CN111772669B - Elbow joint contraction muscle force estimation method based on adaptive long-time and short-time memory network - Google Patents

Elbow joint contraction muscle force estimation method based on adaptive long-time and short-time memory network Download PDF

Info

Publication number
CN111772669B
CN111772669B CN202010830177.XA CN202010830177A CN111772669B CN 111772669 B CN111772669 B CN 111772669B CN 202010830177 A CN202010830177 A CN 202010830177A CN 111772669 B CN111772669 B CN 111772669B
Authority
CN
China
Prior art keywords
muscle
sequence
sound signal
time
subsequence
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.)
Active
Application number
CN202010830177.XA
Other languages
Chinese (zh)
Other versions
CN111772669A (en
Inventor
高理富
陆伟
李泽彬
余田田
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.)
Hefei Institutes of Physical Science of CAS
Original Assignee
Hefei Institutes of Physical Science of CAS
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 Hefei Institutes of Physical Science of CAS filed Critical Hefei Institutes of Physical Science of CAS
Priority to CN202010830177.XA priority Critical patent/CN111772669B/en
Publication of CN111772669A publication Critical patent/CN111772669A/en
Application granted granted Critical
Publication of CN111772669B publication Critical patent/CN111772669B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B7/00—Instruments for auscultation
    • A61B7/02—Stethoscopes
    • A61B7/04—Electric stethoscopes
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B7/00—Instruments for auscultation
    • A61B7/006—Detecting skeletal, cartilage or muscle noise
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
    • G10L21/02—Speech enhancement, e.g. noise reduction or echo cancellation
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/27—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique
    • G10L25/30—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique using neural networks
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/48—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 specially adapted for particular use

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Theoretical Computer Science (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Artificial Intelligence (AREA)
  • Acoustics & Sound (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • Biophysics (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Human Computer Interaction (AREA)
  • Signal Processing (AREA)
  • Multimedia (AREA)
  • Medical Informatics (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Surgery (AREA)
  • Animal Behavior & Ethology (AREA)
  • Public Health (AREA)
  • Veterinary Medicine (AREA)
  • Rheumatology (AREA)
  • Quality & Reliability (AREA)
  • Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)

Abstract

The invention relates to an elbow joint contraction muscle force estimation method based on a self-adaptive long-time memory network, which comprises the following steps: acquiring an elbow joint original muscle tone signal; obtaining a muscle sound signal nonlinear sequence and a muscle sound signal nonstationary sequence; forming a muscle sound signal nonlinear subsequence set; calculating the average absolute value and the root mean square of each muscle sound signal nonlinear subsequence to form an average absolute value characteristic sequence and a root mean square characteristic sequence; constructing and training a long-time memory network model; inputting the test set into a training-finished long-time memory network model to finish the estimation of the muscle force of the elbow joint, and comparing the accuracy of the estimation result. According to the invention, the muscle sound signal of the elbow joint muscle is taken as a research object, so that the defects of low immunity, complex acquisition process, easy damage and the like of other biological signals such as myoelectric signals, electroencephalogram signals and the like are effectively overcome, and compared with the traditional signal denoising algorithm, the method has the advantages of simple calculation, good effect and the like.

