Motion intention analysis method for exoskeleton robot

Abstract

The invention provides a motion intention analysis method for an exoskeleton robot. Firstly, the system synchronously collects hip and knee joint torque signals, plantar pressure and seat reaction force; extracting and normalizing characteristics such as joint torque and change rate thereof, pressure center position and speed, plantar total pressure and the like; the recent multiple frame features form a sliding window to be input into a prediction function, the prediction function is composed of a physical prior branch and a neural residual branch in parallel, the prior branch calculates a first intention probability according to seat reaction attenuation, a joint force generation rate and gravity center forward movement, and the residual branch outputs a second intention probability through a long-short-term memory network; and weighting the two to obtain a final motion intention probability. And the controller adaptively adjusts PID gain according to the probability, and superposes proportional feed-forward torque to generate hip and knee target torque in real time so as to realize compliant-active conversion standing / sitting assisting.

Description

Technical Field

[0001] The present invention relates to the field of robot control technology, and in particular to a motion intention analysis method for an exoskeleton robot. Background Art

[0002] Lower limb exoskeleton robots are widely used for rehabilitation training and weight-bearing assistance. Existing devices often use a single encoder or inertial navigation data, using a finite state machine or fixed threshold to determine user intent, and then using a fixed-gain PD / PID to track a preset trajectory. This solution has three major drawbacks:

[0003] The single perception dimension cannot capture sudden changes in joint torque and center of gravity shifts in real time, resulting in delayed intention recognition.

[0004] The model lacks interpretability and adaptability. Although pure data-driven networks have high accuracy, they are difficult to deploy on resource-constrained embedded systems.

[0005] There is a contradiction between control rigidity and comfort. Fixed gain cannot take into account both the smoothness of sitting still and the large torque requirements of standing up, which can easily cause a "drag feeling" or lack of assistance. Summary of the Invention

[0006] The present invention aims to at least solve the technical problems existing in the prior art, and in particular innovatively proposes a motion intention analysis method for an exoskeleton robot, which is characterized by comprising:

[0007] S1, multimodal data acquisition

[0008] Collect joint torque sensor signals at the exoskeleton hip and knee joints, pressure sensor signals at the soles of the feet, and seat reaction force signals at the seat or backboard, and record them synchronously at a fixed sampling period;

[0009] S2, feature extraction and normalization

[0010] Calculating joint torque and joint torque change rate from the torque signal, calculating the plane coordinates and velocity of the plantar pressure center from the pressure signal, summing all plantar pressures to obtain a total plantar pressure, and normalizing each feature according to a preset maximum value to form a feature vector containing joint, pressure, and seat reaction force information;

[0011] S3, intention probability prediction

[0012] The sliding window is input into a prediction function, which is composed of a physical prior branch and a neural residual branch in parallel. The physical prior branch outputs a first intention probability based on the seat reaction force, the rate of change of the joint torque, and the velocity of the plantar pressure center. The neural residual branch performs temporal learning on the sliding window based on a long short-term memory neural network to output a second intention probability. The two are weighted to obtain the final movement intention probability.

[0013] S4, adaptive control output

[0014] The proportional, integral and differential gains are dynamically adjusted according to the motion intention probability, and a predicted feedforward torque proportional to the intention probability is superimposed to generate a target torque command for controlling the exoskeleton hip and knee joint actuators.

[0015] In a preferred embodiment of the present invention, the multimodal data acquisition includes:

[0016] The human-exoskeleton system is modeled as a planar multi-rigid body system, and its state is described using generalized coordinates; its motion equation can be expressed as a constrained dynamic form:

[0017]

[0018] Where M(q) is the inertia matrix and q is the generalized coordinate;

[0019] is the centrifugal force term;

[0020] is the generalized velocity, is the generalized acceleration;

[0021] G(q) is the gravity term;

[0022] τ exo is the exoskeleton joint driving torque;

[0023] is the Jacobian matrix of the plantar, T is the transpose of the matrix;

[0024] F foot is the contact force on the sole of the foot;

[0025] is the Jacobian matrix of the seat contact;

[0026] F seat is the contact force of the seat.

