CN118551340A

CN118551340A - Electroencephalogram signal analysis method and equipment based on multi-scale electroencephalogram characteristic fusion

Info

Publication number: CN118551340A
Application number: CN202411025981.5A
Authority: CN
Inventors: 李越; 李宝宝; 徐洪凯
Original assignee: Xiaozhou Technology Co ltd
Current assignee: Xiaozhou Technology Co ltd
Priority date: 2024-07-30
Filing date: 2024-07-30
Publication date: 2024-08-27
Anticipated expiration: 2044-07-30
Also published as: CN118551340B

Abstract

The application discloses an electroencephalogram signal analysis method and equipment based on multi-scale electroencephalogram feature fusion, which divide an electroencephalogram target signal processed by an original electroencephalogram signal into a plurality of areas containing electroencephalogram target signals corresponding to brain areas according to brain area distribution knowledge of a human brain; dividing a computing unit corresponding to the pre-trained mental state evaluation model into a plurality of computing subunits; determining a key brain region in the plurality of brain regions, and constructing a first electroencephalogram feature sequence through a computing subunit corresponding to a key region corresponding to the key brain region; acquiring edge local features corresponding to the key brain regions based on a local correlation analysis method so as to optimize weight parameters of computing subunits corresponding to the key brain regions; adjusting the time length and the position of a sliding window of the computing subunit to construct a second electroencephalogram characteristic sequence; and respectively inputting the first electroencephalogram characteristic sequence and the second electroencephalogram characteristic sequence into a psychological state assessment model to obtain a first electroencephalogram state and a second electroencephalogram state, and completing analysis of the original electroencephalogram signals.

Description

Electroencephalogram signal analysis method and equipment based on multi-scale electroencephalogram characteristic fusion

Technical Field

The application relates to the technical field of brain-computer interfaces, in particular to an electroencephalogram signal analysis method and equipment based on multi-scale electroencephalogram feature fusion.

Background

The EEG signal analysis has important application value in the clinical diagnosis and scientific research fields of mental diseases. The mainstream electroencephalogram (EEG) -based analysis method generally adopts means such as power spectrum analysis to analyze electroencephalogram characteristics from a frequency domain, or obtains time domain characteristics through cross-correlation, wavelet transformation and the like, so as to attempt to extract an electroencephalogram index reflecting the overall brain activity state. However, this overall feature extraction approach makes it difficult to take into account structural and functional heterogeneity of the human brain. The effects of different brain regions on specific brain activity states are also different, and it is often difficult for simple overall features to highlight the effects of critical brain regions.

In order to obtain more targeted feature expression, attempts have been made in the prior art to guide feature extraction in combination with brain region location information. For example, the region of interest is predefined and the characteristic channels are selected only from specific brain regions. However, such methods rely on subjective assumptions and do not necessarily accurately locate truly relevant brain regions. There are also researchers trying to learn the dynamic brain region weights for different states through neural network attention models, but the learning process is complex, and it is difficult to explain the exact meaning of the weights. In general, the work of fusing brain region information to guide electroencephalogram feature extraction in the prior art is insufficient.

In addition, most methods only extract the whole characteristics of the electroencephalogram, but the electroencephalogram is used as a high-dimensional complex power system, and the local characteristics of the electroencephalogram also contain effective information. Some students analyze the local characteristics of the brain electricity by adopting local correlation, wavelet local energy and the like, but the local and the whole characteristics are rarely combined, so that the information is incomplete. It should be noted that the brain activity state has a continuous dynamic evolution characteristic, and the static feature is difficult to evaluate the change on the time axis of the brain electrical signal, so that the dynamic feature extraction of multiple time windows is also critical.

Disclosure of Invention

The embodiment of the application provides an electroencephalogram signal analysis method and equipment based on multi-scale electroencephalogram characteristic fusion, which can solve the problem that most methods only extract the whole characteristics of an electroencephalogram signal, but the electroencephalogram is used as a high-dimensional complex power system, and the local characteristics of the electroencephalogram signal also contain effective information. Meanwhile, some students analyze the local characteristics of the brain electricity by adopting local correlation, wavelet local energy and the like, but the local and the whole characteristics are rarely combined, so that the problem of incomplete information is caused.

In a first aspect, an embodiment of the present application provides an electroencephalogram signal analysis method based on multi-scale electroencephalogram feature fusion, where the method includes:

acquiring an original electroencephalogram signal to be analyzed, preprocessing original electroencephalogram signal data, and acquiring an electroencephalogram target signal;

dividing an electroencephalogram target signal into a plurality of areas according to brain area dividing knowledge of a human brain, wherein each area corresponds to one brain area, and each area at least comprises an electroencephalogram target signal corresponding to the brain area;

dividing computing units corresponding to the pre-trained mental state evaluation model to obtain a plurality of computing subunits, wherein each computing subunit corresponds to one region;

determining a key brain region in a plurality of brain regions, wherein the region corresponding to the key brain region is a key region, and constructing a first brain electrical characteristic sequence of the key region through a computing subunit corresponding to the key region;

in the multiple regions, acquiring edge local features corresponding to the key brain regions based on a local correlation analysis method, and optimizing weight parameters of computing subunits corresponding to the key brain regions according to the edge local features;

adjusting the time length and the position of a sliding window of a calculating subunit corresponding to the key brain region, and constructing a second electroencephalogram characteristic sequence of the key brain region through the calculating subunit corresponding to the key brain region;

inputting the first electroencephalogram characteristic sequence and the second electroencephalogram characteristic sequence into a psychological state assessment model respectively, acquiring a first electroencephalogram state corresponding to the first electroencephalogram characteristic sequence and a second electroencephalogram state corresponding to the second electroencephalogram characteristic sequence, and completing analysis of the original electroencephalogram signals according to the first electroencephalogram state and the second electroencephalogram state.

In a second aspect, the present application also provides an electroencephalogram signal analysis apparatus, including:

the signal acquisition unit is used for acquiring an original electroencephalogram signal to be analyzed, preprocessing the original electroencephalogram signal data and acquiring an electroencephalogram target signal;

The signal dividing unit is used for dividing the brain electric target signal into a plurality of areas according to the brain area dividing knowledge of the human brain, each area corresponds to one brain area, and each area at least comprises the brain electric target signal corresponding to the brain area;

The unit dividing unit is used for dividing the calculating units corresponding to the pre-trained psychological state assessment model to obtain a plurality of calculating sub-units, and each calculating sub-unit corresponds to one area;

The key determining unit is used for determining a key brain region in a plurality of brain regions, wherein a region corresponding to the key brain region is a key region, and a first brain electrical characteristic sequence of the key region is constructed through the calculating subunit corresponding to the key region;

the weight optimization unit is used for acquiring edge local features corresponding to the key brain areas in the plurality of areas based on a local correlation analysis method, and optimizing weight parameters of the computing sub-units corresponding to the key brain areas according to the edge local features;

The window adjusting unit is used for adjusting the time length and the position of a sliding window of the calculating subunit corresponding to the key brain region, and constructing a second brain electrical characteristic sequence of the key brain region through the calculating subunit corresponding to the key brain region;

The signal analysis unit is used for inputting the first electroencephalogram characteristic sequence and the second electroencephalogram characteristic sequence into the psychological state assessment model respectively, acquiring a first electroencephalogram state corresponding to the first electroencephalogram characteristic sequence and a second electroencephalogram state corresponding to the second electroencephalogram characteristic sequence, and completing analysis of the original electroencephalogram signal according to the first electroencephalogram state and the second electroencephalogram state.

In a third aspect, the present application further provides a computer device, including a processor and a memory, where the memory is configured to store a computer program, where the computer program when executed by the processor implements the electroencephalogram signal analysis method based on multi-scale electroencephalogram feature fusion according to the first aspect.

In a fourth aspect, the present application also provides a computer readable storage medium storing a computer program, which when executed by a processor, implements the electroencephalogram signal analysis method based on multi-scale electroencephalogram feature fusion according to the first aspect.

Compared with the prior art, the application has at least the following beneficial effects:

1. The accuracy and the comprehensiveness of the EEG signal analysis are improved by multi-scale feature fusion: the method remarkably improves the accuracy and the comprehensiveness of analysis by combining the electroencephalogram characteristics of different time scales and space scales. On the time scale, sliding windows of 256ms and 512ms are adopted simultaneously to extract short-time and long-time electroencephalogram characteristics respectively. Such a dual time scale analysis may capture both instantaneous brain electrical state changes and long-term activity trends, thereby more fully describing the dynamic activity patterns of the brain. On a spatial scale, the electroencephalogram signal is divided into a plurality of regional samples based on a 10-20 system, and each sample contains information of a core electrode and a peripheral electrode. This spatial partitioning allows for interactions between the brain regions that better capture complex brain function network activities. By fusing the multi-scale features, the method can analyze local and global brain electrical activity modes at the same time, and provides a more comprehensive and accurate information basis for understanding complex cognitive processes, emotion changes and brain electrical features of various neuropsychiatric diseases.

2. Brain region adaptive optimization improves the sensitivity of the model to key brain regions: the method introduces brain region self-adaptive optimization technology based on simulated annealing algorithm, and obviously improves the sensitivity of an analysis model to the brain electrical characteristics of the key brain region. By using edge local features, the method can adaptively adjust for specific brain electrical features of each critical brain region. The optimization not only considers the electroencephalogram signal processing in a single brain region, but also optimizes the electroencephalogram information interaction of the brain region and surrounding regions. For example, for different brain regions, the optimized weights may be more sensitive to specific bands of brain electrical activity, consistent with the role of each brain region in different cognitive functions. The self-adaptive optimization technology enables the model to capture the unique contribution of different brain regions under specific cognitive tasks or brain electrical states more accurately, so that the accuracy and the specificity of the overall analysis are improved.