Description

Elbow joint contraction muscle force estimation method based on adaptive long-time and short-time memory network
Technical Field
The invention relates to the technical field of human body biological signal processing, feature extraction and muscle strength estimation, in particular to an elbow joint contraction muscle strength estimation method based on a self-adaptive long-time memory network.
Background
Skeletal muscle drives human joint to move is the main mode of realizing that nervous system and external world carry out the interaction, and when muscle degeneration or unexpected injury appear, can directly produce negative effects to the motor function of limbs. Therefore, the muscle force of the elbow joint can be timely and accurately estimated, and the muscle force estimation device can be used as an input signal of various rehabilitation control devices, and has very important significance in the aspects of preventing muscle degeneration, improving motor functions and the like. Muscle force estimation also has important research foundation in the human neuromuscular field, rehabilitation development, prosthetic control field and the like.
The estimation of the muscle strength is one of the problems existing in the modern biomedical field all the time, although the current foreign experts and scholars research out that the muscle strength can be measured by a sensor directly embedded in a muscle body, the research shows that the muscle strength measured by the method has huge damage to the human body, the manufacturing cost is high, and the method can only be used in clinical medicine. In recent years, with the rapid development of biosensor technology, people are more and more interested in achieving muscle strength estimation through a method based on human body bioelectric signals, and measurement and estimation of muscle strength are achieved by constructing a mapping relation between biological signal characteristics and muscle strength. For this reason, most researchers have chosen bioelectric signals as subjects to study the magnitude of muscle strength.
At present, the muscle force estimation algorithm of the bioelectric signals adopted by experts and scholars at home and abroad mainly has the defects of low instantaneity, poor accuracy, high time complexity, low generalization performance and the like. How to avoid the problems of the above algorithms and improve the performance of the muscle strength estimation algorithm is a technical problem which needs to be solved urgently in the field of muscle strength estimation and even in the field of human body rehabilitation engineering at present.
Disclosure of Invention
The invention aims to provide the elbow joint contraction muscle force estimation method based on the adaptive long-time and short-time memory network, which has the advantages of high precision, low complexity, good real-time performance and no damage.
In order to achieve the purpose, the invention adopts the following technical scheme: an elbow joint contraction muscle force estimation method based on an adaptive long-time and short-time memory network comprises the following sequential steps:
(1) acquiring an elbow joint original muscle sound signal: acquiring the original muscle tone of human elbow joint muscle from 1 st to L th by a muscle tone signal sensorSignal, denoted as R (l) ═ R 1 ，R 2 ，...，R L L is more than or equal to 5000 and less than or equal to 10000;
(2) performing decoupling pretreatment on an original muscle sound signal to obtain a muscle sound signal nonlinear sequence and a muscle sound signal nonstationary sequence;
(3) according to the length of the steady state time in the muscle sound signal nonlinear sequence, segmenting the muscle sound signal nonlinear sequence through a self-adaptive sliding window algorithm to form a muscle sound signal nonlinear subsequence set;
(4) carrying out feature extraction on the muscle sound signal nonlinear subsequence set, calculating the average absolute value and the root mean square of each muscle sound signal nonlinear subsequence to form an average absolute value feature sequence and a root mean square feature sequence, and respectively recording as:
LR MAV (X)＝{LR MAV (T 1 )，LR MAV (T 2 )，...，LR MAV (T m )}
LR RMS (X)＝{LR RMS (T 1 )，LR RMS (T 2 )，...，LR RMS (T m )}；
(5) constructing a long-time and short-time memory network model, and respectively comparing the average absolute value characteristic sequences LR MAV (X) and the root mean square signature LR RMS (X) inputting the front 3/5 data serving as a training set into a long-time memory network model for training;
(6) the remaining 2/5 sets of mean absolute value feature sequences LR MAV (X) and the root mean square signature LR RMS And (X) inputting the data serving as a test set to a trained long-time memory network model to complete the estimation of the muscle force of the elbow joint, and comparing the accuracy of the estimation result.
The step (2) specifically comprises the following steps:
(2a) adopting a sliding average filtering algorithm to perform decoupling and denoising treatment on the elbow joint original muscle sound signal to obtain a muscle sound signal nonlinear sequence and a muscle sound signal nonstationary sequence, wherein the calculation formula of the sliding average filtering calculation method is as follows:
Figure GDA0003731116760000031
Figure GDA0003731116760000032
wherein R is 1 A value representing the 1 st time instant of the original muscle tone signal; r 2 A value representing the 2 nd time instant of the original muscle tone signal; r is τ A value representing the τ th moment of the original muscle tone signal; LR i A value representing the ith moment of the muscle tone signal nonlinear sequence; NR (nitrogen to noise ratio) i A value representing the ith moment of the non-stationary sequence of the muscle sound signal;