[0027] In a preferred embodiment of the present invention, the feature extraction and normalization include:

[0028] The plantar pressure sensor will measure the pressure values ​​​​p at multiple points i (t), its corresponding physical location is r i , so the center of gravity of the sole pressure can be calculated:

[0029]

[0030] Where CoP(t) is the center of pressure of the plantar surface at time t;

[0031] p i (t) is the pressure value of the i-th sensor at time t;

[0032] r i is the position vector of the i-th sensor on the sole of the foot;

[0033] The joint torque sensor outputs the knee and hip joint torque time series τ j (t), the rate of change of torque can be calculated by taking the derivative

[0034] in, is the rate of change of torque of the jth joint;

[0035] τ j (t) is the instantaneous driving torque of the jth joint;

[0036] It means finding the first derivative of time;

[0037] Get the eigenvector:

[0038]

[0039] Where f(t) is the eigenvector;

[0040] CoP x (t) and CoP y (t) is the plantar pressure center in the x and y directions, respectively;

[0041] and are the rates of change of the CoP vector in the x and y directions, respectively;

[0042] τ knee (t) and τ hip (t) are the moments of the knee and hip joints, respectively;

[0043] and are the rates of change of moment at the knee and hip joints, respectively;

[0044] is the total plantar pressure.

[0045] In a preferred embodiment of the present invention, the intention probability prediction includes:

[0046] A parallel two-branch structure prediction model, consisting of a physical prior branch and a neural network residual branch:

[0047] Among them, the physical a priori branch includes:

[0048] Based on the laws of human kinematics and dynamics, the initial probability of movement intention is calculated using a weighted summation method:

[0049] Normalized value of seat reaction force. The lower the value, the more obvious the user's intention to leave the seat.

[0050] The sum of the normalized values ​​of the joint torque change rate. A higher value indicates that the user is actively exerting force.

[0051] Normalized value of the movement speed of the plantar pressure center. A higher value indicates a more obvious trend of center of gravity movement.

[0052] The above three quantities are linearly combined with weights and mapped to the (0, 1) interval by the sigmoid function as the first motion intention probability;

[0053] Among them, the residual branch of the neural network includes:

[0054] The long short-term memory network is used to learn the temporal feature matrix, automatically capturing the complex temporal and nonlinear relationships between features, and outputting the residual intention probability after activation function processing as a correction to the physical prior branch probability;

[0055] The first intention probability output by the physical prior branch and the residual probability output by the neural network residual branch are linearly weighted and summed, and the result is clipped to the range of 0 to 1 to obtain the final motion intention probability;

[0056] Based on the preset threshold, the final movement intention probability is determined as "maintaining the current posture" or "posture change intention" to provide clear analysis results for subsequent exoskeleton control strategies;

[0057] Specifically, the model maps the input feature f(t) to the predicted intent output Π(f(t));

[0058] Among them, Π(*) is the prediction function, and the final output is a specific intent label and its probability value;

[0059] The calculation method is:

[0060] Π(f(t))=σ(g phy (f(t)))+r θ (f(t))-λ phy L phy (t);

[0061] Among them, σ(g phy (f(t))) is the physical prior term, σ(*) is the Sigmoid activation function, which maps the independent variable to the range of 0 to 1;

[0062] g phy (f(t)) represents the physical prior term;

[0063] r θ (f(t)) is the neural residual term;

[0064] λ phy is the physical loss weight;

[0065] L phy (t) is the physical drive loss function, which uses the physical residual of the power system as the loss of motion intention analysis. It is calculated as:

[0066]

[0067] The physical a priori branch is derived as follows:

[0068]

[0069] Where R(t) represents the seat reaction force, R max is the normalized term of the seat reaction force;