3. The dynamic time window analysis realizes continuous tracking and transition detection of the brain electrical state: the method adopts a dynamic time window analysis technology to realize continuous tracking and accurate state transition detection of the brain electrical state. By sliding time windows of different lengths on the time axis, the method is able to continuously output two sets of scores reflecting transient conditions and long-term trends. This dual time scale dynamic analysis enables the method to capture both rapid neural state changes and slow brain electrical pattern evolution. For example, a sustained highly active state of the brain may be identified, or a transition from active to inhibited may be captured. The dynamic analysis method not only can accurately capture the transition time of the brain electrical state, but also can describe the specific process and duration of the transition. This is of great importance for studying the change in attention, mood changes, or dynamic transitions of the brain electrical state during learning. In addition, the method can be applied to clinical researches, such as identifying the precursor of epileptic seizure, monitoring the change of anesthesia depth, or observing the long-term change trend of the brain electrical state of a mental disease patient in the treatment process, and provides important basis for the establishment and evaluation of personalized medical treatment schemes.

It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the application as claimed.

Drawings

Fig. 1 is a flow chart of an electroencephalogram signal analysis method based on multi-scale electroencephalogram feature fusion according to an embodiment of the application;

fig. 2 is a schematic structural diagram of an electroencephalogram signal analysis apparatus according to an embodiment of the present application;

fig. 3 is a schematic structural diagram of a computer device according to an embodiment of the present application.

Detailed Description

In the following description, for purposes of explanation and not limitation, specific details are set forth such as the particular system architecture, techniques, etc., in order to provide a thorough understanding of the embodiments of the present application. It will be apparent, however, to one skilled in the art that the present application may be practiced in other embodiments that depart from these specific details. In other instances, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present application with unnecessary detail.

It should be understood that the terms "comprises" and/or "comprising," when used in this specification and the appended claims, specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof.

It should also be understood that the term "and/or" as used in the present specification and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes such combinations.

As used in the present description and the appended claims, the term "if" may be interpreted in context as "when …" or "once" or "in response to a determination" or "in response to detection. Similarly, the phrase "if a determination" or "if a [ described condition or event ] is detected" may be interpreted in the context of meaning "upon determination" or "in response to determination" or "upon detection of a [ described condition or event ]" or "in response to detection of a [ described condition or event ]".

Furthermore, the terms "first," "second," "third," and the like in the description of the present specification and in the appended claims, are used for distinguishing between descriptions and not necessarily for indicating or implying a relative importance.

Reference in the specification to "one embodiment" or "some embodiments" or the like means that a particular feature, structure, or characteristic described in connection with the embodiment is included in one or more embodiments of the application. Thus, appearances of the phrases "in one embodiment," "in some embodiments," "in other embodiments," and the like in the specification are not necessarily all referring to the same embodiment, but mean "one or more but not all embodiments" unless expressly specified otherwise. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless expressly specified otherwise.

The following describes the technical scheme of the embodiment of the application.

The EEG signal analysis has important application value in the clinical diagnosis and scientific research fields of mental diseases. The electroencephalogram signal is accurately analyzed, so that doctors can be helped to screen, qualitatively and quantitatively analyze mental diseases, and a basis is provided for judging the illness state. Currently, technologies for analyzing brain electrical signals mainly include behavioural detection, physiological signal detection, brain imaging detection, detection based on brain electrical signals, and the like. The brain electrical signal has the advantages of convenience in acquisition, high time resolution and the like, and is one of important ways for analyzing brain activity state changes. However, the electroencephalogram signals have extremely complex dynamic change characteristics in both time and frequency spaces, and it is very difficult to extract key features that truly reflect the state of brain activity therein.

The mainstream electroencephalogram (EEG) -based analysis method generally adopts means such as power spectrum analysis to analyze electroencephalogram characteristics from a frequency domain, or obtains time domain characteristics through cross-correlation, wavelet transformation and the like, so as to attempt to extract an electroencephalogram index reflecting the overall brain activity state. However, this overall feature extraction approach makes it difficult to take into account structural and functional heterogeneity of the human brain. The effects of different brain regions on specific brain activity states are also different, and it is often difficult for simple overall features to highlight the effects of critical brain regions.

To obtain more targeted feature expression, some scholars attempt to guide feature extraction in combination with brain region location information. For example, the region of interest is predefined and the characteristic channels are selected only from specific brain regions. However, such methods rely on subjective assumptions and do not necessarily accurately locate truly relevant brain regions. There are also researchers trying to learn the dynamic brain region weights for different states through neural network attention models, but the learning process is complex, and it is difficult to explain the exact meaning of the weights. In general, the work of fusing brain region information to guide electroencephalogram feature extraction in the prior art is insufficient.

In order to solve the above problems, please refer to fig. 1, fig. 1 is a flow chart of an electroencephalogram signal analysis method based on multi-scale electroencephalogram feature fusion according to an embodiment of the present application. The electroencephalogram signal analysis method based on multi-scale electroencephalogram feature fusion can be applied to computer equipment, wherein the computer equipment comprises, but is not limited to, intelligent mobile phones, notebook computers, tablet computers, desktop computers, physical servers, cloud servers and the like. As shown in fig. 1, the electroencephalogram signal analysis method based on multi-scale electroencephalogram feature fusion of the present embodiment includes steps S101 to S107, which are described in detail as follows:

step S101, acquiring an original electroencephalogram signal to be analyzed, preprocessing the original electroencephalogram signal data, and acquiring an electroencephalogram target signal.

Specifically, the application collects the original electroencephalogram data of the user under different cognitive tasks. And selecting proper electroencephalogram equipment, configuring acquisition parameters, determining a user and a stimulation task, and acquiring electroencephalogram signals according to a standard flow to obtain an original electroencephalogram signal data set containing information of each brain region of the whole brain. The method adopts electroencephalogram (EEG) to collect signals. When the acquisition parameters are determined, the age range of the user needs to be considered, the sampling frequency of the adult user is 250-500Hz, the sampling frequency of the child user is 500-1000Hz, the number of recording channels needs to cover all scalp areas, 32 channels, 64 channels or 128 channels are usually selected, the recording time length needs to fully contain the cognitive task process, and the recording time length is generally controlled to be 10-30 minutes. Taking a 128 channel system as an example, the sampling frequency is set to 500Hz and the recording duration is set to 20 minutes. The EEG sensor is configured by adopting an international 10-20 system arrangement, and the system uniformly divides the sensor into 10% and 20% distance points according to the anatomical position of the scalp, so that electrodes can be configured in a standardized way and all parts of the scalp are covered. Taking a 128 channel system as an example, 128 electrodes would be provided on the scalp, each electrode name corresponding to an area of the scalp. After the electrode configuration is completed, signal quality inspection is required to be performed to confirm that each electrode is well contacted, and the brain electrical signals are clearly available.

In some embodiments, the acquiring the raw electroencephalogram signal to be analyzed includes: obtaining stimulus types and strength parameters corresponding to a plurality of cognitive stimulus tasks; configuring acquisition parameters of preset electroencephalogram acquisition equipment according to the stimulation type and the intensity parameters; and completing acquisition of the original brain electrical signals of the user under a plurality of cognitive stimulation tasks through the adjusted brain electrical acquisition device.

When determining the cognitive stimulation task, proper stimulation types, intensity parameters and the like are set so as to trigger corresponding brain electrical activity modes. For example, assessing brain electrical activity in the attention state, visual search tasks may be set to direct attention to or away from, in particular Steady state visual evoked potential (Steady-State Visual Evoked Potential, SSVEP) paradigms may be used to continuously and rapidly present alphanumeric sequences on a screen, with the user needing to identify target stimuli. Different target stimulus probabilities can be set to guide the attention state, a high target stimulus probability can attract the continuous attention of a user, and a low target stimulus probability can lead to distraction. When parameters are set, the target stimulation probability can be 10% -30% to represent the attention focusing condition, and 2% -5% to represent the attention dispersing condition. The sequence presentation speed is typically 50-100ms per item. An uninterrupted EEG signal was extracted for 5-10min under each condition to collect brain electrical samples under steady attention conditions. By means of the parameter control of the SSVEP task, corresponding EEG data can be effectively acquired in different attentive states.

In addition to assessing brain electrical activity in an attentive state, the method may also be used to assess brain electrical characteristics in other cognitive states, such as mood, stress, fatigue, etc. For example, when assessing brain electrical activity in different emotional states, an international emotion picture system (IAPS) may be employed that collects a large number of images with emotion-inducing effects, which may be presented to induce different emotional responses in the subject, while recording corresponding brain electrical data. The images can be classified into four types of low awakening positive, high awakening positive, low awakening negative and high awakening negative according to the scores, 10-15 images are selected from each type, the images are alternately and rapidly displayed, and each image lasts for 6 seconds. The subject was asked to make subjective scores for each image, recording his emotional experience. EEG data is continuously acquired in the process, and is segmented and marked as an electroencephalogram sample under a corresponding emotional state.

