A flexible ankle exoskeleton based foot drop assist method

Abstract

The application discloses a foot drop auxiliary method based on a flexible ankle exoskeleton and belongs to the field of rehabilitation medical assistance. The method inputs the healthy side ankle joint angle into an adaptive oscillator, learns the related parameters in the adaptive oscillator and predicts the affected side ankle joint angle. Meanwhile, the CNP model predicts the ankle joint angle of the affected side lower leg according to the gait phase estimated by the adaptive oscillator. The two prediction results of the ankle joint angle of the affected side lower leg are weighted and fused to finally obtain the expected ankle joint angle of the affected side. In order to make the prediction output of the fused ankle joint angle more stable and continuous, the result of the weighted fusion is smoothed. The method can dynamically learn the normal walking mode based on the patient's healthy side ankle joint angle in real time, generate the personalized ankle joint angle reference track of the affected side of each moment for different patients, thereby inhibit the abnormal movement mode of the patient and guide the patient to achieve the normal gait in the rehabilitation training process.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of rehabilitation medical assistance, and more particularly, relates to a foot drop assistance method based on a flexible ankle exoskeleton. BACKGROUND

[0002] The manifestation of foot drop is the inability to dorsiflex, and it is a symptom that the toes always touch the ground first when landing during walking or lifting the leg. Therefore, in the swing phase of the gait cycle, the patient cannot complete the dorsiflexion action, and a foot drop gait is formed. The current flexible ankle exoskeleton is mainly used to assist foot drop patients by generating a reference trajectory and pulling the foot back by a steel wire to guide the patient's ankle to track the reference trajectory. Such trajectory generation methods mainly fall into two categories. The first method is to use a trapezoidal trajectory with adjustable amplitude and rising and falling time. In a gait cycle, fixed gait cycle percentage time is set as the rising and falling time of the trapezoidal trajectory, and the trapezoidal amplitude is the position of the pull rope when the patient's foot reaches the standing ankle angle. Although this method can increase the maximum dorsiflexion angle of the patient and lift the patient's foot back, it does not consider the dynamic and subtle angle changes of the patient at every moment, and can only keep the patient relaxed when the patient does not need to lift the foot, and directly pull the patient's foot to the highest when the patient needs to lift the foot, which cannot guide the patient to achieve a normal gait. The second method is to use the average value of the ankle angle of a healthy person walking as the reference trajectory. However, there are differences in the ankle angle of different people walking. This method does not consider the differences between different people and cannot achieve personalized assistance, and cannot achieve the best assistance effect for different patients. SUMMARY

[0003] In view of the above defects or improvement needs of the prior art, the present application provides a foot drop assistance method based on a flexible ankle exoskeleton, which can inhibit the abnormal movement pattern of the patient and guide the patient to achieve a normal gait during rehabilitation training.

[0004] To achieve the above-mentioned purpose, according to the first aspect of the present application, a foot drop assistance method based on a flexible ankle exoskeleton is provided, comprising:

[0005] S1, inputting the ankle angle θ ankle (t) of the healthy side lower leg at time t to an adaptive oscillator to obtain the estimated value φ1(t) of the gait phase at time t and α i (t); wherein when the heel of the healthy side lower leg is detected to touch the ground, ω(t) and φ1(t) in the adaptive oscillator are reset to ω measured and

[0006] The adaptive oscillator is The update law is i∈[0,N],N>1,φ i (t) is the phase of the i th oscillator at time t, α i (t) is the amplitude of the i th sinusoidal component of the adaptive oscillator, ν and η are learning parameters, ω(t) is the fundamental frequency of the gait trajectory; ω measured is the actual walking frequency, δ is the Dirichlet function, t heelstrike is the time of heel strike, is the corresponding gait phase when the heel strikes the ground;

[0007] S2, input the gait phase φ1(t)+π after half a gait cycle into the adaptive oscillator and the ankle angle prediction model respectively, to obtain the first estimated value of the ankle angle after half a gait cycle and the second estimated value

[0008] S3, the weighted calculation is performed on and to obtain and the smoothing processing is performed to obtain the expected angle of the affected side ankle at time t

[0009] According to the second aspect of the present application, an electronic device is provided, comprising: a computer readable storage medium and a processor;

[0010] The computer readable storage medium is used to store executable instructions;

[0011] The processor is used to read the executable instructions stored in the computer readable storage medium, and execute the method as described in the first aspect.

