CN109875565A - A kind of cerebral apoplexy upper extremity exercise function method for automatically evaluating based on deep learning - Google Patents
A kind of cerebral apoplexy upper extremity exercise function method for automatically evaluating based on deep learning Download PDFInfo
- Publication number
- CN109875565A CN109875565A CN201910074666.4A CN201910074666A CN109875565A CN 109875565 A CN109875565 A CN 109875565A CN 201910074666 A CN201910074666 A CN 201910074666A CN 109875565 A CN109875565 A CN 109875565A
- Authority
- CN
- China
- Prior art keywords
- data
- signal
- imu
- patient
- upper extremity
- 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
Links
Abstract
The invention discloses a kind of cerebral apoplexy upper extremity exercise function method for automatically evaluating based on deep learning, comprising the following steps: inertia sensing data and myoelectricity data based on wearable sensing system acquisition patient's upper extremity exercise process；Length normalization method and numerical value normalization pretreatment are carried out to the data of acquisition；Inertia sensing data and myoelectricity data are separately input to two convolutional neural networks and carry out feature extraction, all features are merged, the motor function grade based on Brunnstrom scale is generated, back-propagation algorithm is then based on and model parameter is iterated, training deep learning network model；To the patient for needing to carry out upper extremity exercise assessment of function, data acquisition, pretreatment are carried out to it, and is input to trained deep learning model, automatically generate patient's upper extremity exercise function Brunnstrom evaluation result by stages.Present invention can apply to the precision of automatic evaluation can be improved in hospital environment, community and home environment.
Description
Technical field
The invention belongs to the applied technical fields of wearable device, specifically, being related to a kind of brain based on deep learning
Stroke upper extremity exercise function method for automatically evaluating.
Background technique
Cerebral apoplexy is also known as apoplexy or cerebrovascular accident, is one of main fatal disease in the world.China 40 years old or more people
For mass-brain apoplexy patient up to 12,420,000, annual new hair patient is more than 2,700,000, there is movement in various degree in survival crowd more than 70%
Dysfunction.Clinical research shows that rehabilitation is to reduce patients with cerebral apoplexy disability rate most efficient method, and rehabilitation evaluation is
The important link of rehabilitation understands patient body dysfunction property and severity by evaluation, to formulate rehabilitation goal
Important evidence is provided with therapeutic scheme.
Rehabilitation evaluation is based primarily upon scale score, and most common cerebral apoplexy motor functional evaluation scale is Brunnstrom,
As a result it is divided into I-VI totally 6 grades.One is traditional scale assessment method, patient completes specified comment under evaluation doctor's guidance
Fixed movement, evaluation doctor observe patient's execution process, and patient motion ability is classified or is scored according to scale；It is another
Class is the cerebral apoplexy motor functional evaluation based on machine learning, and patient executes in evaluation action process, is acquired using sensor anti-
The data of patient motion ability are reflected, the time domain and frequency domain character of data are then manually extracted, the feature of extraction is inputted into classifier
(SVM, linear regression, extreme learning machine etc.) constructs automatic evaluation learning model, generates evaluation result corresponding with scale.
Traditional Measuring scale assessing mode needs doctor to participate in whole process, has the following problems: 1) evaluating process, time-consuming, effect
Rate is low；2) evaluation process needs profession evaluation doctor's real-time oversight and guidance；3) evaluation result is big by evaluation teacher's subjective impact, deposits
In subjective differences；4) measuring scale has differences, and there are problems that versatility.
Evaluation based on machine learning is to have ignored signal by manually extracting the time domain of sensing data, frequency domain character
Dependence between difference and signal itself, and because subjective experience causes feature extraction insufficient, influence sorting algorithm
Can, influence the accuracy of evaluation result.
Summary of the invention
In view of this, the present invention is directed to the subjective differences of traditional Measuring scale assessing and the evaluation side based on conventional machines study
The problem of influence of insufficient pair of evaluation result of method feature extraction, provides a kind of cerebral apoplexy upper extremity exercise based on deep learning
Function method for automatically evaluating realizes objectifying for rehabilitation evaluation, and improves evaluation accuracy.
In order to solve the above-mentioned technical problem, the cerebral apoplexy upper extremity exercise function based on deep learning that the invention discloses a kind of
Method for automatically evaluating, comprising the following steps:
S1. inertia sensing data and myoelectricity data based on wearable sensing system acquisition patient's upper extremity exercise process；
S2. length normalization method and numerical value normalization pretreatment are carried out to the data of acquisition；
S3. inertia sensing data and myoelectricity data are separately input to two convolutional neural networks and carry out feature extraction, it is right
All features are merged, and are generated the motor function grade based on Brunnstrom scale, are then based on back-propagation algorithm pair
Model parameter is iterated, training deep learning network model；
S4. to the patient for needing to carry out upper extremity exercise assessment of function, data identical with S1, S2 step are carried out to it and are adopted
Collection, pretreatment, and it is input to trained deep learning model, automatically generate patient's upper extremity exercise function Brunnstrom by stages
Evaluation result.