(2b) the muscle sound signal nonlinear sequence and the muscle sound signal non-stationary sequence which are L in length and obtained after the processing of the sliding average filtering algorithm are respectively recorded as:
LR(L)＝{LR 1 ，LR 2 ，...，LR L }
NR(L)＝{NR 1 ，NR 2 ，...，NR L }
wherein lr (l) represents a myotone signal nonlinear sequence; LR 1 A value representing the moment 1 of the nonlinear sequence of the muscle sound signal; LR 2 A value representing a non-linear sequence of muscle tone signals at time 2; LR L A value representing the Lth time of the nonlinear sequence of the muscle sound signal; nr (l) represents a myotone signal non-stationary sequence; NR (nitrogen to noise ratio) 1 A value representing the 1 st time of a non-stationary sequence of the muscle sound signal; NR (nitrogen to noise ratio) 2 A value representing the 2 nd moment of the non-stationary sequence of the muscle sound signal; NR (nitrogen to noise ratio) L A value representing the lth time instant of the non-stationary sequence of the muscle tone signal.
The step (3) specifically comprises the following steps:
(3a) segmenting a muscle sound signal nonlinear sequence LR (L) by adopting an adaptive sliding window algorithm, wherein the adaptive sliding window algorithm is as follows:
Figure GDA0003731116760000033
wherein n1 represents the starting point of the sub-sequence after being cut;n2 represents the subsequence end point after being sliced; LR (n1) A value representing the starting point of the segmented nonlinear subsequence; LR (n2) A value representing the cut nonlinear subsequence end point; gamma represents the difference between the signal values of the starting point and the ending point of the non-linear subsequence of the muscle sound signal; t represents the size of a window required for segmenting the muscle sound signal nonlinear sequence; alpha and beta are proportionality coefficients;
(3b) the muscle sound signal nonlinear sequence LR (L) is segmented by a self-adaptive sliding window algorithm to obtain a muscle sound signal nonlinear subsequence set DS (M):
DS(M)＝{LR(T 1 )，LR(T 2 )，...，LR(T M )}
wherein, LR (T) 1 ) Representing a non-linear subsequence of the muscle tone signal sliced through a 1 st sliding window; LR (T) 2 ) Representing a non-linear subsequence of the muscle tone signal sliced through the 2 nd sliding window; LR (T) M ) Representing a non-linear subsequence of the M sliding window sliced muscle sound signal; m represents the number of the non-linear subsequences which are segmented into the muscle sound signals, and M is less than or equal to L.
The step (4) specifically comprises the following steps:
(4a) calculating a root mean square and an average absolute value of the set of non-linear subsequences of the muscle tone signal, wherein,
the calculation formula of the root mean square and the average absolute value is shown in formulas (4) and (5):
Figure GDA0003731116760000041
Figure GDA0003731116760000042
among them, LR RMS (T β ) Representing the root mean square of the beta muscle tone signal non-linear subsequence; LR β A value representing the beta time in the non-linear subsequence of the muscle sound signal; LR MAV (T β ) Representing the mean absolute value of the beta-th muscle sound signal non-linear subsequence; epsilon is a counting variable; p represents a window corresponding to a non-linear subsequenceMouth size; m represents the maximum value of the number of signals in the non-linear sub sequence of the muscle sound signal;
(4b) calculating the root mean square and the average absolute value of m muscle sound signal nonlinear subsequence sets in turn according to formulas (4) and (5), forming a root mean square characteristic sequence and an average absolute value characteristic sequence, and recording as:
LR RMS (M)＝{LR RMS (T 1 )，LR RMS (T 2 )，...，LR RMS (T m )} (6)
LR MAV (M)＝{LR MAV (T 1 )，LR MAV (T 2 )，...，LR MAV (T m )} (7)
among them, LR RMS (M) represents a root-mean-square signature sequence consisting of M non-linear subsequences; LR RMS (T 1 ) Represents the root mean square of the 1 st nonlinear subsequence; LR RMS (T 2 ) Represents the root mean square of the 2 nd non-linear subsequence; LR RMS (T m ) Represents the root mean square of the mth nonlinear subsequence; LR MAV (M) represents a mean absolute value signature sequence consisting of the mean absolute values of M non-linear subsequences; LR MAV (T 1 ) Represents the average absolute value of the 1 st non-linear subsequence; LR MAV (T 2 ) Represents the average absolute value of the 2 nd non-linear subsequence; LR MAV (T m ) Representing the mean absolute value of the mth non-linear subsequence.
The step (5) specifically comprises the following steps:
(5a) intercepting root mean square signature sequence LR RMS (M) and the mean absolute value signature sequence LR MAV (M) the first 3/5 sets of data as a training set for a long-and-short memory network;
(5b) inputting data into a long-time and short-time memory network for training according to the idea of a fixed sliding window model, wherein the size of a fixed window is W, the value range is 1 & ltW & lt 5, and the sliding step length is 1;
(5c) the parameters of each layer of the structure of the long-time memory network model are determined as follows:
an input layer: inputting the one-dimensional time sequence, the last moment state quantity and the last moment output quantity to an input layer together, wherein the truncation length of the LSTM neural network is 10;
hiding the layer: building a neural network by using LSTM cells, wherein the number of nodes of a hidden layer in the LSTM is 30;
