, an attention mechanism unit, and a loop unit; the deep residual shrinkage network model is trained based on the residual EEG feature training samples to obtain a deep residual shrinkage network model, wherein the dynamic threshold in the deep residual shrinkage network model is used to eliminate signal artifacts and noise.

[0172] Furthermore, the extraction module is specifically used to obtain reconstruction weights, which include subject weights and feature weights; weights and calculations are performed based on the reconstruction weights, the subject components, and the feature information to obtain reconstructed EEG signals; wherein the subject weights are determined based on the signal variance of the subject components and the signal variance of the feature information, and the sum of the feature weights and the subject weights is 1.

[0173] Furthermore, the main components include frequency main components, time domain main components, and space main components. The extraction module is specifically used to determine the first main weight and the first characteristic weight of the frequency domain main component, and perform weights and calculations based on the first main weight, the first characteristic weight, the main component, and the characteristic information to obtain a first reconstructed EEG signal, and the first characteristic weight is a frequency domain dynamic weight function; determine the second main weight and the second characteristic weight of the time domain main component, and perform weights and calculations based on the second main weight, the second characteristic weight, the main component, and the characteristic information to obtain a second reconstructed EEG signal, and the second main weight is 1, and the second characteristic weight is a time-varying gain function; determine the third main weight and the third characteristic weight of the spatial main component, and perform weights and calculations based on the third main weight, the third characteristic weight, the main component, and the characteristic information to obtain a third reconstructed EEG signal.

[0174] Furthermore, the determination module is also used to determine the mutual information between the reconstructed EEG signal and the standard EEG signal, and to screen the multi-channel EEG signal of the reconstructed EEG signal based on the mutual information, and the mutual information is used to characterize the degree of correlation between the reconstructed EEG signal and the standard EEG signal.

[0175] Furthermore, the processing module is also used to filter the multi-channel EEG signal based on a spatial filter to perform uniform scale decomposition based on the filtered multi-channel EEG signal, and the weight matrix in the spatial filter is determined based on a deep learning network and a common spatial pattern.

[0176] Furthermore, the processing module is specifically used to perform modal decomposition on the multi-channel EEG signal to obtain a modal decomposition result, and filter out a subband set from the modal decomposition result; calculate the energy of each subband in the subband set, and determine the optimal subband based on the dynamic correlation weight and the subband energy; extract features of the optimal subband based on the multi-layer convolution model that has completed model training, and obtain a decomposition result of the subband features.

[0177] Furthermore, the processing module is specifically used to perform feature classification on the decomposition results according to feature types, and generate a view set of the decomposition results after classification according to time, frequency, and time domain, and the feature types include linear type, nonlinear type, and energy type.

[0178] The embodiment of the present invention provides a device for processing EEG signals, which achieves the purpose of accurately collecting EEG signals while adjusting the pressure between the electrode sheet and the head in real time, ensuring that the collected EEG signals can better reflect brain activities. At the same time, the collected EEG signals are extracted and reconstructed, and then the reconstructed EEG signals are decomposed at multiple scales, thereby meeting the analysis requirements of brain activity characteristics in different dimensions, greatly improving the processing accuracy of EEG signals, and realizing accurate identification of brain activities.

[0179] According to an embodiment of the present invention, a storage medium is provided, wherein the storage medium stores at least one executable instruction, and the computer executable instruction can execute the method of processing an electroencephalogram signal in any of the above method embodiments.

[0180] Figure 7 A schematic diagram of the structure of a terminal provided according to an embodiment of the present invention is shown. The specific embodiment of the present invention does not limit the specific implementation of the terminal.

[0181] like Figure 7 As shown, the terminal may include: a processor (processor) 402, a communication interface (Communications Interface) 404, a memory (memory) 406, and a communication bus 408.

[0182] The processor 402 , the communication interface 404 , and the memory 406 communicate with each other via a communication bus 408 .

[0183] The communication interface 404 is used for network communication with other devices such as a client or other servers.

