KR20180059200A - System for estimating joint angle based on neural signals - Google Patents

Abstract

In a living body where a joint angle is determined by the movement of muscles connected to a plurality of distal nerves branched from a proximal nerve, a joint angle detection system for detecting the angle of a joint comprises a collecting electrode body installed on the proximal nerve and collecting the nerve signal of the proximal nerve (proximal nerve signal), and an angle estimator for estimating the angle of the joint by analyzing the proximal nerve signal. It is possible to estimate an exact joint angle.

Description

[0001] The present invention relates to a system for estimating joint angles based on neural signals,

The present invention relates to an angle detection system of a joint, and more particularly, to an angle detection system capable of estimating an angle of a joint based on a nerve signal and using it for angle control of a joint.

Functional neuromuscular stimulation (FNS) is an effective way to rehabilitate the function of weakened or missing muscles. Stimulation of the intact peripheral nerves or muscles allows patients with nerve damage to perform functional movements. The most widely used FNS system in the rehabilitation area is to change the intensity of the stimulus through a predetermined stimulation parameter so that the desired movement can be controlled. However, this traditional approach requires continuous or repetitive stimulation, which can lead to muscle fatigue or tissue damage.

Accordingly, recently, an FNS system using a signal collected using an artificial sensor or a natural sensor as a feedback signal has been proposed. The robustness feedback signal, which is said to be closed loop control, needs to be provided at the precise timing at which the stimulus is applied.

Artificial sensors such as goniometer and foot switches are widely used in closed-loop FNS systems. However, such an artificial sensor has limitations in terms of cosmetics, and there is a limitation in that there is a need for correction due to environmental disturbance.

Instead of using such an artificial sensor, a natural sensor has been used which provides useful information about a particular target sensory or motor organization. For example, there has been a case in which an afferent activity of the skin generated by an external water-soluble sensor in the skin has been used to shake hands or repair a foot drop, and a muscle sphere generated by a self- There is a case in which mental activity is used to control ankle angle.

The collection of neural signals by the cuff-shaped electrodes is known to be an effective approach to enable the provision of the necessary information and feedback control in the FNS system.

For example, in the case of controlling the ankle angle of a patient having a motor function in the future, several methods of using the sensory information extracted by the cuff-type electrode recording have been proposed.

According to one example, a closed loop FNS system using a muscle afferent signal as a feedback signal has been proposed. In order to estimate the ankle angle in these systems, electrodes are implanted in the tibial nerve and the peroneal nerve, respectively, and an attempt is made to estimate ankle angle from two independent nerve signals.

However, according to this conventional technique, in order to measure an afferent signal from a desired muscle, a cuff-shaped electrode should be implanted in the distal branch of the multifibular nerve, but the branch of the nerve is always too short and thin, To the implantation site is too difficult.

Korea Patent No. 10-1418057

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems of the conventional art, and it is an object of the present invention to provide an electrode body on a distal nerve having a relatively large space and to accurately estimate a joint angle using a nerve signal measured from the distal nerve, And an object of the present invention is to provide a joint angle detection system capable of controlling a joint angle.

In order to achieve the above object, according to one aspect of the present invention, in a living body in which an angle of a joint is determined by a movement of muscles connected to a plurality of distal nerves branched from the proximal nerve, A detection system comprising: a collection electrode body disposed on the proximal nerve and collecting neural signals of a proximal nerve (" proximal nerve signal "); and an angle estimator for estimating an angle of the joint by analyzing the proximal nerve signal Angle detection system.

According to one embodiment, the angle estimator separately extracts afferent nerve signals (" distal nerve signals ") from each of the distal nerves to the proximal nerve in the proximal nerve signal, Using the correlation of the angles of the joints and the angle of the joint, we estimate the angle of the joint from the extracted distal nerve signals.

According to one embodiment, the collecting electrode body is a multi-channel electrode body having n channels (n is a natural number larger than 2), and the proximal nerve signal is collected into n channel nerve signals.

According to one embodiment, the motion of the joint involves m (m is a natural number greater than 1) distant nerve, and the angle estimator includes a projection pursuit (PP) for projecting a higher order signal to a lower order signal ) Technique and a Fast ICA (Fast Independent Component Analysis) technique to find orthogonal rotations of multiple signals are used to extract m independent independent nerve signals from each distal nerve .

According to one embodiment, the angle estimator estimates an angle of a current joint through an artificial neural network having input values of extracted extrinsic nerve signals and an estimated joint angle.

According to one embodiment, the artificial neural network is a circular neural network that uses an estimated angle value of a currently jointed joint as an input value.

According to one embodiment, the angle estimator comprises a neural signal separation module for separating the proximal nerve signals into independent distal nerve signals, a noise removal module for removing noise from the distal nerve signals, And an angle estimation module for estimating the angle of the joint from the nerve signals.

According to one embodiment, the joint angle detection system includes an angle controller for generating a joint angle control signal for generating an angle of the joint that follows the predetermined target angle value, using the estimated joint angle value as a feedback signal do.