an output layer: obtaining an output result of the last moment after training, wherein the result is a predicted value of a time sequence of the next moment, and transmitting the state value and the predicted value to a prediction model of the next moment;
the activation function is a ReLU function of the form:
Figure GDA0003731116760000051
wherein z represents an independent variable;
(5d) and (4) segmenting the sequence by a self-adaptive sliding window algorithm, and inputting the subsequence into a self-adaptive long-time and short-time memory network for training.
The step (6) specifically comprises the following steps:
(6a) using the remaining 2/5 groups of data of the root-mean-square characteristic sequence and the average absolute value characteristic sequence as a test set, segmenting according to a fixed sliding window model, inputting the segmented data into a trained adaptive long-time memory network for testing, and obtaining a muscle strength estimation value when the elbow joint contracts;
(6b) the output force is measured by the external handheld force measuring instrument and compared, and the accuracy of the estimation result is verified.
According to the technical scheme, the invention has the beneficial effects that: firstly, the muscle sound signal of the muscle of the elbow joint is taken as a research object, so that the defects of low immunity, complex acquisition process, easy damage and the like of other biological signals such as the muscle sound signal, the brain electrical signal and the like are effectively overcome; secondly, the muscle sound signal decoupling and denoising algorithm provided by the invention realizes the influence of the high-frequency component doped in the muscle sound signal on the estimation result, and has the advantages of simple calculation, good effect and the like compared with the traditional signal denoising algorithm. Thirdly, the adaptive long-time and short-time memory network algorithm greatly reduces the network parameter setting process, compared with the traditional artificial neural network technology, the method has the characteristics of local perception and parameter sharing, greatly reduces the complexity of the model, and reduces the number of weights. Meanwhile, the network has the function of feature extraction, so that corresponding features can be learned from samples effectively, and a complex feature extraction process is avoided. Fourthly, compared with an artificial neural network prediction algorithm, an ARIMA prediction algorithm and an SVR prediction algorithm, the elbow joint contraction muscle strength prediction method based on the long-term and short-term memory network can timely, quickly and accurately achieve nondestructive estimation of the elbow joint muscle strength, and meanwhile, the algorithm has the advantages of short operation time, high estimation precision, good robustness and the like.
Drawings
FIG. 1 is a flow chart of a method of the present invention;
FIG. 2 is a schematic diagram of a structure of an adaptive long-term and short-term memory network according to the present invention;
FIG. 3 is a schematic diagram of the experimental process of muscle strength estimation in the present invention;
FIG. 4 is a diagram of the original muscle tone signal of the present invention;
FIG. 5 is a non-linear sequence chart of the muscle tone signal according to the present invention;
FIG. 6 is a non-stationary sequence diagram of a muscle tone signal according to the present invention;
fig. 7 is a comparison diagram of the muscle strength estimation algorithm of the present invention.
Detailed Description
As shown in fig. 1, an elbow joint contraction muscle force estimation method based on an adaptive long-and-short term memory network includes the following steps:
(1) acquiring an elbow joint original muscle tone signal: collecting original muscle sound signals of human elbow joint muscle from 1 st to L th time by a muscle sound signal sensor, and recording as R (L) ═ R 1 ,R 2 ,…,R L Wherein 5000 < L < 10000, where L is 8000;
(2) performing decoupling pretreatment on an original muscle sound signal to obtain a muscle sound signal nonlinear sequence and a muscle sound signal nonstationary sequence; the effective separation of high-frequency signals and low-frequency signals in the original signals can be realized through decoupling preprocessing, and the interference of the high-frequency signals on the estimation result is avoided;
(3) according to the length of the steady-state time in the muscle sound signal nonlinear sequence, segmenting the muscle sound signal nonlinear sequence through a self-adaptive sliding window algorithm to form a muscle sound signal nonlinear subsequence set; the signal is segmented through a self-adaptive sliding window algorithm, the correlation among adjacent data in the signal is fully mined, the characteristic information of the muscle sound signal is extracted to the maximum extent, and a good foundation is laid for the effectiveness of the signal input into a long-time and short-time memory model;
(4) carrying out feature extraction on the muscle sound signal nonlinear subsequence set, calculating the average absolute value and the root mean square of each muscle sound signal nonlinear subsequence to form an average absolute value feature sequence and a root mean square feature sequence, and respectively recording as:
LR MAV (X)＝{LR MAV (T 1 )，LR MAv (T 2 )，...，LR MAV (T m )}
LR RMS (X)＝{LR RMS (T 1 )，LR RMS (T 2 )，...，LR RMS (T m )}；
in this embodiment, Tm takes a value of 680;
(5) constructing a long-time and short-time memory network model, and respectively comparing the average absolute value characteristic sequences LR MAV (X) and the root mean square signature LR RMS (X) inputting the front 3/5 data serving as a training set into a long-time memory network model for training;