[0184] The processor 402 is used to execute the program 410, and specifically can execute the relevant steps in the above-mentioned EEG signal method embodiment.

[0185] Specifically, the program 410 may include program codes, which include computer operation instructions.

[0186] The processor 402 may be a central processing unit (CPU), or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiment of the present invention. The one or more processors included in the terminal may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0187] The memory 406 is used to store the program 410. The memory 406 may include a high-speed RAM memory, and may also include a non-volatile memory, such as at least one disk memory.

[0188] The program 410 may be specifically configured to enable the processor 402 to perform the following operations:

[0189] When the force feedback sensor determines that the pressure value between the electrode sheet and the detection part matches the preset pressure threshold, the electroencephalogram signal of the detection part is collected by the electrode sheet, and the pressure value is collected by the force feedback sensor in the detection body. The force feedback sensor is also used to drive the motor and the pneumatic push rod in the detection body to adjust the contact angle and contact strength between the electrode sheet and the detection part when the collected pressure value does not match the preset pressure threshold, until the pressure value matches the preset pressure threshold;

[0190] Extracting features of the residual EEG components of the EEG signal based on the deep residual shrinkage network model for which model training has been completed, and reconstructing the feature information after feature extraction with the main components of the EEG signal to obtain a reconstructed EEG signal;

[0191] Uniform scale decomposition is performed on the multi-channel EEG signal of the reconstructed EEG signal to obtain a decomposition result, and view feature processing is performed based on the decomposition result to obtain a multi-view feature set.

[0192] Obviously, those skilled in the art should understand that the above modules or steps of the present invention can be implemented by a general computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, and optionally, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order than here, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Thus, the present invention is not limited to any specific combination of hardware and software.

[0193] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for processing electroencephalogram signals, characterized in that: include: When the force feedback sensor determines that the pressure value between the electrode sheet and the detection part matches the preset pressure threshold, the electroencephalogram signal of the detection part is collected through the electrode sheet, and the force feedback sensor is also used to drive the motor and the pneumatic push rod in the detection body to adjust the contact angle and contact strength between the electrode sheet and the detection part when the collected pressure value does not match the preset pressure threshold, until the pressure value matches the preset pressure threshold; Extracting features of the residual EEG components of the EEG signal based on the deep residual shrinkage network model for which model training has been completed, and reconstructing the feature information after feature extraction with the main components of the EEG signal to obtain a reconstructed EEG signal; Uniform scale decomposition is performed on the multi-channel EEG signal of the reconstructed EEG signal to obtain a decomposition result, and view feature processing is performed based on the decomposition result to obtain a multi-view feature set.

2. The method according to claim 1, characterized in that Before extracting features of the residual EEG components of the EEG signal based on the deep residual shrinkage network model for which model training has been completed, the method further includes: Performing singular spectrum analysis on the EEG signal to obtain multiple signal sub-components; A main component is determined from the multiple signal subcomponents, and the main component is deleted from the EEG signal to obtain a residual EEG component.

3. The method according to claim 2, characterized in that Determining the main component from the plurality of signal subcomponents comprises: Extracting relevant sub-components from the plurality of signal sub-components based on a preset autocorrelation coefficient threshold, and reconstructing the relevant sub-components to obtain the main component; The signal subcomponents include frequency domain subcomponents, time domain subcomponents, space subcomponents, physiological and artifact shadow components, equipment artifact signal subcomponents, and nonlinear subcomponents.

4. The method according to claim 1, characterized in that: The step of reconstructing the feature information after feature extraction and the main components of the EEG signal to obtain the reconstructed EEG signal includes: Obtaining reconstruction weights, wherein the reconstruction weights include subject weights and feature weights; Performing weighted sum calculation based on the reconstruction weight, the main component, and the feature information to obtain a reconstructed EEG signal; The subject weight is determined based on the signal variance of the subject component and the signal variance of the feature information, and the sum of the feature weight and the subject weight is 1.