Optionally, inertia sensing number of the step S1 based on wearable sensing system acquisition patient's upper extremity exercise process
According to myoelectricity data specifically:
S11. the wearable sensor-based system of design acquisition upper extremity exercise performance data, by 2 Inertial Measurement Units (IMU) and
7 channel surface electromyography (sEMG) sensor two parts composition；Inertial Measurement Unit is used to measure three axis of limb motion process
Acceleration, angular speed and magnetometer data realize wireless real-time transmission, sample frequency 20Hz by bluetooth and PC；SEMG acquisition
Motion process skeleton muscle group electrical signals activity, sample frequency 3KHz；
S12. sensor wearing location is designed:
Two Inertial Measurement Units are bundled in respectively in the middle part of Ipsilateral large arm and in the middle part of forearm, and Inertial Measurement Unit X-axis is vertical
In human coronary towards inner, Y-axis perpendicular to it is cross-section downwards, Z axis perpendicular to sagittal plane inwards；7 channel surface myoelectric electrodes point
It is not placed in brachioradialis, the triceps muscle of arm, the bicipital muscle of arm, group, pectoralis major, trapezius muscle at totally 7 before group, deltoid muscle in deltoid muscle；
S13. patient touches the movement of opposite side shoulder with ipsilateral upper limb execution, and sensor acquires the acceleration and myoelectricity number of motion process
According to, and PC is transmitted to by bluetooth.
Optionally, the data to acquisition in the step S2 carry out length normalization method and numerical value normalization pretreatment is specific
Are as follows:
S21. signal length normalize: due to different patients execute touch shoulder movement speed it is different, cause acquisition sensing
Device signal length is different, needs all Length of sample series are unified；Using the method for linear interpolation by all IMU signal lengths
It is normalized to the maximum value L of all sample lengthsimu；All sEMG signal lengths are normalized to by sample using same method
Maximum value Lemg；
S22. binding position otherness is eliminated, it is assumed that IMU signal is X={ xa, xg, xm, wherein xa, xg, xmRespectively plus
Speed, angular speed, magnetometer signals, it is poor to signal bring by following formula elimination Patient height, sensor binding position difference
It is different:
Wherein l indicates that the IMU binds position at a distance from shoulder joint, unit: m；X ' is treated signal；
S23. signal numerical value normalizes, and promotes convergence rate and precision, 0~1 range is mapped the data into, using linear letter
Number method for normalizing, transfer function are as follows:
Wherein max is the maximum value of sample data, and min is the minimum value of sample data；X is original sensor signal number
According to x*It is the result after normalization.
Optionally, inertia sensing data and myoelectricity data are separately input to two convolutional Neural nets in the step S3
Network carries out feature extraction, merges to all features, generates the motor function grade based on Brunnstrom scale, then base
Model parameter is iterated in back-propagation algorithm, training deep learning network model；Specifically:
S31. input data, the patients with cerebral apoplexy for acquiring N different degrees of dyskinesia execute upper limb and evaluate action process
Data as training sample, sample set is denoted as T={ (ximu, xemg, y)1... ..., (ximu, xemg, y)N, wherein ximuWith xemg
Respectively indicate patient IMU sensor signal and surface electromyogram signal matrixing indicate, y indicate the patient by profession evaluation teacher to
Brunnstrom out is by stages as a result, i.e. artificial Measuring scale assessing is as a result, y ∈ { I, II, III, IV, V, VI }, i.e. class categories number C
=6；
S32. convolution sum pond:
A signal sequence TS after pretreatment is expressed as x={ x1, x2..., xL, L ∈ { Limu, Lemg, wherein xtIt is
Value at timestamp t, xt18 dimensional vectors and 7 dimensional vectors are respectively indicated for IMU signal and sEMG signal；Convolutional layer will be in signal
Characteristic carry out depth extraction, use hyperbolic tangent function tanh as the activation primitive of convolutional layer, to obtain non-linear reflect
It penetrates, kth layer convolution algorithm formula is as follows:
Wherein,For k-1 layers of i-th of pond as a result, when first time convolution, as pretreated signal sequence；
Mk-1For kth-level 1 volume product core number, Wk jFor k j-th of convolution kernel of layer, b is corresponding biasing；Every layer of convolution has multiple convolution kernels
Different characteristic for extraction time sequence；
Maximum pondization operation is using the maximum value in the region of pond as its output, is s in step-length thereforeMPNon-overlap area
Between it is upper carry out primary maximum pond, exporting length reduces a sMPThe factor；In l layers of convolution-Chi Huahou, feature vector exists
The output length about L ' ≈ L/s of CNN moduleMP l；
S33. Fusion Features:
After 2 times alternately convolution sum pond, two kinds of signal sequences divide by the abstract characteristics that two CNN are extracted
Other vectorization is expressed as Fimu/emg={ f1, f2..., fL′},L′∈{L′imu, L 'emg}；Each feature vector be original TS more
High-level abstract, length on a timeline is shorter, and two feature vectors are cascaded to form new fusion feature F={ f1,
f2..., fL′imu+L′emgInputted as full articulamentum；
S34. full articulamentum:
