Multi-scene self-adaptive power-assisted control method for hip joint exoskeleton robot

Abstract

The invention provides a hip joint exoskeleton robot multi-scene adaptive power-assisted control method, and the method comprises the steps: achieving the motion scene recognition and motion phase estimation through a deep learning perception model; motion scenes such as standing, walking on the flat ground, going upstairs, going downstairs and squatting up are recognized according to the angles and angular speeds of inertial sensors of the left leg and the right leg of a wearer, and continuous action phases are estimated, and a force trajectory planner plans corresponding force trajectories for different action states according to sensing results. A sine moment trajectory planning strategy of a self-adaptive gait cycle is adopted for providing assistance for a walking state in the scenes of walking on the flat ground, going upstairs and going downstairs, an impedance moment planning strategy is adopted for providing assistance for a rising state in the squatting scene, and direct force control or impedance force control over an exoskeleton joint motor can be achieved through a motor force position switching controller. The method is suitable for various task scenes of the hip joint exoskeleton in the carrying process, walking power-assisted control and squatting and rising impedance power-assisted control of the self-adaptive gait cycle are achieved, and the self-adaptive and compliant effects are achieved.

Description

Technical Field

[0001] This invention belongs to the field of exoskeleton control technology and is a multi-scenario adaptive assist control method for hip joint exoskeleton robots. Background Technology

[0002] my country's manufacturing industry is highly developed, with many labor-intensive sectors requiring heavy-duty handling. In the industrial field, assisted handling exoskeleton robots have become a hot research topic in recent years, experiencing rapid development both domestically and internationally. To ensure effective assistance from hip-jointed assisted handling exoskeleton robots, accurate perception of the handling scenario is crucial, and the exoskeleton needs to provide different assistance effects for various specific movements. Therefore, designing an adaptive, state-sufficient, and well-coordinated assistance control method for handling operations is one of the most challenging tasks that urgently needs to be addressed in the research field of hip-jointed exoskeleton robots.

[0003] Current mainstream hip exoskeleton assistive control methods mainly employ fixed threshold segmentation-based assistive control methods and time-series-based assistive control methods.

[0004] The fixed threshold segmentation-based assist control method analyzes data collected by sensors integrated into the exoskeleton system (such as inertial measurement units, encoders, and plantar pressure sensors). It sets corresponding threshold ranges for different movements (such as walking on flat ground, sitting up, and squatting). When the sensor data reaches the preset threshold, the exoskeleton actuators (such as joint motors and hydraulic actuators) output specific position or torque control signals at the corresponding movement stage, thus achieving staged assistance. This method is simple to implement, has low computational load, and is suitable for movements with obvious threshold characteristics; however, it suffers from poor adaptability. Individual differences in movement and unstructured movements can interfere with movement recognition, leading to premature or delayed assistance and causing discomfort to the wearer.

[0005] Assistive control methods based on temporal patterns model the temporal correlation patterns of continuous motion data, outputting specific position or torque control signals during regular continuous motion phases to achieve assistance. Currently, commonly used methods include time linear interpolation and dynamic motion primitives to model the temporal patterns of continuous motion: Time linear interpolation segments the continuous motion cycle of an individual, calculates the average duration of the cycle, and then performs linear interpolation based on the time data within the cycle to estimate the current motion phase; however, while simple to implement, this method has poor adaptability, as different motion frequencies and abnormal movements can affect recognition; Dynamic motion primitives can learn the motion information of the target trajectory to make the newly generated trajectory have the motion trend of the original trajectory. The core idea is to decompose motion into a set of adjustable dynamic systems. This method is suitable for stage estimation of periodic movements such as walking and climbing stairs, but it is difficult to accurately estimate movements with less obvious periodicity, such as squatting.

[0006] Chinese Patent 2025103401280 discloses a compliant control exoskeleton method and system based on deep learning prediction. It acquires historical gait cycle inertial sensor data of the human body and inputs it into a pre-built deep learning model for training. It then acquires real-time joint angle data during the current gait cycle and uses this data as control signals to compliantly control the exoskeleton, enabling it to adapt to individual gait conditions. However, this solution only performs phase estimation for flat ground walking, lacking recognition of multiple scenarios and estimation of action phases in each scenario. This leads to limitations in the exoskeleton's working scenarios and insufficient adaptive effect. Chinese Patent 2025100704756 discloses a hip-assisted exoskeleton control method based on force-position information fusion. Through multi-modal sensor information perception and fusion, it sets state transition thresholds for kinematic data and interactive force data to accurately identify different movement states. In the walking state, it uses zero-force drag and angular velocity compensation algorithms for assistance; in the standing state, it uses position PD control and support force compensation algorithms for assistance. This solution considers assistance for walking on flat ground and getting up, and uses a threshold discrimination method for recognizing multi-scenario movements. The threshold needs to be manually adjusted for individual cases, resulting in weak perceptual robustness. Chinese Patent 2024115234087 discloses a multi-motion torque-assisted control method for a knee exoskeleton based on activity classification and phase detection. It uses a two-stage artificial neural network to detect activity and gait phase, constructs a multi-activity human dynamics model, and calculates the joint assistance torque required by the exoskeleton. It considers human walking and climbing stairs, and can simultaneously detect activity and phase. However, this solution only considers gait-related activities, and its adaptive effect is only reflected in activity scenario classification; it lacks adaptive functionality in assist torque planning. Summary of the Invention