[0070] and represents the rate of change of joint torque, is the normalized term of the joint torque;

[0071] is the rate of change of the CoP vector, is the normalized term of the rate of change of the CoP vector;

[0072] ΔCop(t) is the displacement difference between the current position of the center of pressure and the initial sitting position;

[0073] γ is the displacement weight, which is used to balance the influence of the displacement gap;

[0074] α1, α2 and α3 are weights, and their constraints are:

[0075] Stage probability update equation:

[0076] I stage (t)=ρ·I stage (t-1)+(1-ρ)·Π(f(t));

[0077] Among them, I stage (t) is the stage probability at time t, which is used to represent the smoothed probability that the user is in a certain posture at time t;

[0078] I stage (t-1) is the stage probability at time t-1;

[0079] ρ is the historical probability memory coefficient;

[0080] The motion intention classification results are obtained based on the stage probability:

[0081]

[0082] in, is the result of motion intention classification;

[0083] θ low is the lower threshold of the stage probability. When it is less than this value, the system determines that the user has no intention to change posture.

[0084] θ mid is the upper threshold. When it is greater than this value, it means that the user has taken a conversion action.

[0085] θ low ≤I stage (t)<θ mid It indicates that the stage probability is between two thresholds, and the system determines that the user is in the transition stage of posture conversion.

[0086] In a preferred embodiment of the present invention, the adaptive control output includes:

[0087] Using the impedance control method, a PID controller is constructed, whose goal is to guide the joint to follow a desired trajectory:

[0088]

[0089] Where τ is the control strategy output;

[0090] q ref (t)-q(t) represents the joint position error, Indicates speed error;

[0091] K p is the proportional gain;

[0092] K d is the differential gain;

[0093] K integrate is the integral gain;

[0094] dλ is the integral differential element.

[0095] The present invention also discloses a computer system, comprising:

[0096] processor;

[0097] a memory for storing processor-executable instructions;

[0098] Wherein, the processor is configured to implement the motion intention analysis method for an exoskeleton robot when executing the executable instructions.

[0099] The present invention also discloses a computer-readable storage medium, comprising:

[0100] a memory having a computer program stored thereon;

[0101] A processor is used to execute the program in the memory to implement the motion intention analysis method for an exoskeleton robot.

[0102] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0103] Improved recognition accuracy and lead time: Seat reaction force, torque rate, and CoP speed form an explicit prior, predicting intent 50–100ms in advance; LSTM residuals compensate for complex nonlinearities, achieving an average accuracy of ≥92% across users.

[0104] On-device interpretability and low computing power: The a priori branch contains only a small amount of addition, subtraction, multiplication, and division. The residual network has less than 7k parameters after pruning and quantization, and can run in real time on the Cortex-M7 at 100Hz.

[0105] Smooth and active self-switching: Intention probability drives continuous scheduling of PID gains and injects feedforward torque; low impedance when static, outputs peak power without overshoot when standing up, achieving comfortable and robust human-machine collaboration.

[0106] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0107] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments with reference to the accompanying drawings, in which:

[0108] Figure 1 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0109] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0110] Based on the human body parameter data obtained by the exoskeleton robot sensors, massive human motion parameters are analyzed through big data, and neural networks are used to learn human motion intentions, and predict the use of the skeleton robot to achieve certain specific behaviors.

