CN113822170B - Method for identifying tremor of hands of body in non-stationary state - Google Patents
Method for identifying tremor of hands of body in non-stationary state Download PDFInfo
- Publication number
- CN113822170B CN113822170B CN202111011247.XA CN202111011247A CN113822170B CN 113822170 B CN113822170 B CN 113822170B CN 202111011247 A CN202111011247 A CN 202111011247A CN 113822170 B CN113822170 B CN 113822170B
- Authority
- CN
- China
- Prior art keywords
- tremor
- time
- phase
- hand
- window
- 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
Links
- 206010044565 Tremor Diseases 0.000 title claims abstract description 185
- 238000000034 method Methods 0.000 title claims abstract description 49
- 230000009471 action Effects 0.000 claims abstract description 41
- 238000013178 mathematical model Methods 0.000 claims abstract description 11
- 230000008030 elimination Effects 0.000 claims abstract description 10
- 238000003379 elimination reaction Methods 0.000 claims abstract description 10
- 230000033001 locomotion Effects 0.000 claims description 33
- 230000006870 function Effects 0.000 claims description 23
- 210000002569 neuron Anatomy 0.000 claims description 22
- 150000001875 compounds Chemical class 0.000 claims description 12
- 230000000737 periodic effect Effects 0.000 claims description 11
- 238000005259 measurement Methods 0.000 claims description 8
- 238000000819 phase cycle Methods 0.000 claims description 5
- 230000001154 acute effect Effects 0.000 claims description 4
- 238000005457 optimization Methods 0.000 claims description 4
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 claims description 3
- 238000012886 linear function Methods 0.000 claims description 3
- 238000010606 normalization Methods 0.000 claims description 3
- 238000013527 convolutional neural network Methods 0.000 description 23
- 230000008569 process Effects 0.000 description 12
- 230000037081 physical activity Effects 0.000 description 9
- 238000012549 training Methods 0.000 description 8
- 238000004458 analytical method Methods 0.000 description 7
- 238000012545 processing Methods 0.000 description 7
- 238000002474 experimental method Methods 0.000 description 6
- 238000000605 extraction Methods 0.000 description 5
- 230000000694 effects Effects 0.000 description 4
- 210000000707 wrist Anatomy 0.000 description 4
- 230000008859 change Effects 0.000 description 3
- 238000007781 pre-processing Methods 0.000 description 3
- 230000000284 resting effect Effects 0.000 description 3
- 238000013459 approach Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000007613 environmental effect Effects 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 230000004927 fusion Effects 0.000 description 2
- 210000004247 hand Anatomy 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 230000004913 activation Effects 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 238000005452 bending Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 230000010365 information processing Effects 0.000 description 1
- 230000007787 long-term memory Effects 0.000 description 1
- 230000035772 mutation Effects 0.000 description 1
- 230000010363 phase shift Effects 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
- 239000000758 substrate Substances 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- 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
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
Abstract
The invention discloses a hand tremor identification method for a body in a non-stationary state, which is implemented according to the following steps: step 1, establishing a hand tremor mathematical model; step 2, data calibration and noise elimination; step 3, eliminating time-varying direct current components and calculating time-varying tremor frequency; and 4, a CNN-LSTM-based hand tremor identification method. The invention solves the problem that the hand tremor action in the non-stationary state of the body cannot be accurately identified in the prior art.
Description
Technical Field
The invention belongs to the technical field of behavior recognition, and particularly relates to a method for recognizing hand tremor in a body non-stationary state.
Background
Zhang et al studied body tremor rate using RFID phase characteristics. Ding applies RFID technology to extract signal characteristics to describe the change of hand motions, and proposes a Multi-Branch one-dimensional CNN (Multi-Branch 1D-CNN) basic framework, and static and dynamic hand motions are directly identified by utilizing RFID signals. The recognition rate in the prior art is low, and the recognition rate based on CNN-LSTM is superior to other models.
Disclosure of Invention
The invention aims to provide a method for identifying hand tremor in a body non-stationary state, which solves the problem that the hand tremor in the body non-stationary state cannot be accurately identified in the prior art.
The technical scheme adopted by the invention is that the hand tremor identification method for the non-stationary state of the body is implemented according to the following steps:
step 1, establishing a hand tremor mathematical model;
step 2, data calibration and noise elimination;
step 3, eliminating time-varying direct current components and calculating time-varying tremor frequency;
and 4, a CNN-LSTM-based hand tremor identification method.
The present invention is also characterized in that,
the step 1 is specifically implemented according to the following steps:
let a hand carrying a passive tag tremors in which tremor frequency r varies with time and tremor amplitude g also varies with time, in which case θ (t) is no longer the ideal periodic curve, equations (1) and (2) are rewritten as:
r (t) represents the distance from the passive tag to the RFID antenna, tremor amplitude g (t), tremor frequency R, reference distance d (t), time t, phi is the acute angle between tremor direction and the line from the RFID antenna to the reference position, θ (t) represents the phase, θ 0 Is the offset introduced by the hardware device, λ is the radio frequency wavelength, where ε is the error caused by ambient noise; the distance R (t) and the phase θ (t) together constitute a mathematical model of hand tremor.