[0007] To address the issues of weak scenario adaptability and poor coordination in existing hip exoskeleton force control methods, this invention provides a multi-scenario adaptive assist control method for hip exoskeleton robots. Based on deep learning perceptrons, force trajectory planners, and motor force-position controllers, this method can effectively improve the control scenario adaptability and human-machine coordination of hip exoskeletons while reducing the complexity of control strategies.

[0008] The technical solution to achieve the purpose of this invention is: a multi-scenario adaptive assist control method for a hip joint exoskeleton robot, comprising the following steps:

[0009] Step 1: Collect kinematic data of the left and right thighs in multiple human scenarios and add labels as the dataset for the deep learning model;

[0010] Step 2: Establish a deep learning model for discrete motion scene classification. Perform sliding window processing on the time series kinematic data in the dataset, and extract the left and right Euler angles and angular velocities in the window as the model input features, and the scene labels as the model output features.

[0011] Step 3: Establish a multi-scene continuous motion phase regression deep learning model, perform sliding window processing on the time series kinematic data in the dataset, extract the left and right Euler angles and angular velocities in the window as model input features, and the phase label as model output features;

[0012] Step 4: Sinusoidal torque trajectory planning for adaptive gait cycle in scenarios of walking on flat ground, going up stairs, and going down stairs; impedance torque trajectory planning for squatting and standing scenarios;

[0013] Step 5: Establish a motor force-position switching controller to realize direct torque control and encoder-based impedance torque control of the joint motor.

[0014] Furthermore, the specific steps of step 1 are as follows:

[0015] Step 1.1: Installation design of inertial sensor, main controller, and actuator;

[0016] Before the human wears the exoskeleton for transport work, two inertial sensors are attached to the left and right thighs respectively to collect the Euler angles and angular velocities of the legs in the sagittal plane. The main controller is used to process the inertial sensor data and run a deep learning algorithm model. The actuator is a motor, which needs to integrate an encoder to obtain its own rotation angle and angular velocity in real time. The motor base is fixedly connected to the waist wear piece of the exoskeleton, and the motor output shaft is fixedly connected to the leg wear piece of the exoskeleton. The main controller and the inertial measurement unit communicate via Bluetooth, and the main controller and the motor drive board communicate via CAN.

[0017] Step 1.2: Collecting individual action data in multiple scenarios;

[0018] When the wearer is carrying, the natural upright position of the human body is defined as the zero position; inertial sensor data of the wearer in multiple motion scenarios such as standing, walking on flat ground, going up and down stairs, and squatting are collected. In the standing scenario, the wearer maintains a natural standing state; in the walking on flat ground, going up and down stairs scenario, the wearer performs gait movements at a natural speed; and in the squatting scenario, the wearer performs knee-bending squatting movements according to a metronome with a fixed frequency.

[0019] Step 1.3: Annotate scene labels for individual action data in multiple scenarios;

[0020] The scene label for standing data is 0, the scene label for walking on flat ground is 1, the scene label for going up stairs is 2, the scene label for going down stairs is 3, and the scene label for squatting and standing data is 4.

[0021] Step 1.4: Phase labeling of individual action data in multiple scenarios;

[0022] The sagittal angle data of the left and right legs in the scenarios of walking on flat ground, going up stairs, and going down stairs are segmented by the extension peak point. The phase labels of adjacent extension peak points are marked with a linear interval value between 0 and 1. The average sagittal angle data of the left and right legs in the scenario of squatting and standing are segmented by the extension peak point and the flexion peak point. The phase labels from the previous extension peak point to the next flexion peak point are marked with a linear interval value between 0 and 1. The phase labels from the previous flexion peak point to the next extension peak point are marked with a linear interval value between 1 and 2.

[0023] Furthermore, the specific steps of step 2 are as follows:

[0024] Step 2.1: Establishment and training of a deep learning model for classifying motion scenarios in multiple individual scenarios;

[0025] A deep learning model for classifying motion scenes is established. The structure of this deep learning model for classifying motion scenes consists of a one-dimensional convolutional layer that captures time-series information, a ReLU activation layer, a max pooling layer, a flattening layer, a Dropout layer, and a fully connected layer that converts the output features of the previous layer into the number of target categories. Convolution and pooling operations can automatically extract local features from sensor time-series data, and Dropout regularization can effectively prevent overfitting during model training.

