Exoskeleton compliant control method and system based on deep learning prediction

Abstract

The invention provides an exoskeleton compliant control method and system based on deep learning prediction. The method comprises the following steps: acquiring IMU (Inertial Measurement Unit) data of a historical gait cycle of a user, and inputting the IMU data into a pre-constructed deep learning model for training; acquiring actual joint angle data of the user in the current gait cycle in real time, performing time alignment on the actual joint angle data and the joint angle data predicted by the trained deep learning model, and performing difference calculation based on the aligned joint angle data to obtain a joint angle offset; the joint angle offset serves as input, the control rate is obtained through fuzzy PID calculation, and the control rate serves as a control signal to softly control the exoskeleton. The method adapts to gait modes of different users, personalized and optimized exoskeleton control is achieved, the deep learning model is continuously optimized according to real-time feedback, and the effect of compliant control is achieved.

Description

Technical Field

[0001] The present invention relates to the field of exoskeleton robots, and particularly to a compliant control exoskeleton method and system based on deep learning prediction. Background Art

[0002] With the aging of the population and the continuous increase of the disabled population, lower limb rehabilitation exoskeletons have been increasingly used to assist patients with movement function disorders in rehabilitation training to help them recover their movement ability. Currently, it has become one of the most promising solutions for lower limb movement function patients to recover their movement ability. Lower limb rehabilitation exoskeletons play an important role in the rehabilitation training of patients suffering from diseases such as spinal cord and spine injuries, stroke injuries, and traumatic brain injuries. Most of the lower limb rehabilitation exoskeletons currently available on the market mainly assist patients in performing simple repetitive movements through pre-set fixed programs, and cannot provide smooth and compliant assistance according to the real-time state. The rehabilitation training methods are limited and cannot achieve sufficient training effects. Summary of the Invention

[0003] Based on this, in view of the problems that the existing control exoskeleton using pre-programmed gait patterns cannot adapt to the changes in muscle tension in real time and lack personalized adaptation, it is necessary to provide a compliant control exoskeleton method and system based on deep learning prediction.

[0004] To solve the above technical problems, the technical solution adopted by the present invention is:

[0005] A compliant control exoskeleton method based on deep learning prediction, comprising the following steps:

[0006] Obtain the historical gait cycle IMU data of the user, and input it into a pre-constructed deep learning model for training to obtain a trained deep learning model;

[0007] Real-time obtain the actual joint angle data of the user in the current gait cycle, align it with the joint angle data θ predicted by the trained deep learning model in terms of time, and perform difference calculation based on the aligned joint angle data to obtain a joint angle offset; PRE Perform time alignment, and based on the aligned joint angle data, perform difference calculation to obtain a joint angle offset;

[0008] Use the joint angle offset as an input to calculate a control rate through fuzzy PID, and use it as a control signal to compliant control the exoskeleton; wherein, the control rate u: e represents the joint angle offset, represents the error change rate, dt represents the gait cycle, K p 、K i and K d represent the initial values of the dynamic adjustment stiffness coefficient, the initial values of the dynamic adjustment steady-state coefficient, and the initial values of the dynamic adjustment damping coefficient, ΔK p 、ΔKi and ΔK d represent the increment of K p , K i and K d .

[0009] Furthermore, for the acquired actual joint angle data, a sliding window method is adopted for processing and fed back into the trained deep learning model to optimize the model.

[0010] Furthermore, the deep learning model includes an LSTM layer for capturing time series information, a fully connected layer for converting the temporal features output by the LSTM layer into specific prediction parameters, and an output layer for mapping the prediction parameters to the prediction target and taking it as the prediction result.

[0011] Furthermore, after obtaining the historical gait cycle IMU data of the user, the processed data is obtained through data synchronization, filtering, segmentation, feature extraction, and standardization, and the processed data is used as the input of the deep learning model.

[0012] Furthermore, the gain coefficient G is the ratio of the standard deviation of the actual joint angle data to the standard deviation of the predicted joint angle data.