According to one embodiment, a joint angle detection system comprises: a stimulation pulse generator for generating a stimulation pulse for stimulating each distal nerve to cause activity of muscles connected thereto based on the joint angle control signal; And an electrical stimulator for causing the corresponding electrical stimulation to be applied to the distal nerve through the stimulating electrode body provided on each of the distal nerves.

According to one embodiment, a joint angle detection system includes a signal processor for pre-processing and transmitting a proximal nerve signal to the angular estimator, wherein the stimulus pulse generator includes a blanking signal corresponding to the period of the stimulation pulse, To the signal processor and the signal processor superimposes the erase signal on the proximal nerve signal to remove the disturbing nerve signal generated by the electrical stimulus.

According to one embodiment, the proximal nerve is a sciatic nerve, the distal nerve is a tibial nerve and a peroneal nerve that diverge from the sciatic nerve, and is used to control the walking and balance of the living body by estimating the angle of the ankle joint of the leg .

According to one embodiment, the collecting electrode body is a cuff-shaped electrode body in which a plurality of electrodes form a plurality of channels, and the outer periphery of the distal nerve is curled.

1 is a conceptual diagram of a joint angle detection system according to an embodiment of the present invention.
Fig. 2 shows a collecting electrode body used in the joint angle detecting system of Fig. 1. Fig.
3 is a conceptual diagram of a data acquisition system according to an example for configuring the motion logic of the joint angle detection system of FIG.
Figure 4 is a graph showing data collected through the data acquisition system of Figure 3;
5 is a functional block diagram of an angle estimator of the joint angle detection system of FIG.
FIG. 6 is a flowchart for explaining the entire process of the PP / Fast ICA technique used in the neural signal separation according to an embodiment.
7 is a graph showing the nerve signals separated by the angle estimator and the ankle joint angle estimated therefrom.
8 is a conceptual diagram of an artificial neural network using joint angle estimation according to an embodiment.
9 is a graph showing the neural signals separated according to various signal separation methods.
10 is a graph showing a stimulus pulse generated in the stimulus pulse generator of the joint angle detection system of FIG.
11 is a conceptual diagram of a verification system for verifying the validity of the joint angle detection system of FIG.
FIG. 12 is a graph showing an ankle angle measured in accordance with an electric stimulus in the verification system of FIG. 11, and a 5-channel neural signal collected at this time.
13 is a graph showing tibial nerve signals and peroneal nerve signals separated from the 5-channel eraser nerve signals and ankle angles estimated from the corresponding nerve signals.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, preferred embodiments of the present invention will be described with reference to the accompanying drawings. Although the present invention has been described with reference to the embodiments shown in the drawings, it is to be understood that the technical idea of the present invention and its essential structure and operation are not limited by the embodiments.

1 is a conceptual diagram of a joint angle detection system 1 according to an embodiment of the present invention.

The joint angle detection system (1) collects and analyzes neural signals ("proximal nerve signals") measured from the proximal nerve in living bodies whose joint angles are determined by the movement of muscles connected to a plurality of distal nerve branches branching from the proximal nerve And a closed loop FNS system that estimates the angle of the joint, and further generates an angle of the desired joint by applying an appropriate electric stimulus to the estimated joint angle as a feedback signal.

According to this embodiment, the joint angle detection system 1 is applied to the legs 8 of New Zealand white rabbits to estimate the angle? Of the ankle joint 7.

As shown in Fig. 1, the legs 8 include a sciatic nerve 2, which is a distal nerve that descends along the thigh, and a proximal nerve branching from the sciatic nerve 2, It includes the tibial nerve (3) and the peroneal nerve (4). The tibial nerve (3) and the peroneal nerve (4) are composed of a dorsi-flexor muscle (5) causing plantar flexion and a plantar-flexor muscle (6) causing plantar flexion .

The angle? Of the ankle joint 7 is determined by the contraction or relaxation movement of the footbath root 5 and the sole bending root 6. Muscle afferent signals are generated when the instep bend muscle (5) and foot bend muscle (6) move, and muscle afferent signals generated by each muscle are transmitted to the tibial nerve (3) and the peroneal nerve (4).

In the present specification, an afferent nerve signal transmitted to the sciatic nerve (2), which is a proximal nerve via the tibial nerve (3) and the peroneal nerve (4), is referred to as a " distal nerve signal ".

According to the present embodiment, in order to estimate the angle? Of the ankle joint 7 and to control the angle of the ankle joint based on the angle?, A collecting electrode body for collecting the proximal nerve signal of the sciatic nerve 2, (10) is transplanted to the sciatic nerve (2).

The joint angle detection system 1 further includes a signal processor 100, an angular estimator (not shown) for estimating the angles of the joints and controlling the angles of the joints by processing the proximal nerve signals collected by the collecting electrode body 10 an angle estimator 200, an angle controller 300, a stimulation pulse generator 400 and an electrical stimulator 500.