[0111] like Figure 1As shown, the present invention discloses a motion intention analysis method for an exoskeleton robot, comprising

[0112] 1. Biomechanical modeling of sit-to-stand transitions

[0113] The human-exoskeleton system is modeled as a planar multi-rigid body system, and its state is described using generalized coordinates (such as hip and knee joint angles). Its motion equation can be expressed in a constrained dynamic form:

[0114]

[0115] Where M(q) is the inertia matrix and q is the generalized coordinate;

[0116] is the centrifugal force term;

[0117] is the generalized velocity, is the generalized acceleration;

[0118] G(q) is the gravity term;

[0119] τ exo is the exoskeleton joint driving torque;

[0120] is the Jacobian matrix of the plantar, T is the transpose of the matrix;

[0121] F foot is the contact force on the sole of the foot;

[0122] is the Jacobian matrix of the seat contact;

[0123] F seat is the contact force of the seat;

[0124] In practice, these contact forces are measured using pressure sensors on the soles of the feet and the seat surface. While seated, the seat contact provides vertical support. As the person prepares to stand up, the seat's reaction force gradually decreases and approaches zero. In fact, the loss of vertical seat support is a key indicator of standing up.

[0125] Therefore, the action is divided into two phases:

[0126] Sitting stage: the sole of the foot is in contact with the seat at the same time;

[0127] Stance phase: Seat contact force disappears, leaving only foot contact.

[0128] If further refinement is required, contact constraints such as ground reaction forces can be incorporated into the sitting phase. This type of joint human-exoskeleton modeling approach accurately reflects the correspondence between sensor readings (such as joint torque and pressure) and the human biomechanical state.

[0129] 2. Signal processing and feature extraction

[0130] Extracting meaningful features from raw sensor signals is a key step in realizing motion intention recognition.

[0131] 2.1 Calculation of pressure center

[0132] The plantar pressure sensor will measure the pressure values ​​​​p at multiple points i (t), its corresponding physical location is r i , so the center of gravity of the pressure on the sole of the foot (i.e., the center of pressure, Center of Pressure, referred to as CoP) can be calculated:

[0133]

[0134] Where CoP(t) is the center of pressure of the plantar surface at time t;

[0135] p i (t) is the pressure value of the i-th sensor at time t;

[0136] r i is the position vector of the i-th sensor on the sole of the foot;

[0137] This formula can be used to obtain the resultant position of the plantar pressure distribution, and further reflect the forward and backward shift of the human body's center of gravity through its change over time (such as the speed of CoP), which is an important indicator for judging the state of readiness to stand up.

[0138] 2.2 Joint torque change rate

[0139] Similarly, the joint torque sensor will output the knee and hip joint torque time series τ j (t), the rate of change of torque can be calculated by taking the derivative

[0140]

[0141] in, is the rate of change of torque of the jth joint;

[0142] τ j (t) is the instantaneous driving torque of the jth joint;

[0143] It means to find the first derivative with time.

[0144] 2.3 Eigenvector

[0145]

[0146] Where f(t) is the eigenvector;

[0147] CoP x (t) and CoP y (t) is the plantar pressure center in the x and y directions, respectively;

[0148] and are the rates of change of the CoP vector in the x and y directions, respectively;

[0149] τ knee (t) and τ hip (t) are the moments of the knee and hip joints, respectively;

[0150] and are the rates of change of moment at the knee and hip joints, respectively;

[0151] is the total plantar pressure.

[0152] The dimensions and options of the feature vector can be expanded or simplified depending on the specific task. Essentially, the mathematical expressions described above transform raw physical signals such as pressure and torque into digital feature structures that facilitate pattern recognition and classification modeling. This step lays a solid mathematical foundation for subsequent motion intention prediction.

[0153] 3. Motion Intention Prediction Framework

[0154] The goal of this invention is to use a deep neural network to map the feature vector f(t) to the user's current movement intention label I(t);

[0155] 3.1 Intent Classifier

[0156] During the training process, the neural network learns the nonlinear mapping relationship between features and intent labels through the backpropagation algorithm; each layer in the network can be expressed in mathematical form (linear transformation + nonlinear activation), so it can be regarded as a "mathematical expression of neural structure."

[0157] In fact, deep learning methods have achieved extremely high accuracy in intention recognition tasks: a review showed that the back-propagation neural network (BPANN) achieved close to 100% accuracy in gait intention prediction, significantly outperforming other traditional methods. Although support vector machines and convolutional neural networks have also been used, BPANN achieved the best overall performance.

