CN113397572A - Surface electromyographic signal classification method and system based on Transformer model - Google Patents

Surface electromyographic signal classification method and system based on Transformer model Download PDF

Info

Publication number
CN113397572A
CN113397572A CN202110839308.5A CN202110839308A CN113397572A CN 113397572 A CN113397572 A CN 113397572A CN 202110839308 A CN202110839308 A CN 202110839308A CN 113397572 A CN113397572 A CN 113397572A
Authority
CN
China
Prior art keywords
data
model
matrix
window
head
Prior art date
Legal status (The legal status 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 status listed.)
Pending
Application number
CN202110839308.5A
Other languages
Chinese (zh)
Inventor
李智军
程琦云
李国欣
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
University of Science and Technology of China USTC
Original Assignee
University of Science and Technology of China USTC
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
Application filed by University of Science and Technology of China USTC filed Critical University of Science and Technology of China USTC
Priority to CN202110839308.5A priority Critical patent/CN113397572A/en
Publication of CN113397572A publication Critical patent/CN113397572A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
    • A61B5/316—Modalities, i.e. specific diagnostic methods
    • A61B5/389—Electromyography [EMG]
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7203—Signal processing specially adapted for physiological signals or for diagnostic purposes for noise prevention, reduction or removal
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A61B5/725—Details of waveform analysis using specific filters therefor, e.g. Kalman or adaptive filters
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems

Landscapes

  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Public Health (AREA)
  • Surgery (AREA)
  • Veterinary Medicine (AREA)
  • Signal Processing (AREA)
  • Artificial Intelligence (AREA)
  • Biophysics (AREA)
  • Pathology (AREA)
  • Biomedical Technology (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Medical Informatics (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Animal Behavior & Ethology (AREA)
  • Physiology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Psychiatry (AREA)
  • Evolutionary Computation (AREA)
  • Fuzzy Systems (AREA)
  • Mathematical Physics (AREA)
  • Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)
  • Image Processing (AREA)

Abstract

The invention provides a surface electromyogram signal classification method and system based on a Transformer model, which comprises the following steps: step S1: collecting corresponding multi-channel surface electromyographic signals according to preset actions, and filtering and removing noise of the collected signals; step S2: converting the filtered signals into a data-tag sequence by using a sliding window technology, and preprocessing the data of each window; step S3: carrying out position coding on the preprocessed window data, and inputting the window data into a coder layer of a Transformer model to extract characteristics; step S4: and integrating the features extracted by the encoder through global pooling, and then obtaining a final classification result through a layer of fully-connected network. The method is based on the deep learning algorithm, solves the problem of low calculation efficiency caused by the fact that the sequence models such as RNN, LSTM, GRU and the like cannot be subjected to parallel calculation, and simultaneously improves the accuracy of the models.

