阅读说明：本技术 基于lda算法的表面肌电信号的通道选取方法 (LDA algorithm-based channel selection method for surface electromyographic signals ) 是由 王念峰 张新浩 张宪民 于 2019-08-28 设计创作，主要内容包括：本发明公开了一种基于LDA算法的表面肌电信号的通道选取方法。对于确定数量的肌电信号通道组合,从胫骨前肌、腓骨长肌和比目鱼肌,腓肠肌外侧肌肉,小腿内侧的腓肠肌内侧肌肉,小腿后侧的腓肠肌内侧肌肉这6个待选的的肌电信号通道中挑出不同的组合,用基于LDA算法的准确率计算算法计算每种通道组合对脚踝的放松、背屈、跖屈、外翻和内翻5种动作模式的平均识别率。通过对每种通道组合的平均识别率进行比较,得到最优的通道组合。通过基于LDA算法的识别率计算,确定最优肌电信号通道组合,相比于通过经验确定通道组合,本发明更加科学和准确,能够使用较少的肌电信号通道实现对脚踝5种动作模式以较高的识别率进行识别。(the invention discloses a channel selection method of a surface electromyogram signal based on an LDA algorithm. For the myoelectric signal channel combinations with the determined number, different combinations are selected from 6 myoelectric signal channels to be selected, namely tibialis anterior muscle, peroneal long muscle, soleus muscle, gastrocnemius lateral muscle, gastrocnemius medial muscle on the inner side of the lower leg and gastrocnemius medial muscle on the rear side of the lower leg, and the average recognition rate of 5 action modes of relaxing, dorsiflexion, plantarflexion, eversion and inversion of the ankle of each channel combination is calculated by using an accuracy calculation algorithm based on an LDA (linear discriminant analysis) algorithm. And comparing the average recognition rate of each channel combination to obtain the optimal channel combination. Compared with the method for determining the channel combination through experience, the method is more scientific and accurate, and can realize the recognition of 5 ankle action modes with higher recognition rate by using fewer myoelectric signal channels.)

1. The method for selecting the channel of the surface electromyographic signal based on the LDA algorithm is characterized by comprising the following steps of:

step S1, cleaning 6 parts of tibialis anterior muscle, peroneal longus muscle, soleus muscle, gastrocnemius lateral muscle, gastrocnemius medial muscle on the inner side of the calf and gastrocnemius medial muscle on the rear side of the calf by alcohol, and fixing surface myoelectric electrodes on the 6 parts, wherein the 6 parts are 6 myoelectric signal channels to be selected;

step S2, the examinee sits on the chair, and the thighs and the calves are correspondingly supported; only the ankle joint is moved, five action modes of relaxation, dorsiflexion, plantarflexion, eversion and inversion are made, myoelectric signal data of each action mode of each myoelectric signal channel is collected by a surface myoelectric sensor and is stored in a computer in a TXT format;

Step S3, selecting m channel combinations with the number of the electromyographic signal channels being m, and selecting n types of m channel combinations for the m electromyographic signal channels, wherein m is 1, …, 6;

step S4, performing data segmentation on the myoelectric signal data of each m-channel combination to obtain a plurality of segments of myoelectric signal data, defining each segment of myoelectric signal data as a data window, and calculating 10 characteristics of absolute mean, variance, root mean square, wavelength, Wilson amplitude, zero-crossing point value, sample entropy, autoregressive coefficient, average frequency and median frequency of the data window;

Step S5, for each m-channel combination, training and identifying by using an LDA algorithm, and counting the average identification rate of each m-channel combination on five action modes of relaxation, dorsiflexion, plantarflexion, eversion and inversion;

and step S6, comparing the average recognition rates of the n types of m-channel combinations, and taking the m-channel combination corresponding to the highest average recognition rate as the optimal m-channel combination.

2. The method for selecting the channel of the surface electromyographic signal based on the LDA algorithm as claimed in claim 1, wherein the collecting frequency range of the electromyographic electrodes in step S2 is 500-1000 Hz.

3. the LDA algorithm-based channel selection method for surface electromyographic signals according to claim 1, wherein the step S4 of performing data segmentation on each type of m-channel combined electromyographic signal data means that each 200 ms electromyographic signal segment is used to divide the collected continuous electromyographic signals into several segments.

4. The LDA algorithm-based surface electromyography signal channel selection method of claim 1, wherein in step S4, each segment of electromyography signal data comprises 100 and 200 electromyography data.

5. the channel selection method for surface electromyography signals based on LDA algorithm of claim 1, wherein in step S4, 10 features are obtained for one m-channel combined electromyography signal data, and the 10 features are randomly combined in different numbers, which results in 1023 feature combinations.