The step 2 is specifically implemented according to the following steps:
step 2.1, decomposing the hand tremor phase θ (t) into:
where θ (t) is the phase, l 0 Is the initial resolution, t represents time, l is the scale factor, k is the pan factor,is a discrete proportional function, ψ l,k (t) is a discrete wavelet function, the two functions are orthogonal to each other, A (l) 0 The approximate and detailed coefficients of each stage, which are filtered during denoising, are defined as:
θ (t) is the phase of the phase,is a discrete proportional function, ψ l,k (t) is a discrete wavelet function, A (l, k) and D (l, k) are approximation and detail coefficients for each stage;
step 2.2, performing threshold selection by using a maximum and minimum threshold algorithm, wherein the definition of the optimal threshold is shown in a formula (8):
wherein,representing the threshold value, σ is the standard deviation, +.>Values referring to the radio frequency carrier wavelength, < >>Obtained by optimizing the objective function as shown in formula (9):
in the above equation, λ is the radio frequency carrier wavelength,referring to the optimized wavelength, d represents one of n coefficients obtained by discrete wavelet transform, and the risk is calculated as R λ (d)＝E(δ λ (d)-d) 2 ，R ideal Is an ideal result, indicating whether the wavelet coefficients are preserved, given by:
R ideal (d):＝min(d 2 ,1) (10)
step 2.3, finally, based on the threshold lambda M And selecting an approximation coefficient and a detail coefficient, reconstructing a noise-free signal by using the approximation coefficient and the detail coefficient, and representing the phase data set after noise elimination as theta.
The step 3 is specifically implemented according to the following steps:
step 3.1, eliminating time-varying direct current components:
step 3.1.1, setting a phase sample θ (t) ∈θ after data calibration and noise removal, where the following relationship exists:
where a is a constant DC component and M (t) is a time-varying DC component due to movement of the reference position of the handIs a tremor component caused by tremor of the hand around a reference position, if the hand tremor motion is a stable motion separated from the "compound motion +.>Expressed by formula (4);
step 3.1.2 using a sliding window based method, the local average value l (t) of time t within window [ t- (W/2), t+ (W/2) ] is defined as:
where W is the window size, the constant DC component a, M (t) is the time-varying DC component,is the tremor component, and in a small time window, assuming that M (t ') is approximately a linear function and θ (t') is approximately a periodic function, the relationship shown in equation (13) is obtained by equation (12):
m (t) is M (t ') in window t' ∈ [ t- (W/2), t+ (W/2)]Mean of the interior, in the formulaIs a curve of hand tremor action after removal of the direct current component, therefore +.>The integral over each period is zero;
step 3.1.3, l (t) is expressed by equation (14):
l(t)≈a+M(t) (14)
l (t) is the time-varying direct current component of the hand tremor to be removed, and the value of l (t), i.e. the value of a+m (t), can be approximated from equation (12) based on the phase data acquired in the time window [ t- (W/2), t+ (W/2) ] assuming S (t) is a subset of the phase sequence in the window, and for each sample θ (t ') acquired at a particular time t ', Δ (t ') is the time interval from that sample to the next sample, the integral in equation (12) is approximated with discrete samples as follows:
wherein the value of W should be determined according to the application environment, and the local average value l (t) is calculated, and then l (t) is subtracted from θ (t), as shown in formula (16):
the phase set after the data denoising and the time-varying direct current component elimination is set asThe phase curve of the hand tremor action is separated by subtracting the phase of the chest label from the phase of the hand label;
step 3.2, calculating time-varying tremor frequency:
step 3.2.1 according to the time windowCalculating the tremor frequency r (t) at time t from the phase data of (c), wherein +.>Is the window size, first, the peak or valley value within the phase curve window needs to be determined, then the peak-to-peak distance is used to estimate the tremor period and calculate the tremor frequency,
assume thatIs derived from phase samples->At->Subset in window, in recognitionAfter all peaks on the window, an estimate of the tremor period, p (t), is obtained using the least squares method, with the peak-to-peak spacing expressed as { v 1 ,v 2 ,...,v m The least squares estimate is given by equation (17):
step 3.2.2 in WindowThe estimation of tremor frequency r (t) at time t can be calculated from equation (18):
r(t)＝1/p(t) (18)
the average tremor frequency over the measurement period is expressed asThe definition is as shown in formula (19):
wherein t is 1 Is the time to start measurement, t n It is the time at which the measurement is ended,
step 3.2.3 assuming n samples belong toThe calculation is shown in formula (20):
wherein Δ (t) is the slave sampleTime interval to next sample.
Step 4 is specifically implemented according to the following steps:
the CNN-LSTM model consists of two parts, wherein the CNN-LSTM model consists of an LSTM layer containing 36 neurons, a full-connection layer, a batch normalization layer and an output layer, the LSTM layer is provided with dropout with a value of 0.5, an Adam optimizer is used for parameter optimization, the initial learning rate is set to be 0.0006, epoch is set to be 400, batch size is set to be 50, and after model parameters are set, the data obtained in the step 3 are input into the model for hand tremor recognition.
The invention has the beneficial effects that the recognition rate of the tremor of hands in the non-stationary state of the body is high, and the recognition rate of the invention is higher than that of a convolutional neural network and a long-term and short-term memory network. The hand tremor action in the body resting state is not affected by the activities of other parts of the body, and the wrist reference position is stable, so that the hand tremor amplitude, tremor rate and movement track are relatively stable without being disturbed by the environment. The hand tremor action in the non-stationary state of the body can be influenced by other physical activities, and the acquired hand tremor data is a 'compound action' comprising the physical activities. The present invention addresses this problem by utilizing multiple labels to separate hand tremor motion data from the "compound motion" due to instability of the hand tremor motion caused by the presence of physical activity. The hand tremor action of the body in a non-stationary state cannot be accurately identified in the prior art, and the method comprehensively utilizes the advantages of CNN extraction characteristics and the advantages of LSTM processing time sequence data to realize the accurate identification of the hand tremor action of the body in the non-stationary state.