Description

Surface electromyographic signal classification method and system based on Transformer model
Technical Field
The invention relates to the field of machine learning, in particular to a surface electromyographic signal classification method and system based on a Transformer model.
Background
The surface electromyographic signals are electrical signals collected on human skin by surface electrodes, and the electrical signals are potential differences generated near muscle fibers by muscle movement. When a human body produces an exercise intention, the intention is generated and encoded in nerve signals by the brain and transmitted to the spinal cord, the nerve signals are transmitted to corresponding limbs (such as lower limbs) through nerve passages after secondary encoding, muscle fibers are contracted by the nerve signals to generate potential differences, and muscles pull the skeleton to complete the exercise. In this process, the movement is intended to be ultimately encoded in the electrical signals generated by the contraction of muscle fibers. By decoding this signal, the original movement intention can be obtained, thereby controlling the external machine. Compared with the method of directly decoding brain signals and nerve signals, the muscle electrical signals are closer to the action implementation stage, the contained information is more accurate, the signal-to-noise ratio is higher, and the acquisition is more convenient.
Machine learning is one of the primary methods of decoding surface muscle electrical signals. The method comprises two stages: feature extraction and classification in feature space. The existing manual feature extraction method has a large difference from the RNN neural network in precision, but the RNN neural network cannot realize parallel calculation in the calculation process and has low speed.
Patent document CN112466326A (application number: 202011470115.9) discloses a method for extracting speech emotion features based on a transform model encoder, which is applicable to the fields of artificial intelligence and speech emotion recognition. Firstly, extracting low-level speech emotion characteristics from an original speech waveform by using a sinnet filter, and then further learning the low-level speech emotion characteristics by using a multilayer transform model encoder; the improved transformer model encoder is added with a layer of s incnet filter, namely a set of parameterized sinc functions with band-pass filters, in front of the conventional transformer model encoder, the sinnet filter is utilized to complete the low-level feature extraction work of the voice original waveform signal, and the network can better capture important narrow-band emotional features, so that frame-level emotional features containing global context information at a deeper level are obtained.
Disclosure of Invention
Aiming at the defects in the prior art, the invention aims to provide a surface electromyogram signal classification method and system based on a Transformer model.
The surface electromyogram signal classification method based on the Transformer model provided by the invention comprises the following steps:
step S1: collecting corresponding multi-channel surface electromyographic signals according to preset actions, and filtering and removing noise of the collected signals;
step S2: converting the filtered signals into a data-tag sequence by using a sliding window technology, and preprocessing the data of each window;
step S3: carrying out position coding on the preprocessed window data, and inputting the window data into a coder layer of a Transformer model to extract characteristics;
step S4: and integrating the features extracted by the encoder through global pooling, and then obtaining a final classification result through a layer of fully-connected network.
Preferably, the step S1 includes: collecting corresponding multichannel surface electromyographic signals according to preset types of movement actions and preset types of rest and relaxation actions, and filtering noise of the collected signals through a band-pass filter.
Preferably, the sliding window technique in step S2 includes: preset group data overlapping is formed between adjacent windows;
the pretreatment comprises the following steps: each channel of each window is individually normalized.
Preferably, the position encoding in step S3 includes:
Figure BDA0003178254360000021
wherein pos represents a location of the data within the time window; i represents the location of the data feature within the current set of data; dinputRepresenting dimensions of input data features.
Preferably, the step S3 of inputting the encoder-layer extraction features of the transform model includes:
the encoder module of the used Transformer comprises a multi-head attention network and a feedforward network;
the multi-head attention network includes: extracting internal features of the input sequence through a multi-head attention mechanism, wherein the formula is as follows:
MultiHead(X)＝Concat(head1,…,headh)WO
Figure BDA0003178254360000023
Figure BDA0003178254360000022
wherein X represents a time window of input; concat represents a splicing function; parameter matrix
Figure BDA0003178254360000031
Figure BDA0003178254360000032
And
Figure BDA0003178254360000033
all are learnable matrices; dmodelRepresenting dimensions of the transform model output features; if the number of the heads of the multi-head is h, the shape of the parameter matrix is dk＝dv＝dmodelH, requirement dkMust be a square number; q represents a query matrix; k represents a matrix of the relevance of the inquired information and other information; v represents a matrix of queried information; superscript T represents matrix transposition; r represents a vector space; dinputCharacteristic dimension representing input information, dkRepresents one dimension of K;
the feedforward network comprises a two-layer fully-connected network, ReLU is used as an activation function, and the calculation formula is as follows:
FFN(x)＝max(0,xW1+b1)W2+b2
parameter matrix
Figure BDA0003178254360000034
And
Figure BDA0003178254360000035
is a learnable matrix. dh iddenRepresenting a hidden layer dimension; b1、b2A deviation term is represented.
Preferably, the step S4 includes: features extracted by the encoder are integrated using global average pooling, and then finally classified via the softmax function of the fully connected network.
The invention provides a surface electromyogram signal classification system based on a Transformer model, which comprises the following components:
module M1: collecting corresponding multi-channel surface electromyographic signals according to preset actions, and filtering and removing noise of the collected signals;
module M2: converting the filtered signals into a data-tag sequence by using a sliding window technology, and preprocessing the data of each window;
module M3: carrying out position coding on the preprocessed window data, and inputting the window data into a coder layer of a Transformer model to extract characteristics;
module M4: and integrating the features extracted by the encoder through global pooling, and then obtaining a final classification result through a layer of fully-connected network.
Preferably, said module M1 comprises: collecting corresponding multichannel surface electromyographic signals according to preset types of movement actions and preset types of rest and relaxation actions, and filtering noise of the collected signals through a band-pass filter.
Preferably, the position encoding in the module M3 includes:
Figure BDA0003178254360000036
wherein pos represents a location of the data within the time window; i represents the location of the data feature within the current set of data; dinputA dimension representing a characteristic of the input data;
the encoder layer extracting features of the transform model input in the module M3 includes:
the encoder module of the used Transformer comprises a multi-head attention network and a feedforward network;
the multi-head attention network includes: extracting internal features of the input sequence through a multi-head attention mechanism, wherein the formula is as follows:
MultiHead(X)＝Concat(head1,…,headh)WO
Figure BDA0003178254360000041
Figure BDA0003178254360000042
wherein X represents a time window of input; concat represents a splicing function; parameter matrix
Figure BDA0003178254360000043
Figure BDA0003178254360000044
And
Figure BDA0003178254360000045
all are learnable matrices; dmodelRepresenting dimensions of the transform model output features; if the number of the heads of the multi-head is h, the shape of the parameter matrix is dk＝dv＝dmodelH, requirement dkMust be a square number; q represents a query matrix; k represents a matrix of the relevance of the inquired information and other information; v represents a matrix of queried information; superscript T represents matrix transposition; r represents a vector space; dinputCharacteristic dimension representing input information, dkRepresents one dimension of K;
the feedforward network comprises a two-layer fully-connected network, ReLU is used as an activation function, and the calculation formula is as follows:
FFN(x)＝max(0,xW1+b1)W2+b2
parameter matrix
Figure BDA0003178254360000046
And
Figure BDA0003178254360000047
is a learnable matrix. dhiddenRepresenting a hidden layer dimension; b1、b2A deviation term is represented.
Preferably, said module M4 comprises: features extracted by the encoder are integrated using global average pooling, and then finally classified via the softmax function of the fully connected network.
Compared with the prior art, the invention has the following beneficial effects:
1. according to the method, the forearm electromyographic signals are classified by using a Transformer algorithm, the characteristics do not need to be designed manually, the process of characteristic selection is omitted, the quality of extracted characteristics is better, and the recognition precision is improved;
2. the invention utilizes a Multi-Head Attention network mechanism in a transform to extract sequence information in a time sequence, and compared with the traditional RNN type network, the accuracy is higher;
3. the Multi-Head Attention network mechanism used in the invention can realize the parallel calculation of a plurality of groups of data, and has higher calculation efficiency and higher speed compared with the traditional RNN type network which can only carry out serial calculation;
4. the method is based on the deep learning algorithm, solves the problem of low calculation efficiency caused by the fact that the sequence models such as RNN, LSTM, GRU and the like cannot be subjected to parallel calculation, and simultaneously improves the accuracy of the models.
Drawings
Other features, objects and advantages of the invention will become more apparent upon reading of the detailed description of non-limiting embodiments with reference to the following drawings:
FIG. 1 is a process framework diagram of the present invention;
FIG. 2 is a network framework of the present invention;
FIG. 3 is a flow chart of a self-attention mechanism operation;
FIG. 4 is a flow chart of the multi-head self-attention mechanism operation.
Detailed Description
The present invention will be described in detail with reference to specific examples. The following examples will assist those skilled in the art in further understanding the invention, but are not intended to limit the invention in any way. It should be noted that it would be obvious to those skilled in the art that various changes and modifications can be made without departing from the spirit of the invention. All falling within the scope of the present invention.
Example 1
The surface electromyogram signal classification method based on the Transformer model provided by the invention comprises the following steps:
step S1: collecting corresponding multi-channel surface electromyographic signals according to preset actions, and filtering and removing noise of the collected signals;
step S2: converting the filtered signals into a data-tag sequence by using a sliding window technology, and preprocessing the data of each window;
step S3: carrying out position coding on the preprocessed window data, and inputting the window data into a coder layer of a Transformer model to extract characteristics;
step S4: and integrating the features extracted by the encoder through global pooling, and then obtaining a final classification result through a layer of fully-connected network.
Specifically, the step S1 includes: according to six types of movement actions and one type of rest relaxing actions, six-channel signals acquired from six different positions of the forearm are acquired corresponding multi-channel surface electromyographic signals, and the acquired signals are filtered by a band-pass filter to filter noise so as to eliminate artifact noise (low frequency) and unnecessary high-frequency noise.
Specifically, the sliding window technique in step S2 includes: 125 under the sliding window band, 35 step length, and 90 groups of data overlap between adjacent windows;
the pretreatment comprises the following steps: each channel of each window is individually normalized.
Specifically, the position encoding in step S3 includes:
Figure BDA0003178254360000061
wherein pos represents a location of the data within the time window; i represents the location of the data feature within the current set of data; dinputRepresenting dimensions of input data features.
Specifically, the step S3 of inputting the encoder layer extraction features of the transform model includes:
the Encoder (Encoder) module of the transform used includes a Multi-Head Attention Network (Multi-Head Attention) and a Feed-Forward Network (Feed Forward Network);
the multi-head attention network includes: extracting internal features of the input sequence through a multi-head attention mechanism, wherein the formula is as follows:
MultiHead(X)＝Concat(head1,…,headh)WO
Figure BDA0003178254360000062
Figure BDA0003178254360000063
wherein X represents a time window of input; concat represents a splicing function; parameter matrix
Figure BDA0003178254360000064
Figure BDA0003178254360000065
And
Figure BDA0003178254360000066
all are learnable matrices; dmodelRepresenting dimensions of the transform model output features; if the number of the heads of the multi-head is h, the shape of the parameter matrix is dk＝dv＝dmodelH, requirement dkMust be a square number; q represents a query matrix; k represents a matrix of the relevance of the inquired information and other information; v represents a matrix of queried information; superscript T represents matrix transposition; r represents a vector space; dinputCharacteristic dimension representing input information, dkRepresents one dimension of K;
the feedforward network comprises a two-layer fully-connected network, ReLU is used as an activation function, and the calculation formula is as follows:
FFN(x)＝max(0,xW1+b1)W2+b2
parameter matrix
Figure BDA0003178254360000067
And
Figure BDA0003178254360000068
is a learnable matrix. dhiddenRepresenting a hidden layer dimension; b1、b2A deviation term is represented.
Specifically, the step S4 includes: features extracted by the encoder are integrated using global average pooling, and then finally classified via the softmax function of the fully connected network.
The invention provides a surface electromyogram signal classification system based on a Transformer model, which comprises the following components:
module M1: collecting corresponding multi-channel surface electromyographic signals according to preset actions, and filtering and removing noise of the collected signals;
module M2: converting the filtered signals into a data-tag sequence by using a sliding window technology, and preprocessing the data of each window;
module M3: carrying out position coding on the preprocessed window data, and inputting the window data into a coder layer of a Transformer model to extract characteristics;
module M4: and integrating the features extracted by the encoder through global pooling, and then obtaining a final classification result through a layer of fully-connected network.
Specifically, the module M1 includes: according to six types of movement actions and one type of rest relaxing actions, six-channel signals acquired from six different positions of the forearm are acquired corresponding multi-channel surface electromyographic signals, and the acquired signals are filtered by a band-pass filter to filter noise so as to eliminate artifact noise (low frequency) and unnecessary high-frequency noise.
Specifically, the sliding window technique in the module M2 includes: 125 under the sliding window band, 35 step length, and 90 groups of data overlap between adjacent windows;
the pretreatment comprises the following steps: each channel of each window is individually normalized.
Specifically, the position encoding in the module M3 includes:
Figure BDA0003178254360000071
wherein pos represents a location of the data within the time window; i represents the location of the data feature within the current set of data; dinputRepresenting dimensions of input data features.
Specifically, the step of inputting the encoder layer extraction features of the transform model in the module M3 includes:
the Encoder (Encoder) module of the transform used includes a Multi-Head Attention Network (Multi-Head Attention) and a Feed-Forward Network (Feed Forward Network);
the multi-head attention network includes: extracting internal features of the input sequence through a multi-head attention mechanism, wherein the formula is as follows:
MultiHead(X)＝Concat(head1,…,headh)WO
Figure BDA0003178254360000072
Figure BDA0003178254360000073
wherein X representsA time window of input; concat represents a splicing function; parameter matrix
Figure BDA0003178254360000074
Figure BDA0003178254360000075
And
