WO2020042511A1 - 基于空间滤波与模版匹配的运动电位脑机接口编解码方法 - Google Patents

基于空间滤波与模版匹配的运动电位脑机接口编解码方法 Download PDF

Info

Publication number
WO2020042511A1
WO2020042511A1 PCT/CN2018/125926 CN2018125926W WO2020042511A1 WO 2020042511 A1 WO2020042511 A1 WO 2020042511A1 CN 2018125926 W CN2018125926 W CN 2018125926W WO 2020042511 A1 WO2020042511 A1 WO 2020042511A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
filter
template
spatial
spatial filtering
Prior art date
Application number
PCT/CN2018/125926
Other languages
English (en)
French (fr)
Inventor
明东
王坤
许敏鹏
Original Assignee
天津大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Priority claimed from CN201810988364.3A external-priority patent/CN109214325B/zh
Priority claimed from CN201810995425.9A external-priority patent/CN109271887A/zh
Application filed by 天津大学 filed Critical 天津大学
Publication of WO2020042511A1 publication Critical patent/WO2020042511A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition

Definitions

  • the invention relates to the field of motion-related potential detection, in particular to a method for encoding and decoding a brain-computer interface of a motion potential based on spatial filtering and template matching.
  • Electroencephalography reflects the bioelectrical activity of brain neurons.
  • the EEG mode can be used to control the Brain-Computer Interface (BCI) output.
  • BCI is a system that directly converts the activity of the central nervous system into artificial output. It can replace, repair, enhance, supplement or improve the normal output of the central nervous system, thereby improving the interaction between the central nervous system and the internal and external environment.
  • Movement-related potentials MRCPs
  • MRCPs are low-frequency potentials that are generated when the body calls exercise-related cognitive resources in the brain when performing suggestive or autonomous exercise.
  • the BCI system based on MRCPs detection can be used in rehabilitation training, intelligent prosthetics, and mechanical exoskeleton control, etc., and has attracted widespread attention from researchers.
  • the coding method is simple. Traditional coding methods mainly include simple limb movement imagination such as left and right hands, feet, tongue, etc. Therefore, the BCI instruction set is small, and it is extremely difficult to implement the large instruction control necessary for communication.
  • MRCPs can be viewed as event-related potentials related to exercise. Due to the low signal-to-noise ratio of EEG signals, it is difficult to effectively extract MRCP waveform features in a single trial. A commonly used method is superposition averaging to improve the signal-to-noise ratio. However, the method of overlay averaging requires the user to output a result after performing the same task multiple times, and the time cost is high. It is difficult to effectively extract the EEG waveform features related to a single task.
  • the scalp EEG signals can be viewed as the superposition of signals from different sources in the brain on the scalp. Features can be extracted using spatial filtering.
  • the main idea of spatial filtering is: by assigning weights to the signals of the original EEG, extracting the signal components of the source of interest, and removing noise signals, thereby improving the signal-to-noise ratio of the characteristic signals.
  • the main idea of template matching is to find the template with the highest matching degree as the recognition mode by matching the feature signal with the template signals of different patterns.
  • the invention provides a motor potential brain-computer interface encoding and decoding method based on spatial filtering and template matching.
  • the invention designs a motion intention encoding method based on sequence motion, and provides a new instruction encoding method, which can be effective
  • the extended motion intent decodes the BCI instruction set.
  • a spatial filtering method based on task-related component analysis and discriminant spatial pattern analysis is used to improve the signal-to-noise ratio of EEG signals. Compared with traditional methods, it can improve motion-related potential detection. Efficiency, further research can open up a new direction for the development of BCI, and it is expected to obtain considerable social and economic benefits, as described in the following description:
  • a motor potential brain-computer interface encoding and decoding method based on spatial filtering and template matching includes the following steps:
  • the pre-processed EEG data is used to construct a task-related component spatial filter, and the EEG data is filtered by the task-related component spatial filter to be the first filtered data;
  • the pre-processed EEG data is used to construct a spatial pattern filter, and the EEG data is filtered by the discriminant spatial pattern filter to be the second filtered data;
  • the test data is subjected to matching analysis of the two template data after passing through the task-related component spatial filter and discriminant spatial pattern filter, respectively, and finally decision classification is performed.
  • Another embodiment is a method for encoding and decoding a motor potential brain-computer interface based on spatial filtering and template matching.
  • the method performs spatial filtering on EEG signals based on task-related component analysis and discriminant spatial pattern analysis, and is realized by template matching.
  • the method includes the following steps:
  • the pre-processed EEG data is used to construct a task-related component spatial filter.
  • the pre-processed EEG data is used to construct a discriminant spatial pattern filter, that is, a DSP filter after the first spatial filtering.
  • the filtered data is subjected to the second spatial filtering to construct template data.
  • the test data is subjected to the spatial filtering and template data to perform matching analysis twice, and finally the decision classification is performed.
  • the invention uses the method based on task-related component analysis combined with discriminant spatial pattern analysis for the first time to achieve accurate recognition of motion-related potentials;
  • the method proposed by the present invention can be used in a brain-computer interface system based on motion intent detection, and the method can further improve the brain-computer interface technology and promote the transformation of the technology into application results;
  • the present invention proposes a sequence coding design, which effectively expands the instruction set for decoding BCI based on motion intent. Compared with the traditional spelling of character spellers (such as the P300-based Oddball paradigm), no external stimulus is required, and users only need Autonomous control of your own motion intent output instructions can truly achieve mind control.
  • the encoding method proposed by the present invention lays a foundation for the large instruction set operation of the MI-BCI system.
  • FIG. 1 is a flowchart of a motor potential brain-computer interface encoding and decoding method based on spatial filtering and template matching;
  • FIG. 2 is another flowchart of a motor potential brain-computer interface encoding and decoding method based on spatial filtering and template matching.
  • the spatial filtering method involved in the embodiments of the present invention includes: task-related component analysis (TRCA) and discriminative spatial pattern analysis (DSP).
  • TRCA aims to extract task-related components from linearly weighted multiple time series.
  • the main idea is to maximize the covariance or correlation between different trials of the same task.
  • DSP discriminative spatial pattern analysis
  • This method is an extension of Fisher's linear discriminant analysis in spatial analysis.
  • the spatial filter is obtained with emphasis on the largest intra-class dispersion projection and the smallest intra-class dispersion projection.
  • the template matching method in the embodiment of the present invention relates to Pearson correlation coefficient, Euclidean distance, and Canonical Correlation Analysis (CCA) method and eigenvalue comparison method.
  • CCA Canonical Correlation Analysis
  • An embodiment of the present invention provides a method for encoding and decoding a motor potential brain-computer interface based on spatial filtering and template matching.
  • Motor-related cortical potentials are patterns of EEG signals that are rich in motor information, strictly time-locked, and phase-locked when people imagine or perform physical movements. Because of its important research value in revealing the human motor nerve mechanism and guiding sports rehabilitation training, it has attracted widespread attention from researchers.
  • the embodiment of the present invention designs a motion-related potential detection method based on spatial filtering, which can improve the recognition accuracy rate.
  • the pre-processed EEG data 1 is used to construct a task-related component spatial filter, and the data 1 is filtered by the task-related component spatial filter to be data 2; Filter, data 1 is discriminated by the discriminant spatial mode filter, and is data 3, and template data is constructed according to data 2 and data 3.