[0158] 3.2 Intent Mapping Function

[0159] A parallel two-branch structure prediction model, consisting of a physical prior branch and a neural network residual branch:

[0160] Among them, the physical a priori branch includes:

[0161] Based on the laws of human kinematics and dynamics, the initial probability of movement intention is calculated using a weighted summation method:

[0162] Normalized value of seat reaction force. The lower the value, the more obvious the user's intention to leave the seat.

[0163] The sum of the normalized values ​​of the joint torque change rate. A higher value indicates that the user is actively exerting force.

[0164] Normalized value of the movement speed of the plantar pressure center. A higher value indicates a more obvious trend of center of gravity movement.

[0165] The above three quantities are linearly combined with weights and mapped to the (0, 1) interval by the sigmoid function as the first motion intention probability;

[0166] Among them, the residual branch of the neural network includes:

[0167] The long short-term memory network is used to learn the temporal feature matrix, automatically capturing the complex temporal and nonlinear relationships between features, and outputting the residual intention probability after activation function processing as a correction to the physical prior branch probability;

[0168] The first intention probability output by the physical prior branch and the residual probability output by the neural network residual branch are linearly weighted and summed, and the result is clipped to the range of 0 to 1 to obtain the final motion intention probability;

[0169] Based on the preset threshold, the final movement intention probability is determined as "maintaining the current posture" or "posture change intention" to provide clear analysis results for subsequent exoskeleton control strategies;

[0170] Specifically, the model maps the input feature f(t) to the predicted intent output Π(f(t));

[0171] Among them, Π(*) is the prediction function, and the final output is a specific intent label and its probability value;

[0172] The calculation method is:

[0173] Π(f(t))=σ(g phy (f(t)))+r θ (f(t))-λ phy L phy (t);

[0174] Among them, σ(gphy (f(t))) is the physical prior term, σ(*) is the Sigmoid activation function, which maps the independent variable to the range of 0 to 1;

[0175] g phy (f(t)) represents the physical prior term;

[0176] r θ (f(t)) is the neural residual term;

[0177] λ phy is the physical loss weight;

[0178] L phy (t) is the physical drive loss function, which uses the physical residual of the power system as the loss of motion intention analysis. It is calculated as:

[0179]

[0180] The physical a priori branch is derived as follows:

[0181]

[0182] Where R(t) represents the seat reaction force, R max is the normalized term of the seat reaction force;

[0183] and represents the rate of change of joint torque, is the normalized term of the joint torque;

[0184] is the rate of change of the CoP vector, is the normalized term of the rate of change of the CoP vector;

[0185] ΔCop(t) is the displacement difference between the current position of the center of pressure and the initial sitting position;

[0186] γ is the displacement weight, which is used to balance the influence of the displacement gap;

[0187] α1, α2 and α3 are weights, and their constraints are:

[0188] Use the intent output to update the stage probability equation:

[0189] I stage (t)=ρ·I stage (t-1)+(1-ρ)·Π(f(t));

[0190] Among them, I stage(t) is the stage probability at time t, which is used to represent the smoothed probability that the user is in a certain posture at time t;

[0191] I stage (t-1) is the stage probability at time t-1;

[0192] ρ is the historical probability memory coefficient;

[0193] The motion intention classification results are obtained based on the stage probability:

[0194]

[0195] in, is the result of motion intention classification;

[0196] θ low is the lower threshold of the stage probability. When it is less than this value, the system determines that the user has no intention to change posture.

[0197] θ mid is the upper threshold. When it is greater than this value, it means that the user has taken a conversion action.

[0198] θ low ≤I stage (t)<θ mid It indicates that the stage probability is between two thresholds, and the system determines that the user is in the transition stage of posture conversion.

[0199] The prediction module runs in real time on the embedded control system in a sliding window manner, continuously judging the user status and is the core link of the entire motion intention analysis system.