[0012] According to the third aspect of the present application, a computer readable storage medium is provided, the computer readable storage medium stores computer instructions, and the computer instructions are used to make the processor execute the method as described in the first aspect.

[0013] According to the fourth aspect of the present application, a computer program product is provided, comprising computer programs or instructions, and the computer programs or instructions are executed by the processor to realize the method as described in the first aspect.

[0014] In general, compared with the prior art, the above technical solutions conceived by the present application can achieve the following beneficial effects:

[0015] 1. The method provided by the present application inputs the healthy side ankle joint angle into the adaptive oscillator to learn the related parameters in the adaptive oscillator and predict the affected side ankle joint angle, simultaneously adopts the CNP model to predict the ankle joint angle of the affected side lower leg according to the gait phase estimated by the adaptive oscillator, and performs weighted fusion on the two prediction results of the ankle joint angle of the affected side lower leg to finally obtain the expected ankle joint angle of the affected side, and in order to reduce the random fluctuations and noises in the prediction and make the ankle joint angle prediction output after fusion more stable and continuous, so as to be more consistent with the actual gait movement mode and avoid discomfort during rehabilitation, the result of weighted fusion is smoothed. The method can dynamically learn the normal walking mode based on the healthy side ankle joint angle of the patient in real time, generate a personalized ankle joint angle reference trajectory of the affected side of each moment for different patients, the trajectory is the trajectory that should be reached by the patient during normal walking, so that the abnormal movement mode of the patient can be inhibited, and the patient can reach the normal gait during the rehabilitation training process.

[0016] 2. Further, the method provided by the present application considers that when the gait changes or external interference occurs, the original oscillator model may lose synchronization, resulting in estimation error, and a phase resetting mechanism based on gait time correction is adopted to correct and adjust the phase of the oscillator, thereby improving the calculation accuracy of the adaptive oscillator.

[0017] 3. Further, the method provided by the present application, when the ankle joint angle prediction results of the affected side lower leg predicted by the adaptive oscillator and the CNP model are weighted and fused, the respective weighting coefficients are calculated according to the prediction uncertainty of the CNP and the stability of the adaptive oscillator, the influence of the adaptive oscillator and the CNP model can be dynamically balanced, another model is more dependent when one method is inaccurate or even fails, thereby improving the fault tolerance, reliability and prediction accuracy of the system in complex walking state. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 A flexible ankle exoskeleton-based foot drop assistance method flowchart is provided for the embodiments of the present application.

[0019] Figure 2 An IMU placement position and coordinate system position diagram when measuring joint angle is provided for the embodiments of the present application.

[0020] Figure 3 A gait phase and joint angle estimation diagram based on an adaptive oscillator is provided for the embodiments of the present application.

[0021] Figure 4 A gait event detection diagram based on a state machine is provided for the embodiments of the present application.

[0022] Figure 5A conditional neural process online learning joint angle and uncertainty estimation schematic diagram provided for an embodiment of the present application;

[0023] Figure 6 A gait detection effect schematic diagram provided for an embodiment of the present application;

[0024] Figure 7 A contralateral trajectory generation effect schematic diagram provided for an embodiment of the present application. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0026] In the trajectory generation method for assisting the patient with foot drop of the existing flexible ankle exoskeleton, the use of fixed reference trajectory does not take into account the differences between different people, and cannot achieve the best assistance effect for different patients, while the use of adjustable reference trajectory only adjusts a few parameters, for example, if a trapezoidal trajectory is used, only the trapezoidal amplitude, the rising moment and the falling moment are adjusted, and the actual normal ankle trajectory is far away. The current foot drop exoskeleton assistance method cannot provide accurate and personalized assistance for patients.

[0027] The ultimate goal of assisting the patient with foot drop should be to enable them to walk like healthy people. Normal movement patterns cannot be established on the basis of abnormal movement patterns. Only by suppressing abnormal movement patterns can normal movement patterns be induced. Therefore, the focus of treatment is to change the abnormal posture and abnormal movement pattern of the patient. Based on this, an embodiment of the present application provides a foot drop assistance method based on a flexible ankle exoskeleton, as shown in Figure 1 , which comprises:

[0028] S1, the ankle joint angle θ ankle (t) of the healthy side lower leg at time t is input to an adaptive oscillator to obtain the estimated value φ1(t) of the gait phase at time t and α i (t);wherein, when the heel of the healthy side lower leg is detected to touch the ground, ω(t) and φ1(t) in the adaptive oscillator are reset to ω measured and

[0029] The adaptive oscillator is The update law is i∈[0,N], N>1, φ i (t) is the phase of the i-th oscillator at time t, and αi (t) is φ i the amplitude of (t), ν and η are learning parameters, ω(t) is the fundamental frequency of the gait trajectory; ω measured is the actual walking frequency, δ is the Dirichlet function, t heelstrike is the time of heel strike, is the corresponding gait phase when the heel strikes the ground.