Convolution sum pondization carries out abstract extraction and expression to the information in IMU or sEMG signal sequence, using connecting entirely
Layer is connect to merge information two kinds of signals and realize classification；Include a fusion feature receiving layer, intermediate hidden layers and defeated
Layer out, output layer neuron quantity are equal to class categories number C；
S35.Softmax layers:
Softmax is an additional data analysis layer, and the classification results that full articulamentum exports are become probability point
Cloth；Full articulamentum neuron number is equal to class categories number, and corresponding output is o1, o2..., oc, after being handled by Softmax
Output are as follows:
Export final result o 'iIndicate that the corresponding Brunnstrom grade of the input signal is the probability of i；
S36. cost function:
Classification results cross entropy is used to calculate error in classification, carries out backpropagation weight instruction for batch stochastic gradient descent
Practice, then cost function is written as:
Wherein y indicates the true generic of sequence n, Oy(n) probability output of the model on true generic is indicated；
Batch is the batch sample number selected in batch stochastic gradient descent；
S37. backpropagation repetitive exercise:
Based on the error amount of obtained the model output and reality output of step S36, the error declined using batch gradient is anti-
The parameter in network is iterated to propagation algorithm realization, after 2000 iteration, model parameter training is completed, to net
Network structure and parameter is saved.
Compared with prior art, the present invention can be obtained including following technical effect:
1) compared with traditional scale assessment method, the IMU and sEMG that the present invention is acquired using wearable sensor first believe
It number combines and upper extremity exercise function is evaluated, be easy to implement objectifying for evaluation, portable wearable system design is not
It is simply possible to use in hospital environment, moreover it is possible to be applied in community and home environment.
2) feature extraction and fusion are carried out to multiclass sensor signal using deep learning method, conventional machines is avoided to learn
The dependence that method extracts manual features, feature extraction is more abundant, and considers the otherness between unlike signal and mutual
The precision of automatic evaluation can be improved in Dependency Specification.
Certainly, it implements any of the products of the present invention it is not absolutely required to while reaching all the above technical effect.
Detailed description of the invention
The drawings described herein are used to provide a further understanding of the present invention, constitutes a part of the invention, this hair
Bright illustrative embodiments and their description are used to explain the present invention, and are not constituted improper limitations of the present invention.In the accompanying drawings:
Fig. 1 is the cerebral apoplexy motor functional evaluation general frame the present invention is based on deep learning；
Fig. 2 is wearable sensors data collection system framework of the present invention；
Fig. 3 is the automatic evaluation network model the present invention is based on deep learning；
Fig. 4 (a) is the 3-axis acceleration signal schematic representation of forearm IMU of the present invention, and Fig. 4 (b) is the bicipital muscle of arm of the present invention
SEMG signal schematic representation.
Specific embodiment
Carry out the embodiment that the present invention will be described in detail below in conjunction with embodiment, whereby to the present invention how application technology hand
Section solves technical problem and reaches the realization process of technical effect to fully understand and implement.
The invention discloses a kind of cerebral apoplexy upper extremity exercise function method for automatically evaluating based on deep learning, such as Fig. 1 institute
Show, comprising the following steps:
S1. inertia sensing data and myoelectricity data based on wearable sensing system acquisition patient's upper extremity exercise process；
Specifically:
S11. the wearable sensor-based system of design acquisition upper extremity exercise performance data, as shown in Fig. 2, by 2 inertia measurements
Unit (IMU) and 7 channel surface electromyography (sEMG) sensor two parts composition.IMU is for measuring the three of limb motion process
Axle acceleration, angular speed and magnetometer data realize wireless real-time transmission, sample frequency 20Hz by bluetooth and PC；SEMG is adopted
Collect motion process skeleton muscle group electrical signals activity, sample frequency 3KHz；
S12. sensor wearing location is designed:
Two IMU sensors are bundled in respectively in the middle part of Ipsilateral large arm and in the middle part of forearm, and sensors X axis is perpendicular to human coronary
Towards inner, Y-axis perpendicular to it is cross-section downwards, Z axis perpendicular to sagittal plane inwards；
7 channel surface myoelectric electrodes be respectively placed in brachioradialis, the triceps muscle of arm, the bicipital muscle of arm, in deltoid muscle before group, deltoid muscle
Group, pectoralis major, trapezius muscle are at totally 7；
S13. patient touches the movement of opposite side shoulder with ipsilateral upper limb execution, and sensor acquires the acceleration and myoelectricity number of motion process
According to, and PC is transmitted to by bluetooth.