Drawings
Figure 1 is a mathematical model of hand tremor;
figure 2 is a hand tremor phase signal in a compound motion;
figure 3 is a graph of hand tremor phase signal separated from compound motion;
FIG. 4 is a frame diagram of the combination of CNN and LSTM;
FIG. 5 is a block diagram of the LSTM model;
FIG. 6 (a) is an Accurcry plot of setting initial parameters, and FIG. 6 (b) is a Loss plot of setting initial parameters;
FIG. 7 is a neuron parameter adjustment graph;
FIG. 8 (a) is an Accurcy curve of the model after setting the model neuron number to 36, and FIG. 8 (b) is a Loss curve of the model after setting the model neuron number to 36;
fig. 9 (a) is a Loss graph with a learning rate of 0.0002, and fig. 9 (b) is a Loss graph with a learning rate of 0.0004;
FIG. 10 is a Batch size parameter adjustment chart;
FIG. 11 (a) is an Accuracy curve of the model after the model is adjusted by the LSTM layer neuron number, the learning rate and the Batch size parameter, and FIG. 11 (b) is a Loss curve of the model after the model is adjusted by the LSTM layer neuron number, the learning rate and the Batch size parameter;
FIG. 12 is a graph of hand tremor motion recognition accuracy, recall and F1-score for various combinations;
FIG. 13 is a graph of recognition accuracy, recall, and F1-score for different model macros averaged.
Detailed Description
The invention will be described in detail below with reference to the drawings and the detailed description.
The hand tremor action in the body resting state is not affected by the activities of other parts of the body, and the wrist reference position is stable, so that the hand tremor amplitude, tremor rate and movement track are relatively stable without being disturbed by the environment. The hand tremor action in the non-stationary state of the body can be influenced by other physical activities, and the acquired hand tremor data is a 'compound action' comprising the physical activities. The invention firstly establishes a mathematical model according to hand tremor actions in an ideal state. Then, the factors influencing the hand tremor action in the environment are analyzed, experiments are set, and data calibration and noise removal work are completed. Finally, a hand tremor recognition method based on CNN-LSTM is provided, and the method comprehensively utilizes the advantages of CNN extraction characteristics and the advantages of LSTM processing time sequence data. According to the invention, the CNN-LSTM model is trained and the parameters are optimized, so that the accurate identification of the hand tremor action of the body in a non-stationary state is realized.
Establishing a mathematical model of hand tremor
The invention establishes the hand tremor model according to the characteristics of hand tremor actions. First, the invention analyzes the phase signal of hand tremor action from a mathematical angle, and then researches the time variability of hand tremor action. The study of the present invention begins with an ideal state of hand tremor action, assuming that the motion trace of the passive tag attached to the hand is periodic in the direction of the double arrow in fig. 1. The midpoint of the double arrow is the reference position of tremor, which represents the reference position of tremor movement (also the reference position of the label movement trace), and not the wrist reference position. Let d be the distance between reference position and RFID antenna, phi be the acute angle between tremble direction and the straight line of RFID antenna to reference position, r be tremble frequency, g be tremble's range.
In an ideal case, the passive tag is a distance gsin (2pi rt) from the reference position. The distance from the passive tag to the RFID antenna can be easily calculated from the figure as shown in equation (1).
The period is (1/r) because the distance of the passive tag from the RFID antenna varies with time, it produces a phase shift in the backscatter signal received by the RFID reader. In conjunction with equation (1), then the phase θ (t) is:
wherein θ is 0 Is the offset introduced by the hardware device, λ is the radio frequency wavelength, and the total distance the signal travels from the RFID reader to the passive tag back to the reader is 2R (t). Obviously, the phase θ (t) thus constructed is a curve having a periodic and continuous variation. In an ideal model, the curves calculated by equations (1) and (2) have a near perfect period.
However, the present invention is directed to hand tremor action with the body in a non-resting state, which has a particular time-variability. This breaks the original assumption of the present invention.
The method for identifying the tremor of the hands in the non-stationary state of the body is implemented according to the following steps:
step 1, establishing a hand tremor mathematical model;
the step 1 is specifically implemented according to the following steps:
let a hand carrying a passive tag tremors in which tremor frequency r varies with time and tremor amplitude g also varies with time, in which case θ (t) is no longer an ideal periodic curve, which may be an extremely complex, periodic and amplitude-varying curve, to reflect the time-varying characteristics described above, equations (1) and (2) are rewritten as:
r (t) represents the distance from the passive tag to the RFID antenna, tremor amplitude g (t), tremor frequency R, reference distance d (t), time t, phi is the acute angle between tremor direction and the line from the RFID antenna to the reference position, θ (t) represents the phase, θ 0 Is the offset introduced by the hardware device, λ is the radio frequency wavelength, where ε is the error caused by ambient noise; such models are difficult to analyze using conventional methods. Therefore, new methods must be studied next to calculate hand tremor rate, which is not based on an ideal model analysis, but directly on experimentally collected phase sample data.
The distance R (t) and the phase θ (t) together constitute a mathematical model of hand tremor.
Step 2, data calibration and noise elimination;
the step 2 is specifically implemented according to the following steps:
in data preprocessing, the effect of passive tags and RFID readers on sample data needs to be considered. Because of environmental interference and product characteristics, the data collected by the experiment contains inherent noise. From the collected phase data, it was found that when the passive tag moves to a position where the real phase value is close to a multiple of 2π, the phase value will shift around this multiple, and the modulo operation in equation (2) may result in a jump in the phase data. For calibration of the phase sample data, if there is a sharp mutation of about 2pi between two consecutive phase values caused by the modulo operation. In this case, the suddenly appearing jump data needs to be removed to ensure continuity between the two phase values. On this basis, data calibration involving phase jumps is completed.