According to this embodiment, the magnetic pole electrode bodies 21 and 22 capable of applying electrical stimulation for the angle control of the ankle joint are provided on the tibial nerve 3 and the peroneal nerve 4, respectively, which are the distal nerves.

The signal processor 100 performs a preprocessing process of amplifying and correcting the proximal nerve signals collected from the collecting electrode unit 10. [ The angle estimator 200 analyzes the proximal nerve signal processed in the signal processor 100 to estimate the angle of the joint 7. The angle controller 300 generates a joint angle control signal for generating an angle of the joint following the predetermined target angle value by using the angle value of the joint estimated by the angle estimator 200 as a feedback signal. The stimulus pulse generator 400 generates a stimulus pulse for stimulating each distal nerve 3, 4 and causing the activity of the muscles 5, 6 connected thereto, based on the joint angle control signal. The electric stimulator 500 allows the electric stimulation corresponding to the stimulation pulse to be applied to the distal nerve 3, 4 via the stimulating electrodes 21, 22 provided on the respective distal nerves 3, 4.

2 shows the collecting electrode unit 10 according to one embodiment.

The collecting electrode unit 10 according to the present embodiment includes a substrate 11 made of polyimide and a cuff-type electrode 10 having ten electrodes 12 made of platinum on the substrate 11. It is an electrode body.

The substrate 11 is self-biased so as to be curled so as to have an inner diameter, and is formed so as to surround the outer periphery of the sciatic nerve 2. The collecting electrode unit 10 according to the present embodiment includes a pair of electrodes 12 arranged in parallel on both sides of the collecting electrode unit 10 to form a single channel to form a total of five channels (ch 1 to ch 5) It is a sieve.

Each electrode 12 has a width of 1 mm and a length of 0.5 mm. When dried, the diameter of the substrate 11 is 1.8 mm, and the distance between the pair of electrodes 12 forming the channel is 9.25 mm.

The collecting electrode body 10 is located at a position approximately 1 cm above the tibial nerve 3 and the branch point of the sciatic nerve 2 branching to the peroneal nerve 4, Proximal nerve signals) are collected. At this time, the proximal nerve signal is collected into five channel nerve signals (see FIG. 3).

The angle estimator 200 according to the present embodiment calculates the angular position of the sciatic nerve (3) and the peroneal nerve (4), which are the distal nerves in the nerve signal (proximal nerve signal) of the sciatic nerve (2) And extracts an afferent nerve signal (distant nerve signal) transmitted to the nerve 2. Further, the angular estimator 200 estimates the angle of the joint from the extracted distal nerve signals using the correlation of the angles of the joint with the learned nerve signals.

In order to construct the angle estimation algorithm of the angle estimator 200, first, a data acquisition system is constructed to collect data.

3 is a conceptual diagram of a data acquisition system according to an example.

3, the collecting electrode body 10 is implanted into the sciatic nerve 2 of the leg 8 of the same kind of rabbit used in the joint angle detecting system 1. [ A cuff-shaped electrode body (not shown) is implanted into the tibial nerve 3 and the peroneal nerve 4, respectively.

The collecting electrode body 10 collects the nerve signals of the sciatic nerve 2 and the cuff-shaped electrode bodies implanted in the tibial nerve 3 and the peroneal nerve 4 receive the tibial nerve 3 and the peroneal nerve 4 To collect neural signals.

A signal processor (100) is connected to the collecting electrode body (10). In the data acquisition system, the signal processor 100 comprises an implantable amplifier module, an external amplifier module, and a converter to amplify and digitize the sciatic nerve signal (proximal nerve signal) collected at the collecting electrode unit 10. Each amplifier module has a gain of 39601 and a -3 dB bandwidth of 425 to 5500 Hz. The implantable amplifier module can be implanted in the subcutaneous space in the hips of the rabbit, and a percutaneous connector is installed across the buttock ridge.

The angle (θ) of the ankle joint was measured using an electronic goniometer (SG 65, Biometrics) (goniometer) 32 located on the lateral side of the shin bone and foot. The final output of the external amplification module, the 5-channel sciatic nerve signal and the angle of the ankle joint, is digitized using an AD converter with a resolution of 16 bits and a sampling period of 25 kHz.

The cuff-shaped electrode bodies implanted in the tibial nerve 3 and the peroneal nerve 4 are short-channel cuff-shaped electrode bodies. During passive ankle joint movement, tibial nerve signals and peroneal nerve signals were measured by each of the cuffed electrode bodies. These proximal nerve signals are used to construct a processing algorithm for multi-channel signals collected from the sciatic nerve.

To create a passive motion of the ankle joint, a computer controlled position control system is used. The shoe (31) was inserted into the foot of the rabbit, and the movement of the ankle joint was caused by moving the shoe using a servo motor (not shown). The motor shaft of the servomotor is aligned with the ankle axis covered with the shoe 31. At this time, the knee fastener 30 for preventing the knee from moving together is fixed to the knee.