(6) the remaining 2/5 sets of mean absolute value signature sequences LR MAV (X) and the root mean square signature LR RMS And (X) inputting the data serving as a test set into a trained long-time memory network model to complete the muscle force estimation of the elbow joint, and comparing the accuracy of the estimation result.
The step (2) specifically comprises the following steps:
(2a) adopting a sliding average filtering algorithm to perform decoupling and denoising treatment on the elbow joint original muscle sound signal to obtain a muscle sound signal nonlinear sequence and a muscle sound signal nonstationary sequence, wherein the calculation formula of the sliding average filtering calculation method is as follows:
Figure GDA0003731116760000071
Figure GDA0003731116760000072
wherein R is 1 A value representing the original muscle tone signal at time 1; r 2 A value representing the 2 nd time instant of the original muscle sound signal; r τ A value representing the τ th moment of the original muscle tone signal; LR i A value representing the ith time of the nonlinear sequence of the muscle sound signal; NR (nitrogen to noise ratio) i A value representing the ith moment of the non-stationary sequence of the muscle sound signal;
(2b) the length of the obtained muscle sound signal nonlinear sequence and the length of the obtained muscle sound signal nonstationary sequence after the processing of the sliding average filtering algorithm are 8000, and the sequences are respectively recorded as:
LR(L)＝{LR 1 ，LR 2 ，...，LR L }
NR(L)＝{NR 1 ，NR 2 ，...，NR L }
wherein lr (l) represents a muscle tone signal nonlinear sequence; LR 1 A value representing the 1 st time of the non-linear sequence of the muscle sound signal; LR 2 A value representing a time instant 2 of the non-linear sequence of the muscle sound signal; LR L A value representing the Lth time of the nonlinear sequence of the muscle sound signal; nr (l) represents a myotone signal non-stationary sequence; NR 1 A value representing the 1 st time of a non-stationary sequence of the muscle sound signal; NR (nitrogen to noise ratio) 2 A value representing the non-stationary sequence 2 time of the muscle sound signal; NR L A value representing the lth time instant of the non-stationary sequence of the muscle tone signal.
The step (3) specifically comprises the following steps:
(3a) segmenting a muscle sound signal nonlinear sequence LR (L) by adopting an adaptive sliding window algorithm, wherein the adaptive sliding window algorithm is as follows:
Figure GDA0003731116760000081
wherein n1 represents the starting point of the sub-sequence after being cut; n2 represents the subsequence end point after being sliced; LR (n1) A value representing the starting point of the segmented nonlinear subsequence; LR (n2) A value representing the cut nonlinear subsequence end point; gamma represents the difference between the signal values of the starting point and the ending point of the non-linear subsequence of the muscle sound signal; t represents the size of a window required for segmenting the muscle sound signal nonlinear sequence; alpha and beta are proportionality coefficients;
(3b) the muscle sound signal nonlinear sequence LR (L) is segmented by a self-adaptive sliding window algorithm to obtain a muscle sound signal nonlinear subsequence set DS (M):
DS(M)＝{LR(T 1 )，LR(T 2 )，...，LR(T M )}
wherein, LR (T) 1 ) Representing a non-linear subsequence of the muscle tone signal sliced through a 1 st sliding window; LR (T) 2 ) Representing a non-linear subsequence of the muscle tone signal sliced through the 2 nd sliding window; LR (T) M ) Representing a non-linear subsequence of the M sliding window sliced muscle sound signal; m represents the number of the non-linear subsequences which are segmented into the muscle tone signals, wherein M is less than or equal to L, and the value of M is 680 in the embodiment;
the step (4) specifically comprises the following steps:
(4a) calculating a root mean square and an average absolute value of the set of non-linear subsequences of the muscle tone signal, wherein,
the calculation formula of the root mean square and the average absolute value is shown in formulas (4) and (5):
Figure GDA0003731116760000091
Figure GDA0003731116760000092
among them, LR RMS (T β ) A root mean square representing the nonlinear subsequence of the beta-th myotone signal; LR β A value representing the beta time in the non-linear subsequence of the muscle sound signal;LR MAV (T β ) Representing the mean absolute value of the beta-th muscle sound signal non-linear subsequence; epsilon is a counting variable; p represents the window size corresponding to the nonlinear subsequence; m represents the maximum value of the number of signals in the non-linear sub sequence of the muscle sound signal;
(4b) calculating the root mean square and the average absolute value of the m muscle sound signal nonlinear subsequence sets according to formulas (4) and (5) to form a root mean square characteristic sequence and an average absolute value characteristic sequence, and recording as:
LR RMS (M)＝{LR RMS (T 1 )，LR RMS (T 2 )，...，LR RMS (T m )} (6)
LR MAV (M)＝{LR MAV (T 1 )，LR MAV (T 2 )，...，LR MAV (T m )} (7)
among them, LR RMS (M) represents a root mean square signature sequence consisting of the root mean square of M non-linear subsequences; LR RMS (T 1 ) Represents the root mean square of the 1 st non-linear subsequence; LR RMS (T 2 ) Represents the root mean square of the 2 nd non-linear subsequence; LR RMS (T m ) Represents the root mean square of the mth nonlinear subsequence; LR MAV (M) represents a mean absolute value signature sequence consisting of the mean absolute values of M non-linear subsequences; LR MAV (T 1 ) Represents the average absolute value of the 1 st non-linear subsequence; LR MAV (T 2 ) Represents the average absolute value of the 2 nd nonlinear subsequence; LR MAV (T m ) Representing the mean absolute value of the mth non-linear subsequence.