Time-varying noise is caused by a variety of factors, including hardware-inherent properties, as well as fluctuations in distance between the RFID antenna and the reference location of the passive tag, tremor frequency, and tremor amplitude. Removing some of the inherent noise in the experimental data that does not fluctuate over time is relatively simple, e.g., collecting phase data of hand tremor motion with 2 passive tags simultaneously, tag 1 Stationary Tag 2 Attached to the finger. Tag is provided with 1 Phase data of (2) is theta 1 ，Tag 2 Phase data of (2) is theta 2 . From equation (4), it can be seen that by θ 2 -θ 1 The inherent noise in the experimental data, which does not fluctuate with time, can be removed, and the rest of the noise needs other methods. Simply using a low pass filter does not work well, although it can remove high frequency noise (e.g. white noise in an experimental environment). However, since most of the data collected by the present invention contains noise that is not high frequency, the use of a low pass filter is ineffective.
After removing noise caused by modulo arithmetic and hardware imperfections, noise still exists in the data. The remaining noise comes from interference from the experimental environment, which may reduce the number of peaks or valleys in the phase curve to affect the performance of the hand tremor recognition method. Conventional approaches that use low pass filters have not been able to effectively address this problem because noise may be in the frequency band of the signal. Thus, the first and second substrates are bonded together,
step 2.1, decomposing the hand tremor phase θ (t) into:
where θ (t) is the phase, l 0 Is the initial resolution, t represents time, l is the scale factor, k is the pan factor,is a discrete proportional function, ψ l,k (t) is a discrete wavelet function, the two functions are orthogonal to each other, A (l) 0 The approximate and detailed coefficients of each stage, which are filtered during denoising, are defined as:
θ (t) is the phase of the phase,is a discrete proportional function, ψ l,k (t) is a discrete wavelet function, A (l, k) and D (l, k) are approximation and detail coefficients for each stage;
step 2.2, after decomposing the phase curve, a threshold λ is set for noise M The invention uses a maximum and minimum threshold algorithm to select a threshold, which is an optimization problem, in which an optimal threshold is selected to achieve the minimization of the maximum mean square error and to obtain the worst risk function for the ideal process. The optimal threshold definition is shown in formula (8):
wherein,representing the threshold value, σ is the standard deviation, +.>Values referring to the radio frequency carrier wavelength, < >>Obtained by optimizing the objective function as shown in formula (9):
in the above equation, λ is the radio frequency carrier wavelength,referring to the optimized wavelength, d represents one of n coefficients obtained by discrete wavelet transform, and the risk is calculated as R λ (d)＝E(δ λ (d)-d) 2 ，R ideal Is an ideal result, indicating whether the wavelet coefficients are preserved, given by:
R ideal (d):＝min(d 2 ,1) (10)
step 2.3, finally, based on the threshold lambda M And selecting an approximation coefficient and a detail coefficient, reconstructing a noise-free signal by using the approximation coefficient and the detail coefficient, and representing the phase data set after noise elimination as theta.
Step 3, eliminating time-varying direct current components and calculating time-varying tremor frequency;
the step 3 is specifically implemented according to the following steps:
step 3.1, eliminating time-varying direct current components:
there is a time-varying dc component in the phase profile, due in part to the instability of hand tremors. Assuming that the hand tremor motion is a periodic motion with the wrist position as a reference point, the reference position of the hand may slightly move during this process.
If other activities exist in the human body, the direct current component contained in the hand tremor data acquired through the passive tag of the hand becomes more remarkable along with the time change. Figure 2 shows measured phase data, using two passive tags attached to the subject's chest and hand, respectively. The experimenter started to be in a stooping state, and hands were tremor during the process of restoring to be upright. Because the motion of other parts of the human body introduces a larger direct current component which changes with time, the phase data acquired at the moment is difficult to accurately describe the hand tremor action. To obtain phase data of hand tremor action, it is necessary to cancel the time-varying dc component after the data calibration and noise removal are completed.
Step 3.1.1, setting a phase sample θ (t) ∈θ after data calibration and noise removal, where the following relationship exists:
where a is a constant DC component, a depends on the initial distance between the passive tag and the RFID antenna and the hardware properties, M (t) is a time-varying DC component due to the movement of the reference position of the hand, andthe present invention, which is a tremor component caused by tremor of the hand around the reference position, hopes to approximately remove a+M (t) from the sample θ (t) to obtain +.>Is a value of (2). If the hand tremor action is a stable movement separated from the "compound action>Using the formula%4) It is indicated that in real life, however, the amplitude, rate and frequency of hand tremor action may vary over time.