The position control system rotates the ankle along the sine wave trajectory on the sagittal plane.

The angle? Of the ankle joint is defined as the inner angle between the shin bone and the foot using the coordinate system shown in Fig. In the neutral position, the angle of the ankle joint is defined as 100 °, and the angles at maximum bending (foot bending) and maximum spreading (sole bending) are defined as 70 ° and 130 °, respectively.

To validate the applicability of the ankle joint angle estimation algorithm presented herein to various joint angle frequencies, the ankle joints were controlled to operate along sinusoidal trajectories at different frequencies (0.3 Hz, 0.5 Hz). These data acquisition systems were applied to 5 average rabbits and data were collected. Each rabbit experimented with 10 sessions at each frequency, and each session lasted 20 seconds.

Figure 4 is a graph showing data collected through the data acquisition system of Figure 3;

As shown in FIG. 4, 5-channel multi-channel neural signals were collected from the sciatic nerve 2 (multi-channel cuff recording) and nerve signals were collected from the tibial nerve 3 and the peroneal nerve 4, respectively (Short channel cuff recording). FIG. 4 is a graph showing the results for rabbit A among 5 rabbits.

The ankle angle plot at the top of the ankle shows the trajectory of the ankle's ankle appearing during passive movement of the ankle where bending and unfolding of the ankle has a sine wave trajectory with a frequency of 0.5 Hz at a range of 60 degrees. .

The short channel cuff recording plot shows the neural signals collected from the tibial nerve 3 and the peroneal nerve 4, respectively. It can be seen that the tibial nerve and peroneal nerve signals have a phase delay of about 90 ° during an ankle joint angle change to a sine wave.

The multi-channel cuff recording plot shows the sciatic nerve signals collected by the 5-channel neural signal. Since no external stimulus is applied to the tibial nerve (3) and the peroneal nerve (4), the sciatic nerve signal can be treated as a superposition of the tibial nerve signal and the peroneal nerve signal.

The angle estimator 200 according to the present embodiment estimates the 5-channel neural signal collected from each channel of FIG. 4 to a maximum of the measured distant neural signals (short channel cuff recording) through a blind source separation process The tibial nerve signal and the peroneal nerve signal, and the extracted tibial nerve signal and the peroneal nerve signal are configured to estimate the angle of the ankle joint as closely as possible to the measured ankle angle shown in Fig.

FIG. 5 is a functional block diagram of the angle estimator 200. FIG.

5, the angular estimator 200 according to the present embodiment may include a neural network that separates the sciatic nerve signals (proximal nerve signals) into independent tibial nerve signals and fibula signals (distal nerve signals) A separation module 210, a noise removal module 220 for removing noise from the distal nerves, and an angle estimation module 230 for estimating the angle of the joint from the noise-canceled distal nerve signals.

According to the present embodiment, the neural signal separation module 210 of the angle estimator 200 includes a projection pursuit (PP) technique for projecting a high dimensional signal into a low dimensional signal, , A non-illustrated linear separation algorithm that performs a blind source separation technique (" PP / Fast ICA ") that applies a fast independent component analysis (Fast ICA) technique to find orthogonal rotations of multiple signals Lt; / RTI >

The PP scheme is a linear transformation technique that can preserve the underlying shape of high-order scattered variance and determine projections that maximize mutually independent orders in the estimated source signals.

The sciatic nerve (2) has independent nerve sources such as the tibial nerve and peroneal nerve in the body of the nerve.

These nerve signals are projected onto the surface area of the torso of the sciatic nerve (2). Therefore, the multi-channel signals collected at each electrode of the collecting electrode unit 10 can be modeled as a linear mixture of signals projected from all the signal sources as shown in Equation (1) below.

[Equation 1]

Here, X is a n-th (n is a natural number greater than 2) measurement signal collected from the multi-channel cuff-shaped electrode unit, S is a m-th source signal (m is a natural number greater than 1) It is a demixing matrix.

If an mxn separation matrix W capable of maximizing the mutual independence of the estimated source signals can be found, the estimated source signal S can be obtained as shown in the following equation (2).

&Quot; (2) "

를 근사화하는 목적 함수가 독립 성분 분석을 위한 빠르고 강인한(fast and robust) 알고리즘으로 개발되었으며, 이를 소위 "FastICA"라고 칭한다. According to the present embodiment, in order to find the separation matrix W, the fact that the objective function measuring the mutual independence of the source signals using the PP technique is equivalent to the objective function for measuring the negative entropy of the variable Respectively. By applying the Fast ICA technique to the PP technique, the problem of shortening the processing time is solved. More specifically, in order to estimate the source signal S,

Is developed as a fast and robust algorithm for independent component analysis and is called the so-called " Fast ICA ".

는

로 정의된다. 여기서, H(S)는 논-가우시어니티(non-Gaussianity)의 척도를 직접 나타내며, 가우시안 랜덤 벡터(Gaussian random vector)에 대해 0으로 설정되는 소스 신호(S)의 엔트로피이다. 또한, S_G는 소스 신호(S)와 동일한 평균과 분산을 가지는 가우시안 랜덤 벡터이다.