[0026] The time-series kinematic data in the dataset is processed using a sliding window method. The Euler angles and angular velocities of the left and right legs in the window are extracted as input features of the model, and discrete scene labels are used as output features of the motion scene classification deep learning model. The scenes include standing, walking on flat ground, going up stairs, going down stairs, and squatting.

[0027] Step 2.2: Use cross-entropy as the loss function and train the model using the Adam optimizer; save the weight parameters of the trained motion data scene classification deep learning model.

[0028] Step 2.3: Real-time recognition of discrete motion scenes;

[0029] The deep learning model for classifying motion data scenarios calls the trained model weight parameters. After receiving the input feature data of the sliding window variable, it performs real-time identification of the current motion scenario and outputs five motion scenarios: standing, walking on flat ground, going up stairs, going down stairs, and squatting.

[0030] Furthermore, the specific steps of step 3 are as follows:

[0031] Step 3.1: Establishment and training of a deep learning model for continuous action phase regression in multiple scenarios for individuals;

[0032] A deep learning model for action phase regression is established. The structure of this deep learning model for action phase regression consists of an LSTM layer that captures time series information and a fully connected layer that converts the output features of the previous layer into the regression target feature values. The LSTM layer can automatically capture the long-term dependence and dynamic change patterns of action phase in the time dimension.

[0033] The time-series kinematic data in the dataset are processed by sliding window. The Euler angles and angular velocities of the left and right legs in the window are extracted as input features, and the phase labels are used as output features. The gait phase labels are divided into linear intervals of 0 to 1 based on the adjacent extension peak points of the legs. The squatting phase labels are divided into linear intervals of 0-1-2 based on the extension peak point - flexion peak point - extension peak point of the legs.

[0034] Among them, the continuous action phase regression deep learning model can be applied to the gait phase estimation task in the scenarios of walking on flat ground, going up and down stairs, and the squatting and standing phase estimation task in the scenario of squatting and standing. In the gait phase estimation task, the output value of the model is a linear interval value between 0 and 1. In the squatting and standing phase estimation task, the output value of the model is a linear interval value between 0, 1 and 2.

[0035] Step 3.2: Based on the results of discrete motion scene recognition in Step 2, select a phase regression deep learning model; use mean squared error as the loss function and train it using the Adam optimizer; save the weight parameters of the trained motion phase regression deep learning model.

[0036] Step 3.3: Call the trained model weight parameters and, after receiving the input feature data of the sliding window variable, estimate the current action phase in real time.

[0037] Furthermore, the specific steps of step 4 are as follows:

[0038] Step 4.1: Sinusoidal torque trajectory planning for adaptive gait cycle in scenarios of walking on flat ground, going up stairs, and going down stairs;

[0039] Multiplying the amplitude gain factor by the reference torque trajectory function yields the sinusoidal torque trajectory function of the adaptive gait period, as shown in the following equation:

[0040] τ1=a(t)·τ std1

[0041] τ2=a(t)·τ std2

[0042] In the formula, τ1 is the sinusoidal torque of the adaptive real-time gait cycle generated by the gait swing phase, τ2 is the sinusoidal torque of the adaptive real-time gait cycle generated by the support phase, and the feedforward torque sent to the motor force-position switching controller provides assistance for the human gait support phase and swing phase.

[0043] Step 4.2: Planning the resistance torque trajectory for the squatting / standing scenario;

[0044] The squatting and standing phase value of 1 serves as the transition phase from squatting to standing. This value is used as the trigger signal for the impedance torque trajectory. At this time, the bottom motor force position switching controller is instructed to output impedance torque with the upright angle θ0 as the position target, high stiffness k1, and low damping k2, providing torque to assist the human body in restoring an upright posture.

[0045] Furthermore, the reference torque trajectory function in step 4.1 is as follows:

[0046]

[0047] In the formula, τ std1 τ std2 These are the reference sinusoidal moments generated during the gait swing phase and the support phase, respectively. These are the linear interval values ​​for the gait swing phase and the support phase, respectively.

[0048] Furthermore, the amplitude gain factor of the adaptive real-time gait cycle in step 4.1 is used to automatically adjust based on the ratio of the human standard gait cycle to the actual gait cycle, as shown in the following formula:

[0049]

[0050] In the formula, a(t) is the amplitude gain factor of the adaptive real-time gait period, a std T is the gain reference value. std The standard gait cycle is denoted as , while the estimated gait cycle is denoted as .