5. The method according to claim 4, characterized in that The main components include frequency main components, time domain main components, and space main components. The weighting and calculation based on the reconstruction weights, the main components, and the feature information to obtain the reconstructed EEG signal includes: Determine a first main body weight and a first characteristic weight of the main component in the frequency domain, and perform weighting and calculation based on the first main body weight, the first characteristic weight, the main component, and the characteristic information to obtain a first reconstructed EEG signal, wherein the first characteristic weight is a frequency domain dynamic weight function; Determine a second subject weight and a second feature weight of the time domain subject component, and perform weighting and calculation based on the second subject weight, the second feature weight, the subject component, and the feature information to obtain a second reconstructed EEG signal, wherein the second subject weight is 1, and the second feature weight is a time-varying gain function; Determine the third subject weight and the third feature weight of the spatial subject component, and perform weighting and calculation based on the third subject weight, the third feature weight, the subject component, and the feature information to obtain a third reconstructed EEG signal.

6. The method according to claim 1, characterized in that Before extracting features of the residual EEG components of the EEG signal based on the deep residual shrinkage network model for which model training has been completed, the method further includes: Constructing a deep residual shrinkage network, wherein the input of the deep residual shrinkage network is a multi-scale input, and a filter unit, an attention mechanism unit, and a recurrent unit are embedded in the deep residual shrinkage network model; The deep residual shrinkage network model is trained based on the residual EEG feature training samples to obtain a deep residual shrinkage network model, wherein the dynamic threshold in the deep residual shrinkage network model is used to eliminate signal artifacts and noise.

7. A device for processing electroencephalogram signals, characterized in that: include: an acquisition module, configured to acquire an electroencephalogram signal of the detection part through the electrode sheet when the force feedback sensor determines that the pressure value between the electrode sheet and the detection part matches a preset pressure threshold, wherein the pressure value is acquired by the force feedback sensor in the detection body, and the force feedback sensor is further configured to drive a motor and a pneumatic push rod in the detection body to adjust a contact angle and a contact strength between the electrode sheet and the detection part when the acquired pressure value does not match the preset pressure threshold, until the pressure value matches the preset pressure threshold; An extraction module, used for extracting features of the residual EEG components of the EEG signal based on the deep residual shrinkage network model that has completed model training, and reconstructing the feature information after feature extraction with the main components of the EEG signal to obtain a reconstructed EEG signal; The processing module is used to perform uniform scale decomposition on the multi-channel EEG signal of the reconstructed EEG signal to obtain a decomposition result, and perform view feature processing based on the decomposition result to obtain a multi-view feature set.

8. A brain electrode cap, characterized in that: include: Cap body, detection body, signal processor, The cap body is provided with a plurality of cavities, the end of the detection body is placed in the cavity, the detection body includes a force feedback sensor, a motor and a pneumatic push rod, and the hemispherical front end of the detection body is provided with an electrode sheet outside. The force feedback sensor exchanges data with the motor and the pneumatic push rod to collect the pressure value between the electrode sheet and the detection part, and sends a control instruction to the motor and the pneumatic push rod when the pressure value does not match the preset pressure threshold; The motor and the pneumatic push rod are used to adjust the contact angle and contact strength between the electrode sheet and the detection part based on the control instruction; The signal processor is used to collect the EEG signal of the detection part through the electrode sheet when the force feedback sensor determines that the pressure value between the electrode sheet and the detection part matches the preset pressure threshold; perform feature extraction on the residual EEG component of the EEG signal based on the deep residual shrinkage network model that has completed model training, and reconstruct the feature information after feature extraction with the main component of the EEG signal to obtain a reconstructed EEG signal; perform uniform scale decomposition on the multi-channel EEG signal of the reconstructed EEG signal to obtain a decomposition result, and perform view feature processing based on the decomposition result to obtain a multi-view feature set.

9. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to claim 1 are implemented.

10. A computer device comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method of claim 1.

Images