S2. the pretreatments such as length normalization method, numerical value normalization are carried out to the data of acquisition；Specifically:
S21. signal length normalize: due to different patients execute touch shoulder movement speed it is different, cause acquisition sensing
Device signal length is different, needs all Length of sample series are unified.Using the method for linear interpolation by all IMU signal lengths
It is normalized to the maximum value L of all sample lengthsimu；All sEMG signal lengths are normalized to by L using same methodemg。
S22. binding position otherness is eliminated, it is assumed that IMU signal is X={ xa, xg, xm, wherein xa, xg, xmRespectively plus
Speed, angular speed, magnetometer signals, it is poor to signal bring by following formula elimination Patient height, sensor binding position difference
It is different:
Wherein l indicates that the IMU binds position (unit: m) at a distance from shoulder joint；X ' is treated signal.
S23. signal numerical value normalizes, and promotes convergence rate and precision, 0~1 range is mapped the data into, using linear letter
Number method for normalizing, transfer function are as follows:
Wherein max is the maximum value of sample data, and min is the minimum value of sample data.X is original sensor signal number
According to x*It is the result after normalization.
S3. feature extraction and classifying is carried out to pretreated signal using deep learning network model shown in Fig. 3.It is used
Property sensing data and myoelectricity data be separately input to two convolutional neural networks (Convolutional neural network,
CNN feature extraction) is carried out, all features are merged, generates the motor function grade based on Brunnstrom scale, then
Model parameter is iterated based on back-propagation algorithm, training deep learning network model；Specifically:
S31. input data, the patients with cerebral apoplexy for acquiring N different degrees of dyskinesia execute upper limb and evaluate action process
Data as training sample, sample set is denoted as T={ (ximu, xemg, y)1... ..., (ximwxemg, y)N, wherein ximuWith xemgPoint
Not Biao Shi patient IMU sensor signal and surface electromyogram signal matrixing indicate, y indicate the patient by profession evaluation teacher provide
Brunnstrom by stages as a result, i.e. artificial Measuring scale assessing is as a result, y ∈ { I, II, III, IV, V, VI }, i.e. class categories number C=
6；
S32. convolution sum pond
A signal sequence TS after pretreatment is expressed as x={ x1, x2..., xL, L ∈ { Limu, Lemg, wherein xt
It is the value at timestamp t, xt18 dimensional vectors and 7 dimensional vectors are respectively indicated for IMU signal and sEMG signal.Convolutional layer is by signal
In characteristic carry out depth extraction, use hyperbolic tangent function tanh as the activation primitive of convolutional layer, it is non-linear to obtain
Mapping, kth layer convolution algorithm formula are as follows:
Wherein,For k-1 layers of i-th of pond as a result, when first time convolution, as pretreated signal sequence；
Mk-1Core number is accumulated for kth-level 1 volume,For k j-th of convolution kernel of layer, b is corresponding biasing.Every layer of convolution can have multiple volumes
Product core is used for the different characteristic of extraction time sequence；
Maximum pondization operation is using the maximum value in the region of pond as its output, is s in step-length thereforeMPNon-overlap area
Between it is upper carry out primary maximum pond, exporting length reduces a sMPThe factor.In l layers of convolution-Chi Huahou, feature vector exists
The output length about L ' ≈ L/s of CNN moduleMP l。
S33. Fusion Features:
After 2 times alternately convolution sum pond, two kinds of signal sequences divide by the abstract characteristics that two CNN are extracted
Other vectorization is expressed as Fimu/emg={ f1, f2..., fL′},L′∈{L′imu, L 'emg}.Each feature vector be original TS more
High-level abstract, length on a timeline is shorter, and two feature vectors are cascaded to form new fusion feature F={ f1,
f2..., fL′imu+L′emgInputted as full articulamentum.
S34. full articulamentum:
Convolution sum pondization carries out abstract extraction and expression to the information in IMU or sEMG signal sequence, using connecting entirely
Layer is connect to merge information two kinds of signals and realize classification.Include a fusion feature receiving layer, intermediate hidden layers and defeated
Layer out, output layer neuron quantity are equal to class categories number C.
S35.Softmax layers:
Softmax is an additional data analysis layer, and the classification results that full articulamentum exports are become probability point
Cloth, full articulamentum neuron number are equal to class categories number, and corresponding output is o1, o2..., oc, after being handled by Softmax
Output are as follows:
Export final result o 'iIndicate that the corresponding Brunnstrom grade of the input signal is the probability of i.
S36. cost function
Classification results cross entropy is used to calculate error in classification, carries out backpropagation weight instruction for batch stochastic gradient descent
Practice, then cost function is written as:
Wherein y indicates the true generic of sequence n, Oy(n) probability output of the model on true generic is indicated.
Batch is the batch sample number selected in batch stochastic gradient descent.
S37. backpropagation repetitive exercise
Based on the error amount of model obtained in the previous step output and reality output, the error declined using batch gradient is reversed
Propagation algorithm realization is iterated the parameter in network, after 2000 iteration, model parameter training is completed, to network
Structure and parameter is saved.