[0051] Furthermore, the force-position switching controller based on motor current loop drive directly acts on the motor, realizing direct torque control and impedance torque control. Essentially, it enables the motor to accurately output the required torque. The formula for calculating the torque directly output by the motor is:

[0052] τ output =k p (p set -p now )+k d (v set -v now )+τ set

[0053] In the formula, τ output p is the torque directly output by the motor. set For the set target location, p now v is the current position detected by the encoder. set For the set target speed, v now k is the current speed detected by the encoder. p k is the stiffness coefficient. d τ is the damping coefficient. set The set feedforward torque.

[0054] Furthermore, the different instruction parameters received by the underlying force-position switching controller specifically include:

[0055] Standing scene, motor remains stationary, k p =0,k d =0,τ set =0;

[0056] The gait swaying phase in scenarios such as walking on flat ground, going up stairs, and going down stairs is controlled by direct torque of the motor, k p =0,k d =0,τ set =τ1;

[0057] Gait support phase for walking on flat ground, going up stairs, and going down stairs; direct torque control of the motor; k p =0,k d =0,τ set =τ2;

[0058] In the squatting-standing scene, during the standing-up phase, the motor's impedance torque is controlled, k p =k1,p set =θ0,k d =k2,p set =0,τ set =0.

[0059] Compared with the prior art, the significant advantages of this invention are:

[0060] (1) This invention introduces the squatting and standing motion during the transportation process into the perception task of the hip joint exoskeleton for the first time. It considers the perception scheme of squatting and standing scenarios and their phases, solves the problem of insufficient working scenarios of the hip joint exoskeleton robot, and realizes full perception of common scenarios and motion phases in the transportation process based on the inertial sensors of the left and right legs.

[0061] (2) This invention proposes for the first time a hierarchical deep learning perception scheme based on convolutional neural networks and LSTM neural networks, which solves the problems of low prediction accuracy and poor adaptability of traditional threshold segmentation and temporal segmentation methods, and achieves high prediction accuracy of action phase based on inertial sensor data.

[0062] (3) This invention proposes for the first time a sinusoidal torque trajectory planning strategy with adaptive gait period, which solves the problem of dynamic planning of torque trajectory under variable gait period and has an effective adaptive effect in assist control. Attached Figure Description

[0063] Figure 1 This is a flowchart of the steps of the hip joint exoskeleton control method provided by the present invention.

[0064] Figure 2 This is a schematic diagram of an inertial sensor worn on the human body.

[0065] Figure 3 This is a flowchart illustrating the specific implementation of the force trajectory planner provided by the present invention.

[0066] Figure 4 This is a block diagram of force-position switching control based on FOC current loop drive.

[0067] Figure 5 A schematic diagram illustrating the results of a deep learning model recognizing motion scenes.

[0068] Figure 6 A schematic diagram illustrating the results of estimating the action phase for a deep learning model.

[0069] Figure 7 This is a schematic diagram of the sinusoidal torque trajectory result of the adaptive gait period obtained by the force trajectory planning method.

[0070] Figure 8 This is a schematic diagram of the impedance torque trajectory obtained by the force trajectory planning method. Detailed Implementation

[0071] The following will refer to the appendices in the embodiments of the present invention. Figure 1-8The technical solutions in the embodiments of the present invention are clearly and completely described herein. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0072] Combination Figure 1 The present invention provides a multi-scenario adaptive assist control method for a hip joint exoskeleton robot, comprising the following steps:

[0073] Step 1: Collect kinematic data of the left and right thighs in multiple human scenarios and add labels as the dataset for the deep learning model;

[0074] The deep learning models include a motion scene classification deep learning model and a motion phase regression deep learning model;

[0075] Step 1.1: Installation design of inertial sensor, main controller, and actuator;

[0076] See Figure 2 Before the exoskeleton is worn by the user for transport tasks, two inertial sensors are attached to the left and right thighs respectively to collect the Euler angles and angular velocities of the legs in the sagittal plane. The main controller processes the inertial sensor data and runs a deep learning algorithm model. The actuator is a brushless DC motor driven by an FOC current loop, which requires an integrated encoder to obtain its own rotation angle and angular velocity in real time. The motor base is fixedly connected to the waist wear piece of the exoskeleton, and the motor output shaft is fixedly connected to the leg wear piece of the exoskeleton. The main controller and the inertial measurement unit communicate via Bluetooth, and the main controller and the motor drive board communicate via CAN.