[0030] It can be understood that the ankle joint angle θ ankle of the healthy lower leg is the included angle between the instep and the lower leg of the healthy lower leg, θ ankle Any existing calculation method can be used, and the embodiments of the present application do not make a unique limitation.

[0031] For example, considering that the flexible exoskeleton cannot accurately and conveniently measure the joint angle using a rotary encoder as the rigid exoskeleton does, although it does not constrain the joint degrees of freedom of the human body as much as the rigid exoskeleton, it can be measured and calculated in combination with an IMU (Inertial Measurement Unit, inertial measurement unit) to measure the ankle joint angle using two IMUs. The IMU is fixed on the healthy lower leg and instep through elastic straps, and communicates wirelessly with a computer at a frequency of 50 Hz through Bluetooth. The rotation matrix of the local coordinate systems {S s} and {S a} of the lower leg and instep IMU relative to the global coordinate system {G} is obtained by fusing the acceleration and gyroscope data of the IMU through an extended Kalman filter algorithm. and

[0032] Considering that the IMU cannot be accurately fixed at the same position every time it is worn, and the IMU coordinate axes cannot be accurately aligned with the joint axes, if calibration is not performed, the measured joint angle will be inaccurate due to the above problems. Preferably, in order to ensure that the global coordinate systems of the two IMUs are consistent, thereby facilitating the calculation of the joint angle, the two IMUs are placed at the same position before use and the coordinate systems are reset. Since the IMU cannot be guaranteed to be fixed at the same position of the human body every time, in order to obtain an accurate ankle joint angle θ ankle , the {S s} and {S a} need to be converted to {B s} and {B a} before the IMU is used to measure the joint angle. The specific calibration method uses prior art, and will not be described again in the embodiments of the present application.

[0033] For the convenience of expression, when there is no need to clearly distinguish the two IMU local coordinate systems, the coordinate system is written as {S}. The direction of the IMU local coordinate system is as follows Figure 2 shown.

[0034] First, construct a local coordinate system fixed on the calf and instep {B s} and {B a}, these two coordinate systems move with the movement of the calf and the instep. Define when a person is standing {B s} and {B a} coordinate systems coincide, and the z-axes of these two coordinate systems are always aligned with the ankle joint axis (e.g. Figure 2 The y-axes of the two coordinate systems are parallel to the direction of gravity when standing. Therefore, when the ankle joint moves in the sagittal plane, the ankle joint angle is the angle of rotation of the two coordinate systems about the z-axis. When standing, since the two local coordinate systems coincide, the ankle joint angle is 0°. The rotation matrix of {S} relative to {B} can be expressed as:

[0035]

[0036] in, and They represent the unit x-axis, y-axis, and z-axis vectors of the ankle joint coordinate system in the IMU coordinate system.

[0037] These coordinate axes are determined by the following steps. First, the wearer is required to stand upright, and the direction opposite to the gravity direction measured by the IMU is - G g stand The y-axis direction of the ankle joint coordinate system is consistent. Therefore, the y-axis vector of the ankle joint coordinate system measured by the IMU is It can be calculated by the following formula

[0038]

[0039] In the formula, - S g stand is the direction vector of the opposite direction of gravity in the IMU local coordinate system, is the rotation matrix from G to S

[0040] Afterwards, the wearer's foot is required to perform plantar flexion and dorsiflexion reciprocating motion in the sagittal plane. The z-axis of the ankle joint coordinate system is the ankle joint rotation axis measured by the IMU. s}、{S a The joint axes (i.e. ankle joint rotation axes) in} are j s 、j a The data collected by the gyroscope are g s (t), g a (t). At any time, the following relationship is satisfied:

[0041] ||g a (t)×j a ||=||g s (t)×j s ||

[0042] With this property, the joint axis vector is solved using the Gauss-Newton method. The following optimization objective function is used