S4. to the patient for needing to carry out upper extremity exercise assessment of function, data identical with S1, S2 step are carried out to it and are adopted
Collection, pretreatment, and it is input to trained deep learning model, automatically generate patient's upper extremity exercise function Brunnstrom by stages
Evaluation result.
Embodiment 1
200 IMU and sEMG sample of signal are acquired by above-mentioned steps S1, S2, deep learning model is constructed by S3
And it completes training process and obtains automatic evaluation deep learning model H.The subject A unknown for upper extremity exercise Performance Level leads to
The sensor signal that shoulder action process is touched in the acquisition of S1, S2 step is crossed, Fig. 4 (a) is the 3-axis acceleration signal hint of forearm IMU
Figure, Fig. 4 (b) are the sEMG signal schematic representation of the bicipital muscle of arm.The institute collected is input in H there are two types of signal, is exported
To A upper extremity exercise function Brunnstrom by stages result be VI, since A is health volunteer in this example, therefore model is correctly
Automatic evaluation, participation of the overall process without doctor are carried out to patient A, it is only necessary to assist in subject itself or 1 family members
Work is bound at sensor.
Above description has shown and described several preferred embodiments of invention, but as previously described, it should be understood that invention is not
It is confined to form disclosed herein, should not be regarded as an exclusion of other examples, and can be used for various other combinations, modification
And environment, and can be carried out within that scope of the inventive concept describe herein by the above teachings or related fields of technology or knowledge
Change.And changes and modifications made by those skilled in the art do not depart from the spirit and scope of invention, then it all should be in the appended power of invention
In the protection scope that benefit requires.
Claims (4)
1. a kind of cerebral apoplexy upper extremity exercise function method for automatically evaluating based on deep learning, which is characterized in that including following step
It is rapid:
S1. inertia sensing data and myoelectricity data based on wearable sensing system acquisition patient's upper extremity exercise process；
S2. length normalization method and numerical value normalization pretreatment are carried out to the data of acquisition；
S3. inertia sensing data and myoelectricity data are separately input to two convolutional neural networks and carry out feature extraction, to all
Feature is merged, and is generated the motor function grade based on Brunnstrom scale, is then based on back-propagation algorithm to model
Parameter is iterated, training deep learning network model；
S4. to the patient for needing to carry out upper extremity exercise assessment of function, data identical with S1, S2 step are carried out to it and acquires, is pre-
Processing, and it is input to trained deep learning model, it automatically generates patient's upper extremity exercise function Brunnstrom and evaluates by stages
As a result.
2. assessment method according to claim 1, which is characterized in that the step S1 is based on wearable sensing system and adopts
Collect the inertia sensing data and myoelectricity data of patient's upper extremity exercise process specifically:
S11. the wearable sensor-based system of design acquisition upper extremity exercise performance data is logical by 2 Inertial Measurement Units (IMU) and 7
Road surface electromyography (sEMG) sensor two parts composition；Three axis that Inertial Measurement Unit is used to measure limb motion process accelerate
Degree, angular speed and magnetometer data realize wireless real-time transmission, sample frequency 20Hz by bluetooth and PC；SEMG acquisition movement
Process skeleton muscle group electrical signals activity, sample frequency 3KHz；
S12. sensor wearing location is designed:
Two Inertial Measurement Units are bundled in respectively in the middle part of Ipsilateral large arm and in the middle part of forearm, and Inertial Measurement Unit X-axis is perpendicular to people
Body coronal-plane inwards, Y-axis perpendicular to it is cross-section downwards, Z axis perpendicular to sagittal plane inwards；7 channel surface myoelectric electrodes are set respectively
Group, pectoralis major, trapezius muscle at totally 7 before group, deltoid muscle in brachioradialis, the triceps muscle of arm, the bicipital muscle of arm, deltoid muscle；
S13. patient touches the movement of opposite side shoulder with ipsilateral upper limb execution, and sensor acquires the acceleration and myoelectricity data of motion process,
And PC is transmitted to by bluetooth.
3. assessment method according to claim 1, which is characterized in that being grown to the data of acquisition in the step S2
Degree normalization and numerical value normalization pretreatment specifically:
S21. signal length normalizes: since different patients execution is touched, shoulder movement speed is different, leads to the sensor letter of acquisition
Number length difference needs all Length of sample series are unified；Using the method for linear interpolation by all IMU signal length normalizings
Turn to the maximum value L of all sample lengthsimu；All sEMG signal lengths are normalized to by Sample Maximal using same method
Value Lemg；
S22. binding position otherness is eliminated, it is assumed that IMU signal is X={ xa, xg, xm, wherein xa, xg, xmRespectively acceleration,
Angular speed, magnetometer signals eliminate Patient height by following formula, sensor binds position difference and gives signal bring difference:
Wherein l indicates that the IMU binds position at a distance from shoulder joint, unit: m；X ' is treated signal；
S23. signal numerical value normalizes, and promotes convergence rate and precision, maps the data into 0~1 range, returned using linear function
One changes method, and transfer function is as follows:
Wherein max is the maximum value of sample data, and min is the minimum value of sample data；X is original sensor signal data, x*
It is the result after normalization.