를 최대화하는 디믹싱 행렬(W)의 값을 계산하기 위해, 회분식(batch-type) 학습 알고리즘을 이용하였다.

The

. Where H (S) is the entropy of the source signal S, which directly represents a measure of non-Gaussianity and is set to zero for a Gaussian random vector. Also, S _G is a Gaussian random vector having the same mean and variance as the source signal S.

A batch-type learning algorithm was used to calculate the value of the demixing matrix W to maximize the sum of the sum of squares.

6 is a flowchart for explaining the entire process of the PP / Fast ICA technique described above.

형태로 주어지는 혼합 신호 데이터를 입력값으로서 수집한다. 본 실시예에서 혼합 신호는 수집 전극체(10)를 통해 수집되는 5-채널 신경 신호이며, N과 n은 5가 된다. (1) First,

And collects the mixed signal data given as a form as an input value. In this embodiment, the mixed signal is a 5-channel neural signal collected through the collecting electrode body 10, and N and n are 5.

의 관계에 의해 데이터를 센터링(centering)하고,

의 관계에 의해 데이터를 백색화(whiten)한다. 여기서, M은 백색화 행렬로

이고, E는 고유 벡터(eigen vector)의 행렬이며, V는 고유 값(eigen value)들의 대각 행렬이다. (2) Next,

The data is centered by the relationship of "

Quot; whiten " the data. Here, M is a whitening matrix

E is a matrix of eigenvectors, and V is a diagonal matrix of eigenvalues.

(3) Next, the number of projection tracking directions ICs is set to m, and a random initial transmission matrix W of nxm size is selected.

의 관계를 이용해 최적화된 투영 방향을 계산해 행렬 W를 업데이트한다. 여기서, G는

으로 가정하며, U는

로 주어진다. (4) Next,

And updates the matrix W by calculating the optimized projection direction. Here, G

, And U is

.

의 관계에 의해 행렬 W를 직교화한다(orthogonalize).(5) Next,

The matrix W is orthogonalized.

(6) Next, the degree of similarity between the previous matrix W (W _pre ) and the updated current matrix W is compared to determine whether or not the matrix W converges.

(7) If it does not converge, repeat from (4).

에 의해 m×n 사이즈의 디믹싱 행렬(W)을 형성한다. (8) If convergence is achieved,

To form a demixing matrix W of size m × n.

에 의해 수집된 신경 신호(혼합 신호 데이터)가 저차인 m차의 독립된 신호로 나타낸다. (9) Thereafter,

(Mixed signal data) collected by the n-th order differential signal is an m-order independent signal of lower order.

FIG. 7 is a graph showing the separated nerve signals and the ankle joint angle estimated therefrom.

In the PP / Fast ICA technique described above, when the value m is set to 2, the 5-channel neural signal can be separated into two independent neural signals as shown in FIG. This 5-channel neural signal can be treated as a signal consisting of a superposition of the tibial nerve signal and the peroneal nerve signal, so that two independent signals can be treated as tibial nerve signals and peroneal nerve signals, respectively.

According to the present embodiment, the noise elimination module 220 of the angle estimator 200 performs a preprocessing process of outputting RBI (rectified bin integration) of the extracted two separated signals.

The RBI of the extracted two neuronal signals is easily extractable and is suitable as a high signal-to-noise (SNR) control method, and the processing is as follows:

First, the wavelet denoising process is performed to find the noise-induced action potential and improve the SNR. The waveform element noise removal process is a method that can be used to reduce background noise that can be approximated as a distributed Gaussian random source. In this embodiment, a 5-level non-invariant waveform element division method is applied. Next, the RBI is performed. The data window is defined to have a 20 ms period. Through the noise elimination module 220, the noise can be removed, and the RBI value of the tibial nerve signal and the peroneal nerve signal, which can be inputted as the input value of the joint angle estimation, can be found. Referring to FIG. 7, two separated neural signals and a gray waveform overlapped with each other are signal values of RBI values of respective neural signals.

According to the present embodiment, the angle estimation module 230 of the angle estimator 200 estimates the angle of the ankle joint as shown in the lower graph of FIG. 7 by using the tibial nerve signal and the peroneal nerve signal for which the noise is removed and the RBI value is determined Lt; / RTI >

The angle estimation module 230 determines the angle of the current joint through the artificial neural network using the extracted distant nerve signals (tibial nerve signal and peroneal nerve signals) and the estimated joint angle as an input value, . In this embodiment, the artificial neural network is a dynamically driven recurrent neural network (DDRNN) that uses the currently estimated joint angle values as feedback values.

8 is a conceptual diagram of DDRNN for joint angle estimation.

DDRNN is configured as shown in Equation (3) to output the angle of the joint.