The step (5) specifically comprises the following steps:
(5a) truncating root mean square characteristic sequence LR RMS (M) and the mean absolute value signature sequence LR MAV (M) the first 3/5 sets of data as a training set for a long-and-short memory network;
(5b) inputting data into a long-time and short-time memory network for training according to the idea of a fixed sliding window model, wherein the size of a fixed window is W, the value range is more than 1 and less than 5, W is 4 in the embodiment, and the sliding step length is 1;
(5c) the parameters of each layer of the structure of the long-time memory network model are determined as follows:
an input layer: inputting the one-dimensional time sequence, the last moment state quantity and the last moment output quantity to an input layer together, wherein the truncation length of the LSTM neural network is 10;
hiding the layer: building a neural network by using LSTM cells, wherein the number of nodes of a hidden layer in the LSTM is 30;
an output layer: obtaining an output result of the last moment after training, wherein the result is a predicted value of a time sequence of the next moment, and transmitting the state value and the predicted value to a prediction model of the next moment;
the activation function is a ReLU function of the form:
Figure GDA0003731116760000101
wherein z represents an independent variable;
(5d) and (4) segmenting the sequence by a self-adaptive sliding window algorithm, and inputting the subsequence into a self-adaptive long-time and short-time memory network for training.
The step (6) specifically comprises the following steps:
(6a) using the remaining 2/5 groups of data of the root-mean-square characteristic sequence and the average absolute value characteristic sequence as a test set, segmenting according to a fixed sliding window model, inputting the segmented data into a trained adaptive long-time memory network for testing, and obtaining a muscle strength estimation value when the elbow joint contracts;
(6b) the output force is measured by the external handheld force measuring instrument and compared, and the accuracy of the estimation result is verified.
Fig. 2 is a structural diagram of a long-term and short-term memory network muscle strength estimation model unit of the present invention, which can effectively determine which input information is forgotten and which information is retained. Forget gate determines last time unit state C t-1 How much to keep the current time C t (ii) a The input gate decides for the current input x t The cell state of (1).
Fig. 3 is a diagram of the muscle strength estimation experiment process of the present invention, which illustrates the muscle strength estimation process of the present invention in detail, by fixing the elbow of the experimenter on the support, sending the collected original muscle sound signal to the computer for the signal preprocessing, feature extraction and muscle strength estimation processes, outputting the estimated value of the muscle strength, and displaying the result on the display. Meanwhile, the output force of the tail end of the joint is measured by adopting external force sensor equipment and is compared with the estimation result, and the accuracy of the estimation result is verified in a cross mode.
FIG. 4 is a time series of 2000 original muscle tone signals collected according to the present invention;
FIG. 5 is a nonlinear sequence of muscle tone signals obtained after processing by a moving average pre-processing algorithm;
FIG. 6 is a non-stationary sequence of the muscle tone signal obtained after the processing by the sliding average pre-processing algorithm;
fig. 7 is a comparison graph of the muscle strength estimation result of the present invention, which is compared with the ARIMA estimation result, the Kalman-filter estimation result, and the wavelet neural network estimation result, respectively, and it can be seen that the muscle strength estimation result of the estimation method of the present invention has the highest accuracy, the wavelet neural network estimation result has the second lowest accuracy, the ARIMA estimation result is the worst, and the model is easy to diverge.
In conclusion, the muscle sound signals of the elbow joint muscles are taken as research objects, so that the defects of low immunity, complex acquisition process, easiness in damage and the like of other biological signals such as myoelectric signals, electroencephalogram signals and the like are effectively overcome. The muscle sound signal decoupling and denoising algorithm provided by the invention realizes the influence of high-frequency components doped in the muscle sound signal on the estimation result, and has the advantages of simple calculation, good effect and the like compared with the traditional signal denoising algorithm. The self-adaptive long-time and short-time memory network algorithm greatly reduces the network parameter setting process, has the characteristics of local perception and parameter sharing compared with the traditional artificial neural network technology, greatly reduces the complexity of the model, and reduces the number of weights. Meanwhile, the network has the function of feature extraction, so that corresponding features can be learned from the sample effectively, and a complex feature extraction process is avoided.