  • the test data passes the task-related component spatial filter, discriminant spatial mode filter, and two template data. Perform matching analysis (ie, the process of performing coherence analysis), and finally perform decision classification (ie, the process of comparing eigenvalues).
  • the embodiment of the present invention uses the TRCA and DSP spatial filtering methods combined with the template matching method for the detection of motion-related potentials for the first time, which will help improve the recognition efficiency and is of great significance for the practical application of BCI based on motion-related potential detection. .
  • FIG. 1 is a flowchart of calculation according to an embodiment of the present invention, which can be used for motion-related potential detection.
  • M samples in each type of training data set x (m) represents the m-th sample
  • N c represents the number of channels for collecting EEG
  • N t represents the length of the intercepted signal.
  • Mean value of all samples in the training set to obtain the template signal of the training samples As shown in formula (1).
  • Data processing mainly includes four steps: data preprocessing, spatial filter construction and data spatial filtering, correlation analysis to extract features and template matching.
  • EEG signals are first pre-processed.
  • sampling frequency of the EEG signal is 1000Hz or higher.
  • Chebyshev filter to filter the signal from 0.5 to 45 Hz.
  • the training set and the test set are processed with zero mean on the time scale.
  • the spatial filtering is divided into two steps.
  • the first step is TRCA filtering to extract task-related components
  • the second step is DSP filtering to maximize the difference between the two types of data.
  • the TRCA algorithm belongs to a spatial filter. The purpose is to maximize the sum of covariance between different trials in the same task through the filter V.
  • the calculation method of the sum of covariance is shown in formula (2) Where N c is the number of channels for collecting EEG, N t is the length of the intercepted signal, N s is the number of samples in the training set, m 1 and m 2 are the numbers of the samples, k 1 and k 2 are the numbers of the leads, Represents the signal of the m 1 sample lead k 1 , and x is the sample set.
  • Cov is the covariance
  • Filters corresponding to leads k 1 and k 2 respectively
  • v is the filter matrix
  • S is the covariance matrix between trials.
  • Q is the covariance matrix between the leads.
  • the optimal solution V is the eigenvector of the matrix Q -1 S.
  • the training set TRCA filter obtained by calculation A total of n sub-filters are obtained.
  • the dimension of the sub-filter can be reduced
  • the selected dimension N ′ c can be determined empirically, and can also be optimized and adjusted according to the algorithm.
  • the reduced sub-filter combination becomes
  • N ′ c is the dimension of the filter after dimensionality reduction
  • DSP is a spatial filtering method, the purpose is to maximize the feature differences between different modes. This method is an extension of Fisher's linear discriminant analysis in spatial analysis.
  • the spatial filter is obtained with emphasis on the largest intra-class dispersion projection and the smallest intra-class dispersion projection.
  • the inter-class divergence matrix S b is shown in formula (6), with The template signals of the two types of training samples are obtained by averaging all the samples of the type i and j training sets.
  • the intra-class divergence matrix S w is shown in formula (7):
  • u is the filter corresponding to each lead.
  • the dimension of the sub-filter can be reduced.
  • the selected dimension N ′ c can be determined empirically, and can also be optimized and adjusted according to the algorithm.
  • the reduced sub-filter combination becomes
  • the correlation analysis method used in the present invention is the Pearson correlation coefficient, the Euclidean distance, and the Canonical Correlation Analysis (CCA) method.
  • the Pearson correlation coefficient is defined as: The Pearson correlation coefficient of two continuous variables X, Y is equal to the covariance cov (X, Y) between them divided by the product of their respective standard deviations ⁇ X ⁇ Y. The value range of the coefficient is [-1,1]. Variables close to 0 are said to have no correlation, and close to 1 or -1 are said to have strong correlation. Corr (*) is used here to represent the Pearson correlation coefficient.
  • Euclidean distance is the abbreviation of Euclidean metric, which refers to the true distance between two points in m-dimensional space.
  • dist (*) means Euclidean distance.
  • the basic principle of the CCA algorithm is: in order to grasp the correlation between the two groups of indicators as a whole, two representative comprehensive variables A and B are extracted from the two groups of variables (respectively for each variable in the two variable groups). Linear combination), using the correlation between the two integrated variables to reflect the overall correlation between the two sets of indicators.
  • the CCA algorithm can be used to project the spatially filtered data to a new space and calculate the correlation.
  • CCA (*) is used here for CCA analysis.
  • the feature vectors ⁇ i [ ⁇ i1 , ⁇ i2 , ⁇ i3 , ⁇ i4 , ⁇ i5 , ⁇ i6 , ⁇ i7 , ⁇ i8 ] T , ⁇ i ⁇ R 8 ⁇ 1 .
  • CCA (*) represents CCA analysis
  • corr (*) represents Pearson correlation coefficient
  • Ai and Bi are linear projection matrices
  • ⁇ i1 is the first feature element of the i-th class feature vector
  • ⁇ i2 is the i-th class.
  • ⁇ i3 is the third feature element of the i-th feature vector
  • ⁇ i4 is the fourth feature element of the i-th feature vector
  • ⁇ i5 is the fifth feature element of the i-th feature vector.
  • ⁇ i6 is the sixth feature element of the i-th type feature vector
  • ⁇ i7 is the seventh feature element of the i-th type feature vector
  • ⁇ i8 is the eighth feature element of the i-th type feature vector.
  • the features extracted in the embodiments of the present invention represent correlations, so a larger feature value indicates that it is more similar to a certain category.
  • weight coefficients ⁇ [ ⁇ 1 , ⁇ 2 , ⁇ 3 , ⁇ 4 , ⁇ 5 , ⁇ 6 , ⁇ 7 , ⁇ 8 ], ⁇ ⁇ R 1 ⁇ 8 are added to optimize the matching result.
  • the recognition result of the test data is a type I motion-related pattern.
  • the weight coefficient ⁇ can be determined empirically, and can also be optimized and adjusted according to the algorithm.
  • the traditional motion intent decoding brain-computer interface system has a small number of instructions in the instruction set, which is extremely difficult to implement the large instruction control necessary for communication.
  • the above-mentioned spatial filtering algorithm can realize fast decoding of motion-related potentials. Based on this, the present invention proposes a new method for encoding EEG signals of exercise intent.
  • the above-mentioned spatial filtering algorithm is used to decode EEG signals, which can effectively extend the motion-based intent. Decodes the instruction set of the brain-computer interface.
  • the technical process is as follows: the user according to the coding paradigm proposed by the present invention autonomously presses the keys, and simultaneously collects EEG signals for decoding and analysis.
  • the brain controls limb movements as contralateral controls, that is, the left functional area of the brain controls the movement of the right limb, and the right functional area of the brain controls the movement of the left limb, so the user is using the left-hand button At this time, the movement-related EEG signal characteristics of the corresponding spatial position on the right side of the brain are more obvious.
  • the right-hand key the movement-related EEG signal characteristics of the corresponding spatial position on the left side of the brain are more obvious.
  • the left-hand key used is defined as code 1 and the right-hand key is defined as code 0.
  • the "0/1" binary coding idea large instruction coding can be realized. Take encoding 4 instructions as an example, 2-bit encoding can be implemented, as shown in Table 1:
  • Table 1 shows the coding method
  • the number of keystrokes that is, the number of coding bits
  • a large instruction set can be controlled.
  • the number of encoded bits is n
  • the number of encoded instructions is 2 n .
  • the interval T between each keystroke depends on the minimum time that the EEG pattern recognition algorithm can effectively detect.
  • EEG signals acquisition Use the 64-lead EEG acquisition system of Neuroscan to collect EEG signals, and use Ag / AgCl electrodes (impedance less than 15000 ohms).
  • the EEG signals of all leads are referenced to the top of the head and the forehead is the ground. -20 international standard lead placement.
  • the EEG sampling frequency is 1000Hz
  • the filtering passband is 0-100Hz
  • a 50Hz notch is used to remove power frequency interference.
  • this method can improve the detection efficiency of exercise-related potentials.