The EEG signal is continuously recorded during the cognitive task of the user, and the duration is controlled to be 10 seconds or longer to collect enough brain electrical data. While the stimulus event flag is set for subsequent analysis. After the recording is completed, EEG data segments containing complete scalp area information and with sufficient duration are extracted, and the EEG data segments are stored and generated into an original EEG data set and transmitted to a subsequent processing step.

Then, preprocessing the original EEG signal data, processing signals by adopting algorithms such as denoising, filtering and the like, and identifying and deleting errors caused by technical noise and human factors to obtain a clean and reliable EEG target signal.

In some embodiments, preprocessing the raw electroencephalogram data comprises: firstly, extracting original electroencephalogram data x (t) of 2 minutes before starting a task for each EEG channel, calculating an average mean (x (t 0: t 1)) of x (t) in the 2-minute time period, taking the average mean as a baseline value of the channel, wherein t0 and t1 are starting and ending points of the 2-minute time period respectively, and subtracting the corresponding baseline value mean (x (t 0: t 1)) from the whole time sequence data x (t) of the channel to obtain baseline corrected electroencephalogram data y (t) =x (t) -mean (x (t 0: t 1)) so as to eliminate the influence of baseline drift of frequency caused by electrochemical effects and the like on the amplitude of the electroencephalogram. Then, an IIR elliptic band-pass filter is designed, the cut-off frequency of the IIR elliptic band-pass filter is lower limit 0.5Hz and upper limit 40Hz, the transition bandwidth is set to be 0.2Hz, and the minimum attenuation of the pass band is required to be not lower than 40dB. The filter can be conveniently designed by using iirfilter functions of MATLAB, the codes are [ b, a ] = iirfilter (4, [ 0.5/(fs/2) 40/(fs/2) ], 'ellip'), 4 is the order of the filter, fs is the sampling frequency of EEG data, and the functions can automatically calculate Num and Den coefficients of the filter according to the specified cutoff frequency, transition bandwidth and other parameters and store the Num and Den coefficients in b and a. After obtaining the filter coefficient, performing phase delay-free filtering on the brain electrical data y after baseline correction by adopting filtfilt functions of MATLAB, wherein the formula is y_filt= filtfilt (b, a, y). Filtfilt is a forward and backward filtering method based on a filter, which can effectively eliminate nonlinear phases of frequency response of a filter bank, thereby avoiding phase distortion of electroencephalogram data and obtaining a filtering result y_filt without phase delay. Through the band-pass filtering, low-frequency drift below 0.5Hz and high-frequency clutter above 40Hz in the electroencephalogram data can be filtered, and the needed electroencephalogram frequency band signals are reserved. And (3) for the missing value in the filtered electroencephalogram data y_filt, if the missing duration is less than 0.5 seconds, performing interpolation by using a linear interpolation method, and if the missing duration exceeds 0.5 seconds, directly rejecting the data segment.

And then performing Independent Component Analysis (ICA) decomposition on the filtered electroencephalogram data by using fastica functions of MATLAB to obtain independent components IC, setting parameters of app reach= 'deflation', numOfICs =64, identifying and removing physiological artifact components such as blinks, body movements and the like according to the characteristics of time domain, frequency domain, topological distribution and the like of the IC, and reconstructing clean EEG data. And calculating the brain electric power spectral density of each time point and channel, marking abnormal points which deviate from the normal value by more than 5 standard deviations, marking abnormal points which are too different from surrounding electrodes based on the topological distribution of adjacent electrodes, and removing the abnormal points.

Finally, the reref function in the EEGLAB tool box is used for re-referencing the electroencephalogram data obtained through the preprocessing step to the average reference of the whole scalp area, the specific method is that for each time point and channel, the average mean (EEG.data, 3) of all channel data of the time point is calculated, wherein EEG.data is input preprocessed three-dimensional matrix data (channel×time×test time), mean (EEG.data, 3) represents the average value calculated in the third dimension (test time) to obtain a two-dimensional matrix (channel×time), then the two-dimensional average matrix is repeated in the third dimension (nbchan represents the channel number) by using repmat function, so that the size of the EEG.data is the same as that of original three-dimensional data, finally, the average matrix is subtracted from EEG.data of each channel and time point, thus the average reference of the whole scalp area is realized, and the formula is EEG.data=EEG.data-repmat (mean (bc1, 3) [1 ] n is eliminated, and the EEG can be obtained independently from the original EEG, and the average electrode has no influence on the scalp area. Where [1 EEG.nbchan ] specifies the number of repetitions in three dimensions, 1 means no repetition, EEG.nbchan means nbchan repetitions, nbchan is the number of channels stored in the EEG data structure.

Step S102, dividing the brain electric target signal into a plurality of areas according to brain area distribution knowledge of a human brain, wherein each area corresponds to one brain area, and each area at least comprises brain electric target signals corresponding to the brain area.

Specifically, the human brain is a complex-structured, functionally diverse organ, and different brain regions are responsible for processing different types of information and performing different cognitive functions. In order to better analyze the brain electrical signals and understand the brain activities reflected by the brain electrical signals, the application needs to divide the collected brain electrical signals into a plurality of area samples according to the anatomical structure and the functional partition of the human brain. The partitioning method not only can keep the integrity of the original signal, but also can reduce the granularity of analysis to a single functional brain region, thereby laying a foundation for the subsequent targeted feature extraction and analysis model construction.

In some embodiments, the electroencephalographic target signal comprises a plurality of electrode measurement signals; the dividing the electroencephalogram target signal into a plurality of areas includes: determining a plurality of brain regions and brain region position information corresponding to the brain regions according to the brain region knowledge of the human brain; acquiring electrode position information of an electroencephalogram acquisition device corresponding to the electroencephalogram target signal; determining electrodes corresponding to each brain region according to the brain region position information and the electrode position information corresponding to each brain region; and constructing a region corresponding to the brain region by using the electrode measurement signals corresponding to each brain region and the electrode measurement signals within a preset range, and completing construction of a plurality of regions.

When brain region division is carried out, the application mainly depends on the international 10-20 system standard. This standard is a widely used method for electroencephalogram (EEG) electrode placement, which distributes electrodes evenly over 10% and 20% of the distance points, depending on the anatomical location of the scalp. This standardized electrode arrangement not only covers all important parts of the scalp, but also ensures that the data between different experiments is comparable. According to the 10-20 system, the present application can divide the human brain roughly into the following main functional areas: frontal lobe region, central region, parietal lobe region, occipital lobe region, temporal lobe region, and some other auxiliary regions.

The frontal lobe region is the portion located at the forefront of the brain and includes electrode locations such as Fp1, fp2, AF3, AF4, F7, F5, F3, F1, fz, F2, F4, F6, F8, and the like. This area is mainly associated with advanced cognitive functions such as execution control, decision making, working memory, etc. The signals of the frontal lobe region are particularly important when analyzing brain electrical activity associated with these advanced cognitive processes.

The frontal lobe region is located behind the frontal lobe and includes the electrode locations FT7, FC5, FC3, FC1, FCz, FC2, FC4, FC6, FT8, etc. This area is involved in a variety of complex cognitive and behavioral processes such as motor planning, problem solving, emotion regulation, etc. Frontal lobe brain electrical activity is of great importance for understanding behavioral control and emotional processing in humans.

The central region spans the middle of the brain and includes electrode locations T7, C5, C3, C1, cz, C2, C4, C6, T8, etc. This region is closely related to sensorimotor integration, corresponding to the motor cortex and somatosensory cortex of the brain. Analysis of the brain electrical signals in the central region may assist the present application in understanding the neural mechanisms of sensory information processing and motor control.

The top leaf region is located behind the central region and includes electrode positions such as CP5, CP3, CP1, CPz, CP2, CP4, CP6, TP7, P5, P3, P1, pz, P2, P4, P6, P8, TP8, and the like. This area participates in spatial processing, attention distribution, sensory integration, and the like. Brain electrical activity of the parietal lobe is of great value for studying spatial cognition and attention mechanisms.

The occipital region is located at the rearmost end of the brain and includes the electrode positions of PO7, PO5, PO3, POz, PO4, PO6, PO8, O1, oz, O2, etc. This area is mainly responsible for the processing of visual information. Analysis of the brain electrical signals of occipital regions may help the present application understand the neural response and visual cognitive processes of visual stimuli.

Temporal lobe regions are located on both sides of the brain, including electrode locations T7, T8, TP7, TP8, FT7, FT8, etc. This region is closely related to the functions of auditory processing, language understanding, and long-term memory. The brain electrical activity of the temporal lobe is of great importance for studying the neural mechanisms of hearing and speech processing.

In addition to these main functional areas, there are some additional electrode positions like AF7, AF8, F9, F10, FT9, FT10, C9, C10, TP9, TP10, etc. These electrodes provide additional spatial resolution, helping to more finely localize brain electrical activity.

Based on the above brain region division, the present application can construct 41 region samples. Each region sample contains not only the core electrode signal of the functional region, but also the signal of the adjacent peripheral electrode as a contextual supplement. This partitioning method ensures that each region sample is sufficient to cover all relevant detailed information of the brain region, while also taking into account interactions between brain regions.