[0043]

[0044] where the residual r t (x)=||g a (t)×j a ||-||g s (t)×j s ||, let j s =[cos(φ s )cos(θ s ),cos(φ s )sin(θ s ),sin(φ s )] T , j a =[cos(φ a )cos(θ a ),cos(φ a )sin(θ a ),sin(φ a )] T , x=[φ s ,θ s ,φ a ,θ a ] T is a four-dimensional variable to be optimized, where φ s , θ s , φ a , θ a are the polar and azimuth angles of j s in the spherical coordinate system and j a in the spherical coordinate system, respectively. Since the x-axis of the {S} coordinate system is close to parallel with the ankle joint rotation axis when the IMU is fixed, the initial parameters x0=[0,0,0,0] T are selected to speed up the convergence. The update formula is as follows:

[0045] x k+1 =x k -(J T J) -1 J T r(x k )

[0046] Where J is the Jacobian matrix of the residual r(x) to the parameter x. The rotation axis j (here j s 、j a After being uniformly expressed as j),

[0047] Finally, the direction of the x-axis of the ankle joint coordinate system in the sensor coordinate system can be expressed as follows:

[0048]

[0049] In finding the rotation matrix and After that, we can calculate Convert the rotation matrix into XYZ Euler angle to obtain the ankle joint angle where r i`j` is the rotation matrix The element in row i and column j in . It can be understood that and Obtained in real time by IMU, changing with time, and is a constant obtained for the calibration step and does not change with time, so, r ij (t) is {B a}Relative to {B s}'s rotation matrix The element in row i and column j in

[0050] The continuous gait phase information is estimated based on N+1 oscillators in parallel, such as Figure 3 As shown in Figure 2. The ankle joint angle is periodic during stable walking. This periodic signal can be decomposed into the superposition of the fundamental frequency (corresponding to the gait period) and its multiples using Fourier analysis. Therefore, only the fundamental frequency ω(t) of the gait trajectory needs to be learned, and the other frequencies are multiples of the fundamental frequency. Specifically,

[0051]

[0052] in is the actual ankle joint trajectory θ ankle (t) and the trajectory estimated by the oscillator The deviation, For example, if N = 5, then i∈[0,5] represents six parallel oscillators. The zeroth oscillator is a simple integrator that learns the input offset, where φ0(t) = φ0(0) = π / 2. ν and η are learning parameters that determine the speed of phase and amplitude synchronization. The minimum non-zero frequency learned by the adaptive oscillator corresponds to the gait period.

[0053] Since the initial conditions of the adaptive oscillator are different each time it learns, the phase estimated by the adaptive oscillator has no actual physical meaning. In addition, when the gait changes or there is external interference, the original oscillator model may lose synchronization, resulting in estimation errors. The phase reset mechanism can quickly adjust the phase of the adaptive oscillator in these cases to resynchronize with the actual gait. Therefore, the phase reset mechanism is used to avoid estimation errors.

[0054]

[0055] Where δ is the Dirichlet function, is the actual gait phase of the human heel landing, t heelstrike Is the moment when the heel touches the ground. For the convenience of calculation, in the embodiment of the present invention, it is defined That is, the gait phase starts from the heel strike, and can also be defined as other moments; ω measured =2π / T measured is the walking frequency actually measured when a person is exercising, T measured is the time interval between two heel strikes. is the gait phase corresponding to the heel landing (0 to 100%), which can actually be any percentage between 0 and 100%. For the convenience of calculation, it is defined as 0 in the embodiment of the present invention. is an additional phase reset reference phase. When the heel touches the ground, the fundamental frequency ω(t) and the phase φ1(t) corresponding to the fundamental frequency in the adaptive oscillator are reset to ω measured and This allows for a continuous gait phase that conforms to actual physical meaning. After the phase is reset, the minimum non-zero frequency learned by the adaptive oscillator is the frequency corresponding to the gait cycle, and the corresponding phase is the gait phase. The continuous gait phase output by the adaptive oscillator is φ1(t).