Step 3.1.2 a simple way to remove the dc component is to subtract θ (t) from the global average of the phase sequence. However, this method is not effective. The better solution to this problem is to use a sliding window based method that calculates the local mean value in the sliding window to achieve the purpose of removing the DC component [31] . Window [ t- (W/2), t+ (W/2)]The local average value l (t) of the internal time t is defined as:
where W is the window size, the constant DC component a, M (t) is the time-varying DC component,is the tremor component, and in a small time window, assuming that M (t ') is approximately a linear function and θ (t') is approximately a periodic function, the relationship shown in equation (13) is obtained by equation (12):
m (t) is M (t ') in window t' ∈ [ t- (W/2), t+ (W/2)]Mean of the interior, in the formulaIs a curve of hand tremor action after removal of the direct current component, therefore +.>The integral over each period is zero;
step 3.1.3 when W is much greater than the period length,value pair +.>The size of the curve becomes negligible. Thus, l (t) is represented by equation (14):
l(t)≈a+M(t) (14)
l (t) is the time-varying direct current component of the hand tremor to be removed, and the value of l (t), i.e. the value of a+m (t), can be approximated from equation (12) based on the phase data acquired in the time window [ t- (W/2), t+ (W/2) ] assuming S (t) is a subset of the phase sequence in the window, and for each sample θ (t ') acquired at a particular time t ', Δ (t ') is the time interval from that sample to the next sample, the integral in equation (12) is approximated with discrete samples as follows:
where the value of W should be determined according to the environment of the application, however, W should be small enough that M (t) may be linear over a time window and W should be greater than the tremor period. After calculating the local average value l (t), l (t) is subtracted from θ (t), as shown in equation (16):
the phase set after the data denoising and the time-varying direct current component elimination is set asThe phase curve of the hand tremor motion is separated by subtracting the phase of the chest label from the phase of the hand label, as shown in fig. 3;
step 3.2, calculating time-varying tremor frequency:
step 3.2.1, the hand tremor action cannot produce perfect periodic phase curves in ideal conditions, hand tremor is time-varying. Not only the intensity of tremors may vary over time, but also the tremor period may be over timeInter-variation. To calculate the time-varying tremor frequency, a sliding window based approach needs to be used. According to time windowCalculating the tremor frequency r (t) at time t from the phase data of (c), wherein +.>Is the window size, even a short number of seconds is suitable when measuring hand tremors, since the frequency of hand tremors is more easily kept consistent in a relatively short time, thus avoiding time-varying factors introduced by the subject itself. First, it is necessary to determine the peak or valley value within the phase curve window, then estimate the tremor period using the peak-to-peak distance and calculate the tremor frequency,
assume thatIs derived from phase samples->At->A simple algorithm to identify peaks is to compare each phase sample to its previous and next samples, a subset of the windows. If the phase sample is larger than the previous and subsequent samples, it will be considered to be peak. The wavelet transform method is used to eliminate the false peaks caused by the environmental noise, and although most of the false peaks can be removed, all the false peaks cannot be removed. After identifying all peaks on the window, an estimate of the tremor period, p (t), is obtained using the least squares method, with the peak-to-peak spacing expressed as { v 1 ,v 2 ,...,v m The least squares estimate is given by equation (17):
step 3.2.2 in WindowThe estimation of tremor frequency r (t) at time t can be calculated from equation (18):
r(t)＝1/p(t) (18)
the average tremor frequency over the measurement period is expressed asThe definition is as shown in formula (19):
wherein t is 1 Is the time to start measurement, t n It is the time at which the measurement is ended,
step 3.2.3 assuming n samples belong toThe calculation is shown in formula (20):
wherein Δ (t) is the slave sampleTime interval to next sample.
And 4, a CNN-LSTM-based hand tremor identification method.
Step 4 is specifically implemented according to the following steps:
the present invention addresses the problem of hand tremor action with physical activity, CNN being inefficient in processing the global time dependence of RFID signal data. Hand tremor recognition can be considered a time series classification problem. From the data preprocessing analysis, the hand tremor action is susceptible to factors such as hand tremor reference point, tremor amplitude and tremor frequency, and the influence is time-varying. Therefore, the study of the present invention needs to consider the correlation between features with a large time span. The time information is captured in the hand tremor action to help model the time-varying hand tremor action, so that the defect of CNN in terms of time sequence processing can be overcome, and the accuracy of the recognition algorithm can be improved.
With reference to fig. 4 and 5, the LSTM solves the problem of gradient disappearance of the common recurrent neural network in the back propagation process by adding different gating units, which makes the LSTM advantageous in processing time series data. The LSTM has the capacity of memorizing long-distance information due to the special structure, so that the problem of dependence among features with larger time span in hand tremor actions can be solved. Based on the analysis, the invention provides a CNN-LSTM-based hand tremor recognition method. The whole idea is to input RFID signal data into CNN for feature extraction, then add LSTM after convolution layer and pooling layer to extract time correlation between features, and further realize accurate identification of hand tremor action.
The CNN-LSTM model of the invention consists of two parts, CNN-based feature extraction and LSTM-based feature fusion, as shown in FIG. 4. The CNN learns local features from original RFID signal data, the LSTM extracts time dependency from the local features, fusion of the local features and global features is realized, deep and advanced features can be extracted by combining the CNN and the LSTM, and hand tremor actions are finely described, so that accurate identification is realized.
The CNN-LSTM model of the invention consists of an LSTM layer containing 36 neurons, a full connection layer, a batch normalization layer and an output layer, and the network structure is shown in figure 5. The LSTM layer is provided with a dropout with a value of 0.5, an Adam optimizer is used for parameter optimization, the initial learning rate is set to be 0.0006, the epoch is set to be 400, the batch size is set to be 50, and after the model parameters are set, the data obtained in the step 3 are input into a model for hand tremor recognition.
Experimental analysis
1 Experimental data set
The present invention is directed to a method for identifying hand tremor in a state where the body is not stationary, and therefore, it is necessary to complete the study using phase data of hand tremor movements accompanying physical activity. The invention collects 3 groups of phase data of 'compound actions', namely phase data of hand tremors in the bending process, hand tremors in the squatting process and hand tremors in the standing process. In the process of collecting each set of data, the body is ensured to have only one physical activity except for the hand tremor action. Each group uses two passive tags, attached to the subject's chest and hand, respectively. The invention can separate the phase data of hand tremor action from the three acquired 'compound actions' by a pair of passive tag groups based on the data preprocessing method in 4.3.
The invention regards 10 consecutive phase sequences as one piece of data, 300 pieces of each set of separated hand tremor phase data, and 900 pieces of data in total. The dataset was randomly split into training samples and test samples (ratio of 7:3) prior to training. And in the experimental process, the convolution kernel size, the learning rate and the Batchsize of the CNN-LSTM hand tremor recognition model are adjusted to obtain an optimal recognition model.
Model 2 training and parameter set-up
The model training process of the invention is similar to the CNN model training process of the third chapter, and the parameters of the model training of the invention mainly comprise: LSTM layer neuron number, learning rate, and Batch size. The initial parameters were set as: the LSTM layer neurons were 30 in number, learning rate was 0.0006, epoch was 400, batch sample Batch size was 50 each time, and initial recognition rate based on this was 93.35%. Fig. 6 (a) and 6 (b) show the Accuracy curve and the Loss curve of the model with initial parameters set.