  • further research on the detection method proposed by the present invention can obtain a perfect brain-computer interface system based on motion intention detection, which is used for the auxiliary output of the disabled and special people, to interact with the outside world, and in the fields of electronic entertainment and industrial control. It is expected to obtain considerable social and economic benefits.
  • An embodiment of the present invention provides a method for encoding and decoding a motor potential brain-computer interface based on spatial filtering and template matching.
  • EEG signals reflect the bioelectrical activity of brain neurons.
  • the EEG mode is closely related to the cognitive activity of the brain. Because of its important research value in revealing the neural mechanism of the brain and implementing brain-machine interface control, it has attracted widespread attention from researchers.
  • the embodiment of the present invention combines TRCA and DSP to design a composite spatial filtering method, which can improve the accuracy of EEG pattern recognition.
  • the pre-processed EEG data 1 is used to construct a task-related component spatial filter, and the data 1 is subjected to the first spatial filtering to be a data 2; the data 2 is used to construct a spatial filter for discriminating spatial patterns; After the second spatial filtering, template data is constructed.
  • the test data is subjected to two spatial filtering and template data for matching analysis (that is, the process of performing coherence analysis), and finally decision classification (that is, the process of comparing eigenvalues).
  • a composite spatial filtering method combining TRCA and DSP and a template matching algorithm are designed for EEG pattern recognition, which will help improve the recognition accuracy rate and lay a strong foundation for the further transformation of EEG-based BCI into application results.
  • Embodiment 1 The solution in Embodiment 1 is further described below with reference to FIG. 2 and a specific calculation formula, as described in detail below:
  • FIG. 2 is a flowchart of calculation according to an embodiment of the present invention, which can be used for motion-related potential detection.
  • M samples in each type of training data set x (m) represents the m-th sample
  • N c represents the number of channels for collecting EEG
  • N t represents the length of the intercepted signal.
  • Mean value of all samples in the training set to obtain the template signal of the training samples This is shown in formula (21).
  • Data processing mainly includes four steps: data preprocessing, spatial filter construction and data spatial filtering, correlation analysis to extract features and template matching.
  • EEG signals are first pre-processed.
  • sampling frequency of the EEG signal is 1000Hz or higher.
  • Chebyshev filter to filter the signal from 0.5 to 45 Hz.
  • the training set and the test set are processed with zero mean on the time scale.
  • the spatial filtering is divided into two steps.
  • the first step is TRCA filtering to extract task-related components
  • the second step is DSP filtering to maximize the difference between the two types of data.
  • the TRCA algorithm belongs to a spatial filter. The purpose is to maximize the sum of covariances between different trials in the same task through the filter V.
  • the calculation method of the sum of covariances is shown in formula (22). Where N c is the number of channels for collecting EEG, N t is the length of the intercepted signal, N s is the number of samples in the training set, m 1 and m 2 are the numbers of the samples, k 1 and k 2 are the numbers of the leads, Signal representing the m 1 sample lead k 1 .
  • Cov is the covariance
  • Filters corresponding to leads k 1 and k 2 respectively
  • v is the filter matrix
  • S is the covariance matrix between trials.
  • Q is the covariance matrix between the leads.
  • the optimal solution V is the eigenvector of the matrix Q -1 S.
  • the training set TRCA filter obtained by calculation A total of n sub-filters are obtained.
  • the dimension of the sub-filter can be reduced.
  • N ′ c is the dimension of the filter after dimensionality reduction
  • the purpose of the DSP filter is to maximize the feature differences between the different modes. This method is an extension of Fisher's linear discriminant analysis in spatial analysis.
  • the spatial filter is obtained with emphasis on the largest intra-class dispersion projection and the smallest intra-class dispersion projection.
  • the intra-class divergence matrix S w is shown in formula (26):
  • the optimal solution U is a matrix Feature vector.
  • the dimension of the sub-filter can be reduced.
  • the selected dimension N ′′ c can be determined empirically, or it can be optimized and adjusted according to the algorithm.
  • the sub-filter combination after dimensionality reduction becomes
  • the correlation analysis method used in the present invention is the Pearson correlation coefficient, the Euclidean distance, and the Canonical Correlation Analysis (CCA) method.
  • the Pearson correlation coefficient is defined as: The Pearson correlation coefficient of two continuous variables X, Y is equal to the covariance cov (X, Y) between them divided by the product of their respective standard deviations ⁇ X ⁇ Y. The value range of the coefficient is [-1,1]. Variables close to 0 are said to have no correlation, and close to 1 or -1 are said to have strong correlation. Corr (*) is used here to represent the Pearson correlation coefficient.
  • Euclidean distance is the abbreviation of Euclidean metric, which refers to the true distance between two points in m-dimensional space.
  • dist (*) means Euclidean distance.
  • the basic principle of the CCA algorithm is: in order to grasp the correlation between the two groups of indicators as a whole, two representative comprehensive variables A and B are extracted from the two groups of variables (respectively for each variable in the two variable groups). Linear combination), using the correlation between the two integrated variables to reflect the overall correlation between the two sets of indicators.
  • the CCA algorithm can be used to project the spatially filtered data to a new space and calculate the correlation.
  • CCA (*) is used here for CCA analysis.
  • the feature vectors [ ⁇ i1 , ⁇ i2 , ⁇ i3 , ⁇ i4 ] T , and ⁇ i ⁇ R 4 ⁇ 1 are obtained .
  • the features extracted in the embodiments of the present invention represent correlations, so a larger feature value indicates that it is more similar to a certain category.
  • weight coefficients ⁇ [ ⁇ 1 , ⁇ 2 , ⁇ 3 , ⁇ 4 ], and ⁇ ⁇ R 1 ⁇ 4 are used to optimize the matching result.
  • the recognition result of the test data is a type I EEG mode.
  • the weight coefficient ⁇ can be determined empirically, and can also be optimized and adjusted according to the algorithm.
  • the traditional motion intent decoding brain-computer interface system has a small number of instructions in the instruction set, which is extremely difficult to implement the large instruction control necessary for communication.
  • the above-mentioned spatial filtering algorithm can realize fast decoding of motion-related potentials. Based on this, the present invention proposes a new method for encoding EEG signals of exercise intent.
  • the above-mentioned spatial filtering algorithm is used to decode EEG signals, which can effectively extend the motion-based intent Decodes the instruction set of the brain-computer interface.
  • the technical process is as follows: the user according to the coding paradigm proposed by the present invention autonomously presses the keys, and simultaneously collects EEG signals for decoding and analysis.
  • the brain controls limb movements as contralateral controls, that is, the left functional area of the brain controls the movement of the right limb, and the right functional area of the brain controls the movement of the left limb, so the user is using the left-hand button At this time, the movement-related EEG signal characteristics of the corresponding spatial position on the right side of the brain are more obvious.
  • the right-hand key the movement-related EEG signal characteristics of the corresponding spatial position on the left side of the brain are more obvious.
  • the left-hand key used is defined as code 1 and the right-hand key is defined as code 0.
  • the "0/1" binary coding idea large instruction coding can be realized. Taking encoding of 4 instructions as an example, 2-bit encoding can be implemented, as shown in Table 1 in Embodiment 2.
  • the number of keystrokes that is, the number of coding bits
  • a large instruction set can be controlled.
  • the number of encoded bits is n
  • the number of encoded instructions is 2 n .
  • the interval T between each keystroke depends on the minimum time that the EEG pattern recognition algorithm can effectively detect.
  • EEG signals acquisition Use the 64-lead EEG acquisition system of Neuroscan to collect EEG signals, and use Ag / AgCl electrodes (impedance less than 15000 ohms).
  • the EEG signals of all leads are referenced to the top of the head and the forehead is the ground. -20 international standard lead placement.
  • the EEG sampling frequency is 1000Hz
  • the filtering passband is 0-100Hz
  • a 50Hz notch is used to remove power frequency interference.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)