[0013] The present invention also relates to a compliant control exoskeleton system based on deep learning prediction, including a data acquisition module, a data fusion module, and a data conversion module.

[0014] The data acquisition module is used to collect joint angle data of different nodes of the lower limb movement of the wearer;

[0015] The data fusion module is used to process the data collected by the data acquisition module, perform time alignment with the joint angle data θ PRE predicted by the deep learning model, and calculate the difference based on the aligned joint angle data to obtain the joint angle offset;

[0016] The data conversion module is used to take the joint angle offset as the input, calculate the control rate through fuzzy PID, and use it as the control signal to compliant control the exoskeleton.

[0017] Furthermore, the specific steps for the data fusion module to process the data are as follows:

[0018] Align the joint angle data of different nodes in time and space, and perform adaptive filtering and wavelet threshold denoising; segment the denoised data according to the gait cycle and extract spatio-temporal features, and then perform normalization standard processing. The processed data is detected for anomalies through an isolation forest to obtain the processed data.

[0019] The present invention also relates to a compliant control exoskeleton device based on deep learning prediction, including an inertial sensing unit and a controller.

[0020] The inertial sensing unit is installed on the lower limbs of the wearer and is used to collect acceleration signals and angular velocity signals during the movement of the wearer.

[0021] The controller is used to control the drive motor of the exoskeleton to drive the joints of the exoskeleton according to instructions; when the controller controls the exoskeleton, the steps of the compliant control exoskeleton method based on deep learning prediction as described above are adopted.

[0022] Compared with the prior art, the beneficial effects of the present invention include:

[0023] The present invention accurately predicts the joint angles of the next gait cycle by using historical IMU data and deep learning algorithms, solves the delay problem in signal transmission and data calculation; calibrates the measured joint angle data with the results predicted by the model to ensure a smooth connection between the movement of the user and the assistance of the exoskeleton, providing a more natural and comfortable experience; the deep learning model adapts to the gait patterns of different users to achieve personalized and optimized exoskeleton control, and continuously updates the calibration parameters and deep learning model according to real-time feedback to continuously optimize the control performance and achieve the effect of compliant control. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The disclosure of the present invention will be described with reference to the accompanying drawings. It should be understood that the drawings are only for illustrative purposes and are not intended to limit the scope of protection of the present invention. In the drawings, the same reference numerals are used to refer to the same components. Among them:

[0025] Figure 1 is a flowchart of a compliant control exoskeleton method based on deep learning prediction introduced by the present invention;

[0026] Figure 2 is a flowchart of fuzzy PID control introduced by the present invention;

[0027] Figure 3 is a schematic diagram of the exoskeleton control method. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0028] It is easy to understand that according to the technical solution of the present invention, without changing the essential spirit of the present invention, those of ordinary skill in the art can propose various structural ways and implementation ways that can be mutually replaced. Therefore, the following detailed embodiments and the accompanying drawings are only illustrative descriptions of the technical solution of the present invention and should not be regarded as the whole of the present invention or as a limitation or restriction on the technical solution of the present invention.

[0029] Embodiment 1

[0030] As Figure 1As shown in the figure, the present invention introduces a compliant control exoskeleton method based on deep learning prediction, including the following steps:

[0031] Step 1: Obtain the historical gait cycle IMU data of the user and input it into a pre-constructed deep learning model for training to obtain a trained deep learning model.

[0032] The deep learning model includes an LSTM layer for capturing time series information, a fully connected layer for converting the temporal features output by the LSTM layer into specific prediction parameters, and an output layer for mapping the prediction parameters to the prediction target and taking it as the prediction result. In addition, the model uses the mean squared error as the loss function and is trained using the Adam optimizer.

[0033] The historical gait cycle IMU data is captured by inertial sensing units located on the lower limbs of the wearer. The model learns the patterns and relationships between the signals of the inertial sensing units and the joint angles within the gait cycle. During real-time operation, the trained deep learning model predicts the joint angles of the next gait cycle based on the current and historical gait cycle IMU data. The model can consider the user-specific gait pattern and dynamically adjust the predicted joint angles.