[0077] Step 1.2: Collecting individual action data in multiple scenarios;

[0078] When the wearer performs a carrying action, the natural upright position of the human body is defined as the zero position. Inertial sensor data of the wearer is collected in multiple motion scenarios such as standing, walking on flat ground, going up and down stairs, and squatting. In the standing scenario, the wearer maintains a natural standing state. In the walking, going up and down stairs scenario, the wearer performs gait movements at a natural speed. In the squatting scenario, the wearer performs knee-bending squatting movements according to a metronome at a fixed frequency. In a specific embodiment, the metronome frequency is selected as 10 beats per minute.

[0079] Step 1.3: Annotate scene labels for individual action data in multiple scenarios;

[0080] The scene label for standing data is 0, the scene label for walking on flat ground is 1, the scene label for going up stairs is 2, the scene label for going down stairs is 3, and the scene label for squatting and standing data is 4.

[0081] Step 1.4: Phase labeling of individual action data in multiple scenarios;

[0082] The sagittal angle data of the left and right legs in the scenarios of walking on flat ground, going up stairs, and going down stairs are segmented by the extension peak point. The phase labels of adjacent extension peak points are marked with a linear interval value between 0 and 1. The average sagittal angle data of the left and right legs in the scenario of squatting and standing are segmented by the extension peak point and the flexion peak point. The phase labels from the previous extension peak point to the next flexion peak point are marked with a linear interval value between 0 and 1. The phase labels from the previous flexion peak point to the next extension peak point are marked with a linear interval value between 1 and 2.

[0083] Step 2: Establish a deep learning model for discrete motion scene classification. Perform sliding window processing on the time series kinematic data in the dataset, and extract the left and right Euler angles and angular velocities in the window as the model input features, and the scene labels as the model output features.

[0084] Step 2.1: Establishment and training of a deep learning model for classifying motion scenarios in multiple individual scenarios;

[0085] A deep learning model for classifying motion scenes is established. The structure of this deep learning model for classifying motion scenes consists of a one-dimensional convolutional layer that captures time-series information, a ReLU activation layer, a max pooling layer, a flattening layer, a Dropout layer, and a fully connected layer that converts the output features of the previous layer into the number of target categories. Convolution and pooling operations can automatically extract local features from sensor time-series data, and Dropout regularization can effectively prevent overfitting during model training.

[0086] The time-series kinematic data in the dataset is processed using a sliding window method. The Euler angles and angular velocities of the left and right legs in the window are extracted as input features of the model, and the discrete scene labels are used as output features of the motion scene classification deep learning model. The scenes include standing, walking on flat ground, going up stairs, going down stairs, and squatting.

[0087] Step 2.2: Use cross-entropy as the loss function and train the model using the Adam optimizer; save the weight parameters of the trained motion data scene classification deep learning model.

[0088] Step 2.3: Real-time recognition of discrete motion scenes;

[0089] The deep learning model for classifying motion data scenarios calls upon the trained model weight parameters. Upon receiving input feature data from a sliding window variable, it performs real-time identification of the current motion scenario and outputs five classification scenarios: standing, walking on flat ground, climbing stairs, descending stairs, and squatting. The scenario identification results are as follows: Figure 5As shown, the scene classification model achieves 100% accuracy in recognizing multiple scenes in the test set.

[0090] Step 3: Establish a multi-scene continuous motion phase regression deep learning model, perform sliding window processing on the time series kinematic data in the dataset, extract the left and right Euler angles and angular velocities in the window as model input features, and the phase label as model output features;

[0091] Step 3.1: Establishment and training of a deep learning model for continuous action phase regression in multiple scenarios for individuals;

[0092] A deep learning model for action phase regression is established. The structure of the deep learning model for action phase regression consists of an LSTM layer that captures time series information and a fully connected layer that converts the output features of the previous layer into the target feature value of the output feature dimension regression. The LSTM layer can automatically capture the long-term dependence and dynamic change pattern of action phase in the time dimension.

[0093] The time-series kinematic data in the dataset are processed by sliding window. The Euler angles and angular velocities of the left and right legs in the window are extracted as input features, and the phase labels are used as output features. The gait phase labels (including walking on flat ground, going up stairs, and going down stairs) are divided into linear intervals of 0 to 1 based on the adjacent extension peak points of the legs. The squatting phase labels are divided into linear intervals of 0-1-2 based on the extension peak point - flexion peak point - extension peak point of the legs.

[0094] Among them, the continuous action phase regression deep learning model can be applied to the gait phase estimation task in the scenarios of walking on flat ground, going up and down stairs, and the squatting and standing phase estimation task in the scenario of squatting and standing. In the gait phase estimation task, the output value of the model is a linear interval value between 0 and 1. In the squatting and standing phase estimation task, the output value of the model is a linear interval value between 0, 1 and 2.

[0095] Step 3.2: Based on the results of discrete motion scene recognition in Step 2, select the corresponding phase regression deep learning model; use mean squared error as the loss function and train it using the Adam optimizer; save the weight parameters of the trained motion phase regression deep learning model.