[0056] In order to confirm the moment when the heel touches the ground during the phase reset, that is, to determine the moment of heel contact, without introducing additional sensors to increase the complexity of the system, as a further preferred embodiment of the present invention, an IMU at the instep is used to detect the heel contact gait event (HS) while measuring the ankle joint angle. The detection of the gait event is achieved through a finite state machine, that is, by defining a finite number of states of the foot and the switching conditions of each state, the state of the foot is determined based on the measurement information of the IMU at the instep, thereby achieving the detection of the heel contact event, such as Figure 4 That is, preferably, a finite state machine is used to detect the heel strike event of the healthy leg, and the states in the finite state machine include heel strike, stance phase, toe strike and swing phase;

[0057] First, get the local coordinate system {S araw acceleration data under the local coordinate system {S} rotation matrix from the local coordinate system {S} to the global coordinate system {G} a rotation matrix from the local coordinate system {S} to the global coordinate system {G} acceleration under the global coordinate system {G} net acceleration a = a g -g, where g is the gravity acceleration vector in the global coordinate system.

[0058] define a factor D representing the state of the foot at time t t :

[0059]

[0060] where a k represents the net acceleration of the instep at time k in the time window from time t-n to time t-1. The length of the time window n = 5 in the algorithm. is the diagonal standard deviation matrix of acceleration, where σ x , σ y and σ z are the standard deviations of the net acceleration of the instep in x, y, z directions respectively in the time period from time t-n+1 to time t.

[0061] In the support phase before toe-off, since the sole is fully in contact with the ground, there is no motion, D t tends to 0, while in the toe-off, the sole changes from static to dynamic, D t will have a large increase. According to this feature, when D t >D h , the toe-off event (TO) occurs, where D h is a threshold constant for TO event detection.

[0062] After the TO event is detected, the foot enters the swing phase. In this phase, the foot swings in the air, and the swing pattern varies significantly due to individual walking habits. This variability can affect the detection of the heel strike (HS) event. Therefore, a fixed delay time T swing is introduced to account for these differences before detecting the HS event. This delay time is determined experimentally and is shorter than the duration of the swing phase. After the delay, the finite state machine will transition from the TO state to the S2 state.

[0063] Due to the different foot gait patterns in the swing phase, it is challenging to accurately detect the HS event using D t for feature extraction. When the heel strike (HS) event occurs, the foot speed suddenly increases and reaches a negative peak. In this paper, the HS event is determined by detecting the negative peak of the foot speed in each gait cycle.

[0064] In S2 state, when the algorithm detects the angular velocity v foot of the foot in the sagittal plane v heelstrike foot is expressed by the x-axis angular velocity of the foot IMU. thres heelstrike is adaptive, and the calculation formula is as follows:

[0065]

[0066] wherein v heelstrike (i) represents the negative peak speed at the time of the occurrence of the HS event, so that the corresponding time t heelstrike (i) can be determined. Since the detection threshold is adaptive, it can be adjusted according to the HS negative peak value of different individuals, without the need to manually change the parameters. After detecting the HS event, the system needs to prepare for detecting the next TO event. At this time, the state machine enters the S1 state, and the foot remains stationary on the ground for a period of time. In order to determine whether the foot has entered the S1 state, the mean μ foot and the variance σ v of v v within 100 ms are used, and the following rules are adopted:

[0067] |μ v |<20

[0068] |σ v |<120

[0069] Once these rules are met, the system begins to prepare to detect the next TO event. Since the foot remains stationary in the S1 state, which is a stable phase in the gait cycle, the above rules are used to determine the starting point of the gait detection algorithm. The algorithm can start at any time in the gait cycle, thereby ensuring high stability. In order to ensure safety, the algorithm also contains an additional rule: if the system exceeds 1.5 gait cycles in the TO detection preparation state without detecting the next TO state, it indicates that the HS or TO event may be missed, the current cycle will be forcibly terminated, and the algorithm will be reset.

[0070] S2, inputting a gait phase φ1(t)+π after half a gait cycle to the adaptive oscillator to obtain a first estimated value of the ankle angle after half a gait cycle inputting φ1(t)+π to the ankle angle prediction model to obtain a second estimated value of the ankle angle after half a gait cycle

[0071] ​Specifically, after obtaining a continuous gait phase φ1(t) through the gait event corrected adaptive oscillator, a second estimated value of the ankle joint angle after half a gait cycle is generated based on an ankle joint angle prediction model

[0072] The ankle joint angle prediction model can adopt any neural network model, and the embodiment of the present application does not make a unique limitation.

[0073] In view of the prediction accuracy, the CNP model is preferably adopted in the embodiment of the present application, which can not only accurately predict the value of the ankle joint angle, but also predict the uncertainty thereof at the same time, thereby providing key support for subsequent information fusion. That is, the ankle joint angle prediction model is a pre-trained CNP model.