6. the method for selecting the channel of the surface electromyogram signal based on the LDA algorithm of claim 1, wherein in step S5, the LDA algorithm is used to train and recognize the feature combination vector data of each feature combination, and to obtain the recognition rate of 5 motion patterns for each feature combination, and then to average the recognition rates of 5 motion patterns for 1023 feature combinations, and use the average as the average recognition rate of the electromyogram signal data of the m-channel combination for 5 motion patterns.

7. The LDA algorithm-based channel selection method for surface electromyographic signals according to claim 6, wherein various features are calculated according to electromyographic data in a data window, and the various features of each feature combination form a vector and are defined as feature combination vectors; training and identifying the feature combination vector of each feature combination by using an LDA algorithm, and dividing the feature combination vector calculated based on all collected electromyographic signal data respectively belonging to 5 action modes into two parts for one feature combination in 1023 feature combinations, wherein the first half is used for training and the second half is used for identifying; the training process comprises the following steps: and training the characteristic combination vectors of which the first half parts respectively belong to 5 action modes by using an LDA algorithm to obtain a dimension reduction matrix for performing dimension reduction operation on the characteristic combination vectors and a label point vector of each action mode category.

8. The method for selecting a channel of a surface electromyogram signal based on an LDA algorithm according to claim 7, wherein the steps of obtaining a dimension reduction matrix for performing a dimension reduction operation on the feature combination vector and a label point vector for each motion pattern category through the LDA algorithm are specifically as follows:

S5.1, calculating an inter-class hash matrix SB of the feature combination vector by using an LDA algorithm:

Wherein, T represents matrix transposition, mi is the number of the feature combination vectors used for training in the ith motion pattern category, namely, the electromyographic signals of the ith motion pattern category are divided, after the feature combination vectors are obtained, the number of the feature combination vectors used for training in the first half is obtained, C is the number of the motion patterns, μ is the average vector of all the feature combination vectors, μ i is the average vector of the feature combination vectors belonging to the ith motion pattern category, and i refers to the number of the motion pattern category;

S5.2, calculating an intra-class hash matrix SW:

wherein, T represents matrix transposition, xj is the jth characteristic combination vector belonging to the ith action mode, and j refers to the serial number of the characteristic combination vector;

s5.3, calculating a projection matrix, wherein J (W) is defined as follows, and the larger J (W), the better the classification of the vector after dimensionality reduction is; obtaining an optimal dimensionality reduction matrix by maximizing J (W):

wherein, T represents matrix transposition, the projection matrix, namely the dimensionality reduction matrix W, can enable J (W) to obtain the maximum value, the method for obtaining W is to obtain the eigenvalue and the eigenvector firstly, sort the eigenvector according to the eigenvalue from big to small, then obtain the first T eigenvectors to form the projection matrix, T represents the data dimensionality after dimensionality reduction, and when the dimensionality of the eigenvector of the eigenvalue combination needing dimensionality reduction is more than the category number of the action mode to be identified, the data dimensionality T after dimensionality reduction is specified to be C-1; when the dimensionality of the feature combination vector needing dimensionality reduction is smaller than the category number of the action mode to be identified, specifying the dimensionality of data subjected to dimensionality reduction as the dimensionality of the feature combination vector minus 1; when the dimension of the feature combination vector needing dimension reduction is 1 dimension, the dimension t of the data after dimension reduction is 1;

And S5.4, multiplying the dimensionality reduction matrix W by the mui to obtain the label point vector of each category.

9. the method for selecting channels of surface electromyographic signals based on the LDA algorithm according to claim 7, wherein the Euclidean distance of feature combination vectors of the same class becomes smaller and the Euclidean distance of feature combination vectors of different classes becomes larger after dimension reduction is performed on the dimension reduction matrix.

10. The method for selecting the channel of the surface electromyogram signal based on the LDA algorithm according to claim 7, wherein the LDA algorithm is used to train and identify the feature combination vector of each feature combination, and the identification process comprises: for the characteristic combination vector used for identification, multiplying the characteristic combination vector by a dimension reduction matrix W for dimension reduction, and obtaining the category of the characteristic combination vector by comparing the Euclidean distance between the characteristic combination vector subjected to dimension reduction and a label point vector; and comparing the action mode category judged by the LDA algorithm with the action mode category to which the characteristic combination vector actually belongs, and counting the recognition rate of the characteristic combination to 5 action modes.