Abstract

一种基于空间滤波与模版匹配的运动电位脑机接口编解码方法，包括：在编码方面，用左手与右手序列按键编码脑电信号。在解码方面，一方面分别构建任务相关成分滤波器与判别空间模式滤波器，训练数据分别经过两个滤波器后构建模板数据，测试数据分别经过两个滤波器后与对应的模板数据匹配分析，最后决策分类；另一方面首先构建任务相关成分滤波器，训练数据经过第一次空间滤波后构建判别空间模式滤波器；训练数据先后经过两个空间滤波器后构建模板数据，测试数据先后经过两个空间滤波器与模板数据匹配分析，最后决策分类。首次结合任务相关成分分析与判别空间模式分析的方法，通过模板匹配可实现对大脑活动不同模式的有效识别。

Description

基于空间滤波与模版匹配的运动电位脑机接口编解码方法 技术领域
本发明涉及运动相关电位检测领域，尤其涉及一种基于空间滤波与模版匹配的运动电位脑机接口编解码方法。
背景技术
脑电(Electroencephalography，EEG)反映了大脑神经元的生物电活动。通过检测EEG模式可用于控制脑-机接口(Brain-Computer Interface，BCI)输出。BCI是一个将中枢神经系统活动直接转化为人工输出的系统，它能够替代、修复、增强、补充或者改善中枢神经系统的正常输出，从而改善中枢神经系统与内外环境之间的交互作用。运动皮质相关电位(Movement-related cortical potentials，MRCPs)是人体在执行提示性或自主性运动时调用大脑中与运动相关的认知资源时所产生的低频电位。基于MRCPs检测的BCI系统可用于康复训练、智能假肢和机械外骨骼控制等多种场合，因而得到了研究者们的广泛关注。然而编码方法简单，传统的编码方法主要包括左、右手，双脚，舌等简单肢体运动想象，因此BCI的指令集小，对于实现交流必须的大指令控制具有极大的难度。
MRCPs可以看作是与运动相关的事件相关电位。由于脑电信号的信噪比低，单试次的MRCP波形特征很难有效提取。常用的方法为叠加平均，提高信噪比。但是，叠加平均的方法需要使用者多次执行同一任务后输出一个结果，时间成本较高。单试次的运动任务相关EEG波形特征很难有效提取。头皮脑电信号可以看作是大脑中不同的源产生的信号在头皮的叠加。可以利用空间滤波的方法提取特征。空间滤波的主要思想是：通过对原始脑电各路信号分配权重，提取感兴趣源信号成分，剔除噪声信号，从而提高特征信号的信噪比。模板匹配的主要思想是通过将特征信号与不同模式的模板信号进行匹配，寻找匹配度最高的模板作为识别模式。
发明内容
本发明提供了一种基于空间滤波与模版匹配的运动电位脑机接口编解码方法，在编码方面本发明设计一种基于序列运动的运动意图编码方法，提供一种新的指令编码方法，能够有效的扩展运动意图解码BCI的指令集，在解码方面利用一种基于任务相关成分分析与判别空间模式分析的空间滤波方法提高脑电信号信噪比，与传统方法相比，可提高运动相 关电位检测效率，进一步的研究可为BCI的发展开辟新的发展方向，有望获得可观的社会效益和经济效益，详见下文描述：
一种基于空间滤波与模版匹配的运动电位脑机接口编解码方法，所述方法包括以下步骤：
利用预处理后的脑电数据构建任务相关成分空间滤波器，脑电数据经过任务相关成分空间滤波器滤波后为第一滤波数据；
利用预处理后的脑电数据构建空间模式滤波器，脑电数据经过判别空间模式滤波器滤波后为第二滤波数据；
根据第一滤波数据和第二滤波数据分别构建两个模板数据，测试数据分别经过任务相关成分空间滤波器、判别空间模式滤波器后与两个模板数据分别进行匹配分析，最后进行决策分类。
另一实施例，一种基于空间滤波与模版匹配的运动电位脑机接口编解码方法，所述方法基于任务相关成分分析与判别空间模式分析，对脑电信号进行空间滤波，且通过模板匹配实现对大脑活动不同模式的有效识别，所述方法包括以下步骤：
利用预处理后的脑电数据构建任务相关成分空间滤波器，预处理后的脑电数据经过第一次空间滤波后用于构建判别空间模式滤波器，即DSP滤波器；
滤波后的数据经过第二次空间滤波后构建模板数据，测试数据经过两次空间滤波与模板数据进行匹配分析，最后进行决策分类。
本发明提供的技术方案的有益效果是：
1、本发明首次利用基于任务相关成分分析结合判别空间模式分析的方法，可实现运动相关电位准确识别；
2、本发明提出的方法可用于基于运动意图检测的脑-机接口系统，利用该方法能进一步完善脑-机接口技术，促进该技术向应用成果转化；
3、本发明提出序列编码的设计，有效的扩展了基于运动意图解码BCI的指令集，与传统范式的字符拼写器(如基于P300的Oddball范式)相比，不需要外界刺激，使用者仅需要自主的控制自己的运动意图输出指令，可真正实现意念控制。本发明提出的编码方式，为MI-BCI系统大指令集操作奠定基础。
附图说明
图1为一种基于空间滤波与模版匹配的运动电位脑机接口编解码方法的流程图；
图2为一种基于空间滤波与模版匹配的运动电位脑机接口编解码方法的另一流程图。
具体实施方式
为使本发明的目的、技术方案和优点更加清楚，下面对本发明实施方式作进一步地详细描述。
本发明实施例中涉及的空间滤波方法包括：任务相关成分分析(Task-related component analysis,TRCA)与判别空间模式分析(Discriminative spatial pattern,DSP)。TRCA旨在从线性加权的多重时间序列中提取任务相关成分。其主要思想是最大化同一类任务不同试次之间的协方差或者相关性。DSP的目的是使不同模式之间的特征差异最大化。该方法是Fisher线性判别分析思想在空间分析方面的拓展，在强调类间离差投影最大和类内离差投影最小的情况下获得空间滤波器。本发明实施例的模板匹配方法涉及皮尔森相关系数、欧式距离和典型相关分析(Canonical correlation analysis,CCA)方法和特征值比较方法。
实施例1
本发明实施例提供了一种基于空间滤波与模版匹配的运动电位脑机接口编解码方法。运动相关皮质电位(MRCPs)是人们在想象或执行身体运动时产生的富含运动信息、有严格锁时和锁相的脑电信号模式。因其在揭示人体运动神经机制、指导运动康复训练等方面具有重要的研究价值，受到了研究者们的广泛关注。本发明实施例设计了基于空间滤波的运动相关电位检测方法，可以提高识别正确率。
其技术流程是：利用预处理后的脑电数据1构建任务相关成分空间滤波器，数据1经过任务相关成分空间滤波器滤波后为数据2，利用预处理后的脑电数据1构建判别空间模式滤波器，数据1经过判别空间模式滤波器滤波后为数据3，分别根据数据2和数据3构建模板数据，测试数据分别经过任务相关成分空间滤波器、判别空间模式滤波器后与两个模板数据进分别行匹配分析(即进行相干性分析的过程)，最后进行决策分类(即特征值比较的过程)。
综上所述，本发明实施例首次将TRCA、DSP空间滤波方法结合模板匹配方法用于运动相关电位的检测，将有利于提高识别效率，对基于运动相关电位检测的BCI走向实用化 具有重要意义。
实施例2
下面结合图1、以及具体的计算公式对实施例1中的方案进行进一步地介绍，详见下文描述：
图1为本发明实施例计算的流程图，可用于运动相关电位检测。假设X i＝{x (m)} i为第i类运动模式相关脑电训练数据的集合(i＝1,2,...,n；m＝1,2,...,M)，其中每一类训练数据集合中有M个样本
Figure PCTCN2018125926-appb-000001
x (m)表示第m个样本，
Figure PCTCN2018125926-appb-000002
为测试样本，其中N c表示采集脑电的通道数，N t表示截取信号长度。对训练集所有样本求均值得到训练样本的模板信号
Figure PCTCN2018125926-appb-000003
如公式(1)所示。
数据处理主要包括了数据预处理、构建空间滤波器并对数据空间滤波、相关性分析提取特征和模板匹配四个步骤。
Figure PCTCN2018125926-appb-000004
一、数据预处理
脑电信号首先进行预处理。通常脑电信号的采样频率为1000Hz或更高，为在保证信号质量的前提下节约计算成本，首先降采样到200Hz。使用切比雪夫滤波器对信号做0.5～45Hz的滤波。且对训练集和测试集都在时间尺度上进行了零均值处理。
二、空间滤波
在本发明实施例中，空间滤波分为两个步骤，首先是TRCA滤波，提取任务相关成分，第二步是DSP滤波，最大化两类数据的差异。
1、构建TRCA空间滤波器
TRCA算法属于一种空间滤波器，目的是通过滤波器V最大化同一个任务内不同试次间的协方差之和，该协方差之和的计算方法如公式(2)所示，
Figure PCTCN2018125926-appb-000005
其中N c表示采集脑电的通道数，N t表示截取信号长度，N s表示训练集样本个数，m 1与m 2表示样本的编号，k 1与k 2表示导联的编号，
Figure PCTCN2018125926-appb-000006
表示第m 1个样本导联k 1的信号，x为样本集。
Figure PCTCN2018125926-appb-000007
其中，Cov为协方差；
Figure PCTCN2018125926-appb-000008
分别为k 1，k 2导联对应的滤波器；v为滤波器矩阵；S为试次间的协方差矩阵。
为了得到限定的解，设置约束条件如公式(3)：
Figure PCTCN2018125926-appb-000009
其中，Q为导联间的协方差矩阵。
因此最终的优化函数如公式(4)：
Figure PCTCN2018125926-appb-000010
其中，最优解V是矩阵Q -1S的特征向量。
对于本发明实施例，训练集
Figure PCTCN2018125926-appb-000011
通过计算得到TRCA滤波器
Figure PCTCN2018125926-appb-000012
Figure PCTCN2018125926-appb-000013
共得到n个子滤波器。
为得到最优滤波器，可对子滤波器降维
Figure PCTCN2018125926-appb-000014
选取的维度N′ c可根据经验确定，也可根据算法优化调整。对降维后的子滤波器组合变为
Figure PCTCN2018125926-appb-000015