Figure BDA0003178254360000076
all are learnable matrices; dmodelRepresenting dimensions of the transform model output features; if the number of the heads of the multi-head is h, the shape of the parameter matrix is dk＝dv＝dmodelH, requirement dkMust be a square number; q represents a query matrix; k represents a matrix of the relevance of the inquired information and other information; v represents a matrix of queried information; superscript T represents matrix transposition; r represents a vector space; dinputCharacteristic dimension representing input information, dkRepresents one dimension of K;
the feedforward network comprises a two-layer fully-connected network, ReLU is used as an activation function, and the calculation formula is as follows:
FFN(x)＝max(0,xW1+b1)W2+b2
parameter matrix
Figure BDA0003178254360000081
And
Figure BDA0003178254360000082
is a learnable matrix. dh iddenRepresenting a hidden layer dimension; b1、b2A deviation term is represented.
Specifically, the module M4 includes: features extracted by the encoder are integrated using global average pooling, and then finally classified via the softmax function of the fully connected network.
Example 2
Example 2 is a preferred example of example 1
The invention is further described below with reference to the accompanying drawings. The following examples are only for illustrating the technical solutions of the present invention more clearly, and the protection scope of the present invention is not limited thereby.
As shown in fig. 1 to 4, in a surface electromyographic signal classification method based on a transform model, firstly, electromyographic signals of forearms are collected according to seven designed actions, wherein the seven actions are respectively as follows: fist making, stretching, inward folding of forearm, outward folding of wrist towards forearm, inward folding of elbow, outward folding of elbow and natural relaxation. Through the seven actions, corresponding forearm surface electromyographic signals are collected, and the sampling frequency is 1024 Hz. For the collected signals, a band-pass filter is arranged for eliminating artifact noise and high-frequency noise with small information quantity, wherein the artifact noise does not exceed 20Hz generally, and the high-frequency noise exceeds 500Hz generally, and the passing frequency of the band-pass filter is 20-500 Hz.
For the filtered data, we use sliding window technique to convert the collected time series into "data-label" pairs, and normalize each channel in each time window, with the following formula:
Figure BDA0003178254360000083
where min is the minimum value taken for the channel and max is the maximum value taken for the channel.
For the data after windowing and normalization, a position code is added to the data so that a subsequent model can conveniently learn position information in a sequence during learning, and the position code formula is as follows:
Figure BDA0003178254360000084
where pos is the location of the data within the time window, i is the location of the data feature within the set of data, dinputFor the set model output dimension, from the data we have collected, there is dinput＝6
For data added with position information, we input it into Encoder module of Transformer, the data first passes through Multi-Head Attention layer, and the calculation formula of this part is as follows:
MultiHead(X)＝Concat(head1,…,headh)WO
Figure BDA0003178254360000091
Figure BDA0003178254360000092
where X is the input time window, Concat is the splicing function, the parameter matrix
Figure BDA0003178254360000093
Figure BDA0003178254360000094
And
Figure BDA0003178254360000095
are all learnable matrices, dmodelFor the dimension of the model output features, let d heremodel256. If the number of the multi-head is h is 4, the parameter matrix shape is dk＝dv＝dmodelWhen d is 64kThe requirement of being an average is satisfied.
Inputting the features extracted by the Multi-Head Attention layer into a Feed Forward Network layer, wherein the Feed Forward Network layer consists of two layers of fully connected networks, and ReLU is used as an activation function, and the calculation formula is as follows:
FFN(x)＝max(0,xW1+b1)W2+b2
parameter matrix
Figure BDA0003178254360000096
And
Figure BDA0003178254360000097
for learnable matrices, d is set herehidden＝256。
We use global tiepooling to integrate features extracted by the Encoder layer, followed finally by a fully connected network layer, using the softmax function as the activation function to complete the final classification.
Those skilled in the art will appreciate that, in addition to implementing the systems, apparatus, and various modules thereof provided by the present invention in purely computer readable program code, the same procedures can be implemented entirely by logically programming method steps such that the systems, apparatus, and various modules thereof are provided in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers, embedded microcontrollers and the like. Therefore, the system, the device and the modules thereof provided by the present invention can be considered as a hardware component, and the modules included in the system, the device and the modules thereof for implementing various programs can also be considered as structures in the hardware component; modules for performing various functions may also be considered to be both software programs for performing the methods and structures within hardware components.
The foregoing description of specific embodiments of the present invention has been presented. It is to be understood that the present invention is not limited to the specific embodiments described above, and that various changes or modifications may be made by one skilled in the art within the scope of the appended claims without departing from the spirit of the invention. The embodiments and features of the embodiments of the present application may be combined with each other arbitrarily without conflict.