4. assessment method according to claim 1, which is characterized in that in the step S3 by inertia sensing data and flesh
Electric data are separately input to two convolutional neural networks and carry out feature extraction, merge to all features, generation is based on
The motor function grade of Brunnstrom scale, is then based on back-propagation algorithm and is iterated to model parameter, training depth
Learning network model；Specifically:
S31. input data, the patients with cerebral apoplexy for acquiring N different degrees of dyskinesia execute the number of upper limb evaluation action process
According to as training sample, sample set is denoted as T={ (ximu, xemg, y)1..., (ximu, xemg, y)N, wherein ximuWith xemgPoint
Not Biao Shi patient IMU sensor signal and surface electromyogram signal matrixing indicate, y indicate the patient by profession evaluation teacher provide
Brunnstrom by stages as a result, i.e. artificial Measuring scale assessing is as a result, y ∈ { I, II, III, IV, V, VI }, i.e. class categories number C=
6；
S32. convolution sum pond:
A signal sequence TS after pretreatment is expressed as x={ x1, x2..., xL, L ∈ { Limu, Lemg, wherein xtIt is the time
Stab the value at t, xt18 dimensional vectors and 7 dimensional vectors are respectively indicated for IMU signal and sEMG signal；Convolutional layer is by the spy in signal
It levies data and carries out depth extraction, use hyperbolic tangent function tanh as the activation primitive of convolutional layer, to obtain Nonlinear Mapping, the
K layers of convolution algorithm formula are as follows:
Wherein,For k-1 layers of i-th of pond as a result, when first time convolution, as pretreated signal sequence；Mk-1
Core number is accumulated for kth-level 1 volume,For k j-th of convolution kernel of layer, b is corresponding biasing；Every layer of convolution has multiple convolution kernels to use
In the different characteristic of extraction time sequence；
Maximum pondization operation is using the maximum value in the region of pond as its output, is s in step-length thereforeMPNonoverlapping intervals on
Primary maximum pond is carried out, output length reduces a sMPThe factor；In l layers of convolution-Chi Huahou, feature vector is in CNN mould
The output length about L ' ≈ L/s of blockMP l；
S33. Fusion Features:
After 2 times alternately convolution sum pond, the abstract characteristics that two kinds of signal sequences are extracted by two CNN respectively to
Quantization means are Fimu/emg={ f1, f2..., fL′},L′∈{L′imu, L 'emg}；Each feature vector is the more advanced of original TS
It does not make a summary, length on a timeline is shorter, and two feature vectors are cascaded to form new fusion featureIt is inputted as full articulamentum；
S34. full articulamentum:
Convolution sum pondization carries out abstract extraction and expression to the information in IMU or sEMG signal sequence, utilizes full articulamentum
Information two kinds of signals is merged and realizes classification；Comprising a fusion feature receiving layer, intermediate hidden layers and output layer,
Output layer neuron quantity is equal to class categories number C；
S35.Softmax layers:
Softmax is an additional data analysis layer, and the classification results that full articulamentum exports are become a probability distribution；Entirely
Articulamentum neuron number is equal to class categories number, and corresponding output is o1, o2..., oc, output after being handled by Softmax
Are as follows:
Export final result o 'iIndicate that the corresponding Brunnstrom grade of the input signal is the probability of i；
S36. cost function:
Classification results cross entropy is used to calculate error in classification, carries out backpropagation weight training for batch stochastic gradient descent,
Then cost function is written as:
Wherein y indicates the true generic of sequence n, Oy(n) probability output of the model on true generic is indicated；Batch is
The batch sample number selected in batch stochastic gradient descent；
S37. backpropagation repetitive exercise:
Based on the error amount of obtained the model output and reality output of step S36, the error declined using batch gradient is reversely passed
It broadcasts algorithm realization to be iterated the parameter in network, after 2000 iteration, model parameter training is completed, to network knot
Structure and parameter are saved.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910074666.4A CN109875565A (en) | 2019-01-25 | 2019-01-25 | A kind of cerebral apoplexy upper extremity exercise function method for automatically evaluating based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910074666.4A CN109875565A (en) | 2019-01-25 | 2019-01-25 | A kind of cerebral apoplexy upper extremity exercise function method for automatically evaluating based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109875565A true CN109875565A (en) | 2019-06-14 |