Figure PCTCN2018125926-appb-000016
对训练样本的模板信号及测试数据经过子空间滤波后，得到
Figure PCTCN2018125926-appb-000017
以及
Figure PCTCN2018125926-appb-000018
其中，N′ c为降维后滤波器的维度，
Figure PCTCN2018125926-appb-000019
为模板信号的转置，
Figure PCTCN2018125926-appb-000020
为降维后的第i类TRCA滤波器。
2、构建DSP空间滤波器
DSP是一种空间滤波方法，目的是使不同模式之间的特征差异最大化。该方法是Fisher线性判别分析思想在空间分析方面的拓展，在强调类间离差投影最大和类内离差投影最小 的情况下获得空间滤波器。其中类间散度矩阵S b如公式(6)所示，
Figure PCTCN2018125926-appb-000021
和
Figure PCTCN2018125926-appb-000022
为对第i类和第j类训练集所有样本求均值得到两类训练样本的模板信号。
Figure PCTCN2018125926-appb-000023
类内散度矩阵S w如公式(7)所示：
Figure PCTCN2018125926-appb-000024
其中，
Figure PCTCN2018125926-appb-000025
为第i类训练集，
Figure PCTCN2018125926-appb-000026
为第j类训练集。
因此最终的目标函数如公式(8)：
Figure PCTCN2018125926-appb-000027
其中，最优解U是矩阵
Figure PCTCN2018125926-appb-000028
的特征向量，u为每个导联对应的滤波器。
对于本发明实施例，两类训练集
Figure PCTCN2018125926-appb-000029
和
Figure PCTCN2018125926-appb-000030
通过计算后得到DSP滤波器
Figure PCTCN2018125926-appb-000031
对每两类训练集计算DSP滤波器，则得到
Figure PCTCN2018125926-appb-000032
个子滤波器。为得到最优滤波器，可对子滤波器降维
Figure PCTCN2018125926-appb-000033
选取的维度N′ c可根据经验确定，也可根据算法优化调整。对降维后的子滤波器组合变为
Figure PCTCN2018125926-appb-000034
Figure PCTCN2018125926-appb-000035
对训练样本的模板信号
Figure PCTCN2018125926-appb-000036
以及测试数据进行空间滤波，得到
Figure PCTCN2018125926-appb-000037
以及
Figure PCTCN2018125926-appb-000038
其中，
Figure PCTCN2018125926-appb-000039
为第i类训练数据和第j类数据通过计算后得到的降维后的子滤波器；
Figure PCTCN2018125926-appb-000040
为包括所有子滤波器的滤波器组。
三、相关性分析
在本发明中使用的相关性分析的方法为皮尔森相关系数、欧式距离和典型相关分析(Canonical correlation analysis,CCA)方法。
皮尔森相关系数的定义为：两个连续变量X,Y的皮尔森相关性系数等于它们之间的协方差cov(X,Y)除以它们各自标准差的乘积σ Xσ Y。系数的取值范围为[-1,1]，接近0的变量 被称为无相关性，接近1或者-1被称为具有强相关性。在这里使用corr(*)表示皮尔森相关系数。
欧式距离是欧几里得度量的简称，指在m维空间中两个点之间的真实距离，在本发明实施例中，计算了测试样本与两类训练模板的距离，认为距离越近，相关性越强。在这里dist(*)表示欧式距离。
CCA算法的基本原理是：为了从总体上把握两组指标之间的相关关系，分别在两组变量中提取有代表性的两个综合变量A和B(分别为两个变量组中各变量的线性组合)，利用这两个综合变量之间的相关关系来反映两组指标之间的整体相关性。应用CCA算法可以将空间滤波后的数据投影到新的空间并计算相关性。在这里使用CCA(*)表示CCA分析。
以下为计算特征的详细说明。
首先计算经过TRCA空间滤波后测试数据与各模板的皮尔森相关系数：
Figure PCTCN2018125926-appb-000041
计算经过TRCA空间滤波后测试数据与各模板的欧式距离：
Figure PCTCN2018125926-appb-000042
对经过TRCA空间滤波后测试数据与各模板进行CCA分析并在新的投影空间计算皮尔森相关系数：
Figure PCTCN2018125926-appb-000043
Figure PCTCN2018125926-appb-000044
Figure PCTCN2018125926-appb-000045
计算经过DSP空间滤波后测试数据与各模板的皮尔森相关系数：
Figure PCTCN2018125926-appb-000046
计算经过DSP空间滤波后测试数据与各模板的欧式距离：
Figure PCTCN2018125926-appb-000047
对经过DSP空间滤波后测试数据与各模板进行CCA分析并在新的投影空间计算皮尔森相关系数：
Figure PCTCN2018125926-appb-000048
Figure PCTCN2018125926-appb-000049
Figure PCTCN2018125926-appb-000050
由此，测试数据与每一类训练数据的模板匹配计算后得到了特征向量ρ i＝[ρ i1，ρ i2，ρ i3，ρ i4，ρ i5，ρ i6，ρ i7，ρ i8] T，ρ i∈R 8×1。
其中，CCA(*)表示CCA分析，corr(*)表示皮尔森相关系数，Ai ，B i为线性投影矩阵，ρ i1为第i类特征向量的第一个特征元素，ρ i2为第i类特征向量的第二个特征元素，ρ i3为第i类特征向量的第三个特征元素，ρ i4为第i类特征向量的第四个特征元素，ρ i5为第i类特征向量的第五个特征元素，ρ i6为第i类特征向量的第六个特征元素，ρ i7为第i类特征向量的第七个特征元素，ρ i8为第i类特征向量的第八个特征元素。
四、特征值比较
如前文所述，本发明实施例中提取的特征表征的是相关性，因此特征值越大表示与某一类越相似。通过比较特征值的大小判定测试样本与哪一类训练样本更为匹配。在这里，加入权重系数ω＝[ω 1，ω 2，ω 3，ω 4，ω 5，ω 6，ω 7，ω 8]，ω∈R 1×8，优化匹配结果。
Figure PCTCN2018125926-appb-000051
选取最大的ωρ i，则测试数据的识别结果为第I类运动相关模式。权重系数ω可根据经验确定，也可根据算法优化调整。
实际应用时会发现传统的运动意图解码脑-机接口系统的指令集中指令个数较少，这对于实现交流必需的大指令控制具有极大的难度。上述空间滤波算法可实现对运动相关电位的快速解码，基于此，本发明提出了一种新的运动意图脑电信号编码方法，利用上述空间滤波算法对脑电信号解码，可有效扩展基于运动意图解码脑-机接口的指令集。
其技术流程是：使用者根据本发明提出的编码范式自主序列按键，同时采集脑电信号，用于解码分析。
具体技术方案：
编码范式：由于大脑对于肢体运动的控制为对侧控制，即左侧大脑的运动功能区控制右侧肢体的运动，右侧大脑的运动功能区控制左侧肢体的运动，故用户在使用左手按键时，大脑右侧对应空间位置运动相关脑电信号特征较为明显，用户在使用右手按键时，大脑左侧对应空间位置运动相关脑电信号特征较为明显。在本发明中，将用于的左手按键定义为编码1，右手按键定义为编码0，利用“0/1”二进制编码思想，可实现大指令编码。以编码 4个指令为例，2位编码即可实现，如表1所示：
表1编码方法示意
Figure PCTCN2018125926-appb-000052
由此，通过增加按键次数即编码位数可实现大指令集的控制。具体说来，当编码的位数为n时，编码的指令个数为2 n个。每次按键之间的间隔时间T取决于脑电模式识别算法能够有效检测的最短时间。
信号采集：使用Neuroscan公司的64导联脑电采集系统采集脑电信号，使用Ag/AgCl电极(阻抗小于15000欧姆)，所有导联的脑电信号以头顶为参考，以前额为地，根据10-20国际标准导联位置摆放。脑电采样频率为1000Hz，滤波通带为0～100Hz，并采用50Hz陷波器去除工频干扰。
综上所述，与传统方法相比，本方法可提高运动相关电位检测效率。另外本发明提出的检测方法进一步研究可以得到完善的基于运动意图检测的脑-机接口系统，用于残疾人、特殊人群辅助输出，与外界进行信息交互，并在电子娱乐、工业控制等领域，有望获得可观的社会效益和经济效益。
实施例3
本发明实施例提供了一种基于空间滤波与模版匹配的运动电位脑机接口编解码方法。脑电信号反映了大脑神经元的生物电活动。脑电振荡模式与大脑认知活动密切相关。由于其在揭示大脑神经机制、实现脑-机接口控制等方面具有重要的研究价值，受到了研究者们的广泛关注。本发明实施例结合TRCA与DSP，设计了复合空间滤波方法，可提高脑电模式识别的准确性。
其技术流程是：利用预处理后的脑电数据1构建任务相关成分空间滤波器，数据1经过第一次空间滤波后为数据2，利用数据2构建判别空间模式空间滤波器，数据2经过第二次空间滤波后构建模板数据，测试数据经过两次空间滤波与模板数据进行匹配分析(即进行相干性分析的过程)，最后进行决策分类(即特征值比较的过程)。
本发明实施例，设计了结合TRCA、DSP的复合空间滤波方法以及模板匹配的算法用 于EEG模式识别，将有利于提高识别准确率，对基于EEG的BCI进一步向应用成果转化奠定了有力基础。
实施例4
下面结合图2、以及具体的计算公式对实施例1中的方案进行进一步地介绍，详见下文描述：
图2为本发明实施例计算的流程图，可用于运动相关电位检测。假设X i＝{x (m)} i为第i类运动模式相关脑电训练数据的集合(i＝1,2,...,n；m＝1,2,...,M)，其中每一类训练数据集合中有M个样本