Technical Field

The invention relates to the field of pattern recognition, in particular to a surface electromyographic signal channel selection method based on an LDA algorithm. The method can be used for selecting the position of the myoelectric electrode in ankle action pattern recognition or selecting the position of the myoelectric electrode in other action patterns.

background

Surface Electromyography (SEMG), which is simply called surface electromyography or electromyography, is a bioelectric signal generated during muscle activity collected by a surface electrode. When the ankle moves correspondingly, the muscles of the lower leg generate action potentials to induce muscle contraction, and the signals are conducted through subcutaneous tissues and can be finally detected by the surface electrodes on the skin surface. The surface electromyogram signal contains information of corresponding action, and the detection of the surface electromyogram signal belongs to non-invasive detection and is more convenient.

rehabilitation exercises of the ankle have important significance for recovery of injury of the ankle. Ankle rehabilitation generally allows the rehabilitee to use the ankle to perform several actions, including relaxation, dorsiflexion, plantarflexion, eversion and inversion. The action of the ankle is judged through the myoelectric signal, and the action information can be provided for corresponding display or rehabilitation auxiliary equipment, so that the rehabilitation training process is simpler and more visual.

The LDA algorithm is called linear discriminant analysis, and is a typical pattern recognition algorithm. The characteristic combination vectors of the electromyographic signals of different action modes are trained by an LDA algorithm, so that different action modes can be effectively judged.

Humans possess a complex muscular system, with many muscles on the lower leg being associated with ankle movement. In practical use, it is very important to realize recognition with a high recognition rate by using fewer electromyographic signals on muscles, so that an electromyographic signal channel is often selected. In the research and application of pattern recognition of electromyographic signals, there are not many researches on how to select effective muscles, and the selection is performed based on the experience of human body kinematics.

disclosure of Invention

Aiming at the technical problems, the invention provides a surface electromyogram signal channel selection method based on an LDA algorithm. The ankle relaxation, dorsiflexion, plantarflexion, eversion and inversion are considered as 5 recognition modes. Through statistical comparison of the motion pattern recognition rate, the optimal electromyographic signal channel combination is selected from 6 channels to be selected, namely the main muscles of the lower leg, namely tibialis anterior muscle, peroneus longus muscle, soleus muscle, the lateral muscles of the gastrocnemius muscle, the medial muscles of the gastrocnemius muscle on the inner side of the lower leg and the medial muscles of the gastrocnemius muscle on the rear side of the lower leg.

The purpose of the invention is realized by at least one of the following technical solutions.

the method for selecting the channel of the surface electromyographic signal based on the LDA algorithm comprises the following steps:

Step S1, cleaning 6 parts of tibialis anterior muscle, peroneal longus muscle, soleus muscle, gastrocnemius lateral muscle, gastrocnemius medial muscle on the inner side of the calf and gastrocnemius medial muscle on the rear side of the calf by alcohol, and fixing surface myoelectric electrodes on the 6 parts, wherein the 6 parts are 6 myoelectric signal channels to be selected;

Step S2, the examinee sits on the chair, and the thighs and the calves are correspondingly supported; only the ankle joint is moved, five action modes of relaxation, dorsiflexion, plantarflexion, eversion and inversion are made, a surface electromyographic sensor, namely an electromyographic electrode is used for collecting electromyographic signal data of each action mode of each electromyographic signal channel, and the electromyographic signal data are stored in a computer in a TXT format;

Step S3, selecting m channel combinations with the number of the electromyographic signal channels being m, and then selecting the m electromyographic signal channels, wherein n channel combinations are provided in total, where m is 1, …, 6;

Step S4, performing data segmentation on the myoelectric signal data of each m-channel combination to obtain a plurality of segments of myoelectric signal data, defining each segment of myoelectric signal data as a data window, and calculating 10 characteristics of an absolute mean value (MAV), a Variance (VAR), a Root Mean Square (RMS), a Wavelength (WL), a Wilson amplitude value (WAMP), a zero-crossing point value (ZC), a Sample Entropy (SE), an autoregressive coefficient (AR), an average frequency (MNF) and a median frequency (MDF) of the data window;

Step S5, for each m-channel combination, training and identifying by using an LDA algorithm, and counting the average identification rate of each m-channel combination on five action modes of relaxation, dorsiflexion, plantarflexion, eversion and inversion;

And step S6, comparing the average recognition rates of the n types of m-channel combinations, and taking the m-channel combination corresponding to the highest average recognition rate as the optimal m-channel combination.

further, in step S2, the collection frequency range of the myoelectric electrode is 500-1000 Hz.

further, in step S4, the data division of the electromyographic signal data for each m-channel combination means that the electromyographic signal data for each 200 milliseconds is divided into several segments, and the collected continuous electromyographic signals are divided into several segments.

further, in step S4, there are 100 and 200 pieces of electromyographic data for each segment of electromyographic signal data.