Claims (1)

1. An elbow joint contraction muscle force estimation method based on a self-adaptive long-time memory network is characterized by comprising the following steps: the method comprises the following steps in sequence:
(1) acquiring an elbow joint original muscle tone signal: collecting original muscle sound signals of human elbow joint muscles from 1 st to L th time through a muscle sound signal sensor, and recording as R (L) { R } 1 ，R 2 ，...，R L L is more than or equal to 5000 and less than or equal to 10000;
(2) performing decoupling pretreatment on an original muscle sound signal to obtain a muscle sound signal nonlinear sequence and a muscle sound signal non-stationary sequence;
(3) according to the length of the steady-state time in the muscle sound signal nonlinear sequence, segmenting the muscle sound signal nonlinear sequence through a self-adaptive sliding window algorithm to form a muscle sound signal nonlinear subsequence set;
(4) carrying out feature extraction on the muscle sound signal nonlinear subsequence set, calculating the average absolute value and the root mean square of each muscle sound signal nonlinear subsequence to form an average absolute value feature sequence and a root mean square feature sequence, and respectively recording as:
LR MAV (X)＝{LR MAV (T 1 )，LR MAV (T 2 )，...，LR MAV (T m )}
LR RMS (X)＝{LR RMS (T 1 )，LR RMS (T 2 )，...，LR RMS (T m )}；
(5) constructing a long-time and short-time memory network model, and respectively comparing the average absolute value characteristic sequences LR MAV (X) and the root mean square signature LR RMS (X) inputting the front 3/5 data serving as a training set into a long-time and short-time memory network model for training;
(6) the remaining 2/5 sets of mean absolute value feature sequences LR MAV (X) and the root mean square signature LR RMS (X) inputting the data serving as a test set into a trained long-time and short-time memory network model to complete the muscle force estimation of the elbow joint, and comparing the accuracy of an estimation result;
the step (2) specifically comprises the following steps:
(2a) decoupling and denoising an elbow joint original muscle sound signal by adopting a sliding average filtering algorithm to obtain a muscle sound signal nonlinear sequence and a muscle sound signal nonstationary sequence;
(2b) and (3) processing by a sliding average filtering algorithm to obtain a muscle sound signal nonlinear sequence and a muscle sound signal nonstationary sequence with the length of L, and respectively recording as:
LR(L)＝{LR 1 ，LR 2 ，...，LR L }
NR(L)＝{NR 1 ，NR 2 ，...，NR L }
wherein lr (l) represents a muscle tone signal nonlinear sequence; LR 1 A value representing the 1 st time of the non-linear sequence of the muscle sound signal; LR 2 A value representing a time instant 2 of the non-linear sequence of the muscle sound signal; LR L A value representing the Lth time of the nonlinear sequence of the muscle sound signal; nr (l) represents a myotone signal non-stationary sequence; NR (nitrogen to noise ratio) 1 A value representing the 1 st time of a non-stationary sequence of the muscle sound signal; NR 2 A value representing the 2 nd moment of the non-stationary sequence of the muscle sound signal; NR (nitrogen to noise ratio) L A value representing the Lth moment of the non-stationary sequence of the muscle sound signal;
the step (3) specifically comprises the following steps:
(3a) segmenting a muscle sound signal nonlinear sequence LR (L) by adopting an adaptive sliding window algorithm;
(3b) the muscle sound signal nonlinear sequence LR (L) is segmented by a self-adaptive sliding window algorithm to obtain a muscle sound signal nonlinear subsequence set DS (M):
DS(M)＝{LR(T 1 )，LR(T 2 )，...，LR(T M )}
wherein, LR (T) 1 ) Representing a non-linear subsequence of the muscle tone signal sliced through a 1 st sliding window; LR (T) 2 ) Representing a non-linear subsequence of the muscle tone signal sliced through the 2 nd sliding window; LR (T) M ) Representing a non-linear subsequence of the M sliding window sliced muscle sound signal; m represents the number of the nonlinear subsequences which are cut into the muscle sound signals, and M is less than or equal to L;
the step (4) specifically comprises the following steps:
(4a) calculating a root mean square and an average absolute value of the set of non-linear subsequences of the muscle tone signal, wherein,
the calculation formula of the root mean square and the average absolute value is shown in formulas (4) and (5):
Figure FDA0003714456110000021
Figure FDA0003714456110000022
among them, LR RMS (T β ) Representing the root mean square of the beta muscle tone signal non-linear subsequence; LR β A value representing the beta time in the non-linear subsequence of the muscle sound signal; LR MAV (T β ) Representing the mean absolute value of the beta-th muscle sound signal non-linear subsequence; epsilon is a counting variable; ρ represents the window size corresponding to the non-linear subsequence; m represents the maximum value of the number of signals in the non-linear sub sequence of the muscle sound signal;
(4b) calculating the root mean square and the average absolute value of the m muscle sound signal nonlinear subsequence sets according to formulas (4) and (5) to form a root mean square characteristic sequence and an average absolute value characteristic sequence, and recording as:
LR RMS (M)＝{LR RMS (T 1 )，LR RMS (T 2 )，...，LR RMS (T m )} (6)
LR MAV (M)＝{LR MAV (T 1 )，LR MAV (T 2 )，...，LR MAV (T m )} (7)
among them, LR RMS (M) represents a root mean square signature sequence consisting of the root mean square of M non-linear subsequences; LR RMS (T 1 ) Represents the root mean square of the 1 st non-linear subsequence; LR RMS (T 2 ) Represents the root mean square of the 2 nd non-linear subsequence; LR RMS (T m ) Represents the root mean square of the mth non-linear subsequence; LR MAV (M) represents a flat formed by the mean absolute values of M non-linear subsequencesAbsolute value average characteristic sequence; LR MAV (T 1 ) Represents the average absolute value of the 1 st non-linear subsequence; LR MAV (T 2 ) Represents the average absolute value of the 2 nd nonlinear subsequence; LR MAV (T m ) Represents the average absolute value of the mth non-linear subsequence;
the step (5) specifically comprises the following steps:
(5a) intercepting root mean square signature sequence LR RMS (M) and the mean absolute value signature sequence LR MAV (M) the first 3/5 sets of data as a training set for a long-and-short term memory network;
(5b) inputting data into a long-time and short-time memory network for training according to the idea of a fixed sliding window model, wherein the size of a fixed window is W, the value range is 1 & ltW & lt 5, and the sliding step length is 1;
(5c) the method comprises the following steps of determining parameters of each layer of a long-time and short-time memory network model structure as follows:
an input layer: inputting the one-dimensional time sequence, the last moment state quantity and the last moment output quantity to an input layer together, wherein the truncation length of the LSTM neural network is 10;
hiding the layer: building a neural network by using LSTM cells, wherein the number of nodes of a hidden layer in the LSTM is 30;
an output layer: obtaining an output result of the last moment after training, wherein the result is a predicted value of a time sequence of the next moment, and transmitting the state value and the predicted value to a prediction model of the next moment;
the activation function is a ReLU function of the form:
Figure FDA0003714456110000041
wherein z represents an independent variable;
(5d) after the sequence is segmented by a self-adaptive sliding window algorithm, a subsequence is input to a self-adaptive long-time and short-time memory network for training;
the step (6) specifically comprises the following steps:
(6a) using the remaining 2/5 groups of data of the root-mean-square characteristic sequence and the average absolute value characteristic sequence as a test set, segmenting according to a fixed sliding window model, inputting the segmented data into a trained adaptive long-time memory network for testing, and obtaining a muscle strength estimation value when the elbow joint contracts;
(6b) the output force is measured by the external hand-held force measuring instrument and compared, and the accuracy of the estimation result is verified.
CN202010830177.XA 2020-08-18 2020-08-18 Elbow joint contraction muscle force estimation method based on adaptive long-time and short-time memory network Active CN111772669B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010830177.XA CN111772669B (en) 2020-08-18 2020-08-18 Elbow joint contraction muscle force estimation method based on adaptive long-time and short-time memory network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010830177.XA CN111772669B (en) 2020-08-18 2020-08-18 Elbow joint contraction muscle force estimation method based on adaptive long-time and short-time memory network