Claims (10)

1. A surface electromyogram signal classification method based on a Transformer model is characterized by comprising the following steps:
step S1: collecting corresponding multi-channel surface electromyographic signals according to preset actions, and filtering and removing noise of the collected signals;
step S2: converting the filtered signals into a data-tag sequence by using a sliding window technology, and preprocessing the data of each window;
step S3: carrying out position coding on the preprocessed window data, and inputting the window data into a coder layer of a Transformer model to extract characteristics;
step S4: and integrating the features extracted by the encoder through global pooling, and then obtaining a final classification result through a layer of fully-connected network.
2. The method for classifying surface electromyogram signals based on a fransformer model according to claim 1, wherein the step S1 comprises: collecting corresponding multichannel surface electromyographic signals according to preset types of movement actions and preset types of rest and relaxation actions, and filtering noise of the collected signals through a band-pass filter.
3. The method for classifying surface electromyographic signals based on a fransformer model according to claim 1, wherein the sliding window technique in the step S2 comprises: preset group data overlapping is formed between adjacent windows;
the pretreatment comprises the following steps: each channel of each window is individually normalized.
4. The method for classifying surface electromyographic signals based on a fransformer model according to claim 1, wherein the position encoding in the step S3 comprises:
Figure FDA0003178254350000011
wherein pos represents a location of the data within the time window; i represents the location of the data feature within the current set of data; dinputRepresenting dimensions of input data features.
5. The method for classifying surface electromyogram signals based on a fransformer model according to claim 1, wherein the step S3 of inputting the encoder layer extraction features of the fransformer model comprises:
the encoder module of the used Transformer comprises a multi-head attention network and a feedforward network;
the multi-head attention network includes: extracting internal features of the input sequence through a multi-head attention mechanism, wherein the formula is as follows:
MultiHead(X)＝Concat(head1,…,headh)WO
headi＝Attention(XWi Q,XWi K,XWi V)
Figure FDA0003178254350000021
wherein X represents a time window of input; concat represents a splicing function; parameter matrix
Figure FDA0003178254350000022
Figure FDA0003178254350000023
And
Figure FDA0003178254350000024
all are learnable matrices; dmodelRepresenting dimensions of the transform model output features; if the number of the heads of the multi-head is h, the shape of the parameter matrix is dk＝dv＝dmodelH, requirement dkMust be a square number; q represents a query matrix; k represents a matrix of the relevance of the inquired information and other information; v represents a matrix of queried information; superscript T represents matrix transposition; r represents a vector space; dinputCharacteristic dimension representing input information, dkRepresents one dimension of K;
the feedforward network comprises a two-layer fully-connected network, ReLU is used as an activation function, and the calculation formula is as follows:
FFN(x)＝max(0,xW1+b1)W2+b2
parameter matrix
Figure FDA0003178254350000025
And
Figure FDA0003178254350000026
is a learnable matrix. dh iddenRepresenting a hidden layer dimension; b1、b2A deviation term is represented.
6. The method for classifying surface electromyogram signals based on a fransformer model according to claim 1, wherein the step S4 comprises: features extracted by the encoder are integrated using global average pooling, and then finally classified via the softmax function of the fully connected network.
7. A surface electromyogram signal classification system based on a Transformer model is characterized by comprising:
module M1: collecting corresponding multi-channel surface electromyographic signals according to preset actions, and filtering and removing noise of the collected signals;
module M2: converting the filtered signals into a data-tag sequence by using a sliding window technology, and preprocessing the data of each window;
module M3: carrying out position coding on the preprocessed window data, and inputting the window data into a coder layer of a Transformer model to extract characteristics;
module M4: and integrating the features extracted by the encoder through global pooling, and then obtaining a final classification result through a layer of fully-connected network.
8. The transform model-based surface electromyographic signal classification system according to claim 7, wherein the module M1 comprises: collecting corresponding multichannel surface electromyographic signals according to preset types of movement actions and preset types of rest and relaxation actions, and filtering noise of the collected signals through a band-pass filter.
9. The transform model-based surface electromyogram signal classification system according to claim 7, wherein the position coding in the module M3 comprises:
Figure FDA0003178254350000031
wherein pos represents a location of the data within the time window; i represents the location of the data feature within the current set of data; dinputA dimension representing a characteristic of the input data;
the encoder layer extracting features of the transform model input in the module M3 includes:
the encoder module of the used Transformer comprises a multi-head attention network and a feedforward network;
the multi-head attention network includes: extracting internal features of the input sequence through a multi-head attention mechanism, wherein the formula is as follows:
MultiHead(X)＝Concat(head1,…,headh)WO
headi＝Attention(XWi Q,XWi K,XWi V)
Figure FDA0003178254350000032
wherein X represents a time window of input; concat represents a splicing function; parameter matrix
Figure FDA0003178254350000033
Figure FDA0003178254350000034
And
Figure FDA0003178254350000035
all are learnable matrices; dmodelRepresenting dimensions of the transform model output features; if the number of the heads of the multi-head is h, the shape of the parameter matrix is dk＝dv＝dmodelH, requirement dkMust be a square number; q represents a query matrix; k represents a matrix of the relevance of the inquired information and other information; v represents a matrix of queried information; superscript T represents matrix transposition; r represents a vector space; dinputCharacteristic dimension representing input information, dkRepresents one dimension of K;
the feedforward network comprises a two-layer fully-connected network, ReLU is used as an activation function, and the calculation formula is as follows:
FFN(x)＝max(0,xW1+b1)W2+b2
parameter matrix
Figure FDA0003178254350000036
And
Figure FDA0003178254350000037
is a learnable matrix. dh iddenRepresenting a hidden layer dimension; b1、b2A deviation term is represented.
10. The transform model-based surface electromyographic signal classification system according to claim 7, wherein the module M4 comprises: features extracted by the encoder are integrated using global average pooling, and then finally classified via the softmax function of the fully connected network.
CN202110839308.5A 2021-07-23 2021-07-23 Surface electromyographic signal classification method and system based on Transformer model Pending CN113397572A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110839308.5A CN113397572A (en) 2021-07-23 2021-07-23 Surface electromyographic signal classification method and system based on Transformer model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110839308.5A CN113397572A (en) 2021-07-23 2021-07-23 Surface electromyographic signal classification method and system based on Transformer model