Figure PCTCN2018125926-appb-000053
x (m)表示第m个样本，
Figure PCTCN2018125926-appb-000054
为测试样本，其中N c表示采集脑电的通道数，N t表示截取信号长度。对训练集所有样本求均值得到训练样本的模板信号
Figure PCTCN2018125926-appb-000055
如公式(21)所示。
数据处理主要包括了数据预处理、构建空间滤波器并对数据空间滤波、相关性分析提取特征和模板匹配四个步骤。
Figure PCTCN2018125926-appb-000056
一、数据预处理
脑电信号首先进行预处理。通常脑电信号的采样频率为1000Hz或更高，为在保证信号质量的前提下节约计算成本，首先降采样到200Hz。使用切比雪夫滤波器对信号做0.5～45Hz的滤波。且对训练集和测试集都在时间尺度上进行了零均值处理。
二、空间滤波
在本发明实施例中，空间滤波分为两个步骤，首先是TRCA滤波，提取任务相关成分，第二步是DSP滤波，最大化两类数据的差异。
TRCA算法属于一种空间滤波器，目的是通过滤波器V最大化同一个任务内不同试次间的协方差之和，该协方差之和的计算方法如公式(22)所示，
Figure PCTCN2018125926-appb-000057
其中N c表示采集脑电的通道数，N t表示截取信号长度，N s表示训练集样本个数，m 1与m 2表示样本的编号，k 1与k 2表示导联的编号，
Figure PCTCN2018125926-appb-000058
表示第m 1个样本导联k 1的信号。
Figure PCTCN2018125926-appb-000059
其中，Cov为协方差；
Figure PCTCN2018125926-appb-000060
分别为k 1，k 2导联对应的滤波器；v为滤波器矩阵；S为试次间的协方差矩阵。
为了得到限定的解，设置约束条件如公式(23)：
Figure PCTCN2018125926-appb-000061
其中，Q为导联间的协方差矩阵。
因此最终的优化函数如公式(24)：
Figure PCTCN2018125926-appb-000062
其中，最优解V是矩阵Q -1S的特征向量。
对于本发明实施例，训练集
Figure PCTCN2018125926-appb-000063
通过计算得到TRCA滤波器
Figure PCTCN2018125926-appb-000064
Figure PCTCN2018125926-appb-000065
共得到n个子滤波器。
为得到最优滤波器，可对子滤波器降维
Figure PCTCN2018125926-appb-000066
对训练样本的模板信号及测试数据经过子空间滤波后，得到
Figure PCTCN2018125926-appb-000067
以及
Figure PCTCN2018125926-appb-000068
其中，N′ c为降维后滤波器的维度，
Figure PCTCN2018125926-appb-000069
为模板信号的转置，
Figure PCTCN2018125926-appb-000070
为降维后的第i类TRCA滤波器。
DSP滤波器的目的是使不同模式之间的特征差异最大化。该方法是Fisher线性判别分析思想在空间分析方面的拓展，在强调类间离差投影最大和类内离差投影最小的情况下获得空间滤波器。
Figure PCTCN2018125926-appb-000071
和
Figure PCTCN2018125926-appb-000072
为对第i类和第j类训练集模板
Figure PCTCN2018125926-appb-000073
和
Figure PCTCN2018125926-appb-000074
经过对应的TRCA滤波器滤波后的新模板，
Figure PCTCN2018125926-appb-000075
与
Figure PCTCN2018125926-appb-000076
为第i类和第j类训练数据集和经过TRCA滤波后的新集合，集合中的样本
Figure PCTCN2018125926-appb-000077
其中类间散度矩阵S b如公式(25) 所示：
Figure PCTCN2018125926-appb-000078
类内散度矩阵S w如公式(26)所示：
Figure PCTCN2018125926-appb-000079
因此最终的目标函数如公式(27)：
Figure PCTCN2018125926-appb-000080
最优解U是矩阵
Figure PCTCN2018125926-appb-000081
的特征向量。
对于本发明实施例，两类新训练数据通过计算后得到DSP滤波器：
Figure PCTCN2018125926-appb-000082
且
Figure PCTCN2018125926-appb-000083
对每两类训练集计算DSP滤波器，则得到
Figure PCTCN2018125926-appb-000084
个子滤波器。
为得到最优滤波器，可对子滤波器降维
Figure PCTCN2018125926-appb-000085
选取的维度N″ c可根据经验确定，也可根据算法优化调整。对降维后的子滤波器组合变为
Figure PCTCN2018125926-appb-000086
Figure PCTCN2018125926-appb-000087
对训练样本的模板信号
Figure PCTCN2018125926-appb-000088
以及测试数据进行空间滤波，得到
Figure PCTCN2018125926-appb-000089
以及
Figure PCTCN2018125926-appb-000090
三、相关性分析
在本发明中使用的相关性分析的方法为皮尔森相关系数、欧式距离和典型相关分析(Canonical correlation analysis,CCA)方法。
皮尔森相关系数的定义为：两个连续变量X,Y的皮尔森相关性系数等于它们之间的协方差cov(X,Y)除以它们各自标准差的乘积σ Xσ Y。系数的取值范围为[-1,1]，接近0的变量被称为无相关性，接近1或者-1被称为具有强相关性。在这里使用corr(*)表示皮尔森相关系数。
欧式距离是欧几里得度量的简称，指在m维空间中两个点之间的真实距离，在本发明实施例中，计算了测试样本与两类训练模板的距离，认为距离越近，相关性越强。在这里dist(*)表示欧式距离。
CCA算法的基本原理是：为了从总体上把握两组指标之间的相关关系，分别在两组变量中提取有代表性的两个综合变量A和B(分别为两个变量组中各变量的线性组合)，利用这两个综合变量之间的相关关系来反映两组指标之间的整体相关性。应用CCA算法可以将空间滤波后的数据投影到新的空间并计算相关性。在这里使用CCA(*)表示CCA分析。
以下为计算特征的详细说明。
首先计算经过复合空间滤波后测试数据与各模板的皮尔森相关系数：
Figure PCTCN2018125926-appb-000091
计算经过复合空间滤波后测试数据与各模板的欧式距离：
Figure PCTCN2018125926-appb-000092
对经过复合空间滤波后测试数据与各模板进行CCA分析并在新的投影空间计算皮尔森相关系数：
Figure PCTCN2018125926-appb-000093
Figure PCTCN2018125926-appb-000094
Figure PCTCN2018125926-appb-000095
由此，测试数据与每一类训练数据的模板匹配计算后得到了特征向量[ρ i1，ρ i2，ρ i3，ρ i4] T，ρ i∈R 4×1。
四、特征值比较
如前文所述，本发明实施例中提取的特征表征的是相关性，因此特征值越大表示与某一类越相似。通过比较特征值的大小判定测试样本与哪一类训练样本更为匹配。在这里，加入权重系数ω＝[ω 1，ω 2，ω 3，ω 4]，ω∈R 1×4优化匹配结果。
Figure PCTCN2018125926-appb-000096
选取最大的ωρ i，则测试数据的识别结果为第I类脑电模式。权重系数ω可根据经验确定，也可根据算法优化调整。
实际应用时会发现传统的运动意图解码脑-机接口系统的指令集中指令个数较少，这对于实现交流必需的大指令控制具有极大的难度。上述空间滤波算法可实现对运动相关电位的快速解码，基于此，本发明提出了一种新的运动意图脑电信号编码方法，利用上述空间滤波算法对脑电信号解码，可有效扩展基于运动意图解码脑-机接口的指令集。
其技术流程是：使用者根据本发明提出的编码范式自主序列按键，同时采集脑电信号，用于解码分析。
具体技术方案：
编码范式：由于大脑对于肢体运动的控制为对侧控制，即左侧大脑的运动功能区控制右侧肢体的运动，右侧大脑的运动功能区控制左侧肢体的运动，故用户在使用左手按键时，大脑右侧对应空间位置运动相关脑电信号特征较为明显，用户在使用右手按键时，大脑左侧对应空间位置运动相关脑电信号特征较为明显。在本发明中，将用于的左手按键定义为编码1，右手按键定义为编码0，利用“0/1”二进制编码思想，可实现大指令编码。以编码4个指令为例，2位编码即可实现，如实施例2中的表1所示。
由此，通过增加按键次数即编码位数可实现大指令集的控制。具体说来，当编码的位数为n时，编码的指令个数为2 n个。每次按键之间的间隔时间T取决于脑电模式识别算法能够有效检测的最短时间。
信号采集：使用Neuroscan公司的64导联脑电采集系统采集脑电信号，使用Ag/AgCl电极(阻抗小于15000欧姆)，所有导联的脑电信号以头顶为参考，以前额为地，根据10-20国际标准导联位置摆放。脑电采样频率为1000Hz，滤波通带为0～100Hz，并采用50Hz陷波器去除工频干扰。
本领域技术人员可以理解附图只是一个优选实施例的示意图，上述本发明实施例序号仅仅为了描述，不代表实施例的优劣。
以上所述仅为本发明的较佳实施例，并不用以限制本发明，凡在本发明的精神和原则之内，所作的任何修改、等同替换、改进等，均应包含在本发明的保护范围之内。