[0200] 4. Control strategy and state machine mapping

[0201] This paper uses a mid-level controller to map the predicted motion intention I(t) into specific exoskeleton execution control instructions. This controller is constructed as a finite state machine, where each state corresponds to a human posture or movement phase, and defines the transition conditions between states. Based on previous research, we define the following four states:

[0202] Sit

[0203] Standing Up

[0204] Stand

[0205] Sitting Down

[0206] The predicted motion intention will trigger the transition of the state machine. For example:

[0207] When in the sit state and I(t) predicts the intention to be "Stand", it moves to the StandingUp state;

[0208] When in the Stand state and I(t) predicts the intention to be "Stand", it transitions to the StandingUp state;

[0209] The trigger event of state transition can not only come from the output I(t) of the intention recognition module, but also be combined with sensor thresholds (such as changes in plantar pressure, changes in joint angular velocity, etc.) for dual judgment, thereby improving robustness and safety.

[0210] 4.1 Control Strategy

[0211] Using the impedance control method, a PID controller is constructed, whose goal is to guide the joint to follow a desired trajectory:

[0212]

[0213] Where τ is the control strategy output;

[0214] q ref (t)-q(t) represents the joint position error, Indicates speed error;

[0215] K p is the proportional gain;

[0216] K d is the differential gain;

[0217] K integrate is the integral gain;

[0218] dλ is the integral differential element.

[0219] This control strategy enables the joint to move in a controlled manner toward the target trajectory and automatically adjusts the output torque according to the actual deviation, thereby achieving a smooth and active standing movement.

[0220] In different states, the control law can be set as follows:

[0221] Stand: Maintains the target angle, applying only minimal gravity compensation or damping to stabilize the stance.

[0222] Sit (sitting posture): also maintain the current angle, turn off the active torque output or apply extremely low stiffness;

[0223] SittingDown: The target trajectory is reversed, and it descends smoothly from standing to sitting. The control law is the same, but the trajectory direction is opposite.

[0224] This state machine combined with impedance control strategy enables natural, safe, and user-intentioned standing / sitting assistance. Its logical structure is clear and feasible, and it dynamically adjusts trajectory and control gain based on the user's physiological state (such as strength and speed), thereby improving rehabilitation training effectiveness and wearing comfort.

[0225] 5. Closed-loop system structure

[0226] The present invention constitutes a complete closed-loop control system, in which every link in the system works closely together to form a continuous feedback loop between human body status, sensor observation, movement intention prediction and exoskeleton control.

[0227] 5.1 Control Model

[0228] Use differential equations to formally describe the closed-loop system:

[0229]

[0230] Where x(t) is the system state vector, is the time derivative of the state, u(t) is the joint driving torque output by the controller;

[0231] y(t) is the observed value of the system state obtained through sensors, including joint torque, plantar pressure and encoder reading;

[0232] f(t) is the eigenvector, Φ(*) represents the signal processing module, which is used to process the data obtained by the sensor to obtain the eigenvector;

[0233] It is the predicted movement intention, which generates the intention prediction result based on the feature vector through the prediction function;

[0234] C(*) is an execution instruction generation function that generates new execution instructions based on the predicted intention and the current system status.

[0235] The feedback system performs an update process every sampling period (e.g. 10ms to 50ms) in real-time control, namely:

[0236] Acquisition system status;

[0237] Extract features and predict intent;

[0238] Generate control instructions based on intention and status;

[0239] The controller outputs the execution torque;

[0240] Cause the system state to change and close the loop again.

[0241] 5.2 Structural modules of closed-loop systems

[0242] The closed-loop system consists of the following modules:

[0243] Controlled object (Plant): The dynamic system of human body + exoskeleton, represented by x(t);

[0244] Sensor module (Sensors): reads the system state and outputs the observation value y(t);

[0245] Signal processing module (Feature Extractor): converts y(t) into feature vector f(t);

[0246] Intent Predictor: Predicts motion intention based on features

[0247] Controller module (Controller): Generates torque output u(t) based on intention and current state.