[0034] For the historical gait cycle IMU data, preprocessing is required, that is, the processed data is obtained through data synchronization, filtering, segmentation, feature extraction, and normalization, and the processed data is used as the input of the deep learning model.

[0035] Step 2: Real-time obtain the actual joint angle data of the user in the current gait cycle, align it with the joint angle data θ predicted by the trained deep learning model PRE in time, and calculate the joint angle offset based on the aligned joint angle data.

[0036] Align the actual joint angle data and the joint angle data θ predicted by the deep learning model PRE in time to ensure that they have the same time step and corresponding relationship. Calculate the offset of the joint angle data, that is, the average difference between the sensor real-time monitoring data and the model prediction data. The sensor used in this embodiment is an inertial sensing unit. The offset is calculated by the following formula:

[0037] Offset = Sensor real-time monitoring data - Model predicted joint angle data

[0038] Apply the calculated offset to the joint angle data predicted by the model. For each time step, add the offset corresponding to the time step to the joint angle data predicted by the model;

[0039] In addition, use the joint angle offset δ for the predicted joint angle data θPRE Calibration is performed to obtain calibrated joint angle data θ’ = θ PRE *G + δ; where G represents the gain coefficient. The calibrated joint angle data θ’ can be fed back to the deep learning model for optimization.

[0040] The gain coefficient is used to adjust the amplification or reduction ratio of the joint angle data to better match the real-time monitoring data of the sensor.

[0041] Gain coefficient = standard deviation (real-time monitoring data of the sensor) / standard deviation (joint angle data predicted by the model).

[0042] Considering the real-time performance and stability during the calibration fusion process, a sliding window method is used to process real-time data. Each time the calibration parameters and fusion data are updated, only the latest part of the data is used for calculation to maintain real-time performance.

[0043] Throughout the process, the key to real-time calibration and fusion lies in accurately calculating the calibration parameters and applying them to the results of model prediction to achieve consistency with the actual monitoring data. This can improve the accuracy and compliance of the exoskeleton control system and achieve better assistive control effects.

[0044] Step 3: Use the joint angle offset as the input to calculate the control rate through fuzzy PID and use it as the control signal to compliant control the exoskeleton; where the control rate u: e represents the joint angle offset, represents the error change rate, dt represents the gait cycle, K p 、K i and K d represent the initial values of the dynamically adjusted stiffness coefficient, the initial values of the dynamically adjusted steady-state coefficient, and the initial values of the dynamically adjusted damping coefficient, ΔK p 、ΔK i and ΔK d represent the increments of K p 、K i and K d .

[0045] As Figure 2 shown, the error e and the error change rate de / dt are used as inputs. Fuzzy PID uses fuzzy rules for fuzzy inference and continuously adjusts the parameters K p 、K i and K d according to the actual situation to obtain the inputs ΔK p 、ΔK i and ΔK d . The three output variables obtained from fuzzy inference are input into ΔK p 、ΔK i and ΔK dIntroduced into the PID control.

[0046] K p For quickly correcting joint angle deviations (such as a 10° hip joint lag during the gait cycle) and counteracting human inertia (such as increasing the support torque when quickly lifting the leg). K i For eliminating the residual torque caused by mechanical transmission backlash and compensating for the long-term offset of the wearer's weight (such as continuously assisting when carrying a heavy load). K d For smoothly starting and stopping assistance (such as braking when the knee joint extension end speed exceeds the safety threshold) and suppressing high-frequency jitter (such as torque fluctuations caused by sensor noise).

[0047] The following is a detailed description in combination with the flat-ground assisted walking scenario.

[0048] (1) Gait phase: Heel strike

[0049] K p Suddenly increases to near the upper limit value, K i For preventing impact, K d For suppressing overshoot of the swinging leg.

[0050] (2) Gait phase: Single-leg support

[0051] K p Maintains at a stable value, K i For compensating for gravity, K d For compliant damping.

[0052] (3) Gait phase: Single-leg support

[0053] K p Descends to the lower limit value, K i For accumulating the kinetic energy error of the swinging leg, K d For pre-tensioning control.