&Quot; (3) "

는 과거 출력값들(

)과, 현재 및 과거의 입력값(

)들에 의해 도출된다. 본 실시예에 따른 DDRNN의 출력 값은 관절의 각도 값이다. Here, F is a nonlinear mapping by the neural network, and the current output value

Lt; RTI ID = 0.0 >

), Current and past input values (

Lt; / RTI > The output value of the DDRNN according to the present embodiment is an angle value of the joint.

According to the present embodiment, DDRNN corresponds to the RBI value of the tibial nerve signal and the peroneal nerve signal having the input delay of 200 ms (since the RBI value is extracted at a cycle of 20 ms, each of the 10 tibial nerve signals The neural signal values are collected, ie, p = 10), and the input layer into which the input values are input is composed of 20 neurons. In addition, there is a hidden layer of one layer consisting of 50 hidden neurons, and the output layer is composed of one neuron outputting the estimated value of the joint angle. Such a layer structure is not limited to the illustrated example, and it is possible to construct a hidden layer of multiple layers, or to adjust the number of input neurons and hidden neurons.

The weight given to each neuron constituting the DDRNN is designed by analyzing the relationship between the measured angle measured with the data acquisition system (FIG. 3) and the fibular and tibial nerve signals as shown in FIG.

의 신경 신호 값이 입력되고, 시간 지연를 가지고 관절 각도 값이 입력된다. The output signal is fed back to the input layer with an output delay (q = 10) of 200 ms. That is, input neurons

And a joint angle value is input with a time delay.

In each input neuron, the neural signal value and the joint angle value are combined through a given weight to extract a predetermined result, and 20 results are input to 50 hidden neurons again and processed by the weight given to each hidden neuron. The 50 resultant values output from the hidden neuron are input to one output neuron and finally output an estimated value of the joint angle.

The output value of the current joint angle is fed back to the input neuron so that the angle of the joint can be estimated in real time in response to the detection of the nerve signal.

Referring again to FIG. 7, after a predetermined delay (400 ms in this embodiment) has elapsed until an input value input to the input neuron and an output value output from the output neuron are completely circulated, DDRNN It can be seen that the angle value of the estimated joint follows the measured angle value of the joint considerably.

To confirm the accuracy of angular values of joints estimated by DDRNN from tibial nerve signals and peroneal nerve signals extracted from extracted 5-channel nerve signals by using PP / FastICA technique and extracted nerve signals, and correlation coeffi- cients (CCs).

In the data acquisition system of FIG. 3, the 5-channel neural signal collected from the collection electrode unit 10 is separated and extracted into two neural signals through the angle estimator 200 according to the present embodiment, And the result as shown in FIG. 7 is obtained.

At the same time, the peroneal nerve signal and the tibia signal as shown in Fig. 4 are measured from the short channel cuff electrode body provided on the tibial nerve and the peroneal nerve, and the angle of the ankle joint is measured through a goniometer.

In order to confirm the signal separation accuracy, CCs for determining the similarity between the two extracted neural signals as shown in FIG. 7 and the measured peroneal nerve signals and the tibial nerve signals shown in FIG. 4 were performed. In order to confirm the accuracy of the joint angle estimation, CCs for determining the degree of similarity between the peroneal nerve signal and the tibial nerve signal were performed on the joint angle estimated as shown in FIG. 7 and the measured joint angle shown in FIG.

The CCs are calculated using the following equation (4).

&Quot; (4) "

When calculating the accuracy of signal separation, x represents the measured neural signal and y represents the neural signal separated from the 5-channel neural signal. Similarly, when calculating the accuracy of the joint angle estimation, x represents the measured joint angle and y represents the joint angle estimated by the angle estimator.

Table 1 and Table 2 below show the accuracy results of signal separation. In the table, the value of each number means the value of R ² .

[Table 1]

[Table 2]

In order to confirm the superiority of the PP / Fast ICA technique for signal separation according to the present embodiment, a method of factor analysis (FA) and principal component analysis (PCA) We separated the 5-channel neural signal into two independent neural signals. FIG. 9 shows a result of separating a 5-channel neural signal into two independent signals through the PP / Fast ICA method, the FA method, and the PCA method according to the present embodiment.

The separated neuronal signals were compared with the measured neural signals in FIG. 4 to calculate the CCs value. Table 1 and Table 2 also show signal separation accuracy according to the above method.

FA and PCA assume that the source data is normalized to zero mean and has Gaussian variance. Furthermore, only the covariance between the observed variables is used for estimation.

FA is an attempt to describe the covariance relationship between variables by some unobservable random elements called factors. Assuming that the elements are not related, they imply independence in the case of Gaussian data. The PCA determines the direction of maximum change along the orthogonal direction. The PCA is performed using the eigenvalues of the data covariance matrix. When having Gaussian data, determining independent elements is straightforward because unrelated elements are always independent. FA and PCA are valid only when the source data has Gaussian data.

On the other hand, the PP technique applied in this embodiment finds potential non-Gaussian structures from the observations and has the advantage of preserving the characteristics of the original data of the higher dimension. This PP technique finds one projection at one point as the extracted signal becomes non-Gaussian.