Specifically, the application can divide the forehead lobe area into 8 brain area samples, and the brain electrical target signal of each sample comprises signals of 2-3 core electrodes and peripheral electrodes thereof. For example, the first forehead lobe sample may contain signals of Fp1, fp2, AF3, AF7, AF8, AF4 and its peripheral electrodes. The top lobe areas on the left and right sides are each divided into 11 brain area samples, each sample containing 5-6 core electrodes and their peripheral electrode signals. The midline region may be divided into 6 brain region samples, including FCz, cz, CPz, pz, POz, oz, each consisting of a midline electrode and its neighboring electrodes. The occipital lobe region can be divided into 4 brain region samples, which are composed of core electrodes such as PO7, PO3z, PO4z, PO8 and the like and peripheral electrode signals thereof. Other electrode signals can be respectively marked into nearest neighbor related brain area samples according to the spatial positions of the electrode signals.

An important feature of this brain region division method is that it not only considers the activity of individual brain regions, but also reflects the interaction between brain regions by including peripheral electrode information. This is critical for understanding complex patterns of brain electrical activity, as the various regions of the brain do not operate in isolation, but rather cooperate, interact with each other. For example, while occipital lobe areas are the primary visual processing centers when performing visual tasks, frontal and parietal lobe areas may also participate in advanced visual cognitive processes. By the partitioning method of the present application, such trans-regional collaborative activities can be better captured and analyzed.

In practice, the present application requires the establishment of a mapping relationship between a brain region and electrodes. This can be accomplished by creating a data structure that contains a list of core and peripheral electrodes for each brain region sample. The application can then write the corresponding function, extract each brain region and its context signal from the original EEG data, and generate a data set of 41 region samples.

The detailed brain region division and signal extraction process can enable the brain signal acquisition device to better capture the spatial distribution characteristics of brain signals. The method provides richer and more targeted information for subsequent feature extraction and analysis, so that the method can more fully describe and understand the complex brain electrical activity mode. By the method, the brain information can be kept, and the brain information can be focused on the activity of a specific brain region, so that more accurate and targeted brain electrical signal analysis is realized.

Step S103, dividing the computing units corresponding to the pre-trained mental state evaluation model to obtain a plurality of computing subunits, wherein each computing subunit corresponds to one region.

Specifically, the application groups the calculation units of the trained electroencephalogram analysis model to obtain a plurality of calculation subunits, and each group of calculation subunits corresponds to a specific brain region sample. The establishment of the corresponding relation enables the model to capture and analyze the characteristics of the brain electrical signals from different brain areas more accurately. For example, the calculation subunit associated with the forehead lobe will process exclusively the brain electrical signals from the forehead lobe area, while the calculation subunit associated with the occipital lobe is focused on processing visually related brain electrical activity.

Step S104, determining a key brain region in the brain regions, wherein the region corresponding to the key brain region is the key region, and constructing a first brain electrical characteristic sequence of the key region through a computing subunit corresponding to the key region.

Specifically, from the multiple regions divided in the steps, the time domain global features and the frequency domain global features of the key brain region are extracted based on a time-frequency analysis method to form a first electroencephalogram feature sequence, and the first electroencephalogram feature sequence is realized on a computing subunit of the region.

In some embodiments, the mental state assessment model comprises an input layer, a plurality of convolution layers, at least one fully connected layer, and an output layer, the computing unit corresponding to a plurality of convolution kernels within the convolution layers; the dividing the computing units corresponding to the pre-trained mental state evaluation model to obtain a plurality of computing subunits, including: dividing the convolution kernel of the first layer of the convolution layer, and dividing the convolution kernel into the computing subunits corresponding to the brain regions, namely a forehead lobe region, a top lobe region, a midline region and a occipital lobe region; obtaining the size of a feature map output by a first layer of the convolution layer, and dividing the convolution kernel of a second layer of the convolution layer into a plurality of calculation subunits corresponding to the brain areas according to the size of the feature map; a shared identifier is added to the convolution kernels of the remaining convolution layers, such that each of the compute subunits can use the convolution kernel with the shared identifier.

Assume that the application uses an electroencephalogram analysis model based on a Convolutional Neural Network (CNN). The basic structure of this model includes: an input layer of size n×t (N is the number of channels, T is the number of time points); the 4 convolutions are used for extracting the space-time characteristics of the brain electrical signals; a 2-layer full link layer for integrating these features; finally, an output layer is provided for giving the final analysis result. The weight parameters of the entire network total about 370 tens of thousands.

To divide this network into 41 groups of computing units, corresponding to the 41 brain region samples divided before the present application, the present application may take the following grouping strategy:

First, the present application groups the first layer convolution kernels. In particular, the present application may assign the first 16 convolution kernels to 8 sets of computing units of the frontal lobe region, 2 convolution kernels per set. This is set up because the frontal lobe area is closely related to advanced cognitive functions, requiring more computing resources to handle its complex signal patterns. Next, the present application may assign 22 convolution kernels to 11 sets of computing units of the left top leaf, 2 convolution kernels each; likewise, another 22 convolution kernels are assigned to the 11 sets of computing units of the right top leaf. The top leaf area participates in important functions such as spatial processing and attention distribution, and thus a considerable number of calculation units are also required.

For the midline region, the present application may assign 12 convolution kernels to 6 groups of computing units, each group of 2 convolution kernels. The midline region, although spatially small, connects the left and right hemispheres and plays an important role in many cognitive processes. The pincushion region may assign 8 convolution kernels to 4 groups of computing units, each group of 2 convolution kernels. While occipital lobes are primarily responsible for visual processing, their signal patterns may be complex, requiring sufficient computing resources. Finally, the present application can assign the remaining 4 convolution kernels to 2 sets of computing units of other regions, 2 convolution kernels per set.

This allocation takes into account the functional importance and signal complexity of the different brain regions, ensuring that each critical brain region has sufficient computational resources to handle its own signal patterns. Meanwhile, the grouping method also maintains certain flexibility, and allows information interaction between different brain areas.

For the second layer of convolution layers, the application can distribute the convolution kernels to each group of computing units proportionally according to the size of the first layer of output feature maps. This ensures that information from different brain regions can be adequately processed and integrated in a deeper network hierarchy.

Notably, starting from the third layer of convolutional layers, the application can no longer be grouped strictly by brain regions, as the dimensions of the feature map have been greatly reduced, but allows the computing units of these layers to share information from different brain regions. This design reflects the feature that advanced cognitive processes in the brain often involve the co-operation of multiple brain regions.

An important advantage of this grouping of computing units method is that it allows the model structure to correspond to the functional anatomy of the human brain. This not only improves the interpretation of the model, but also may improve its performance. For example, in analyzing brain electrical activity associated with working memory, the model may focus more on the signals of the forehead and top lobe regions, as these regions play a critical role in working memory.

In practical implementation, the application needs to assign a unique index number to 41 brain region samples. For example, the present application may number a first sample of the frontal lobe region 0, a second sample 1, and so on up to number 40. Then, when defining the convolutional layer of the CNN model, the present application needs to bind the index of the convolutional kernel with the number of the corresponding brain region. For example, a convolution kernel with index 0-1 would be responsible for extracting the features of forehead lobe region 1 with index 0, a convolution kernel with index 2-3 would be responsible for extracting the features of forehead lobe region 2 with index 1, and so on.

For the fully connected layer, the application also needs to divide the weight vector of the neuron into 41 subvectors according to a certain proportion and connect with the calculation unit group of the corresponding brain region. This design ensures that information from different brain regions remains somewhat independent even at the higher layers of the network.

In the model training stage, the application can divide the parameters of the whole network according to the index determined before by utilizing the sub-network segmentation function provided by the deep learning framework, such as tf.slice operation of TensorFlow, so as to obtain 41 sub-network parameter sets. Each subset corresponds to a brain region and is responsible for processing the brain electrical signals of the brain region. Then, the application can construct a calculation map of the whole model, and each sub-network parameter set can calculate the output of the corresponding brain region respectively. The parallelization calculation mode can not only improve training efficiency, but also enable the model to process information from different brain areas simultaneously.

In this way, the application finally obtains 41 groups of computing units, each group focusing on extracting and analyzing the brain electrical characteristics of a specific brain region. The design lays a foundation for optimizing and adjusting the calculation unit in the subsequent step, so that the method can enhance the influence of the method in the whole analysis aiming at the characteristics of different brain areas. For example, the application may adjust the weights of the computational units corresponding to the critical brain regions according to the particular analysis task to increase the sensitivity of the model to these regional signals.

Illustratively, before said determining a critical brain region among a plurality of said brain regions, further comprising: acquiring a weight vector corresponding to the full connection layer, and dividing the weight vector into a plurality of sub-vectors; each of the sub-vectors is coupled to one of the computing sub-units.

For the full connection layer, the application adopts a special grouping mode. Specifically, the weight vector of each full connection layer is divided into 41 subvectors according to a certain proportion, and each subvector is respectively connected with the calculation unit group of the corresponding brain region. This design allows the model to integrate information from different brain regions while still maintaining a distinction of contributions to the individual brain regions.

In some embodiments, the constructing, by the computing subunit corresponding to the key region, the first electroencephalogram feature sequence of the key region includes: performing short-time Fourier transform on the brain electricity target signal corresponding to the key region to obtain a time-frequency diagram corresponding to the key region; extracting time domain global features and frequency domain global features corresponding to the key regions according to the time-frequency diagram; and splicing the time domain global features and the frequency domain global features to obtain the first electroencephalogram feature sequence.