(1) LSTM layer neuron count
The number of the LSTM layer neurons is a very important parameter, and when the number of the neurons is too small, the learning ability and the information processing ability of the network are poor, so that the recognition rate is low. And when the number of neurons is too large, the network structure becomes more complex. In order to verify the influence of the number of the LSTM layer neurons on the recognition result, the experiment of the invention respectively verifies the recognition rates when the number of the neurons is 6, 12, 18, 24, 30, 36, 42 and 48 on the premise of not changing the initial learning rate and the Batch size, and the result is shown in fig. 7. After changing the number of neurons, fig. 8 (a) and 8 (b) show the Accuracy curve and the Loss curve of the new model.
Since the neurons set to 36 have good performance, and increasing the number of neurons only slows down the learning speed of the network, the invention sets the LSTM layer neurons to 36.
(2) Learning rate determination
The invention sets learning rates of 0.0001, 0.0002, 0.0004, 0.0006, 0.0008, 0.001, 0.005 and 0.01 when training the model. The recognition rates of the different learning rates are shown in table 1.
Table 1 recognition rates corresponding to different learning rates
Tab.1 The Recognition Rate Corresponding to Different Learning Rate
| Learning rate | Recognition rate |
| 0.0001 | 0.9447 |
| 0.0002 | 0.9487 |
| 0.0004 | 0.9491 |
| 0.0006 | 0.9456 |
| 0.0008 | 0.9380 |
| 0.001 | 0.9335 |
| 0.005 | 0.9392 |
| 0.01 | 0.9314 |
In the experimental process, as shown in fig. 9 (a) and 9 (b), the loss value of the model with the learning rate of 0.0002 is better than that of the model with the learning rate of 0.0004, and under the condition that the recognition rate is not greatly different, the invention selects to set the learning rate to 0.0002 and completes the subsequent parameter adjustment.
(3) Batch size determination
FIG. 10 shows the recognition rates for different Batch sizes, 10, 20, 30, 40, 50, 60, 70, 80, 90 and 100, respectively, with the highest recognition rate seen for a Batch size of 60.
Fig. 11 (a) and 11 (b) are an Accuracy curve and a Loss curve of the CNN-LSTM-based hand tremor recognition model according to the present invention, which have been adjusted by LSTM layer neuron number, learning rate and Batch size parameter, and the recognition rate of the model is 94.96% at the highest. Figure 12 shows the recognition accuracy, recall and F1-score of hand tremor motion separated from 3 "compound motions".
After the above-mentioned series of parameter adjustment and model training results are compared, specific parameters of the CNN-LSTM-based hand tremor recognition model of the present invention are shown in table 2.
Table 2 list of model parameters
| Parameters (parameters) | Value of |
| Convolution kernel size | 3 |
| Pool size | 3 |
| Activation function | Relu |
| LSTM layer number | 1 |
| Number of neurons per layer | 36 |
| Dropout | 0.5 |
| Learning rate | 0.0002 |
| Number of samples in batch | 60 |
| epoch | 400 |
3 comparison and analysis of experiments
In order to verify the effectiveness of the CNN-LSTM-based hand tremor identification method provided by the invention, the CNN, LSTM and CNN-LSTM models are used for experimental comparison and analysis respectively, so that the superiority of the CNN-LSTM model is embodied.
TABLE 3 identification rates of different models on the same dataset
| Method | Recognition rate |
| CNN | 93.80％ |
| LSTM | 93.02％ |
| CNN-LSTM | 94.96％ |
As can be seen from Table 3, the recognition rate of all the methods reaches more than 90%, and the recognition rate of the method is highest and is 94.96%, which is superior to other methods. Fig. 13 shows Precision, recall and F1-score when Macro average (Macro avg) evaluations were performed on 3 models.
The invention uses hand tremor motion data separated from 3 kinds of compound motions to respectively test the good and bad performances of CNN, LSTM and CNN-LSTM hand tremor recognition models. It can be seen from fig. 13 that the evaluation criteria of the method of the present invention are all higher than those of the first two methods.
During the course of the experiment it was found that the experimental results for LSTM were not always lower than for CNN. Analysis of the data and recognition models revealed that the recognition accuracy of CNN and LSTM was related to the change in hand tremor reference point and tremor rate of the subject. In the case where the hand tremor reference point is frequently changed and the hand tremor rate fluctuation is small, the recognition accuracy of CNN is low with respect to LSTM. In the case of frequent changes in hand tremor rate, the recognition accuracy of LSTM is low relative to CNN.
The CNN-LSTM model does not generate the condition of large fluctuation of accuracy rate when processing the data in the two cases, which shows that the stability of the CNN-LSTM model is superior to the stability of the CNN-LSTM model and the data in the two cases. The CNN-LSTM model combines the advantages of CNN feature extraction and LSTM processing time sequence data, deep features can be extracted, correlation among features with larger time span is reserved, and therefore a more accurate recognition result is obtained. Taken together with the above comparative experiments, it can be seen that the CNN-LSTM model performs better than the CNN and LSTM models, with better performance on the dataset.
Knot (S)
The invention provides and realizes a hand tremor identification method in a non-stationary state of the body, and solves the problem that other physical activities of the body cannot be limited when the hand tremors are detected. First, a mathematical model of hand tremor in an ideal state is established. Then, the discrete wavelet transform is used to decompose the phase signal of hand tremor action and the maximum and minimum threshold algorithm is used to complete the threshold selection. Then, the time-varying direct current component is removed by using a method of calculating a local mean value within the sliding window. Finally, a CNN-LSTM based hand tremor recognition model is presented. The optimal model is obtained through parameter tuning, and experimental comparison is completed with the CNN model and the LSTM model, and the recognition accuracy is 94.96%.