[0054] If in the stair climbing assistance scenario, when a 20° knee joint unexpected bending is detected: K p Rises from the steady value to the upper limit value within a certain time, K i Locks the current value to avoid integral saturation, K d Generates a reverse damping torque.

[0055] Converts the control rate into a control signal for controlling the joint drive and movement of the exoskeleton device. The control signal ensures smooth and compliant coordination between the natural movements of the user and the assistance of the exoskeleton. Continuously monitors the user's movement and receives real-time feedback from the sensors. These feedbacks are used to optimize the deep learning model and improve the accuracy of joint angle prediction and overall control performance in the next gait cycle.

[0056] In this embodiment, by using historical IMU data and deep learning algorithms to accurately predict joint angles in the next gait cycle, the problem of delay in signal transmission and data calculation is solved; the measured joint angle data is calibrated with the results predicted by the model to ensure a smooth connection between the user's movement and the assistance of the exoskeleton, providing a more natural and comfortable experience; the deep learning model adapts to the gait patterns of different users to achieve personalized and optimized exoskeleton control, and continuously updates the calibration parameters and the deep learning model according to real-time feedback to continuously optimize the control performance and achieve the effect of compliant control.

[0057] Embodiment 2

[0058] The present invention introduces a compliant control exoskeleton system based on deep learning prediction, including a data acquisition module, a data fusion module, and a data conversion module.

[0059] The data acquisition module is used to collect joint angle data of different nodes of the wearer's lower limb movement.

[0060] The data fusion module is used to process the data collected by the data acquisition module, perform time alignment with the joint angle data θ predicted by the deep learning model, PRE calculate the difference based on the aligned joint angle data to obtain the joint angle offset.

[0061] The specific steps for processing the collected data are as follows: the joint angle data of different nodes are aligned in time and space, and adaptive filtering and wavelet threshold denoising are performed; the denoised data is segmented according to the gait cycle and spatio-temporal features are extracted, and then normalized standard processing is performed. The processed data is subjected to anomaly detection by an isolation forest to obtain the processed data.

[0062] The data conversion module is used to take the joint angle offset as an input, calculate the control rate through fuzzy PID, and use it as a control signal to compliant control the exoskeleton.

[0063] Embodiment 3

[0064] This embodiment introduces a compliant control exoskeleton device based on deep learning prediction, including an inertial sensing unit and a controller.

[0065] The inertial sensing unit is installed on the wearer's lower limb and is used to collect acceleration signals and angular velocity signals during the wearer's movement;

[0066] The controller is used to control the driving motor of the exoskeleton to drive the joints of the exoskeleton according to instructions; when the controller controls the exoskeleton, it adopts the steps of the compliant control exoskeleton method based on deep learning prediction as described above, as Figure 3 shown.

[0067] In practical applications, six IMU nodes are deployed on both lower limbs, that is, they are respectively installed on the thighs / calves / feet, and multi-node clock synchronization is achieved through the CAN bus. Moreover, a quaternion fusion algorithm is adopted to eliminate the coordinate system deviation of each IMU.

[0068] Embodiment 4

[0069] This embodiment provides a computer terminal, which includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the steps of the compliant control exoskeleton method based on deep learning prediction in Embodiment 1.

[0070] When the compliant control exoskeleton method based on deep learning prediction in Embodiment 1 is applied, it can be applied in the form of software. For example, it can be designed as an independently running program and installed on a computer terminal, which can be a computer, a smart phone, etc. It can also be designed as an embedded running program and installed on a computer terminal, such as installed on a single-chip microcomputer.

[0071] Embodiment 5

[0072] This embodiment provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the steps of the compliant control exoskeleton method based on deep learning prediction in Embodiment 1.

[0073] When the compliant control exoskeleton method based on deep learning prediction in Embodiment 1 is applied, it can be applied in the form of software. For example, it can be designed as an independently running program on a computer-readable storage medium. The computer-readable storage medium can be a USB flash drive, designed as a USB key, and a program for triggering the start of the entire method through the USB flash drive is designed.