First, the critical brain area to be analyzed is determined: firstly, the target brain electrical state such as attention, emotion, pressure, fatigue and the like is clarified, and related neuroscience research literature is consulted to know the association relation between different brain electrical states and specific brain areas. For example, the attention state is mainly related to brain electrical activity in regions of the forehead, parietal cortex, etc.; positive emotion is related to the enhancement of brain electrical activity in areas such as amygdala, ventral premenstrual cortex, etc.; negative estrus preliminary provision is associated with increased brain electrical activity in the areas of the islets, anterior cingulate gyrus, etc. Candidate brain regions closely related to the target brain electrical state can be preliminarily determined through literature research. And then comparing the candidate brain regions with 41 brain region samples divided in the steps to find out the region samples containing the candidate brain regions. For example, if the study shows that the attention status is closely related to the forehead lobe region and the top lobe region, out of 41 region samples, 8 region samples containing forehead lobes and 22 region samples containing top lobes are preliminary key region candidates. The specific brain region covered by each candidate region sample is then carefully examined, ensuring that it fully encompasses the target brain region, and that neighboring regions are properly expanded as context supplements. For example, the forehead lobe area 2 nd area sample contains AF3, AF4, F7, F5, F3, F1 electrodes, can completely cover the forehead She Fu lateral area, and also includes some of the orbit area and orbit forehead area electrodes, so it can be determined as one of the critical brain area samples. According to the complexity of the brain electrical state and the number of relevant brain regions reported in the literature, 5-10 most critical regions are screened from the candidate region samples to be used as final analysis objects, and the importance of each region is weighed during screening. For example, attention state studies indicate that the forehead and top lobes are two core regions, so they must be considered as critical regions; and the correlation between occipital leaves and attention is low, which can be temporarily ignored.

For each critical brain region, a Short Time Fourier Transform (STFT) is first performed on its brain electrical signal to obtain a time-frequency graph representation. Let the input brain electrical signal be x (t), the window function be w (t), the calculation formula of STFT be STFT (t, f) = jc x (τ) w (τ -t) e (-j 2 pi f tau) dτ, where t is time and f is frequency. A suitable window length is selected, e.g. 256 time points, corresponding to a 500Hz sampling rate window length of 512ms, and the step size is set to 64 time points, i.e. 128ms. And performing STFT (space time Fourier transform) on the brain electrical data of each key brain region to obtain the time-frequency representation of the brain electrical data.

Then, two types of global features, namely a time domain and a frequency domain, are extracted from the time-frequency diagram. In the time domain, for each time segment, the Hjorth parameters, namely three parameters of Activity (A), mobility (M) and Complexity (C), are calculated from the time axis of the time-frequency diagram, and represent the variance of the electroencephalogram signal, the variance ratio of the derivative and the entropy of the variance ratio of the derivative respectively, so that the global statistical characteristic of the time domain signal can be well described. The calculation formula is A=var (x), M= (var (x '))/(var (x)), C= (M-1))/(var (x' ')/var (x'))/(1/2), wherein var is a variance calculation, and x '' are a first derivative and a second derivative respectively. And calculating three parameters of Hjorth for each window section by adopting a sliding window with the length of 1s and overlapping by 50 percent, and taking the three parameters as the time domain global characteristic of the time point.

In the Frequency domain, parameters such as a relative power spectrum (Relative Power Spectrum, RPS), a spectrum entropy (Spectral Entropy, SE), a spectrum edge Frequency (SPECTRAL EDGE Frequency, SEF) and the like are calculated as global features for the spectrum of each time segment. The RPS is calculated by dividing the frequency into five frequency bands of delta (1-4 Hz), theta (4-8 Hz), alpha (8-13 Hz), beta (13-30 Hz) and gamma (30-50 Hz), then calculating the power P_delta, P_theta, P_alpha, P_beta and P_gamma of each frequency band respectively, normalizing the power to be the relative power ratio of RPS_delta=P_delta/sum (P), RPS_theta=P_theta/sum (P), and so on, wherein sum (P) is the total power. SE reflects the disorder degree of the distribution of the brain electrical spectrum, the calculation formula is se= -sum (p.log 2 (P))/log 2 (N), and N is the frequency point number. Let the minimum frequency required to contain a% of the total energy be f_a, sef_a=f_a/f_s is 100%, and SEF95% is typically used for a=95%. And respectively extracting RPS, SE, SEF and other frequency domain parameters for each time segment by adopting the sliding window which is the same as the time domain, and taking the frequency domain parameters as the frequency domain global characteristics of the time point.

And splicing the global features of the time domain (Hjorth parameters) and the frequency domain (RPS, SE, SEF) of each time point into a feature vector to form the global electroencephalogram features corresponding to the time point. And splicing the global features of all time points according to time sequences to finally form a first global brain electrical feature sequence of the key brain region. And (3) adopting the same time-frequency analysis and feature extraction flow to the other key brain regions to obtain a global electroencephalogram feature sequence corresponding to each region. And finally, cascading all the key brain region sequences to obtain a complete first brain electrical characteristic sequence containing global time domain frequency domain information.

Step S105, obtaining edge local features corresponding to the key brain regions in the multiple regions based on the local correlation analysis method, and optimizing weight parameters of the computing subunits corresponding to the key brain regions according to the edge local features.

Specifically, the critical brain region samples obtained in the above steps need to be segmented, and divided into a plurality of overlapped local window segments, the window length is set to 256 time points (512 ms corresponding to a 500Hz sampling rate), and the step length is set to 64 time points (128 ms). Then, within each partial window, for this window center point t_c, the pearson correlation coefficient R (t_c, t_i) between it and other time points t_i (i=1, 2,..256) within the window is calculated, forming a 256×256 partial correlation matrix R centered on t_c. For each local window, elements R (1, 1), R (2, 2) R (256) on the diagonal line are spliced in sequence, so that a 256-dimensional local autocorrelation characteristic vector can be obtained, and the 256-dimensional local autocorrelation characteristic vector is used for describing the autocorrelation change condition of local electroencephalogram signals in the window.

In addition, a cross-correlation coefficient between the window center point t_c and the edge point t_e needs to be calculated, where t_e is n edge time points nearest to t_c (n=8 is typically taken). Taking the forehead leaf area 2 as an example, the area sample includes electrodes such as AF3, AF4, F7, F5, F3, F1, etc., when calculating the local edge cross-correlation between the center point F3 and the edge points, AF3, AF4, F7, F5, F1 can be selected as an edge point set, the cross-correlation coefficient r (F3, t_e) between F3 and each edge point can be calculated respectively, and then the average value of the cross-correlation coefficients can be taken as the edge cross-correlation value of the local window position. The edge cross-correlation values are spliced into an n-dimensional local edge cross-correlation feature vector which is cascaded with the previous 256-dimensional local autocorrelation feature vector to obtain a 256+n-dimensional local correlation feature vector of the window position, and the 256+n-dimensional local correlation feature vector is used for describing the correlation of the brain electrical signals between the window center point and the internal and edge points. In this way, for each local window position, a corresponding local correlation feature vector can be extracted, and then spliced in the time dimension to form a local correlation feature sequence on the whole time sequence, wherein the feature sequence comprises local autocorrelation information in the key brain region and local edge cross-correlation information between the key brain region and an edge region. And (3) extracting corresponding local autocorrelation and edge cross correlation sequences by adopting the same local correlation analysis method for other critical brain region samples.

It should be emphasized that the above-mentioned local correlation analysis process needs to use a sliding window manner to dynamically update and calculate in the time dimension, so as to effectively capture the time variation of the local features. Specifically, for each critical brain region sample, an initial window position is set first, and a local autocorrelation matrix and an edge cross correlation coefficient are calculated in the window to obtain a local autocorrelation value and an edge cross correlation value of the time point. And then sliding the window to the next time position according to the set step length, and repeating the calculation process to obtain a new local correlation characteristic value. The window is continuously slid until the whole time sequence is covered, and finally the local autocorrelation sequence and the edge cross correlation sequence which dynamically change along with time can be obtained.

Through the local correlation analysis, not only the local autocorrelation characteristics in the key brain region can be extracted, but also the local edge correlation between the region and the peripheral region can be analyzed, so that the defect that important information of the peripheral region is ignored only when the central brain region is concerned is overcome. The local autocorrelation sequence reflects the fine dynamic activation mode of the neuron clusters in the key brain region, the edge cross correlation sequence describes the interaction condition between the brain region and the peripheral region, and the two mutually complement each other to jointly form a more complete and rich local brain electrical characteristic representation. In addition, because sliding window calculation is adopted, the obtained local correlation sequence itself contains time dynamic change information, and the time-varying characteristic of the electroencephalogram signal can be captured well.

An advantage of this local correlation analysis method is that it is able to capture subtle changes in the electroencephalogram signal and local features that may be averaged or ignored in the global analysis. For example, in certain cognitive tasks, a transient but significant pattern of activation may occur in a particular brain region, which may last for only a few hundred milliseconds. Through local correlation analysis, the present application can detect such transient activation patterns and correlate them to task performance or other physiological indicators.

In addition, edge cross-correlation analysis may reveal functional connection patterns between different brain regions. For example, in some cognitive tasks, the forehead and top lobes may exhibit strong synchronous activity, which may reflect the activation of the attention network. By analyzing the edge cross-correlation between these brain regions, the present application can quantify the strength of this functional connection and observe its change over time.

And (3) adjusting and optimizing the weight parameters of the computing subunits corresponding to the key region by using the local characteristics of the edge of the key brain region obtained by the steps, and enhancing the influence of the brain electrical characteristics of the region on the final result.