Claims (1)
1. The method for identifying the tremor of the hands in the non-stationary state of the body is characterized by comprising the following steps of:
step 1, establishing a hand tremor mathematical model;
the step 1 is specifically implemented according to the following steps:
let a hand carrying a passive tag tremors in which tremor frequency r varies with time and tremor amplitude g also varies with time, in which case θ (t) is no longer the ideal periodic curve, equations (1) and (2) are rewritten as:
r (t) represents the distance from the passive tag to the RFID antenna, tremor amplitude g (t), tremor frequency R, reference distance d (t), time t, phi is the acute angle between tremor direction and the line from the RFID antenna to the reference position, θ (t) represents the phase, θ 0 Is the offset introduced by the hardware device, λ is the radio frequency wavelength, where ε is the error caused by ambient noise; the distance R (t) and the phase theta (t) together form a hand tremor mathematical model;
step 2, data calibration and noise elimination;
the step 2 is specifically implemented according to the following steps:
step 2.1, decomposing the hand tremor phase θ (t) into:
where θ (t) is the phase, l 0 Is the initial resolution, t represents time, l is the scale factor, k is the pan factor, φ l0,k (t) is a discrete scaling function, ψ l,k (t) is a discrete wavelet function, the two functions are orthogonal to each other, A (l) 0 The approximate and detailed coefficients of each stage, which are filtered during denoising, are defined as:
θ (t) is the phase, φ l0,k (t) is a discrete scaling function, ψ l,k (t) is a discrete wavelet function, A (l, k) and D (l, k) are eachApproximation and detail coefficients of the first order;
step 2.2, performing threshold selection by using a maximum and minimum threshold algorithm, wherein the definition of the optimal threshold is shown in a formula (8):
wherein,representing the threshold value, σ is the standard deviation, +.>Values referring to the radio frequency carrier wavelength, < >>Obtained by optimizing the objective function as shown in formula (9):
in the above equation, λ is the radio frequency carrier wavelength,referring to the optimized wavelength, d represents one of n coefficients obtained by discrete wavelet transform, and the risk is calculated as R λ (d)＝E(δ λ (d)-d) 2 ，R ideal Is an ideal result, indicating whether the wavelet coefficients are preserved, given by:
R ideal (d):＝min(d 2 ,1) (10)
step 2.3, finally, based on the threshold lambda M Selecting an approximation coefficient and a detail coefficient, and reconstructing a noise-free signal by using the approximation coefficient and the detail coefficient, wherein the phase data set after noise elimination is expressed as theta;
step 3, eliminating time-varying direct current components and calculating time-varying tremor frequency;
the step 3 is specifically implemented according to the following steps:
step 3.1, eliminating time-varying direct current components:
step 3.1.1, setting a phase sample θ (t) ∈θ after data calibration and noise removal, where the following relationship exists:
where a is a constant DC component and M (t) is a time-varying DC component due to movement of the reference position of the handIs a tremor component caused by tremor of the hand around a reference position, if the hand tremor motion is a stable motion separated from the "compound motion +.>Expressed by formula (4);
step 3.1.2 using a sliding window based method, the local average value l (t) of time t within window [ t- (W/2), t+ (W/2) ] is defined as:
where W is the window size, the constant DC component a, M (t) is the time-varying DC component,is the tremor component, and in a small time window, assuming that M (t ') is approximately a linear function and θ (t') is approximately a periodic function, the relationship shown in equation (13) is obtained by equation (12):
m (t) is M (t ') in window t' ∈ [ t- (W/2), t+ (W/2)]Mean of the interior, in the formulaIs a curve of hand tremor action after removal of the direct current component, therefore +.>The integral over each period is zero;
step 3.1.3, l (t) is expressed by equation (14):
l(t)≈a+M(t) (14)
l (t) is the time-varying direct current component of the hand tremor to be removed, and the value of l (t), i.e. the value of a+m (t), can be approximated from equation (12) based on the phase data acquired in the time window [ t- (W/2), t+ (W/2) ] assuming S (t) is a subset of the phase sequence in the window, and for each sample θ (t ') acquired at a particular time t ', Δ (t ') is the time interval from that sample to the next sample, the integral in equation (12) is approximated with discrete samples as follows:
wherein the value of W should be determined according to the application environment, and the local average value l (t) is calculated, and then l (t) is subtracted from θ (t), as shown in formula (16):
the phase set after the data denoising and the time-varying direct current component elimination is set asSubtracting chest label from phase of hand labelPhase curves of hand tremor actions are separated from the phases of the hand tremor actions;
step 3.2, calculating time-varying tremor frequency:
step 3.2.1 according to the time windowCalculating the tremor frequency r (t) at time t from the phase data of (c), wherein +.>Is the window size, first, the peak or valley value within the phase curve window needs to be determined, then the peak-to-peak distance is used to estimate the tremor period and calculate the tremor frequency,
assume thatIs derived from phase samples->At->A subset of the windows, after identifying all peaks on the window, using least squares to obtain an estimate of the tremor period p (t), representing the peak-to-peak spacing as { v } 1 ,v 2 ,...,v m The least squares estimate is given by equation (17):
step 3.2.2 in WindowThe estimation of tremor frequency r (t) at time t can be calculated from equation (18):
r(t)＝1/p(t) (18)
throughout the measuring periodIs expressed as the average tremor frequency ofThe definition is as shown in formula (19):
wherein t is 1 Is the time to start measurement, t n It is the time at which the measurement is ended,
step 3.2.3 assuming n samples belong toThe calculation is shown in formula (20):
wherein Δ (t) is the slave sampleTime interval to next sample;