Family
ID=66926916
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910074666.4A Pending CN109875565A (en) | 2019-01-25 | 2019-01-25 | A kind of cerebral apoplexy upper extremity exercise function method for automatically evaluating based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109875565A (en) |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110210366A (en) * | 2019-07-05 | 2019-09-06 | 青岛理工大学 | Assemble rundown process sample acquisition system, deep learning network and monitoring system |
| CN110569968A (en) * | 2019-10-18 | 2019-12-13 | 上海大学 | Method and system for evaluating entrepreneurship failure resilience based on electrophysiological signals |
| CN110633736A (en) * | 2019-08-27 | 2019-12-31 | 电子科技大学 | Human body falling detection method based on multi-source heterogeneous data fusion |
| CN110931103A (en) * | 2019-11-01 | 2020-03-27 | 深圳市迈步机器人科技有限公司 | Control method and system of rehabilitation equipment |
| CN111296994A (en) * | 2019-12-20 | 2020-06-19 | 石狮市森科智能科技有限公司 | Intelligent gesture interaction control system |
| CN111543989A (en) * | 2020-05-12 | 2020-08-18 | 中国人民解放军总医院第八医学中心 | Proprioception information detection system and proprioception recovery system |
| CN111544004A (en) * | 2020-05-15 | 2020-08-18 | 中国科学院自动化研究所 | System, method and device for detecting motion function of stroke patient |
| CN111815487A (en) * | 2020-06-28 | 2020-10-23 | 珠海中科先进技术研究院有限公司 | Health education assessment method, device and medium based on deep learning |
| CN111950460A (en) * | 2020-08-13 | 2020-11-17 | 电子科技大学 | Muscle strength self-adaptive stroke patient hand rehabilitation training action recognition method |
| CN111986801A (en) * | 2020-07-14 | 2020-11-24 | 珠海中科先进技术研究院有限公司 | Rehabilitation evaluation method, device and medium based on deep learning |
| CN112085169A (en) * | 2020-09-11 | 2020-12-15 | 西安交通大学 | Autonomous learning and evolution method for limb exoskeleton assisted rehabilitation brain-myoelectricity fusion sensing |
| CN112185558A (en) * | 2020-09-22 | 2021-01-05 | 珠海中科先进技术研究院有限公司 | Mental health and rehabilitation evaluation method, device and medium based on deep learning |
| CN112184010A (en) * | 2020-09-27 | 2021-01-05 | 北京斯年智驾科技有限公司 | Tooth brushing effect evaluation method, device, system, electronic device and storage medium |
| CN112674760A (en) * | 2020-12-22 | 2021-04-20 | 山东农业大学 | Wearable sensor-based Parkinson upper limb movement detection method |
| CN112932508A (en) * | 2021-01-29 | 2021-06-11 | 电子科技大学 | Finger activity recognition system based on arm electromyography network |
| CN113476817A (en) * | 2021-07-07 | 2021-10-08 | 鹤壁市人民医院 | Cognitive rehabilitation training system based on neural network algorithm |
| CN115399791A (en) * | 2022-06-28 | 2022-11-29 | 天津大学 | Stroke lower limb function assessment method and system based on myoelectric motion multi-data fusion |
| CN115844383A (en) * | 2022-12-27 | 2023-03-28 | 国家康复辅具研究中心 | Limb movement function evaluation system and method fusing brain function indexes |
-
2019
- 2019-01-25 CN CN201910074666.4A patent/CN109875565A/en active Pending
Non-Patent Citations (1)
| Title |
|---|
| 王刚 主编: "《临床康复医学》", 31 July 2017 * |
Cited By (21)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110210366B (en) * | 2019-07-05 | 2021-04-27 | 青岛理工大学 | Assembling and screwing process sample acquisition system, deep learning network and monitoring system |
| CN110210366A (en) * | 2019-07-05 | 2019-09-06 | 青岛理工大学 | Assemble rundown process sample acquisition system, deep learning network and monitoring system |
| CN110633736A (en) * | 2019-08-27 | 2019-12-31 | 电子科技大学 | Human body falling detection method based on multi-source heterogeneous data fusion |
| CN110569968A (en) * | 2019-10-18 | 2019-12-13 | 上海大学 | Method and system for evaluating entrepreneurship failure resilience based on electrophysiological signals |
| CN110931103A (en) * | 2019-11-01 | 2020-03-27 | 深圳市迈步机器人科技有限公司 | Control method and system of rehabilitation equipment |
| CN111296994A (en) * | 2019-12-20 | 2020-06-19 | 石狮市森科智能科技有限公司 | Intelligent gesture interaction control system |
| CN111543989A (en) * | 2020-05-12 | 2020-08-18 | 中国人民解放军总医院第八医学中心 | Proprioception information detection system and proprioception recovery system |
| CN111544004A (en) * | 2020-05-15 | 2020-08-18 | 中国科学院自动化研究所 | System, method and device for detecting motion function of stroke patient |
| CN111815487A (en) * | 2020-06-28 | 2020-10-23 | 珠海中科先进技术研究院有限公司 | Health education assessment method, device and medium based on deep learning |
| CN111815487B (en) * | 2020-06-28 | 2024-02-27 | 珠海中科先进技术研究院有限公司 | Deep learning-based health education assessment method, device and medium |