For each determined critical brain region, the weight of the neural network computing unit corresponding to the region needs to be optimized and adjusted by utilizing the edge cross-correlation brain electrical characteristic sequence of the region, so that the output of the neural network computing unit can be well fit with the edge cross-correlation sequence when the original brain electrical signal is input, and the influence of the brain electrical characteristic of the critical brain region on the final brain electrical state evaluation result is enhanced. The scheme adopts a simulated annealing algorithm to carry out optimization adjustment, and comprises the following specific steps:

Firstly, some parameters required by an initial temperature T0, a termination temperature Tend, the temperature reduction during each iteration, a cooling coefficient alpha and the like are required to be initialized, the setting of the parameters has important influence on the optimization performance of the algorithm, and the optimal value is required to be determined through multiple experiments according to a specific electroencephalogram signal analysis problem.

Generally, the initial temperature T0 needs to be set high enough to ensure that the algorithm can perform sufficient search in the whole electroencephalogram feature solution space at high temperature, so as to avoid sinking into local minima; the ending temperature Tend needs to be set with a lower value, and when the temperature drops to the value, the algorithm can end iteration and output the optimal solution, because the temperature is low enough at this time, the algorithm can not jump out of the current minimum value; the smaller the temperature reduction amount is in each iteration, the slower and smoother the cooling process is, which is beneficial to the algorithm to fully search the electroencephalogram characteristics in each temperature range, but the convergence time of the algorithm is increased; the temperature reduction coefficient alpha usually takes a value of 0.7-0.99, and when the alpha is larger, the temperature reduction rate is slower, so that the algorithm is favorable for carrying out fine search locally.

Depending on the complexity of the specific electroencephalogram analysis problem, the ideal setting value of the parameter is usually determined through continuous debugging and testing, but generally, the initial temperature T0 is about 1000, the end temperature Tend is about 1e-6, the cooling amount can be calculated according to 1% -5% of the initial temperature, and the cooling coefficient alpha is about 0.95. After the initialization is completed, an initial weight parameter set theta0 is given and used as a searching starting point of an algorithm, and the initial solution can be an electroencephalogram signal analysis model weight parameter obtained through pre-training or a group of weight parameter values generated randomly.

In each iteration of the algorithm, firstly, the current weight parameter set theta is calculated, when the key brain region electroencephalogram original signal x is input, the current weight parameter set theta corresponds to a loss function value L0=L (f (x; theta) between the output f (x; theta) of the neural network calculation unit and the edge cross-correlation electroencephalogram characteristic sequence y of the region, and the loss function L can adopt a common mean square error loss function or other suitable loss function forms. Then, a new candidate weight parameter set theta_new=theta+delta_theta is generated by adding a random disturbance term delta_theta to the current weight parameter set theta, and then a loss function value L_new corresponding to the new parameter set is calculated.

Next, it is necessary to determine whether to accept the new candidate weight parameter set theta_new as the initial solution for the next iteration, where the criteria used are: when L_new < L0, it is necessary to accept the new solution theta_new to replace the old solution; when l_new > =l0, then a new solution is accepted with a certain probability p=exp (- (l_new-L0)/T), where T is the current temperature. It can be seen that when the loss function value of the new solution is smaller than the current solution, it is necessarily accepted, which ensures that the algorithm does not reject a new solution that is better than the current solution; when the loss value of the new solution is larger, a certain probability is still accepted, but the probability is smaller and smaller along with the reduction of the temperature T, so that at high temperature, the algorithm has a larger opportunity to accept the solution with increased loss so as to jump out of the local minimum trap, and at low temperature, the algorithm is more prone to accept only the new solution with reduced loss.

If it is decided to accept the new solution theta new, it is taken as the initial solution for the next iteration; if the new solution is rejected, the current solution theta is kept unchanged as the initial solution for the next iteration. At the end of each iteration, the current optimal solution also needs to be perturbed with a certain probability to prevent the algorithm from prematurely converging into a suboptimal solution. After the above iterative process is completed, the temperature T needs to be properly reduced according to a preset cooling mode, for example, according to a cooling coefficient alpha, so as to obtain a new current temperature, and a next iterative cycle is performed based on the new temperature.

Repeating the iterative process until the temperature T is reduced to a preset termination temperature Tend or other termination conditions are met, and outputting the current obtained optimal weight parameter set theta_best as a global optimal solution by the algorithm. the theta-best is the weight of the calculation unit of the neural network after being specially optimized and adjusted aiming at the critical brain region, so that when an electroencephalogram original signal of the region is input, the output f (x; theta-best) of the calculation unit can be well fit with the edge cross-correlation electroencephalogram characteristic sequence y of the region, thereby enhancing the mapping capability of the relation between the critical brain region and the target electroencephalogram state and improving the sensitivity of the whole model to the electroencephalogram activity of the brain region.

An important advantage of this optimization method is that it enables adaptive adjustment for specific brain electrical characteristics of each critical brain region. For example, for the frontal lobe region, the optimized weights may be more sensitive to high frequency beta band brain electrical activity, consistent with the role of the frontal lobe in attention and executive control. While for occipital regions, the optimized weights may be more focused on alpha band brain electrical activity, reflecting the importance of occipital in vision processing and relaxation states.

In addition, this optimization method also takes into account interactions between brain regions. By using the edge cross-correlation electroencephalogram feature sequence as an optimization target, the application not only optimizes the electroencephalogram signal processing in a single brain region, but also optimizes the electroencephalogram information interaction between the brain region and surrounding regions. This is particularly important for capturing complex brain function network activities, as many cognitive processes and brain electrical states involve the cooperative work of multiple brain regions.

Step S106, the time length and the position of a sliding window of a calculating subunit corresponding to the key brain area are adjusted, and a second brain electrical characteristic sequence of the key brain area is constructed through the calculating subunit corresponding to the key brain area.

Specifically, in the above steps, based on the critical brain area determined in the previous step, the time-frequency analysis and global feature extraction of the electroencephalogram signal are performed again by using a sliding time window of 512ms and a step length parameter of 256ms, so as to obtain a second electroencephalogram feature sequence. Unlike the above steps in which 256ms window length and 128ms step length are used to extract the first global feature sequence, the longer time window and the larger step length set here aim to capture the electroencephalogram information of a longer time course, so that the slow evolution process of the electroencephalogram state can be better described. The occurrence and transition of the brain electrical state is not usually a transient process, but is accompanied by a gradual adjustment of the brain's internal thermodynamics, which requires a certain time to complete, gradually transitioning from one state to another. Therefore, the adoption of a longer window length of 512ms is beneficial to acquiring dynamic characteristics of brain activities in a wider time range, so that the whole process of the evolution of the brain electrical state can be reflected more completely.

Meanwhile, a larger step length of 256ms is selected, so that the calculation and storage overhead is reduced, and the analysis efficiency is improved. In the above step, a smaller step size of 128ms is used in order to obtain finer time resolution within a fixed 256ms time scale. However, in the above steps, the goal is to extract dynamic electroencephalogram features of longer duration, so that the requirement on time resolution can be properly relaxed, and a larger step size can be adopted to reduce the calculation amount. It should be noted that, although the step size is increased, since the time window length is doubled correspondingly, the time sampling density equivalent to the above steps can be obtained in practice within the same time range, thereby ensuring the time coverage rate of the electroencephalogram feature extraction.

After selecting the 512ms window length and 256ms step size, a time-frequency analysis will be performed on the raw brain signal data for each critical brain region. Similar to the above steps, the brain electrical data of each critical brain region is first subjected to a Short Time Fourier Transform (STFT) to obtain a time-frequency representation thereof within a 512ms time window. In calculating the STFT, a hamming window function of length 512ms and overlapping 256ms is required to ensure that the time window can be slid continuously over the entire time axis in 256ms steps without discontinuities. The STFT conversion formula is as described above, assuming that the input electroencephalogram signal is x (t), and the window function w (t) is a hamming window with a length of 512ms, the STFT result at this time point is: STFT (t, f) = ≡x (τ) w (τ -t) e++j 2 pi f τ) dτ.

Where t is the current point in time and f is the frequency component. By this transformation, a short-time fourier spectrum centered on t and having a time width of 512ms can be obtained, reflecting the time-frequency distribution characteristics of the electroencephalogram signal in this time period.

Next, global electroencephalographic features of the time and frequency domains need to be extracted from the time-frequency representation. In terms of time domain, the Hjorth parameter is calculated, and the Hjorth parameter comprises Activity, mobility and Complexity indexes which respectively represent the variance, the first derivative variance ratio and the normalized second derivative variance ratio of the electroencephalogram time sequence, so that the statistical time domain characteristic of the electroencephalogram can be well described. The calculation formula is as follows:

Activity = var(x)；

Mobility = (var(x'))/(var(x))；

Complexity = (Mobility*(Mobility-1))/(var(x'')/var(x'))^(1/2)；

In the frequency domain, the Relative Power Spectrum (RPS), spectral Entropy (SE) and Spectral Edge Frequency (SEF) will be extracted as global electroencephalogram features. The RPS is calculated by dividing the frequency into five frequency bands of δ (1-4 Hz), θ (4-8 Hz), α (8-13 Hz), β (13-30 Hz) and γ (30-50 Hz), then calculating the power p_δ, p_θ, p_α, p_β and p_γ of each frequency band in the current 512ms time window, and normalizing them to a relative power ratio, such as rps_δ=p_δ/sum (P). SE reflects the disorder degree of the distribution of the brain electrical spectrum, and the calculation formula is se= -sum (p.log2 (P))/log 2 (N), wherein N is the number of frequency points, and P is the power of each frequency component. SEF represents the percentage of the minimum frequency required to contain a% of the total energy over the entire sampling frequency, typically taking a = 95% using SEF95% as the frequency domain feature.