[0074] The pre-training process of the CNP model is as follows:

[0075] It is assumed that there are N sampling points in a gait cycle, each sampling point including the gait phase φ1(t) and the joint angle θ ankle (t) obtained above. The data set collected in a gait cycle In the training process, M points are randomly selected as context points, and the remaining N-M+1 points are selected as test points. That is, the gait phase in the data set is taken as a sample, and the corresponding ankle joint angle is taken as a label to supervise the training of the CNP model.

[0076] The CNP model includes an encoder and a decoder, and the encoder and the decoder each include an input layer, a hidden layer and an output layer connected in sequence.

[0077] In the embodiment of the present application, in order to improve the calculation efficiency, the hidden layer of the encoder and the decoder each includes a fully connected layer and an activation function layer connected in sequence.

[0078] The number of hidden layers in the encoder and the decoder can be set according to the ability to process complex data and the calculation efficiency. The more the number of layers, the stronger the ability to process complex data, and the lower the calculation efficiency. The selection mainly considers the balance between the model expression ability and the calculation efficiency.

[0079] In the embodiment of the present application, the number of hidden layers in the encoder is preferably set to 3, and the number of hidden layers in the decoder is preferably set to 2. The 3-layer hidden layer encoder can better process features in layers, so that it can cope with more complex input data and more effectively extract relevant information in the latent space. The 2-layer hidden layer decoder can reduce the calculation overhead of the model, while ensuring that the decoder has sufficient ability to generate high-quality output results. The smaller number of hidden layers can also reduce the risk of overfitting, so that the decoder is more robust.

[0080] First, the context points are used to train the encoder A global representation vector r is mapped by the neural network and the aggregation operation:

[0081]

[0082] The encoder Encoder processes the input data as follows: the input layer accepts the concatenated gait phase φ1(t) and ankle angle θ ankle (t) at time t; there are 3 hidden layers with 128 hidden units in each, which are fully connected layers with ReLU activation function, and the last linear layer outputs a fixed-dimension vector. Each context point is mapped to a high-dimensional space, and finally mapped to the representation r t of the context point. The specific formula is as follows:

[0083]

[0084] The final aggregation operation is realized by taking the average value of the representation of the context point:

[0085]

[0086] The decoder is represented as follows:

[0087]

[0088] The structure of the decoder is similar to that of the encoder, and its input is the global representation vector r and the concatenation of the gait phase φ * (t) of the test point. The 2 hidden layers of the decoder are each composed of a fully connected layer with ReLU activation function. The specific formula is as follows:

[0089] h o (1) = ReLU(W5·concat(r, φ * (t))+b5)

[0090] h o (2) = ReLU(W6h o (1) +b6)

[0091] O = W7h o (2) +b7

[0092] The original output of the decoder is a vector O = [o μ , o σ ] T . The first half of O is directly taken as the output joint angle, i.e. The latter half of o is passed through a softplus activation function to ensure the variance is non-negative. To guarantee the range of uncertainty of the output, the CNP model uses the following formula to calculate the uncertainty:

[0093] σ CNP = 0.1 + 0.9 · softplus(o σ )

[0094] The final output is a conditional normal distribution where is the predicted ankle angle value, σ CNP (t) is the predicted uncertainty. There is a fully connected layer between the hidden layer and the output layer.

[0095] Preferably, during the training process, the model is trained by maximizing the log-likelihood estimate, using the Adam optimizer, using the negative value of the log-likelihood as the loss function:

[0096]

[0097] The variance σ CNP is not directly optimized during the training process because there is no real label. The training process mainly relies on maximizing the likelihood function, indirectly learning the variance by optimizing the log-likelihood of the target value. In the loss function, if the prediction error is large (i.e. is large), the model tends to increase the variance σ CNP (t), thereby reducing the result of the loss function. If the prediction error is small, the model tends to reduce the variance and increase the confidence of the prediction.

[0098] In the prediction phase of the model, the decoder receives the global representation generated from the historical context points and the new gait phase as input to predict the new ankle angle.