In other words, since FA and PCA find independent elements based on Gaussian data, the accuracy of separating the tibial nerve signals from the peroneal nerve signals is less than that of PP techniques.

Actually, as shown in [Table 1] and [Table 2], the accuracy of signal separation by the FA and PCA methods is 0.3 ~ 0.4 on average, which is lower than that of the PP / Fast ICA method of 0.8 or more. This can be seen even when the signal patterns of FIGS. 4 and 9 are visually compared.

[Table 3] and [Table 4] show the accuracy results of the joint angle estimation.

[Table 3]

[Table 4]

In order to confirm the superiority of the joint angle estimation using DDRNN for signal separation according to the present embodiment, a joint artificial neural network RBN (radial basis network) and TDNN (time-delayed neural network) And compared with accuracy. [Table 3] and [Table 4] also show the accuracy of joint angle estimation through other neural networks.

As shown in Table 3, it can be seen that the joint angle estimation result using the DDRNN according to the present embodiment is the highest as compared with the artificial neural network using the other concept. On the other hand, as shown in Table 4, when the signal is separated through PP / Fast ICA and the joint angle is estimated through DDRNN as in the present embodiment, the accuracy of the joint angle estimation is at the level of R ² = 0.995 It was confirmed that the measured value reached almost. This confirms that the combination of the PP / FastICA method and the DDRNN results in a tremendous synergy in the calculation of joint angles.

1, the angle? Of the joint 7 estimated and calculated in the angle estimator 200 is provided to the angle controller 300, and the angle controller 300 uses the estimated angle value as a feedback signal And a joint angle control signal for generating an angle of the joint following the predetermined target joint angle value. For example, the joint angle control signal may be a signal for determining the magnitude and period of the stimulus pulse, and the information for generating the joint angle control signal may be related to the relationship between the angle of the electrical stimulation to the pre- Is derived from the database.

The stimulus pulse generator 400 generates a stimulus pulse for stimulating each distal nerve 3, 4 and causing the activity of the muscles 5, 6 connected thereto, based on the joint angle control signal.

10 is a graph showing a stimulus pulse generated in the stimulus pulse generator 400. FIG. The plot located at the top of FIG. 10 represents the stimulation pulse.

10, the stimulation pulses have a repetition period of 50 Hz and a pulse width of 0.4 ms, and a trapezoidal shape (rise time: 0.5 s, duration: 0.2 s, drop time: 0.5 s). Pulses are provided to stimulate the tibial and peroneal nerves over 20 s.

Referring again to FIG. 1, the electrical stimulator 500 receives electric stimuli corresponding to the stimulation pulses through the stimulating electrodes 21, 22 provided on each of the distal nerves 3, 4 to the distal nerves 3, 4 . When electrical stimulation is applied, tibial nerve signals and peroneal nerve signals transmitted from the tibial nerve (3) and the peroneal nerve (4) to the sciatic nerve (2) are transmitted through a kind of electrical stimulation Disturbing nerve signals are mixed. The 5-channel neuron signal contains this disturbed nerve signal, and the disturbed nerve signal is a factor that reduces the accuracy of the joint angle estimation.

Accordingly, according to the present embodiment, the stimulus pulse generator 400 generates and transmits a blanking signal corresponding to the period of the stimulation pulse to the signal processor 100, and the signal processor 100 generates the erase signal Overlapping the proximal nerve signal removes disturbing nerve signals caused by electrical stimulation.

Referring to FIG. 10, the erase signal turns the nerve signal to a ground state for a period of 6 ms, thereby eliminating the influence on each stimulus. Thus, as shown at the bottom of FIG. 10, a multi-channel eraser nerve signal with 14 millisecond disturbances removed is obtained.

To verify the validity of estimating the angles of the joints by applying the angular estimation method described above to such erase nerve signals, a verification system was constructed.

11 is a conceptual diagram of a verification system according to an example.

As shown in Fig. 11, in addition to the articulation angle detection system of Fig. 1, joints appearing by applying electric stimulation to the tibial nerve 3 and the peroneal nerve 4, using a goniometer 32, The angle was measured.

12 is a graph showing an ankle angle measured according to an electric stimulus and a 5-channel neural signal (multi-channel cuff recording) collected at this time.

Trapezoidal electrical stimulation was applied to the tibial nerve (3) and the peroneal nerve (4), respectively, and the ankle angle was controlled to vary in the range of 60 °. It can be seen that the ankle angle measured by the nonlinear characteristic of the neuromuscular system shows a shape without symmetry.

As described above, the erase signal was used to remove the disturbance signal, and the 5-channel neural signal was obtained as the eraser nerve signal from which the disturbance signal was removed. The joint angles were estimated using the PP / FastICA method and DDRNN described above for these 5-channel neural signals and compared with the measured ankle angles.