[0074] The technical scope of the present invention is not limited to the content described above. Those skilled in the art can make various deformations and modifications to the above embodiments without departing from the technical idea of the present invention, and these deformations and modifications should all fall within the protection scope of the present invention.

Claims

1. A compliant control exoskeleton method based on deep learning prediction, characterized in that, It includes the following steps: Obtain the historical gait cycle IMU data of the user, and input it into a pre-constructed deep learning model for training to obtain a trained deep learning model; Obtain the actual joint angle data of the user in the current gait cycle in real time, and perform time alignment with the joint angle data θ predicted by the trained deep learning model. Calculate the difference based on the aligned joint angle data to obtain the joint angle offset; PRE ​ The joint angle offset is used as the input, and the control rate is obtained through fuzzy PID calculation and used as the control signal to compliant control the exoskeleton. Among them, the control rate u: e represents the joint angle offset, represents the error change rate, dt represents the gait cycle, K p 、K i and K d represent the initial values of the dynamically adjusted stiffness coefficient, the initial values of the dynamically adjusted steady-state coefficient, and the initial values of the dynamically adjusted damping coefficient, ΔK p 、ΔK i and ΔK d represent the increments of K p 、K i and K d respectively.

2. The compliant control exoskeleton method based on deep learning prediction according to claim 1, wherein For the obtained actual joint angle data, use the sliding window method to process and feedback it into the trained deep learning model to optimize the model.

3. The compliant control exoskeleton method based on deep learning prediction according to claim 2, wherein The deep learning model includes an LSTM layer for capturing time series information, a fully connected layer for converting the temporal features output by the LSTM layer into specific prediction parameters, and an output layer for mapping the prediction parameters to the prediction target and taking it as the prediction result.

4. The compliant control exoskeleton method based on deep learning prediction according to claim 1, characterized in that After obtaining the historical gait cycle IMU data of the user, the processed data is obtained through data synchronization, filtering, segmentation, feature extraction, and normalization, and the processed data is used as the input of the deep learning model.

5. The compliant control exoskeleton method based on deep learning prediction according to claim 1, wherein The gain coefficient G is the ratio of the standard deviation of the actual joint angle data to the standard deviation of the predicted joint angle data.

6. A compliant control exoskeleton system based on deep learning prediction, characterized in that, It includes: A data acquisition module, which is used to collect the joint angle data of different nodes of the wearer's lower limb movement; A data fusion module, which is used to process the data collected by the data acquisition module and perform time alignment with the joint angle data θ predicted by the deep learning model, and calculate the difference based on the aligned joint angle data to obtain the joint angle offset; PRE Perform time alignment, calculate the difference based on the aligned joint angle data, and obtain the joint angle offset; A data conversion module, which is used to use the joint angle offset as the input to calculate the control rate through fuzzy PID and use it as the control signal to compliant control the exoskeleton.

7. The compliant control exoskeleton system based on deep learning prediction according to claim 6, characterized in that, The specific steps for the data fusion module to process the data are as follows: Align the joint angle data of different nodes in time and space, and perform adaptive filtering and wavelet threshold denoising; Segment the denoised data according to the gait cycle and perform spatio-temporal feature extraction, and then perform normalization standard processing. The processed data is detected for anomalies through an isolation forest to obtain the processed data.

8. A compliant control exoskeleton device based on deep learning prediction, characterized in that, It includes: An inertial sensing unit, which is installed on the wearer's lower limb and is used to collect the acceleration signal and angular velocity signal during the wearer's movement; A controller, which is used to control the driving motor of the exoskeleton to make the joints of the exoskeleton drive according to the instructions; when the controller controls the exoskeleton, it adopts the steps of the compliant control exoskeleton method based on deep learning prediction described in any one of claims 1-5.

9. A computer terminal, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the compliant control exoskeleton method based on deep learning prediction described in any one of claims 1-5.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the compliant control exoskeleton method based on deep learning prediction described in any one of claims 1-5.

Images