[0099] After completing this pre-training process, the CNP model is already able to predict ankle angles based on gait phases. Further, considering that there is a certain difference between the gait cycle of the object used to train the CNP model and the gait cycle of the actual user (i.e. the rehabilitation object), and the gait trajectories in different gait cycles of the same collection object are also not exactly the same. Based on this, preferably, the pre-trained CNP model is further adjusted to adapt to a new gait cycle (i.e. the new gait cycle of the actual user). That is, preferably, in step S2, before inputting the gait phase φ1(t)+π after half a gait cycle into the ankle angle prediction model, it further comprises:

[0100] The gait phases and ankle angles of the plurality of sampling points in the last gait cycle of the healthy side lower leg are used to fine-tune the ankle angle prediction model.

[0101] Specifically, assume that the data set The initial model trained is Upon receiving the data set of the next gait cycle (i.e. the time between two heel strike events) A subset is randomly selected from is trained together with At this time, the model is further fine-tuned using a smaller learning rate. The adjustment process is shown in Figure 5 .

[0102] S3, weighted calculation of and is performed to obtain and smoothing processing is performed to obtain the expected angle of the affected ankle at time t

[0103] Specifically, the gait phase describes the current state in a periodic process, and the most commonly used mathematical description of periodic things is to use periodic functions such as sine, cosine, etc., whose period is 2π. Therefore, it can be understood that the range of the entire gait cycle is 0 to 2π. When a person walks normally, the angles of the left and right ankle joints should be symmetrical, only with a phase difference of π for half a gait cycle. Therefore, knowing the mapping relationship between the gait phase of the healthy ankle joint and the ankle joint angle and the current gait phase of the healthy ankle joint, the ankle joint angle that the affected side should reach can be obtained by adding or subtracting the phase difference. That is, in steps S1 and S2, the adaptive oscillator and the CNP model learned the mapping relationship between the gait phase and the ankle joint angle, respectively. Therefore, the ankle joint angles predicted by the adaptive oscillator and the CNP model after half a gait cycle are and are weighted to obtain and are weighted to obtain and smoothing processing is performed to obtain that is, the expected angle of the affected ankle joint is finally obtained.

[0104] The weighting coefficients of can be set according to the accuracy of the adaptive oscillator and the CNP model.

[0105] As a further preferred embodiment of the present application, the weighting coefficient w CNP (φ1(t)+π) and w AO (φ1(t)+π) are defined, which are calculated based on the prediction uncertainty σ CNP (φ1(t)+π) obtained by the CNP and the self-defined stability constant σ AO (φ1(t)+π) of the adaptive oscillator: ​

[0106]

[0107] where, for convenience of comparison with the uncertainty of CNP, the adaptive oscillator stability constant is expressed as follows:

[0108]

[0109] α is a scaling factor to adjust the degree of influence of the error on the uncertainty. By default, α = 1.

[0110] Finally, the fused ankle joint angle prediction value can be expressed as follows:

[0111]

[0112] In order to reduce the random fluctuations and noise in the prediction, make the fused ankle joint angle prediction output more smooth and continuous, so as to be more consistent with the actual gait movement mode, and avoid causing discomfort during rehabilitation, the existing smoothing method is used to smooth the . For example, a moving average filter with a window size of k is used to smooth the output, and the smoothed output can be expressed as:

[0113]

[0114] By pulling the ankle joint of the foot drop patient to the desired angle at every moment, the foot drop patient can be guided to restore normal gait when walking.

[0115] The above algorithm was tested on a healthy subject.

[0116] A plantar pressure sensor was installed at the heel and toe of the subject, and the actual heel strike (HS) and toe-off (TO) events were identified by threshold detection. The plantar pressure signals during walking are shown in Figure 6 . The values of these plantar pressure signals do not represent the actual physical pressure, but are used for qualitative judgment of gait events. The heel strike (HS) and toe-off (TO) events calculated by the threshold method are represented by black circles, while the gait events detected by the algorithm proposed in the present application are marked with red triangles. As can be seen from the figure, the algorithm can accurately detect the HS and TO events and distinguish between the stance phase and the swing phase.

[0117] Figure 7 The measured healthy ankle joint trajectory θ ankle and the smoothed expected ankle joint trajectory of the affected side The overall root mean square error (RMSE) was 2.59°. Although there were some differences between the oscillator-fitted trajectory and the actual trajectory during plantar flexion (when the ankle joint angle was negative), these differences did not affect the assistance effect because the assistance was only provided during dorsiflexion. The figure shows that the foot drop assistance method proposed by the application can effectively capture the characteristics of the healthy side trajectory and generate a reasonable reference trajectory for the affected side.