[0248] This structure highlights the innovative point of the present invention, which is the organic integration of human biomechanical modeling, multimodal sensing, feature extraction, intent recognition, and control strategies into a closed-loop system with a rigorous mathematical description. This enables the exoskeleton to adaptively assist in standing up and sitting down movements based on the user's current state and predicted intentions, forming a "human-machine fusion" intelligent control system.

[0249] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.

Claims

1. A motion intention analysis method for an exoskeleton robot, characterized in that: include: S1, multimodal data acquisition Collect joint torque sensor signals at the exoskeleton hip and knee joints, pressure sensor signals at the soles of the feet, and seat reaction force signals at the seat or backboard, and record them synchronously at a fixed sampling period; S2, feature extraction and normalization Calculating joint torque and joint torque change rate from the torque signal, calculating the plane coordinates and velocity of the plantar pressure center from the pressure signal, summing all plantar pressures to obtain a total plantar pressure, and normalizing each feature according to a preset maximum value to form a feature vector containing joint, pressure, and seat reaction force information; S3, intention probability prediction The sliding window is input into a prediction function, which is composed of a physical prior branch and a neural residual branch in parallel. The physical prior branch outputs a first intention probability based on the seat reaction force, the rate of change of the joint torque, and the velocity of the plantar pressure center. The neural residual branch performs temporal learning on the sliding window based on a long short-term memory neural network to output a second intention probability. The two are weighted to obtain the final movement intention probability. S4, adaptive control output The proportional, integral and differential gains are dynamically adjusted according to the motion intention probability, and a predicted feedforward torque proportional to the intention probability is superimposed to generate a target torque command for controlling the exoskeleton hip and knee joint actuators.

2. The motion intention analysis method for an exoskeleton robot according to claim 1, characterized in that: The multimodal data acquisition includes: The human-exoskeleton system is modeled as a planar multi-rigid body system, and its state is described using generalized coordinates; its motion equation can be expressed as a constrained dynamic form: Where M(q) is the inertia matrix and q is the generalized coordinate; is the centrifugal force term; is the generalized velocity, is the generalized acceleration; G(q) is the gravity term; τ exo is the exoskeleton joint driving torque; is the Jacobian matrix of the plantar, T is the transpose of the matrix; F foot is the contact force on the sole of the foot; is the Jacobian matrix of the seat contact; F seat is the contact force of the seat.

3. The motion intention analysis method for an exoskeleton robot according to claim 2, characterized in that: The feature extraction and normalization include: The plantar pressure sensor will measure the pressure values ​​​​p at multiple points i (t), its corresponding physical location is r i , so the center of gravity of the sole pressure can be calculated: Where CoP(t) is the center of pressure of the plantar surface at time t; p i (t) is the pressure value of the i-th sensor at time t; r i is the position vector of the i-th sensor on the sole of the foot; The joint torque sensor outputs the knee and hip joint torque time series τ j (t), the rate of change of torque can be calculated by taking the derivative in, is the rate of change of torque of the jth joint; τ j (t) is the instantaneous driving torque of the jth joint; It means finding the first derivative of time; Get the eigenvector: Where f(t) is the eigenvector; CoP x (t) and CoP y (t) is the plantar pressure center in the x and y directions, respectively; and are the rates of change of the CoP vector in the x and y directions, respectively; τ knee (t) and τ hip (t) are the moments of the knee and hip joints, respectively; and are the rates of change of moment at the knee and hip joints, respectively; is the total plantar pressure.