Publications (1)

Publication Number Publication Date
CN113397572A true CN113397572A (en) 2021-09-17

Family

ID=77687607

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110839308.5A Pending CN113397572A (en) 2021-07-23 2021-07-23 Surface electromyographic signal classification method and system based on Transformer model

Country Status (1)

Country Link
CN (1) CN113397572A (en)

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113901893A (en) * 2021-09-22 2022-01-07 西安交通大学 Electrocardiosignal identification and classification method based on multiple cascade deep neural network
CN114626424A (en) * 2022-05-16 2022-06-14 天津大学 Data enhancement-based silent speech recognition method and device
CN114863912A (en) * 2022-05-05 2022-08-05 中国科学技术大学 Silent voice decoding method based on surface electromyogram signals
CN116070985A (en) * 2023-04-06 2023-05-05 江苏华溯大数据有限公司 Dangerous chemical vehicle loading and unloading process identification method
CN116127364A (en) * 2023-04-12 2023-05-16 上海术理智能科技有限公司 Integrated transducer-based motor imagery decoding method and system
CN116434343A (en) * 2023-04-25 2023-07-14 天津大学 Video motion recognition method based on high-low frequency double branches
CN116485729A (en) * 2023-04-03 2023-07-25 兰州大学 Multistage bridge defect detection method based on transformer

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109924977A (en) * 2019-03-21 2019-06-25 西安交通大学 A kind of surface electromyogram signal classification method based on CNN and LSTM
US10743809B1 (en) * 2019-09-20 2020-08-18 CeriBell, Inc. Systems and methods for seizure prediction and detection
CN111616706A (en) * 2020-05-20 2020-09-04 山东中科先进技术研究院有限公司 Surface electromyogram signal classification method and system based on convolutional neural network
CN112466326A (en) * 2020-12-14 2021-03-09 江苏师范大学 Speech emotion feature extraction method based on transform model encoder
CN113033657A (en) * 2021-03-24 2021-06-25 武汉理工大学 Multi-user behavior identification method based on Transformer network

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109924977A (en) * 2019-03-21 2019-06-25 西安交通大学 A kind of surface electromyogram signal classification method based on CNN and LSTM
US10743809B1 (en) * 2019-09-20 2020-08-18 CeriBell, Inc. Systems and methods for seizure prediction and detection
CN111616706A (en) * 2020-05-20 2020-09-04 山东中科先进技术研究院有限公司 Surface electromyogram signal classification method and system based on convolutional neural network
CN112466326A (en) * 2020-12-14 2021-03-09 江苏师范大学 Speech emotion feature extraction method based on transform model encoder
CN113033657A (en) * 2021-03-24 2021-06-25 武汉理工大学 Multi-user behavior identification method based on Transformer network