step 4, a hand tremor identification method based on CNN-LSTM;
the step 4 is specifically implemented according to the following steps:
the CNN-LSTM model consists of two parts, wherein the CNN-LSTM model consists of an LSTM layer containing 36 neurons, a full-connection layer, a batch normalization layer and an output layer, the LSTM layer is provided with dropout with a value of 0.5, an Adam optimizer is used for parameter optimization, the initial learning rate is set to be 0.0006, epoch is set to be 400, batch size is set to be 50, and after model parameters are set, the data obtained in the step 3 are input into the model for hand tremor recognition.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111011247.XA CN113822170B (en) | 2021-08-31 | 2021-08-31 | Method for identifying tremor of hands of body in non-stationary state |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111011247.XA CN113822170B (en) | 2021-08-31 | 2021-08-31 | Method for identifying tremor of hands of body in non-stationary state |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113822170A CN113822170A (en) | 2021-12-21 |
| CN113822170B true CN113822170B (en) | 2024-02-06 |
Family
ID=78913875
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111011247.XA Active CN113822170B (en) | 2021-08-31 | 2021-08-31 | Method for identifying tremor of hands of body in non-stationary state |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113822170B (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US4632341A (en) * | 1985-02-06 | 1986-12-30 | The United States Of America As Represented By The Secretary Of The Air Force | Stabilizing force feedback in bio-actuated control systems |
| CN102272327A (en) * | 2008-11-07 | 2011-12-07 | 赛昆塔公司 | Methods of monitoring conditions by sequence analysis |
| RU2690001C1 (en) * | 2017-12-29 | 2019-05-30 | Федеральное государственное бюджетное образовательное учреждение высшего образования "Московский государственный университет имени М.В. Ломоносова" (МГУ) | Method of processing vector signals for pattern recognition based on wavelet analysis |
| CN112075940A (en) * | 2020-09-21 | 2020-12-15 | 哈尔滨工业大学 | Tremor detection system based on bidirectional long-time and short-time memory neural network |
| CN112580588A (en) * | 2020-12-29 | 2021-03-30 | 西北工业大学 | Intelligent flutter signal identification method based on empirical mode decomposition |
-
2021
- 2021-08-31 CN CN202111011247.XA patent/CN113822170B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US4632341A (en) * | 1985-02-06 | 1986-12-30 | The United States Of America As Represented By The Secretary Of The Air Force | Stabilizing force feedback in bio-actuated control systems |
| CN102272327A (en) * | 2008-11-07 | 2011-12-07 | 赛昆塔公司 | Methods of monitoring conditions by sequence analysis |
| RU2690001C1 (en) * | 2017-12-29 | 2019-05-30 | Федеральное государственное бюджетное образовательное учреждение высшего образования "Московский государственный университет имени М.В. Ломоносова" (МГУ) | Method of processing vector signals for pattern recognition based on wavelet analysis |
| CN112075940A (en) * | 2020-09-21 | 2020-12-15 | 哈尔滨工业大学 | Tremor detection system based on bidirectional long-time and short-time memory neural network |
| CN112580588A (en) * | 2020-12-29 | 2021-03-30 | 西北工业大学 | Intelligent flutter signal identification method based on empirical mode decomposition |
Non-Patent Citations (2)
| Title |
|---|
| 武文豪 ; 吴曦 ; 沈林勇 ; 钱晋武 ; .基于可穿戴传感器的帕金森病患者肢体末端震颤的检测.生物医学工程研究.2017,(第01期),全文. * |
| 艾玲梅 ; 王珏 ; .基于EMD奇异值分解诊断震颤的新方法.生物医学工程学杂志.2009,(第06期),全文. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113822170A (en) | 2021-12-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106407905B (en) | Wireless aware action identification method based on machine learning | |
| EP3790463A1 (en) | Multiclass classification method for the estimation of eeg signal quality | |
| Azami et al. | Spike detection approaches for noisy neuronal data: assessment and comparison | |
| CN114048773A (en) | Behavior identification method and system based on transfer learning and WiFi | |
| Biagetti et al. | Reduced complexity algorithm for heart rate monitoring from PPG signals using automatic activity intensity classifier | |
| Higashi et al. | Smoothing of spatial filter by graph Fourier transform for EEG signals | |
| CN109768838B (en) | Interference detection and gesture recognition method based on WiFi signal | |
| Tao et al. | Second-Order Adaptive Synchrosqueezing ${S} $ Transform and Its Application in Seismic Ground Roll Attenuation | |
| CN112107310A (en) | ECG identity recognition method based on IWT and AGA-BP models | |
| CN109567798A (en) | Daily behavior recognition methods based on myoelectricity small echo coherence and support vector machines | |
| CN113822170B (en) | Method for identifying tremor of hands of body in non-stationary state | |
| Sreeshakthy et al. | A survey on emotion classification from EEG signal using various techniques and performance analysis | |
| Pravin et al. | A novel ecg signal denoising filter selection algorithm based on conventional neural networks | |
| CN108898062B (en) | Hand motion recognition method based on improved signal segment extraction algorithm | |
| Khalefa et al. | Fingerprint image enhancement by develop mehtre technique | |
| Azami et al. | Automatic signal segmentation based on singular spectrum analysis and imperialist competitive algorithm | |
| Mihandoost et al. | EEG signal analysis using spectral correlation function & GARCH model | |
| Keković et al. | Detection of change points in time series with moving average filters and wavelet transform: Application to EEG signals | |
| Shcherbakova et al. | Determination of characteristic points of electrocardiograms using multi-start optimization with a wavelet transform | |
| Awad et al. | Fingerprint singularity detection: A comparative study | |
| Sorvillo et al. | Single beat ECG-based Identification System: Development and robustness test in different working conditions | |
| Zhang et al. | E-nose algorithms and challenges | |
| CN110327043B (en) | Event-related potential waveform map solving method based on sparse modeling | |
| CN113836995B (en) | Age group identification method and device | |
| Sattar et al. | Time-scale-based segmentation for degraded PCG signals using NMF |
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 |