4. The motion intention analysis method for an exoskeleton robot according to claim 3, characterized in that: The intention probability prediction includes: A parallel two-branch structure prediction model, consisting of a physical prior branch and a neural network residual branch: Among them, the physical a priori branch includes: Based on the laws of human kinematics and dynamics, the initial probability of movement intention is calculated using a weighted summation method: Normalized value of seat reaction force. The lower the value, the more obvious the user's intention to leave the seat. The sum of the normalized values ​​of the joint torque change rate. A higher value indicates that the user is actively exerting force. Normalized value of the movement speed of the plantar pressure center. A higher value indicates a more obvious trend of center of gravity movement. The above three quantities are linearly combined with weights and mapped to the (0, 1) interval by the sigmoid function as the first motion intention probability; Among them, the residual branch of the neural network includes: The long short-term memory network is used to learn the temporal feature matrix, automatically capturing the complex temporal and nonlinear relationships between features, and outputting the residual intention probability after activation function processing as a correction to the physical prior branch probability; The first intention probability output by the physical prior branch and the residual probability output by the neural network residual branch are linearly weighted and summed, and the result is clipped to the range of 0 to 1 to obtain the final motion intention probability; Based on the preset threshold, the final movement intention probability is determined as "maintain current posture" or "posture change intention" to provide clear analysis results for subsequent exoskeleton control strategies; Specifically, the model maps the input feature f(t) to the predicted intent output Π(f(t)); Among them, Π(*) is the prediction function, and the final output is a specific intent label and its probability value; The calculation method is: Π(f(t))=σ(g phy (f(t)))+r θ (f(t))-λ phy L phy (t); Among them, σ(g phy (f(t))) is the physical prior term, σ(*) is the Sigmoid activation function, which maps the independent variable to the range of 0 to 1; g phy (f(t)) represents the physical prior term; r θ (f(t)) is the neural residual term; λ phy is the physical loss weight; L phy (t) is the physical drive loss function, which uses the physical residual of the power system as the loss of motion intention analysis. It is calculated as: The physical a priori branch is derived as follows: Where R(t) represents the seat reaction force, R max is the normalized term of the seat reaction force; and represents the rate of change of joint torque, is the normalized term of the joint torque; is the rate of change of the CoP vector, is the normalized term of the rate of change of the CoP vector; ΔCop(t) is the displacement difference between the current position of the center of pressure and the initial sitting position; γ is the displacement weight, which is used to balance the influence of the displacement gap; α1, α2 and α3 are weights, and their constraints are: Stage probability update equation: I stage (t)=ρ·I stage (t-1)+(1-ρ)·Π(f(t)); Among them, I stage (t) is the stage probability at time t, which is used to represent the smoothed probability that the user is in a certain posture at time t; I stage (t-1) is the stage probability at time t-1; ρ is the historical probability memory coefficient; The motion intention classification results are obtained based on the stage probability: in, is the result of motion intention classification; θ low is the lower threshold of the stage probability. When it is less than this value, the system determines that the user has no intention to change posture. θ mid is the upper threshold. When it is greater than this value, it means that the user has taken a conversion action. θ low ≤I stage (t)<θ mid It indicates that the stage probability is between two thresholds, and the system determines that the user is in the transition stage of posture conversion.

5. The motion intention analysis method for an exoskeleton robot according to claim 1, characterized in that: The adaptive control output includes: Using the impedance control method, a PID controller is constructed, whose goal is to guide the joint to follow a desired trajectory: Where τ is the control strategy output; q ref (t)-q(t) represents the joint position error, Indicates speed error; K p is the proportional gain; K d is the differential gain; K integrate is the integral gain; dλ is the integral differential element.

6. A computer system, characterized in that: include: processor; a memory for storing processor-executable instructions; Wherein, the processor is configured to implement a motion intention analysis method for an exoskeleton robot according to any one of claims 1 to 5 when executing the executable instructions.

7. A computer-readable storage medium, characterized in that include: a memory having a computer program stored thereon; A processor is used to execute the program in the memory to implement a motion intention analysis method for an exoskeleton robot according to any one of claims 1 to 5.

Images