Claims (12)

  1. 一种基于空间滤波与模版匹配的运动电位脑机接口编解码方法，其特征在于，所述方法包括以下步骤：
    利用预处理后的脑电数据构建任务相关成分空间滤波器，脑电数据经过任务相关成分空间滤波器滤波后为第一滤波数据；
    利用预处理后的脑电数据构建空间模式滤波器，脑电数据经过判别空间模式滤波器滤波后为第二滤波数据；
    根据第一滤波数据和第二滤波数据分别构建两个模板数据，测试数据分别经过任务相关成分空间滤波器、判别空间模式滤波器后与两个模板数据分别进行匹配分析，最后进行决策分类。
  2. 根据权利要求1所述的一种基于空间滤波与模版匹配的运动电位脑机接口编解码方法，其特征在于，所述利用预处理后的脑电数据构建任务相关成分空间滤波器具体为：
    训练集
    Figure PCTCN2018125926-appb-100001
    通过计算得到任务相关成分分析滤波器；
    Figure PCTCN2018125926-appb-100002
    共得到n个子滤波器；其中N c表示滤波器的维度，N t表示截取信号长度；
    对子滤波器降维
    Figure PCTCN2018125926-appb-100003
    对降维后的子滤波器组合变为
    Figure PCTCN2018125926-appb-100004
    Figure PCTCN2018125926-appb-100005
    对训练样本的模板信号及测试数据经过子空间滤波后，得到
    Figure PCTCN2018125926-appb-100006
    以及
    Figure PCTCN2018125926-appb-100007
    其中，N′ c为降维后滤波器的维度，
    Figure PCTCN2018125926-appb-100008
    为模板信号的转置，
    Figure PCTCN2018125926-appb-100009
    为降维后的第i类TRCA滤波器。
  3. 根据权利要求2所述的一种基于空间滤波与模版匹配的运动电位脑机接口编解码方法，其特征在于，所述利用预处理后的脑电数据构建空间模式滤波器具体为：
    对每两类训练集计算DSP滤波器，则得到
    Figure PCTCN2018125926-appb-100010
    个子滤波器，对子滤波器降维
    Figure PCTCN2018125926-appb-100011
    对降维后的子滤波器组合变为
    Figure PCTCN2018125926-appb-100012
    Figure PCTCN2018125926-appb-100013
    对训练样本的模板信号
    Figure PCTCN2018125926-appb-100014
    以及测试数据进行空间滤波，得到
    Figure PCTCN2018125926-appb-100015
    以及
    Figure PCTCN2018125926-appb-100016
    其中，
    Figure PCTCN2018125926-appb-100017
    为第i类训练数据和第j类数据通过计算后得到的降维后的子滤波器；
    Figure PCTCN2018125926-appb-100018
    为包括所有子滤波器的滤波器组。
  4. 根据权利要求3所述的一种基于空间滤波与模版匹配的运动电位脑机接口编解码方法，其特征在于，所述测试数据分别经过任务相关成分空间滤波器、判别空间模式滤波器后与两个模板数据分别进行匹配分析的过程为：对经过TRCA空间滤波后测试数据与各模板进行CCA分析并在新的投影空间计算皮尔森相关系数的过程；
    具体计算公式如下：
    Figure PCTCN2018125926-appb-100019
    Figure PCTCN2018125926-appb-100020
    Figure PCTCN2018125926-appb-100021
    计算经过DSP空间滤波后测试数据与各模板的皮尔森相关系数：
    Figure PCTCN2018125926-appb-100022
    计算经过DSP空间滤波后测试数据与各模板的欧式距离：
    Figure PCTCN2018125926-appb-100023
    对经过DSP空间滤波后测试数据与各模板进行CCA分析并在新的投影空间计算皮尔森相关系数：
    Figure PCTCN2018125926-appb-100024
    Figure PCTCN2018125926-appb-100025
    Figure PCTCN2018125926-appb-100026
    由此，测试数据与每一类训练数据的模板匹配计算后得到了特征向量[ρ i1，ρ i2，ρ i3，ρ i4，ρ i5，ρ i6，ρ i7，ρ i8] T，ρ i∈R 8×1
    其中，CCA(*)表示CCA分析，corr(*)表示皮尔森相关系数，A i，B i为线性投影矩阵，ρ i1为第i类特征向量的第一个特征元素，ρ i2为第i类特征向量的第二个特征元素，ρ i3为第i类特征向量的第三个特征元素，ρ i4为第i类特征向量的第四个特征元素，ρ i5为第i类特征向量的第五个特征元素，ρ i6为第i类特征向量的第六个特征元素，ρ i7为第i类特 征向量的第七个特征元素，ρ i8为第i类特征向量的第八个特征元素。
  5. 根据权利要求4所述的一种基于空间滤波与模版匹配的运动电位脑机接口编解码方法，其特征在于，所述决策分类的过程即为特征值比较的过程，具体为：
    通过比较特征值的大小判定测试样本与哪一类训练样本更为匹配，加入权重系数，ω＝[ω 1，ω 2，ω 3，ω 4，ω 5，ω 6，ω 7，ω 8]，ω∈R 1×8优化匹配结果：
    Figure PCTCN2018125926-appb-100027
    选取最大的ωρ i，则测试数据的识别结果为第I类运动相关模式。
  6. 根据权利要求1-5中任一权利要求所述的一种基于空间滤波与模版匹配的运动电位脑机接口编解码方法，其特征在于，所述方法还包括：用左手按键定义为编码1，用右手按键定义为编码0，利用“0/1”二进制编码思想序列编码，当编码的位数为n时，编码的指令个数为2 n个。
  7. 一种基于空间滤波与模版匹配的运动电位脑机接口编解码方法，其特征在于，所述方法基于任务相关成分分析与判别空间模式分析，对脑电信号进行空间滤波，且通过模板匹配实现对大脑活动不同模式的有效识别，所述方法包括以下步骤：
    利用预处理后的脑电数据构建任务相关成分空间滤波器，预处理后的脑电数据经过第一次空间滤波后用于构建判别空间模式滤波器，即DSP滤波器；
    滤波后的数据经过第二次空间滤波后构建模板数据，测试数据经过两次空间滤波与模板数据进行匹配分析，最后进行决策分类。
  8. 根据权利要求7所述的一种基于空间滤波与模版匹配的运动电位脑机接口编解码方法，其特征在于，所述利用预处理后的脑电数据构建任务相关成分空间滤波器具体为：
    训练集
    Figure PCTCN2018125926-appb-100028
    通过计算得到任务相关成分分析滤波器：
    Figure PCTCN2018125926-appb-100029
    共得到n个子滤波器；其中N c表示滤波器的维度，N t表示截取信号长度；
    对子滤波器降维
    Figure PCTCN2018125926-appb-100030
    对训练样本的模板信号及测试数据经过子空间滤波后，得到
    Figure PCTCN2018125926-appb-100031
    以及
    Figure PCTCN2018125926-appb-100032
    其中，N′ c为降维后滤波器的维度，由经验确定，
    Figure PCTCN2018125926-appb-100033
    为模板信号的转置，
    Figure PCTCN2018125926-appb-100034
    为降维后的第i类TRCA滤波器；R为实数集。
  9. 根据权利要求8所述的一种基于空间滤波与模版匹配的运动电位脑机接口编解码方法，其特征在于，所述预处理后的脑电数据经过第一次空间滤波后用于构建判别空间模式滤波器具体为：
    两类新训练数据通过计算后得到DSP滤波器：
    Figure PCTCN2018125926-appb-100035
    且
    Figure PCTCN2018125926-appb-100036
    其中U i，j表示第i类训练数据和第j类数据通过计算后得到的DSP滤波器；
    对每两类训练集计算DSP滤波器，则得到
    Figure PCTCN2018125926-appb-100037
    个子滤波器；
    对子滤波器降维
    Figure PCTCN2018125926-appb-100038
    其中N″ c为滤波器降维后的维度，对降维后的子滤波器组合变为
    Figure PCTCN2018125926-appb-100039
    Figure PCTCN2018125926-appb-100040
    对训练样本的模板信号
    Figure PCTCN2018125926-appb-100041
    以及测试数据进行空间滤波，得到
    Figure PCTCN2018125926-appb-100042
    以及
    Figure PCTCN2018125926-appb-100043
  10. 根据权利要求9所述的一种基于空间滤波与模版匹配的运动电位脑机接口编解码方法，其特征在于，所述测试数据经过两次空间滤波与模板数据进行匹配分析的过程即为：对经过复合空间滤波后测试数据与各模板进行典型相关分析，即CCA，并在新的投影空间计算皮尔森相关系数的过程；
    具体计算公式如下：
    Figure PCTCN2018125926-appb-100044
    Figure PCTCN2018125926-appb-100045
    Figure PCTCN2018125926-appb-100046
    由此，测试数据与每一类训练数据的模板匹配计算后得到了特征向量：
    ρ i＝[ρ i1，ρ i2，ρ i3，ρ i4] T，ρ i∈R 4×1；
    其中，CCA(*)表示CCA分析，corr(*)表示皮尔森相关系数，A i，B i为线性投影矩阵，ρ ii为第i类特征向量的第一个特征元素，ρ i2为第i类特征向量的第二个特征元素，ρ i3为 第i类特征向量的第三个特征元素，ρ i4为第i类特征向量的第四个特征元素。
  11. 根据权利要求10所述的一种基于空间滤波与模版匹配的运动电位脑机接口编解码方法，其特征在于，所述决策分类的过程即为特征值比较的过程，具体为：
    通过比较特征值的大小判定测试样本与哪一类训练样本更为匹配，加入权重系数ω＝[ω 1，ω 2，ω 3，ω 4]，ω∈R 1×4优化匹配结果；
    Figure PCTCN2018125926-appb-100047
    其中，选取最大的ωρ i，则测试数据的识别结果为第I类脑电模式。
  12. 根据权利要求7-11中任一权利要求所述的一种基于空间滤波与模版匹配的运动电位脑机接口编解码方法，其特征在于，所述方法还包括：用左手按键定义为编码1，用右手按键定义为编码0，利用“0/1”二进制编码思想序列编码，当编码的位数为n时，编码的指令个数为2 n个。
PCT/CN2018/125926 2018-08-28 2018-12-30 基于空间滤波与模版匹配的运动电位脑机接口编解码方法 WO2020042511A1 (zh)