And combining the time domain Hjorth parameter with indexes such as the frequency domain RPS, SE, SEF and the like to form a global electroencephalogram characteristic vector of the time point. By sliding the 512ms time window with 256ms step length, the global electroencephalogram characteristic sequence on the whole time sequence can be obtained, and the electroencephalogram dynamic change condition of the key brain region under the time scale is reflected. And executing the same time-frequency analysis and feature extraction process on other key brain regions, and finally obtaining global electroencephalogram feature sequences corresponding to the key regions. And splicing the sequences according to a certain sequence to form a second electroencephalogram global characteristic sequence in the steps.

It is worth mentioning that the feature sequence extracted under the longer time window of 512ms can better capture the slowly evolving dynamics of the electroencephalogram state compared with the first feature sequence obtained based on the 256ms window in the above steps. For example, in analyzing cognitive load changes, the brain electrical state tends not to change drastically in a short time, but rather to exhibit a slow rising or falling trend that is more easily observed on a 512ms time scale. If the features extracted by the 256ms window only depend on the above steps, the overall trend may be ignored due to the too short time scale. By supplementing the characteristic sequence under the window length of 512ms through the steps, the dynamic evolution process of the longer time interval of the electroencephalogram state can be effectively captured, so that more comprehensive information support is provided for subsequent time evolution analysis.

Step S107, inputting the first electroencephalogram feature sequence and the second electroencephalogram feature sequence into a psychological state assessment model respectively, obtaining a first electroencephalogram state corresponding to the first electroencephalogram feature sequence and a second electroencephalogram state corresponding to the second electroencephalogram feature sequence, and completing analysis of the original electroencephalogram signals according to the first electroencephalogram state and the second electroencephalogram state.

Specifically, since the input of the Convolutional Neural Network (CNN) needs to be three-dimensional tensor data, it is necessary to convert the one-dimensional global electroencephalogram feature sequence into a two-dimensional feature map first. Assuming that the length of the first global electroencephalogram feature sequence is T1, and the feature dimension is M, the first global electroencephalogram feature sequence comprises N key brain regions, then the first global electroencephalogram feature sequence can be remodeled into a three-dimensional tensor of [ N, T1, M ] to serve as an input feature map of CNN. Similarly, for a second global electroencephalogram feature sequence of length T2, it can also be converted into an input tensor of [ N, T2, M ].

In some embodiments, the analyzing the original electroencephalogram signal according to the first and second electroencephalogram states includes: respectively obtaining a first output score and a second output score corresponding to the first electroencephalogram state and the second electroencephalogram state; if the first output score and the second output score are both larger than a first preset threshold, determining that the user corresponding to the original electroencephalogram signal is in a continuous high-activation state; if the first output score and the second output score are smaller than a second preset threshold value, determining that the user is in a relatively resting or relaxed state; and if the first output score is smaller than the second preset threshold value and the second output score is larger than the first preset threshold value, determining that the user is in a transition period of activating the steering inhibition, and if the first output score is smaller than the second preset threshold value and the second output score is smaller than the first preset threshold value, determining that the user is in a transition period of activating the resting steering.

After the electroencephalogram features are input to the CNN, a convolution operation needs to be performed on the feature map of each brain region separately. According to the unit grouping of the above steps, a calculation unit group corresponding to each brain region can be obtained, including a convolution kernel parameter set of the group. And then, performing convolution operation on the characteristic map of the brain region by utilizing the optimized and adjusted parameter weight in the steps, so as to extract the characteristic local brain electrical mode characteristic of the region. After convolution operation, the feature map of each region is converted into a new feature map sequence, and the local mode distribution of the brain electrical signals of the region in time and channel dimension is reflected.

After the first layer convolution, the output feature map sequences of all regions need to be combined to form the input of the next layer. The merging process is to splice the feature map sequences of all the regions in the first dimension (brain region dimension) to form a large feature map sequence containing the information of all the regions. And then inputting the large sequence into a second-layer convolution unit, and continuously extracting a higher-level brain electrical characteristic mode. The subsequent third and fourth convolutions and the two full-connection layers are all performed on the whole large feature map sequence.

Finally, the output layer of the model outputs a scalar value representing the prediction score of the brain electrical state under the current time window. Because two global electroencephalogram feature sequences with different time scales are input, the model can evaluate the states of the two sequences and output corresponding scores. In particular, for the feature sequence of the first 256ms window, the score y1 (t) of the model output may reflect the instantaneous level of the brain electrical state at time t. If y1 (t) is high, this point in time indicates that the brain is in a particular activation state; if y1 (t) is low, this indicates that the brain may be in a relatively resting or inhibited state at this point in time.

For the characteristic sequence of the second 512ms window, the score y2 (t) output by the model can reflect the overall change trend of the electroencephalogram state in a longer time period. For example, if y2 (t) is in an upward trend, it is indicated that the brain's activity level is gradually increasing during this period of time; if y2 (t) is in a decreasing trend, it is indicated that brain activity may be gradually coming to rest.

By comparing the two output scores of y1 (t) and y2 (t), a comprehensive determination can be made of the electroencephalogram state at time t:

1. if both y1 (t) and y2 (t) are higher, indicating that the instantaneous brain electrical activity at time t is stronger, and also consistent with the brain electrical trend over longer periods of time, it may be indicative of the brain being in a sustained highly active state, such as a highly focused or intense emotional experience.

2. If both y1 (t) and y2 (t) are lower, indicating that the instantaneous brain electrical activity at time t is weaker, and also consistent with the brain electrical trend over longer periods of time, it may be indicative that the brain is in a relatively resting or relaxed state.

3. If y1 (t) is higher and y2 (t) is lower and trend downward, this indicates that the instantaneous brain electrical activity at time t is stronger, but there is a divergence from the trend of gradual decrease in brain electrical activity over a longer period of time, which may mean that the brain state is turning from active to inactive, in the transitional phase of the transition.

4. If y1 (t) is lower and y2 (t) is higher and in an upward trend, this indicates that the instantaneous brain activity at time t is weaker, but there is a divergence from the trend of increasing brain activity gradually over a longer period of time, which may mean that the brain state is turning from rest to active, also in the transitional phase of transition.

Through the analysis, the instantaneous electroencephalogram state at a specific time point can be estimated, and the long-time dynamic evolution trend of the electroencephalogram state can be obtained, so that the time change process of the electroencephalogram state is comprehensively grasped. This method is particularly suitable for analyzing complex cognitive processes such as attention switching, mood changes, or brain electrical state transitions during learning.

Furthermore, since the evaluation is not based on a static output at a single point in time, but slides a time window along the time axis, the time-series outputs of y1 (t) and y2 (t) are continuously obtained, and thus the variation history of the brain electrical state can be dynamically tracked. For example, if both y1 (t) and y2 (t) remain higher and more consistent over a period of time, indicating that the brain is in a sustained highly active state, which may correspond to a prolonged task of concentration or a sustained emotional experience; if a significant divergence of y1 (t) and y2 (t) occurs before and after a certain time node and lasts for a period of time, the brain's activity state is explained to undergo a transition process from activation to inhibition or from inhibition to activation, and the time node and evolution history of the transition can be accurately captured.

The analysis method with double time scales provides a new visual angle for electroencephalogram signal research. The brain dynamic activity model can capture the instantaneous brain electrical state and reflect the long-term trend of brain electrical activity, thereby describing the brain dynamic activity mode more comprehensively. This is important for understanding complex cognitive processes, affective changes, and the brain electrical characteristics of various neuropsychiatric disorders. For example, in studying the brain electrical characteristics of Attention Deficit Hyperactivity Disorder (ADHD) patients, such methods may reveal unique patterns of brain electrical state changes in the patient during attention maintenance, providing new insight into clinical diagnosis and treatment.

Furthermore, the provided solution has the following beneficial effects:

In order to execute the electroencephalogram signal analysis method based on multi-scale electroencephalogram feature fusion corresponding to the embodiment of the method, corresponding functions and technical effects are achieved. Referring to fig. 2, fig. 2 shows a block diagram of an electroencephalogram signal analysis apparatus 200 according to an embodiment of the present application. For convenience of explanation, only a portion related to this embodiment is shown, and the electroencephalogram signal analysis apparatus 200 provided in the embodiment of the present application includes:

a signal acquisition unit 201, configured to acquire an original electroencephalogram signal to be analyzed, and perform preprocessing on the original electroencephalogram signal data to acquire an electroencephalogram target signal;

The signal dividing unit 202 is configured to divide the brain electrical target signal into a plurality of areas according to brain area distribution knowledge, each area corresponds to a brain area, and each area at least contains brain electrical target signals corresponding to the brain area;

The unit dividing unit 203 is configured to divide the computing units corresponding to the pre-trained mental state evaluation model, and obtain a plurality of computing sub-units, where each computing sub-unit corresponds to one of the areas;

A key determining unit 204, configured to determine a key brain region among the brain regions, where a region corresponding to the key brain region is a key region, and construct a first electroencephalogram feature sequence of the key region through the computing subunit corresponding to the key region;

The weight optimization unit 205 is configured to obtain, in a plurality of the regions, edge local features corresponding to the key brain regions based on a local correlation analysis method, and optimize weight parameters of a computation subunit corresponding to the key brain regions according to the edge local features;

A window adjusting unit 206, configured to adjust a sliding window time length and a position of a computation subunit corresponding to the key brain region, and construct a second electroencephalogram feature sequence of the key brain region through the computation subunit corresponding to the key brain region;

The signal analysis unit 207 is configured to input the first electroencephalogram feature sequence and the second electroencephalogram feature sequence into the mental state evaluation model, obtain a first electroencephalogram state corresponding to the first electroencephalogram feature sequence and a second electroencephalogram state corresponding to the second electroencephalogram feature sequence, and complete analysis of the original electroencephalogram signal according to the first electroencephalogram state and the second electroencephalogram state.