| CN111986801A (en) * | 2020-07-14 | 2020-11-24 | 珠海中科先进技术研究院有限公司 | Rehabilitation evaluation method, device and medium based on deep learning |
| CN111950460A (en) * | 2020-08-13 | 2020-11-17 | 电子科技大学 | Muscle strength self-adaptive stroke patient hand rehabilitation training action recognition method |
| CN112085169A (en) * | 2020-09-11 | 2020-12-15 | 西安交通大学 | Autonomous learning and evolution method for limb exoskeleton assisted rehabilitation brain-myoelectricity fusion sensing |
| CN112185558A (en) * | 2020-09-22 | 2021-01-05 | 珠海中科先进技术研究院有限公司 | Mental health and rehabilitation evaluation method, device and medium based on deep learning |
| CN112184010A (en) * | 2020-09-27 | 2021-01-05 | 北京斯年智驾科技有限公司 | Tooth brushing effect evaluation method, device, system, electronic device and storage medium |
| CN112674760A (en) * | 2020-12-22 | 2021-04-20 | 山东农业大学 | Wearable sensor-based Parkinson upper limb movement detection method |
| CN112932508A (en) * | 2021-01-29 | 2021-06-11 | 电子科技大学 | Finger activity recognition system based on arm electromyography network |
| CN113476817A (en) * | 2021-07-07 | 2021-10-08 | 鹤壁市人民医院 | Cognitive rehabilitation training system based on neural network algorithm |
| CN113476817B (en) * | 2021-07-07 | 2022-06-07 | 鹤壁市人民医院 | Cognitive rehabilitation training system based on neural network algorithm |
| CN115399791A (en) * | 2022-06-28 | 2022-11-29 | 天津大学 | Stroke lower limb function assessment method and system based on myoelectric motion multi-data fusion |
| CN115844383A (en) * | 2022-12-27 | 2023-03-28 | 国家康复辅具研究中心 | Limb movement function evaluation system and method fusing brain function indexes |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109875565A (en) | A kind of cerebral apoplexy upper extremity exercise function method for automatically evaluating based on deep learning | |
| CN102622605B (en) | Surface electromyogram signal feature extraction and action pattern recognition method | |
| CN102138860B (en) | Intelligentized rehabilitation training equipment for hand functions of patients suffering from cerebral injury | |
| CN103054585B (en) | Biological motion information based upper limb shoulder elbow wrist joint motion function evaluation method | |
| CN104207793B (en) | A kind of grip function assessment and training system | |
| CN104382595A (en) | Upper limb rehabilitation system and method based on myoelectric signal and virtual reality interaction technology | |
| Shao et al. | Single-channel SEMG using wavelet deep belief networks for upper limb motion recognition | |
| CN104997523A (en) | Upper limb rehabilitation robot rehabilitation training motion function assessment method | |
| Kuang et al. | Extreme learning machine classification method for lower limb movement recognition | |
| CN108742613A (en) | Orient coupling analytical method between the flesh of coherence partially based on transfer entropy and broad sense | |
| Wu et al. | Upper limb motion recognition based on LLE-ELM method of sEMG | |
| CN107618018A (en) | A kind of manipulator behavior speed proportional control method based on myoelectricity | |
| CN103908361A (en) | Method for acquiring and operating artificial limb joint movement coupling drive signals | |
| Li et al. | Quantitative assessment of hand motor function for post-stroke rehabilitation based on HAGCN and multimodality fusion | |
| Zhang et al. | The design of a hemiplegic upper limb rehabilitation training system based on surface EMG signals | |
| Wang et al. | Lower limb motion recognition based on surface electromyography signals and its experimental verification on a novel multi-posture lower limb rehabilitation robots☆ | |
| KR100994408B1 (en) | Method and device for deducting pinch force, method and device for discriminating muscle to deduct pinch force | |
| Veer | A flexible approach for segregating physiological signals | |
| Raikova et al. | The influence of the way the muscle force is modeled on the predicted results obtained by solving indeterminate problems for a fast elbow flexion | |
| Peng et al. | An sEMG-driven neuromusculoskeletal model of upper limb for rehabilitation robot control | |
| Veer et al. | Analysis and recognition of operations using SEMG from upper arm muscles | |
| CN113974612A (en) | Automatic assessment method and system for upper limb movement function of stroke patient | |
| Guo et al. | A novel fuzzy neural network-based rehabilitation stage classifying method for the upper limb rehabilitation robotic system | |
| CN109522810B (en) | Myoelectric prosthetic hand gesture recognition method based on community voting mechanism | |
| CN106371590A (en) | High-performance motor imagery online brain-computer interface system based on OpenVIBE |
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 |
Application publication date: 20190614 |
|
| RJ01 | Rejection of invention patent application after publication |