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
QIZHENG GU ET AL.: "Automatic Generation of Electromyogram Diagnosis Report", 《2020 IEEE INTERNATIONAL CONFERENCE ON BIOINFORMATICS AND BIOMEDICINE》 *

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113901893A (en) * 2021-09-22 2022-01-07 西安交通大学 Electrocardiosignal identification and classification method based on multiple cascade deep neural network
CN113901893B (en) * 2021-09-22 2023-09-15 西安交通大学 Electrocardiosignal identification and classification method based on multi-cascade deep neural network
CN114863912A (en) * 2022-05-05 2022-08-05 中国科学技术大学 Silent voice decoding method based on surface electromyogram signals
CN114863912B (en) * 2022-05-05 2024-05-10 中国科学技术大学 Silent voice decoding method based on surface electromyographic signals
CN114626424A (en) * 2022-05-16 2022-06-14 天津大学 Data enhancement-based silent speech recognition method and device
CN114626424B (en) * 2022-05-16 2022-09-13 天津大学 Data enhancement-based silent speech recognition method and device
CN116485729A (en) * 2023-04-03 2023-07-25 兰州大学 Multistage bridge defect detection method based on transformer
CN116485729B (en) * 2023-04-03 2024-01-12 兰州大学 Multistage bridge defect detection method based on transformer
CN116070985A (en) * 2023-04-06 2023-05-05 江苏华溯大数据有限公司 Dangerous chemical vehicle loading and unloading process identification method
CN116127364A (en) * 2023-04-12 2023-05-16 上海术理智能科技有限公司 Integrated transducer-based motor imagery decoding method and system
CN116434343A (en) * 2023-04-25 2023-07-14 天津大学 Video motion recognition method based on high-low frequency double branches
CN116434343B (en) * 2023-04-25 2023-09-19 天津大学 Video motion recognition method based on high-low frequency double branches

Similar Documents

Publication Publication Date Title
CN113397572A (en) Surface electromyographic signal classification method and system based on Transformer model
CN108491077B (en) Surface electromyographic signal gesture recognition method based on multi-stream divide-and-conquer convolutional neural network
CN110610172B (en) Myoelectric gesture recognition method based on RNN-CNN architecture
CN109620651B (en) Intelligent auxiliary rehabilitation equipment based on synchronous brain and muscle electricity
CN107736894A (en) A kind of electrocardiosignal Emotion identification method based on deep learning
CN109598222B (en) EEMD data enhancement-based wavelet neural network motor imagery electroencephalogram classification method
CN112466326A (en) Speech emotion feature extraction method based on transform model encoder
Montazerin et al. ViT-HGR: Vision transformer-based hand gesture recognition from high density surface EMG signals
Godoy et al. Electromyography-based, robust hand motion classification employing temporal multi-channel vision transformers
CN113158964A (en) Sleep staging method based on residual learning and multi-granularity feature fusion
CN113180659A (en) Electroencephalogram emotion recognition system based on three-dimensional features and cavity full convolution network
CN113128353B (en) Emotion perception method and system oriented to natural man-machine interaction
Huang et al. Classify motor imagery by a novel CNN with data augmentation
CN110598628A (en) Electromyographic signal hand motion recognition method based on integrated deep learning
Roy et al. Hand movement recognition using cross spectrum image analysis of EMG signals-A deep learning approach
Ye et al. Attention bidirectional LSTM networks based mime speech recognition using sEMG data
CN111950460A (en) Muscle strength self-adaptive stroke patient hand rehabilitation training action recognition method
CN116628420A (en) Brain wave signal processing method based on LSTM neural network element learning
CN116225222A (en) Brain-computer interaction intention recognition method and system based on lightweight gradient lifting decision tree
CN114743569A (en) Speech emotion recognition method based on double-layer fusion deep network
CN113642528B (en) Hand movement intention classification method based on convolutional neural network
Ye et al. Upper Limb Motion Recognition Using Gated Convolution Neural Network via Multi-Channel sEMG
CN111883178B (en) Double-channel voice-to-image-based emotion recognition method
CN114343679A (en) Surface electromyogram signal upper limb action recognition method and system based on transfer learning
Bo et al. Hand gesture recognition using semg signals based on cnn

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20210917