[0096] Step 3.3: Call the trained model weight parameters. After receiving the input feature data from the sliding window variable, perform real-time estimation of the current action phase. For scenarios of walking on flat ground, going up stairs, and going down stairs, output a linear interval value between 0 and 1; for the squatting / standing scenario, output a linear interval value between 0 and 2. The phase estimation results are as follows: Figure 6 As shown, the average loss of the phase regression model for estimating the action phase in multiple scenarios in the test set is 0.029.

[0097] Step 4: Sinusoidal torque trajectory planning for adaptive gait cycle in scenarios of walking on flat ground, going up stairs, and going down stairs; and impedance torque trajectory planning for scenarios of squatting and standing up.

[0098] Based on the output of step 3, a sinusoidal torque trajectory with adaptive gait cycle is planned for walking on flat ground, going up stairs, and going down stairs, and an impedance torque trajectory is planned for squatting and standing movements.

[0099] Step 4.1: Sinusoidal torque trajectory planning for adaptive gait cycle in scenarios of walking on flat ground, going up stairs, and going down stairs;

[0100] The gait phase linear interval value between 0 and 1 is divided into two parts, 0 to ψ and ψ to 1, using ψ as the dividing point. These represent the gait swing phase and the gait support phase, respectively. In this specific embodiment, ψ = 0.4 is selected. The two sets of linear interval values ​​are mapped to the ranges between -π and 0 and between 0 and π, respectively. The sine value of the two mapped sets of linear interval values ​​is calculated to obtain the planned reference torque trajectory function, as shown in the following formula:

[0101]

[0102] In the formula, τ std1 τ std2 These are the reference sinusoidal moments generated during the gait swing phase and the support phase, respectively. These are the linear interval values ​​for the gait swing phase and the support phase, respectively, and π is the constant value of pi in radians.

[0103] An adaptive real-time gait cycle amplitude gain factor is introduced, which can be automatically adjusted based on the ratio of the human standard gait cycle to the actual gait cycle, as shown in the following formula:

[0104]

[0105] In the formula, a(t) is the amplitude gain factor of the adaptive real-time gait period, a std T is the gain reference value. std T represents the standard gait cycle of the human body. gait (t) represents the gait period estimated in real time; in the specific embodiment, T is selected. std =1.2, T gait (t) represents the gait period estimated in real time;

[0106] Multiplying the amplitude gain factor by the reference torque trajectory function yields the sinusoidal torque trajectory function of the adaptive gait period, as shown in the following equation:

[0107] τ1=a(t)·τ std1

[0108] τ2=a(t)·τ std2

[0109] In the formula, τ1 and τ2 are the sinusoidal torques generated by the adaptive real-time gait cycle in the gait swing phase and support phase, respectively. These torques are sent to the motor force-position switching controller as feedforward torques to provide assistance for the human gait support and swing phases. The sinusoidal torque trajectory results are as follows: Figure 7 As shown, the sinusoidal torque trajectory output when the actual gait period is 1.0, 1.2, and 1.4 can be adaptively adjusted according to the corresponding ratio.

[0110] Step 4.2: Planning the resistance torque trajectory for the squatting / standing scenario;

[0111] The squatting-to-standing phase value of 1 serves as the transition phase from a squatting to a standing posture. This value is used as the trigger signal for the impedance torque trajectory. At this time, the underlying motor force-position switching controller outputs an impedance torque with an upright angle θ0 as the position target, high stiffness k1, and low damping k2, providing torque to assist the human body in restoring an upright posture. In the specific embodiment, θ0 = 0, k1 = 5, and k2 = 1 are selected. The impedance torque trajectory result is as follows. Figure 8 As shown, the output impedance torque trajectory has different torque values ​​at different joint angles, with the torque being 0 when the upright angle θ0 is reached.

[0112] Step 5: Establish a motor force-position switching controller, which can realize direct torque control and encoder-based impedance torque control of the joint motor;

[0113] See Figure 4 The force-position switching controller based on motor current loop drive directly acts on a brushless DC motor driven by FOC current loop. It can realize direct torque control and impedance torque control. Essentially, it enables the motor to accurately output the required torque. The formula for calculating the direct output torque of the motor is:

[0114] τ output =k p (p set -p now )+k d (v set -v now )+τ set

[0115] In the formula, τ output p represents the torque directly output by a brushless DC motor driven by the FOC current loop. set For the set target location, p now v is the current position detected by the encoder. set For the set target speed, v now k is the current speed detected by the encoder. p k is the stiffness coefficient. d τ is the damping coefficient. setThe set feedforward torque;

[0116] The underlying force-position switching controller can receive parameters sent by the torque trajectory planner to achieve the execution effect of direct force control or impedance force control, as shown in Table 1:

[0117] Table 1. Different command parameters received by the bottom-level force-position switching controller

[0118]

[0119] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. A multi-scenario adaptive assist control method for a hip joint exoskeleton robot, characterized in that, Includes the following steps: Step 1: Collect kinematic data of the left and right thighs in multiple human scenarios and add labels as the dataset for the deep learning model; Step 2: Establish a deep learning model for discrete motion scene classification. Perform sliding window processing on the time series kinematic data in the dataset, and extract the left and right Euler angles and angular velocities in the window as the model input features, and the scene labels as the model output features. Step 3: Establish a multi-scene continuous motion phase regression deep learning model, perform sliding window processing on the time series kinematic data in the dataset, extract the left and right Euler angles and angular velocities in the window as model input features, and the phase label as model output features; Step 4: Sinusoidal torque trajectory planning for adaptive gait cycle in scenarios of walking on flat ground, going up stairs, and going down stairs; impedance torque trajectory planning for squatting and standing scenarios; Step 5: Establish a motor force-position switching controller to realize direct torque control and encoder-based impedance torque control of the joint motor.

2. The multi-scenario adaptive assist control method for a hip joint exoskeleton robot according to claim 1, characterized in that, The specific steps of step 1 are as follows: Step 1.1: Installation design of inertial sensor, main controller, and actuator; Before the human wears the exoskeleton for transport work, two inertial sensors are attached to the left and right thighs respectively to collect the Euler angles and angular velocities of the legs in the sagittal plane. The main controller is used to process the inertial sensor data and run a deep learning algorithm model. The actuator is a motor, which needs to integrate an encoder to obtain its own rotation angle and angular velocity in real time. The motor base is fixedly connected to the waist wear piece of the exoskeleton, and the motor output shaft is fixedly connected to the leg wear piece of the exoskeleton. The main controller and the inertial measurement unit communicate via Bluetooth, and the main controller and the motor drive board communicate via CAN. Step 1.2: Collecting individual action data in multiple scenarios; When the wearer is carrying, the natural upright position of the human body is defined as the zero position; inertial sensor data of the wearer in multiple motion scenarios such as standing, walking on flat ground, going up and down stairs, and squatting are collected. In the standing scenario, the wearer maintains a natural standing state; in the walking on flat ground, going up and down stairs scenario, the wearer performs gait movements at a natural speed; and in the squatting scenario, the wearer performs knee-bending squatting movements according to a metronome with a fixed frequency. Step 1.3: Annotate scene labels for individual action data in multiple scenarios; The scene label for standing data is 0, the scene label for walking on flat ground is 1, the scene label for going up stairs is 2, the scene label for going down stairs is 3, and the scene label for squatting and standing data is 4. Step 1.4: Phase labeling of individual action data in multiple scenarios; The sagittal angle data of the left and right legs in the scenarios of walking on flat ground, going up stairs, and going down stairs are segmented by the extension peak point. The phase labels of adjacent extension peak points are marked with a linear interval value between 0 and 1. The average sagittal angle data of the left and right legs in the scenario of squatting and standing are segmented by the extension peak point and the flexion peak point. The phase labels from the previous extension peak point to the next flexion peak point are marked with a linear interval value between 0 and 1. The phase labels from the previous flexion peak point to the next extension peak point are marked with a linear interval value between 1 and 2.

3. The multi-scenario adaptive assist control method for a hip joint exoskeleton robot according to claim 1, characterized in that, The specific steps of step 2 are as follows: Step 2.1: Establishment and training of a deep learning model for classifying motion scenarios in multiple individual scenarios; A deep learning model for motion scene classification is established, which sequentially includes a one-dimensional convolutional layer for capturing time-series information, a ReLU activation layer, a max pooling layer, a flattening layer, a Dropout layer, and a fully connected layer that converts the output features of the previous layer into the number of target categories. The time-series kinematic data in the dataset is processed by a sliding window, and the Euler angles and angular velocities of the left and right legs in the window are extracted as input features of the model. The discrete scene labels are used as output features of the deep learning model for motion scene classification. The scenarios include standing, walking on flat ground, going up and down stairs, and squatting. Step 2.2: Use cross-entropy as the loss function and train the model using the Adam optimizer; save the weight parameters of the trained motion data scene classification deep learning model. Step 2.3: Real-time recognition of discrete motion scenes; The deep learning model for classifying motion data scenarios calls the trained model weight parameters. After receiving the input feature data of the sliding window variable, it performs real-time identification of the current motion scenario and outputs five motion scenarios: standing, walking on flat ground, going up stairs, going down stairs, and squatting.