Further, in step S4, 10 kinds of features can be obtained for one m-channel combined electromyogram signal data, and the 10 kinds of features are randomly combined in different numbers, so that 1023 kinds of feature combinations can be obtained in total.

Further, in step S5, the feature combination vector data of each feature combination is trained and recognized by the LDA algorithm, and the recognition rates of 5 motion patterns for each feature combination are obtained, and then the recognition rates of 5 motion patterns for 1023 feature combinations are averaged, and the average is used as the average recognition rate of 5 motion patterns for the electromyographic signal data of the m-channel combination.

Further, various features are calculated according to the electromyographic data in the data window, and various features of each feature combination form a vector which is defined as a feature combination vector; training and identifying the feature combination vector of each feature combination by using an LDA algorithm, and dividing the feature combination vector calculated based on all collected electromyographic signal data respectively belonging to 5 action modes into two parts for one feature combination in 1023 feature combinations, wherein the first half is used for training and the second half is used for identifying; the training process comprises the following steps: and training the characteristic combination vectors of which the first half parts respectively belong to 5 action modes by using an LDA algorithm to obtain a dimension reduction matrix for performing dimension reduction operation on the characteristic combination vectors and a label point vector of each action mode category.

Further, the steps of obtaining a dimension reduction matrix for performing dimension reduction operation on the feature combination vector and a label point vector of each action mode category through an LDA algorithm are specifically as follows:

S5.1, calculating an inter-class hash matrix SB of the feature combination vector by using an LDA algorithm:

wherein, T represents matrix transposition, mi is the number of the feature combination vectors used for training in the ith motion pattern category, namely, the electromyographic signals of the ith motion pattern category are divided, after the feature combination vectors are obtained, the number of the feature combination vectors used for training in the first half is obtained, C is the number of the motion patterns, μ is the average vector of all the feature combination vectors, μ i is the average vector of the feature combination vectors belonging to the ith motion pattern category, and i refers to the number of the motion pattern category;

S5.2, calculating an intra-class hash matrix SW:

wherein, T represents matrix transposition, xj is the jth characteristic combination vector belonging to the ith action mode, and j refers to the serial number of the characteristic combination vector;

S5.3, calculating a projection matrix, wherein J (W) is defined as follows, and the larger J (W), the better the classification of the vector after dimensionality reduction is; obtaining an optimal dimensionality reduction matrix by maximizing J (W):

Wherein, T represents matrix transposition, the projection matrix, namely the dimensionality reduction matrix W, can enable J (W) to obtain the maximum value, the method for obtaining W is to obtain the eigenvalue and the eigenvector firstly, sort the eigenvector according to the eigenvalue from big to small, then obtain the first T eigenvectors to form the projection matrix, T represents the data dimensionality after dimensionality reduction, and when the dimensionality of the eigenvector of the eigenvalue combination needing dimensionality reduction is more than the category number of the action mode to be identified, the data dimensionality T after dimensionality reduction is specified to be C-1; when the dimensionality of the feature combination vector needing dimensionality reduction is smaller than the category number of the action mode to be identified, specifying the dimensionality of data subjected to dimensionality reduction as the dimensionality of the feature combination vector minus 1; when the dimension of the feature combination vector needing dimension reduction is 1 dimension, the dimension t of the data after dimension reduction is 1;

And S5.4, multiplying the dimensionality reduction matrix W by the mui to obtain the label point vector of each category.

further, after the dimension reduction is performed on the dimension reduction matrix, the Euclidean distance of the feature combination vectors of the same category becomes smaller, and the Euclidean distance of the feature combination vectors of different categories becomes larger.

Further, the feature combination vector of each feature combination is trained and identified by utilizing an LDA algorithm, and the identification process is as follows: for the characteristic combination vector used for identification, multiplying the characteristic combination vector by a dimension reduction matrix W for dimension reduction, and obtaining the category of the characteristic combination vector by comparing the Euclidean distance between the characteristic combination vector subjected to dimension reduction and a label point vector; and comparing the action mode category judged by the LDA algorithm with the action mode category to which the characteristic combination vector actually belongs, and counting the recognition rate of the characteristic combination to 5 action modes.

compared with the prior art, the invention has the advantages that:

The optimal electromyographic signal channel combination is determined through the identification rate calculation based on the LDA algorithm, compared with the method that the channel combination is determined through experience, the method is more scientific and accurate, and 5 ankle action modes can be identified with higher identification rate by using fewer electromyographic signal channels.

Drawings

Fig. 1 is a flowchart of a method for selecting a channel of a surface electromyogram signal based on an LDA algorithm in an embodiment of the present invention.

Detailed Description

specific embodiments of the present invention will be described in further detail with reference to the accompanying drawings and examples.