Applications Claiming Priority (4)

Application Number Priority Date Filing Date Title
CN201810988364.3A CN109214325B (zh) 2018-08-28 2018-08-28 一种基于空间滤波与模版匹配的运动相关电位检测方法
CN201810988364.3 2018-08-28
CN201810995425.9 2018-08-29
CN201810995425.9A CN109271887A (zh) 2018-08-29 2018-08-29 一种用于脑电模式识别的复合空间滤波与模板匹配方法

Publications (1)

Publication Number Publication Date
WO2020042511A1 true WO2020042511A1 (zh) 2020-03-05

Family

ID=69644903

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/125926 WO2020042511A1 (zh) 2018-08-28 2018-12-30 基于空间滤波与模版匹配的运动电位脑机接口编解码方法

Country Status (1)

Country Link
WO (1) WO2020042511A1 (zh)

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112883914A (zh) * 2021-03-19 2021-06-01 西安科技大学 一种多分类器结合的矿用机器人意念感知与决策方法
CN113221831A (zh) * 2021-05-31 2021-08-06 南方科技大学 解码标签预测方法、模型建立方法、装置及电子设备
CN113591598A (zh) * 2021-07-07 2021-11-02 河北工业大学 一种基于相关性分析的脑-机接口跨负荷线性判别方法
CN113792588A (zh) * 2021-08-05 2021-12-14 深兰科技（上海）有限公司 一种脑电波处理装置、方法、计算机设备及存储介质
CN113935380A (zh) * 2021-10-22 2022-01-14 北京理工大学 一种基于模板匹配的自适应运动想象脑机接口方法与系统
CN114145752A (zh) * 2021-10-22 2022-03-08 杭州电子科技大学 一种基于小波变换的多模态脑机接口数据融合方法
CN114145756A (zh) * 2021-12-15 2022-03-08 电子科技大学中山学院 协作机器人控制方法、装置及计算机可读存储介质
CN115969389A (zh) * 2021-10-15 2023-04-18 中国科学院沈阳自动化研究所 一种基于个体脑电信号迁移的运动想象意图识别方法
CN117807417A (zh) * 2023-08-09 2024-04-02 上海韶脑传感技术有限公司 一种基于et-trca的ssvep脑电识别方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101219048A (zh) * 2008-01-25 2008-07-16 北京工业大学 想象单侧肢体运动的脑电特征的提取方法
CN102940490A (zh) * 2012-10-19 2013-02-27 西安电子科技大学 基于非线性动力学的运动想象脑电信号特征提取方法
CN105824418A (zh) * 2016-03-17 2016-08-03 天津大学 一种基于非对称视觉诱发电位的脑-机接口通讯系统
CN106200984A (zh) * 2016-07-21 2016-12-07 天津大学 运动想象脑‑机接口模型建模方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101219048A (zh) * 2008-01-25 2008-07-16 北京工业大学 想象单侧肢体运动的脑电特征的提取方法
CN102940490A (zh) * 2012-10-19 2013-02-27 西安电子科技大学 基于非线性动力学的运动想象脑电信号特征提取方法
CN105824418A (zh) * 2016-03-17 2016-08-03 天津大学 一种基于非对称视觉诱发电位的脑-机接口通讯系统
CN106200984A (zh) * 2016-07-21 2016-12-07 天津大学 运动想象脑‑机接口模型建模方法

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
LIAO, XIANG: "Combining Spatial Filters for the Classification of Sing- le-Trial EEG in a Finger Movement Task", IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, vol. 54, no. 5, 31 May 2007 (2007-05-31), pages 821 - 831, XP011176949 *
NAKANISHI, MASAKI: "Independent Component Analysis-Based Spatial Filte- ring Improves Template-Based SSVEP Detection", 2017 39 TH ANNUAL INTERNATIONAL CONFERENCE OF THE IEEE ENGINEERING IN MEDICI- NE AND BIOLOGY SOCIETY (EMBC, 31 December 2017 (2017-12-31), pages 3620 - 3623, XP033152816 *

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112883914B (zh) * 2021-03-19 2024-03-19 西安科技大学 一种多分类器结合的矿用机器人意念感知与决策方法
CN112883914A (zh) * 2021-03-19 2021-06-01 西安科技大学 一种多分类器结合的矿用机器人意念感知与决策方法
CN113221831A (zh) * 2021-05-31 2021-08-06 南方科技大学 解码标签预测方法、模型建立方法、装置及电子设备
CN113591598A (zh) * 2021-07-07 2021-11-02 河北工业大学 一种基于相关性分析的脑-机接口跨负荷线性判别方法
CN113591598B (zh) * 2021-07-07 2024-07-12 河北工业大学 一种基于相关性分析的脑-机接口跨负荷线性判别方法
CN113792588A (zh) * 2021-08-05 2021-12-14 深兰科技（上海）有限公司 一种脑电波处理装置、方法、计算机设备及存储介质
CN113792588B (zh) * 2021-08-05 2024-04-09 深兰科技（上海）有限公司 一种脑电波处理装置、方法、计算机设备及存储介质
CN115969389A (zh) * 2021-10-15 2023-04-18 中国科学院沈阳自动化研究所 一种基于个体脑电信号迁移的运动想象意图识别方法
CN114145752B (zh) * 2021-10-22 2024-03-29 杭州电子科技大学 一种基于小波变换的多模态脑机接口数据融合方法
CN114145752A (zh) * 2021-10-22 2022-03-08 杭州电子科技大学 一种基于小波变换的多模态脑机接口数据融合方法
CN113935380A (zh) * 2021-10-22 2022-01-14 北京理工大学 一种基于模板匹配的自适应运动想象脑机接口方法与系统
CN114145756A (zh) * 2021-12-15 2022-03-08 电子科技大学中山学院 协作机器人控制方法、装置及计算机可读存储介质
CN117807417A (zh) * 2023-08-09 2024-04-02 上海韶脑传感技术有限公司 一种基于et-trca的ssvep脑电识别方法

Similar Documents

Publication Publication Date Title
WO2020042511A1 (zh) 基于空间滤波与模版匹配的运动电位脑机接口编解码方法
Zhang et al. Bayesian learning for spatial filtering in an EEG-based brain–computer interface
WO2017084416A1 (zh) 一种基于运动想象脑-机接口的反馈系统
Qi et al. RSTFC: A novel algorithm for spatio-temporal filtering and classification of single-trial EEG
CN113158793B (zh) 一种基于多特征融合的多类运动想象脑电信号识别方法
CN111265212A (zh) 一种运动想象脑电信号分类方法及闭环训练测试交互系统
CN114533086B (zh) 一种基于空域特征时频变换的运动想象脑电解码方法
CN103425249A (zh) 基于正则化csp和src的脑电信号分类识别方法及其遥控系统
Liu et al. FBMSNet: A filter-bank multi-scale convolutional neural network for EEG-based motor imagery decoding
CN105956624A (zh) 基于空时频优化特征稀疏表示的运动想象脑电分类方法
Chen et al. Multiattention adaptation network for motor imagery recognition
CN109325586B (zh) 一种用于脑电信号去噪的系统
Ge et al. A double-partial least-squares model for the detection of steady-state visual evoked potentials
CN110175510B (zh) 基于脑功能网络特征的多模式运动想象识别方法
Shi et al. A binary harmony search algorithm as channel selection method for motor imagery-based BCI
Wang et al. Sparse Bayesian learning for end-to-end EEG decoding
Meng et al. A motor imagery EEG signal classification algorithm based on recurrence plot convolution neural network
Malekmohammadi et al. An efficient hardware implementation for a motor imagery brain computer interface system
CN110074779A (zh) 一种脑电信号识别方法及装置
Jia et al. Joint spatial and temporal features extraction for multi-classification of motor imagery EEG
Zhang et al. An amplitudes-perturbation data augmentation method in convolutional neural networks for EEG decoding
CN111931646A (zh) 基于经验模态分解算法的运动想象脑电信号的识别系统
Zheng et al. A power spectrum pattern difference-based time-frequency sub-band selection method for MI-EEG classification
Wang et al. Motor imagination eeg recognition algorithm based on dwt, CSP and extreme learning machine
CN109214325B (zh) 一种基于空间滤波与模版匹配的运动相关电位检测方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 18932139

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 18932139

Country of ref document: EP

Kind code of ref document: A1