Publications (2)

Publication Number Publication Date
CN111772669A CN111772669A (en) 2020-10-16
CN111772669B true CN111772669B (en) 2022-08-19

Family

ID=72762871

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010830177.XA Active CN111772669B (en) 2020-08-18 2020-08-18 Elbow joint contraction muscle force estimation method based on adaptive long-time and short-time memory network

Country Status (1)

Country Link
CN (1) CN111772669B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114343648B (en) * 2022-01-07 2024-06-04 中山大学附属第一医院 Muscle strength assessment method, system and computer readable storage medium
CN114931390B (en) * 2022-05-06 2023-06-13 电子科技大学 Muscle strength estimation method based on fatigue analysis
CN115238753B (en) * 2022-09-21 2022-12-06 西南交通大学 Self-adaptive SHM data cleaning method based on local outlier factor

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107961007A (en) * 2018-01-05 2018-04-27 重庆邮电大学 A kind of electroencephalogramrecognition recognition method of combination convolutional neural networks and long memory network in short-term
CN108478215A (en) * 2018-01-25 2018-09-04 深圳市德力凯医疗设备股份有限公司 EEG Noise Cancellation, storage medium based on wavelet analysis and device
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
CN110427924A (en) * 2019-09-09 2019-11-08 桂林电子科技大学 A kind of heart impact signal based on LSTM more classifying identification methods automatically
CN111110268A (en) * 2019-11-26 2020-05-08 中国科学院合肥物质科学研究院 Human body muscle sound signal prediction method based on random vector function connection network technology

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8376968B2 (en) * 2009-05-15 2013-02-19 The Hong Kong Polytechnic University Method and system for quantifying an intention of movement of a user
US10870002B2 (en) * 2018-10-12 2020-12-22 DePuy Synthes Products, Inc. Neuromuscular sensing device with multi-sensor array

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107961007A (en) * 2018-01-05 2018-04-27 重庆邮电大学 A kind of electroencephalogramrecognition recognition method of combination convolutional neural networks and long memory network in short-term
CN108478215A (en) * 2018-01-25 2018-09-04 深圳市德力凯医疗设备股份有限公司 EEG Noise Cancellation, storage medium based on wavelet analysis and device
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
CN110427924A (en) * 2019-09-09 2019-11-08 桂林电子科技大学 A kind of heart impact signal based on LSTM more classifying identification methods automatically
CN111110268A (en) * 2019-11-26 2020-05-08 中国科学院合肥物质科学研究院 Human body muscle sound signal prediction method based on random vector function connection network technology

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
"Feasibility of using an artificial neural network model to estimate the elbow flexion force from mechanomyography";Wonkeun Youn et al.;《Journal of Neuroscience Methods》;20110115;第194卷(第2期);图1，图7，第2.1节第1-2段至第4节第4段 *

Also Published As

Publication number Publication date
CN111772669A (en) 2020-10-16

Similar Documents

Publication Publication Date Title
CN111772669B (en) Elbow joint contraction muscle force estimation method based on adaptive long-time and short-time memory network
CN108304917B (en) P300 signal detection method based on LSTM network
CN113288183B (en) Silent voice recognition method based on facial neck surface myoelectricity
CN112418013A (en) Complex working condition bearing fault diagnosis method based on meta-learning under small sample
CN102499797B (en) Artificial limb control method and system
CN111553307A (en) Gesture recognition system fusing bioelectrical impedance information and myoelectric information
CN114469124B (en) Method for identifying abnormal electrocardiosignals in movement process
CN108509823A (en) The detection method and device of QRS complex
CN115346561B (en) Depression emotion assessment and prediction method and system based on voice characteristics
CN110710984B (en) Ankle moment prediction method of recursion cerebellum model based on surface electromyogram signal
CN112998676A (en) Continuous blood pressure measuring method based on photoelectric array enhanced signal multi-feature extraction
CN112244851A (en) Muscle movement recognition method and surface electromyogram signal acquisition device
CN112307996B (en) Fingertip electrocardio identity recognition device and method
CN111110268B (en) Human body muscle sound signal prediction method based on random vector function connection network technology
CN113974607A (en) Sleep snore detecting system based on impulse neural network
CN111242100B (en) Action identification method based on GST and VL-MOBPNN
Ison et al. Beyond user-specificity for emg decoding using multiresolution muscle synergy analysis
CN111554319B (en) Multichannel heart-lung sound abnormality recognition system and device based on low-rank tensor learning
CN111481193B (en) Fall risk assessment and early warning method and system
CN116172531B (en) Blood pressure change trend estimation method based on wavelet dispersion transformation
CN112735480A (en) Vocal cord pathological change detection device based on neural network
CN116982952A (en) Noninvasive blood pressure measurement method and system based on fractional derivative
CN116956063A (en) Gesture recognition method and system based on multi-mode optimal data selection and enhancement
Chen et al. Recognition of american sign language gestures based on electromyogram (emg) signal with xgboost machine learning
CN115374815A (en) Automatic sleep staging method based on visual Transformer

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
GR01 Patent grant
GR01 Patent grant