[0118] An electronic device is provided, including: a computer readable storage medium and a processor;

[0119] The computer readable storage medium is configured to store executable instructions.

[0120] The processor is configured to read the executable instructions stored in the computer readable storage medium, and execute the method according to any one of the above embodiments.

[0121] A computer readable storage medium is provided, and the computer readable storage medium stores computer instructions, and the computer instructions are configured to cause a processor to execute the method according to any one of the above embodiments.

[0122] A computer program product is provided, and the computer program product includes a computer program or instructions, and the computer program or instructions are executed by a processor to implement the method according to any one of the above embodiments.

[0123] Those skilled in the art can easily understand that the above description is only a preferred embodiment of the present application, and is not intended to limit the present application, and any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A foot drop assist method based on a flexible ankle exoskeleton, characterized in that: include: S1, will t Ankle joint angle of the unaffected calf at any given moment Input to the adaptive oscillator to obtain t Estimated value of gait phase at moment and Wherein, when it is detected that the heel of the healthy leg touches the ground, the adaptive oscillator and Reset to and ; The adaptive oscillator is , the update law is , , N >1, For the i An oscillator in t The phase of the moment, is the amplitude of the i-th sinusoidal component of the adaptive oscillator, , and are all learning parameters, is the fundamental frequency of the gait trajectory; is the actual walking frequency, , is the Dirichlet function, For the moment when the heel touches the ground, is the gait phase corresponding to heel strike; S2, the gait phase after half a gait cycle Input into the adaptive oscillator and ankle joint angle prediction model respectively to obtain the first estimated value of the ankle joint angle after half a gait cycle and the second estimate ; S3, yes and Perform weighted calculation to obtain , and smooth it using a sliding average filter with a window size of k to obtain t Expected angle of the affected ankle at any moment .

2. The method according to claim 1, wherein A finite state machine is used to detect a heel strike event of the healthy leg, wherein the states of the finite state machine include heel strike, stance phase, toe strike, and swing phase; Among them, in the support phase state, if , then the toe touchdown event occurs, after a fixed delay time The hind leg enters the swing phase, when the angular velocity of the foot in the sagittal plane Cross the heel strike detection threshold twice, first from top to bottom, then from bottom to top When the heel touches the ground, if the The mean and variance of satisfy , the foot enters the support phase; , Indicates time tn At the time t -1 in the time window k The net acceleration of the instep at the moment, , 、 and Separate moments tn +1 to the time t During the time period x 、 y 、 z The standard deviation of the net acceleration of the instep along the axis, Threshold for toe strike event detection.

3. The method according to claim 1, wherein The ankle joint angle prediction model is a CNP model, including an encoder and a decoder, wherein the encoder and the decoder both include an input layer, a hidden layer and an output layer connected in sequence, and the hidden layer both include a fully connected layer and an activation function layer connected in sequence.

4. The method according to claim 3, wherein In step S2, the gait phase after half a gait cycle is Before inputting the ankle joint angle prediction model, it also includes: The ankle joint angle prediction model is fine-tuned using the gait phases and ankle joint angles of multiple sampling points in the previous gait cycle of the healthy side calf.

5. The method according to claim 3, wherein When training the CNP model, the negative value of the log-likelihood is used as the loss function.

6. The method according to claim 3, wherein and The weight coefficients are and ; in, , is the prediction uncertainty of the CNP model, , is the prediction variance of the decoder output in the CNP model, is the stability constant of the adaptive oscillator, , a is the scaling factor.

7. The method according to claim 1, wherein The calculation formula is: ;in, for relatively The rotation matrix Middle i Rank j Elements of the column, i =1, j =1,2; and are the local coordinate systems fixed on the healthy side calf and instep respectively, and The z-axis is always parallel to the ankle joint axis, and the y-axis is parallel to the direction of gravity when the healthy calf is standing; , and The local coordinate systems of the calf and instep IMUs are and Relative global coordinate system The rotation matrix of for relatively The rotation matrix of for relatively The rotation matrix of .

8. An electronic device, characterized in that: include: Computer-readable storage medium and processor; The computer-readable storage medium is used to store executable instructions; The processor is configured to read the executable instructions stored in the computer-readable storage medium and execute the method according to any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to execute the method according to any one of claims 1 to 7.

10. A computer program product comprising a computer program or instructions, characterized in that When the computer program or instruction is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images