4. The multi-scenario adaptive assist control method for a hip joint exoskeleton robot according to claim 1, characterized in that, The specific steps of step 3 are as follows: Step 3.1: Establishment and training of a deep learning model for continuous action phase regression in multiple scenarios for individuals; A deep learning model for action phase regression is established, which includes an LSTM layer that captures time series information and a fully connected layer that converts the output features of the previous layer into the target feature value for the output feature dimension regression. The time-series kinematic data in the dataset are processed by sliding window. The Euler angles and angular velocities of the left and right legs in the window are extracted as input features, and the phase labels are used as output features. The gait phase labels are divided into linear intervals of 0 to 1 based on the adjacent extension peak points of the legs. The squatting phase labels are divided into linear intervals of 0-1-2 based on the extension peak point - flexion peak point - extension peak point of the legs. Among them, the continuous action phase regression deep learning model is applied to both the gait phase estimation task in the scenarios of walking on flat ground, going up and down stairs, and the squatting phase estimation task in the scenario of squatting and standing. Step 3.2: Based on the results of discrete motion scene recognition in Step 2, select a phase regression deep learning model; use mean squared error as the loss function and train it using the Adam optimizer; save the weight parameters of the trained motion phase regression deep learning model. Step 3.3: Call the trained model weight parameters and, after receiving the input feature data of the sliding window variable, estimate the current action phase in real time.

5. The multi-scenario adaptive assist control method for a hip joint exoskeleton robot according to claim 1, characterized in that, The specific steps of step 4 are as follows: Step 4.1: Sinusoidal torque trajectory planning for adaptive gait cycle in scenarios of walking on flat ground, going up stairs, and going down stairs; Multiplying the amplitude gain factor by the reference torque trajectory function yields the sinusoidal torque trajectory function of the adaptive gait period, as shown in the following equation: τ1=a(t)·τ std1 τ2=a(t)·τ std2 In the formula, τ1 is the sinusoidal torque of the adaptive real-time gait cycle generated by the gait swing phase, τ2 is the sinusoidal torque of the adaptive real-time gait cycle generated by the support phase, and the feedforward torque sent to the motor force-position switching controller provides assistance for the human gait support phase and swing phase. Step 4.2: Planning the resistance torque trajectory for the squatting / standing scenario; The squatting and standing phase value of 1 serves as the transition phase from squatting to standing. This value is used as the trigger signal for the impedance torque trajectory. At this time, the bottom motor force position switching controller is instructed to output impedance torque with the upright angle θ0 as the position target, high stiffness k1, and low damping k2, providing torque to assist the human body in restoring an upright posture.

6. The multi-scenario adaptive assist control method for a hip joint exoskeleton robot according to claim 5, characterized in that, The reference torque trajectory function in step 4.1 is as follows: In the formula, τ std1 τ std2 These are the reference sinusoidal moments generated during the gait swing phase and the support phase, respectively. These are the linear interval values ​​for the gait swing phase and the support phase, respectively.

7. The multi-scenario adaptive assist control method for a hip joint exoskeleton robot according to claim 5, characterized in that, The amplitude gain factor of the adaptive real-time gait cycle in step 4.1 is used to automatically adjust based on the ratio of the human standard gait cycle to the actual gait cycle, as shown in the following formula: In the formula, a(t) is the amplitude gain factor of the adaptive real-time gait period, a std T is the gain reference value. std The standard gait cycle is denoted as , while the estimated gait cycle is denoted as .

8. The multi-scenario adaptive assist control method for a hip joint exoskeleton robot according to claim 1, characterized in that, The force-position switching controller based on motor current loop drive directly acts on the motor, realizing direct torque control and impedance torque control. Essentially, it enables the motor to accurately output the required torque. The formula for calculating the direct torque output by the motor is: t output =k p (p set -p now )+k d (v set -v now )+τ set In the formula, τ output p is the torque directly output by the motor. set For the set target location, p now v is the current position detected by the encoder. set For the set target speed, v now k is the current speed detected by the encoder. p k is the stiffness coefficient. d τ is the damping coefficient. set The set feedforward torque.

9. The multi-scenario adaptive assist control method for a hip joint exoskeleton robot according to claim 8, characterized in that, The different instruction parameters received by the underlying force-position switching controller specifically include: Standing scene, motor remains stationary, k p =0,k d =0,τ set =0; The gait swaying phase in scenarios such as walking on flat ground, going up stairs, and going down stairs is controlled by direct torque of the motor, k p =0,k d =0,τ set =τ1; Gait support phase for walking on flat ground, going up stairs, and going down stairs; direct torque control of the motor; k p =0,k d =0,τ set =τ2; In the squatting-standing scene, during the standing-up phase, the motor's impedance torque is controlled, k p =k1,p set =θ0,k d =k2,p set =0,τ set =0.

Images