13 is a graph showing tibial nerve signals and peroneal nerve signals separated from the 5-channel eraser nerve signals and ankle angles estimated from the corresponding nerve signals. In Fig. 13, the gray line indicates the RBI value. When the measured ankle angle is compared with the estimated ankle angle, R ² = 0.981, indicating a very high degree of similarity. This is to verify that the joint estimation method according to the present embodiment can be effectively applied to a system that constitutes the FNS system and performs electrical stimulation.

The joint angle detection system 1 according to the present embodiment collects nerve signals using a multi-channel electrode body that is implanted in a distal nerve such as a sciatic nerve, separates the nerve signals, and detects and controls joint angles. It is very advantageous in securing space.

Since the articulated angle is predicted using the implantable multi-channel electrode body, it is possible to construct the FNS system without a device that can seriously harm beauty such as a goniometer.

In addition, we found that using a combination of PP / FastICA and DDRNN can estimate joint angles with very high accuracy. The PP / FastICA technique and DDRNN have also proven to be useful in FNS systems where electrical stimulation is real-time.

As in the present embodiment, when the joint angle detection system 1 is used to control the angle of the ankle joint by estimating the angle of the ankle joint of the leg, the walking and balance of the living body can be effectively maintained.

However, the articulation angle detection system (1) is not limited to being applied to the legs, and the angle of the joint is determined by movement of muscles connected to a plurality of distal nerves that branch from the proximal nerve, such as the sciatic nerve, the tibial nerve, and the peroneal nerve By applying to various bio-organs to detect and control the angle of the joint, it can help with physical activity.

In addition, the separated signals are not limited to two, and m separate signals (m is a natural number greater than 1) are extracted from the proximal nerve signals collected over n channels, so that they are operated by m independent distal nerves The angle of the joint will be estimated.

Claims (12)

A joint angle detection system for detecting an angle of a joint in a living body in which an angle of a joint is determined by movement of muscles connected to a plurality of distal nerves branched from the proximal nerve,
A collecting electrode body installed on the proximal nerve to collect nerve signals of the proximal nerve (" proximal nerve signal ");
And an angle estimator for analyzing the proximal nerve signal to estimate an angle of the joint. The method according to claim 1,
Wherein the angle estimator comprises:
(&Quot; distal nerve signals ") from each of the distal nerves in the proximal nerve signal to the proximal nerve,
Wherein the angle of the joint is estimated from the extracted distant nerve signals using the correlation between the learned nerve signals and the angles of the joints. 3. The method of claim 2,
Wherein the collecting electrode body is a multi-channel electrode body having n channels (n is a natural number greater than 2)
Wherein the proximal nerve signal is collected into n channel neural signals. The method of claim 3,
The movement of the joint involves m (n is a natural number greater than 1) distal nerves,
Wherein the angle estimator comprises:
Projection tracking (PP) technique for projecting higher order signals to lower order signals and Fast Independent Component Analysis (Fast ICA) for finding orthogonal rotations of multiple signals are applied together. And extracting m independent independent nerve signals induced from each distal nerve. 3. The method of claim 2,
Wherein the angle estimator comprises:
Wherein the angle of the current joint is estimated through the artificial neural network having the extracted distant nerve signals and the estimated joint angle as the input values. 6. The method of claim 5,
Wherein the artificial neural network is a circular neural network that feeds back an estimated angle value of a joint and uses the angle as an input value. 3. The method of claim 2,
Wherein the angle estimator comprises:
A neural signal separation module for separating the proximal nerve signal into independent distal nerve signals;
A noise removal module for removing noise from the distal nerve signals;
And an angle estimation module for estimating an angle of the joint from the noise-removed distal nerve signals. The method according to claim 1,
And an angle controller for generating a joint angle control signal for generating an angle of the joint following the predetermined target angle value by using the estimated joint angle value as a feedback signal. 9. The method of claim 8,
A stimulation pulse generator for generating a stimulation pulse for stimulating each distal nerve based on the joint angle control signal to induce activity of muscles connected thereto;
Further comprising an electric stimulator for applying an electric stimulus corresponding to the stimulation pulse to the distal nerve through a stimulating electrode body provided on each of the distal nerves. 10. The method of claim 9,
Further comprising: a signal processor for preprocessing the proximal nerve signal and delivering it to the angular estimator,
Wherein the stimulus pulse generator generates a blanking signal corresponding to the period of the stimulation pulse and transmits the blanking signal to the signal processor,
Wherein the signal processor superimposes the erase signal on the proximal nerve signal to remove the disturbing nerve signal generated by the electrical stimulation. The method according to claim 1,
The proximal nerve is the sciatic nerve,
The distal nerves are tibial nerve and peroneal nerve branching from the sciatic nerve,
And the angle of the ankle joint of the leg is estimated to control the walking and balancing of the living body. The method according to claim 1,
Wherein the collecting electrode body is a cuff-shaped electrode in which a plurality of electrodes form a plurality of channels, and the outer periphery of the distal nerve is curled.

Images