The electroencephalogram signal analysis device 200 can implement the electroencephalogram signal analysis method based on multi-scale electroencephalogram feature fusion according to the method embodiment. The options in the method embodiments described above are also applicable to this embodiment and will not be described in detail here. The rest of the embodiments of the present application may refer to the content of the above method embodiments, and in this embodiment, no further description is given.

Fig. 3 is a schematic structural diagram of a computer device according to an embodiment of the present application. As shown in fig. 3, the computer device 3 of this embodiment includes: at least one processor 30 (only one is shown in fig. 3), a memory 31 and a computer program 32 stored in the memory 31 and executable on the at least one processor 30, the processor 30 implementing the steps in any of the method embodiments described above when executing the computer program 32.

The computer device 3 may be a smart phone, a tablet computer, a desktop computer, a cloud server, or other computing devices. The computer device may include, but is not limited to, a processor 30, a memory 31. It will be appreciated by those skilled in the art that fig. 3 is merely an example of the computer device 3 and is not meant to be limiting as the computer device 3, and may include more or fewer components than shown, or may combine certain components, or different components, such as may also include input-output devices, network access devices, etc.

The Processor 30 may be a central processing unit (Central Processing Unit, CPU), the Processor 30 may also be other general purpose processors, digital signal processors (DIGITAL SIGNAL processors, DSP), application SPECIFIC INTEGRATED Circuit (ASIC), off-the-shelf Programmable gate array (Field-Programmable GATE ARRAY, FPGA) or other Programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or the like. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like.

The memory 31 may in some embodiments be an internal storage unit of the computer device 3, such as a hard disk or a memory of the computer device 3. The memory 31 may in other embodiments also be an external storage device of the computer device 3, such as a plug-in hard disk, a smart memory card (SMART MEDIA CARD, SMC), a Secure Digital (SD) card, a flash memory card (FLASH CARD) or the like, which are provided on the computer device 3. Further, the memory 31 may also include both an internal storage unit and an external storage device of the computer device 3. The memory 31 is used for storing an operating system, application programs, boot loader (BootLoader), data, other programs etc., such as program codes of the computer program etc. The memory 31 may also be used for temporarily storing data that has been output or is to be output.

In addition, the embodiment of the present application further provides a computer readable storage medium, where a computer program is stored, where the computer program is executed by a processor to implement the steps in any of the above-mentioned method embodiments.

Embodiments of the present application provide a computer program product which, when run on a computer device, causes the computer device to perform the steps of the method embodiments described above.

In several embodiments provided by the present application, it will be understood that each block in the flowchart or block diagrams may represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved.

The functions, if implemented in the form of software functional modules and sold or used as a stand-alone product, may be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application may be embodied essentially or in a part contributing to the prior art or in the form of a software product stored in a storage medium, comprising several instructions for causing a computer device to perform all or part of the steps of the method according to the embodiments of the present application. And the aforementioned storage medium includes: a usb disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, or an optical disk, or other various media capable of storing program codes.

The foregoing embodiments have been provided for the purpose of illustrating the general principles of the present application, and are not to be construed as limiting the scope of the application. It should be noted that any modifications, equivalent substitutions, improvements, etc. made by those skilled in the art without departing from the spirit and principles of the present application are intended to be included in the scope of the present application.

Claims

1. An electroencephalogram signal analysis method based on multi-scale electroencephalogram feature fusion is characterized by comprising the following steps:

acquiring an original electroencephalogram signal to be analyzed, preprocessing the original electroencephalogram signal data, and acquiring an electroencephalogram target signal;

Dividing the brain electricity target signal into a plurality of areas according to brain area dividing knowledge, wherein each area corresponds to one brain area, and each area at least comprises brain electricity target signals corresponding to the brain areas;

Determining a key brain region in a plurality of brain regions, wherein a region corresponding to the key brain region is a key region, and constructing a first brain electrical characteristic sequence of the key region through the computing subunit corresponding to the key region;

In a plurality of areas, acquiring edge local features corresponding to the key brain areas based on a local correlation analysis method, and optimizing weight parameters of computing subunits corresponding to the key brain areas according to the edge local features;

adjusting the time length and the position of a sliding window of a computing subunit corresponding to the key brain region, and constructing a second electroencephalogram characteristic sequence of the key brain region through the computing subunit corresponding to the key brain region;

and respectively inputting the first electroencephalogram characteristic sequence and the second electroencephalogram characteristic sequence into the psychological state assessment model, acquiring a first electroencephalogram state corresponding to the first electroencephalogram characteristic sequence and a second electroencephalogram state corresponding to the second electroencephalogram characteristic sequence, and completing analysis of the original electroencephalogram signal according to the first electroencephalogram state and the second electroencephalogram state.

2. The method according to claim 1, wherein the obtaining the original electroencephalogram signal to be analyzed comprises:

obtaining stimulus types and strength parameters corresponding to a plurality of cognitive stimulus tasks;

configuring acquisition parameters of preset electroencephalogram acquisition equipment according to the stimulation type and the intensity parameters;

and completing acquisition of the original brain electrical signals of the user under a plurality of cognitive stimulation tasks through the adjusted brain electrical acquisition equipment.

3. The method of claim 1, wherein the brain electrical target signal comprises a plurality of electrode measurement signals; the dividing the electroencephalogram target signal into a plurality of areas includes:

Determining a plurality of brain regions and brain region position information corresponding to the brain regions according to the brain region knowledge of the human brain;

acquiring electrode position information of an electroencephalogram acquisition device corresponding to the electroencephalogram target signal;

determining electrodes corresponding to each brain region according to the brain region position information and the electrode position information corresponding to each brain region;

and constructing a region corresponding to the brain region by using the electrode measurement signals corresponding to each brain region and the electrode measurement signals within a preset range, and completing construction of a plurality of regions.

4. The method of claim 1, wherein the mental state assessment model comprises an input layer, a plurality of convolution layers, at least one fully connected layer, and an output layer, the computing unit corresponding to a plurality of convolution kernels within the convolution layers; the dividing the computing units corresponding to the pre-trained mental state evaluation model to obtain a plurality of computing subunits, including:

dividing the convolution kernel of the first layer of the convolution layer, and dividing the convolution kernel into the computing subunits corresponding to the brain regions, namely a forehead lobe region, a top lobe region, a midline region and a occipital lobe region;

obtaining the size of a feature map output by a first layer of the convolution layer, and dividing the convolution kernel of a second layer of the convolution layer into a plurality of calculation subunits corresponding to the brain areas according to the size of the feature map;

a shared identifier is added to the convolution kernels of the remaining convolution layers, such that each of the compute subunits can use the convolution kernel with the shared identifier.

5. The method of claim 4, further comprising, prior to said determining a critical brain region among a plurality of said brain regions:

Acquiring a weight vector corresponding to the full connection layer, and dividing the weight vector into a plurality of sub-vectors;

each of the sub-vectors is coupled to one of the computing sub-units.

6. The method according to claim 1, wherein the constructing the first electroencephalogram feature sequence of the key region by the computing subunit corresponding to the key region includes:

Performing short-time Fourier transform on the brain electricity target signal corresponding to the key region to obtain a time-frequency diagram corresponding to the key region;

extracting time domain global features and frequency domain global features corresponding to the key regions according to the time-frequency diagram;

and splicing the time domain global features and the frequency domain global features to obtain the first electroencephalogram feature sequence.

7. The method of claim 1, wherein the analyzing the raw electroencephalogram signal based on the first and second electroencephalogram states comprises:

respectively obtaining a first output score and a second output score corresponding to the first electroencephalogram state and the second electroencephalogram state;

if the first output score and the second output score are both larger than a first preset threshold, determining that the user corresponding to the original electroencephalogram signal is in a continuous high-activation state;

If the first output score and the second output score are smaller than a second preset threshold value, determining that the user is in a relatively resting or relaxed state;

if the first output score is greater than the first preset threshold value and the second output score is less than the second preset threshold value, determining that the user is in a transition period for activating steering inhibition;

and if the first output score is smaller than the second preset threshold value and the second output score is larger than the first preset threshold value, determining that the user is in a transition period of resting turning activation.

8. An electroencephalogram signal analysis apparatus, comprising:

9. A computer device, the computer device comprising a memory and a processor;

the memory is used for storing a computer program;

the processor is configured to execute the computer program and implement the steps of the electroencephalogram signal analysis method based on multi-scale electroencephalogram feature fusion according to any one of claims 1 to 7 when the computer program is executed.

10. A computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program which, when executed by a processor, causes the processor to implement the steps of the electroencephalogram signal analysis method based on multi-scale electroencephalogram feature fusion as claimed in any one of claims 1 to 7.