An intelligent walking aid robot system based on human-machine force interaction and a control method thereof

Abstract

This invention relates to the field of rehabilitation and assistive robotics technology, and discloses an intelligent walking assistance robot system and method based on human-machine force-sensory interaction. The system includes a perception layer, an estimation layer, a decision-making layer, a control layer, and an execution layer. The perception layer collects multi-dimensional human-machine interaction forces, user movement postures, and environmental information. The estimation layer accurately estimates the interaction forces based on a floating base full exoskeleton dynamic model and performs gravity and dynamic force compensation, while simultaneously identifying terrain features. The decision-making layer identifies the user's movement intentions through a multi-modal fusion model and employs a sequential probability ratio test algorithm for proactive fall risk protection. The control layer adaptively adjusts impedance parameters through a virtual mass controller and uses a reinforcement learning algorithm for online personalized updates. The execution layer provides assistance. This invention solves the problems of unnatural interaction, lack of personalized adaptation, passive safety protection, and poor environmental adaptability in existing walking assistance robots, achieving natural, safe, and personalized intelligent walking assistance.

Description

Technical Field

[0001] This invention relates to the field of rehabilitation and assistive robot technology, specifically to an intelligent walking assistance robot system and its control method based on human-machine force perception interaction. Background Technology

[0002] Exoskeleton robotics is a cutting-edge field that intersects rehabilitation engineering and robotics, aiming to provide active and compliant walking assistance for elderly people with reduced mobility or patients in rehabilitation. Achieving natural, safe, and personalized human-machine collaborative walking is the core goal that this field has long pursued.

[0003] Currently, mainstream exoskeleton-based mobility robots typically sense human-robot interaction forces by installing torque sensors at the joints or force sensors on the soles of the feet, using this information as input to the control system. Based on the framework of "admittance control" or "impedance control," the system adjusts the joint output according to the measured interaction forces, thereby providing assistance to the user. However, in practical applications, existing technologies still face a series of technical bottlenecks that urgently need to be addressed: First, the "unnaturalness" of human-computer interaction is prominent. Most existing systems rely on simplified local dynamics models (such as linear inverted pendulum models) for force estimation and control. These models struggle to accurately describe and compensate in real-time for the gravitational and inertial forces generated by the exoskeleton's mechanical body during a complete gait cycle (especially during dynamic switching between single-leg and double-leg support). As a result, the system suffers from significant "body dynamics disturbances," leading to persistently high residual interaction torque. Users often feel the need to exert extra effort to "counter" or "accommodate" the robot's movements while walking, resulting in a stiff and cumbersome collaborative experience that deviates from the human-centered design principle of assistance.

[0004] Secondly, the system lacks dynamic and personalized adaptability. Most control strategies employ pre-set, fixed impedance or admittance parameters. However, there are significant differences in weight, gait characteristics, muscle strength levels, and rehabilitation stages among different users, and the condition of the same user can fluctuate on different training days. Fixed control parameters cannot adapt to this intra- and inter-individual variability, resulting in either insufficient or overshooting assistance, making it difficult to achieve precise, adaptive, "tailor-made" assistance, thus limiting rehabilitation efficacy and user experience.

[0005] Third, the active safety protection mechanism is weak. Existing systems rely heavily on setting static thresholds for single kinematic parameters (such as trunk tilt angle) to determine the severity of falls and other emergencies. This method is slow to react and prone to false alarms due to fluctuations in daily movements. More importantly, its protective measures are often limited to passive emergency braking (joint locking), failing to predict and apply active balance correction torque in the early stages of imbalance risk, thus failing to effectively reduce the probability of actual falls.

[0006] Fourth, the system lacks adaptability to complex terrain. Existing systems typically design and optimize their control algorithms for flat laboratory surfaces. When faced with slopes, steps, or uneven surfaces common in real-world environments, the system lacks the ability to smoothly and in real-time switch control strategies based on terrain characteristics. This can lead to a mismatch between the assistive behavior and the environment, not only reducing the assistive effect but also potentially introducing new risks of instability.

[0007] In summary, those skilled in the art have been seeking a solution for assistive robots that can deeply integrate high-precision force perception, personalized adaptive control, intelligent safety decision-making, and multi-environment adaptability, in order to overcome the limitations of existing technologies in terms of natural interaction, accurate adaptation, proactive protection, and scenario generalization. Summary of the Invention

[0008] Based on the deficiencies of the prior art pointed out in the background section, this application aims to solve the following technical problems: How can we reduce human-computer interaction torque through more accurate dynamic estimation and compensation, achieve a natural and smooth collaborative walking experience, and overcome the stiffness of interaction?

[0009] How can control parameters adapt to changes in different users and their capabilities to achieve personalized and precise assistance, thus solving the problem of poor adaptability of fixed parameters?

[0010] How to achieve early and highly reliable prediction of fall risk and proactively apply corrective interventions, upgrading safety protection from passive braking to active protection, and reducing the actual risk of falls.

[0011] How can the system be made able to identify complex terrain and smoothly switch control strategies to improve the stability and safety of assistance in different environments and overcome the limitations of scene adaptability?

[0012] To address the aforementioned technical challenges, this application presents the following overall technical concept: Constructing a five-layer collaborative intelligent closed-loop system encompassing perception, estimation, decision-making, control, and execution. This system utilizes multi-dimensional force and motion perception, employing a full exoskeleton dynamics model to achieve high-precision interactive force estimation and compensation, and integrates multi-modal information to intelligently recognize user intentions. Based on this, it generates personalized assistance through an online adaptive control algorithm, while simultaneously integrating active safety protection and multi-environmental adaptation mechanisms, ultimately achieving natural, safe, and personalized intelligent mobility assistance.

[0013] Based on this concept, the specific technical solution of this application is as follows: In the first aspect, this application provides an intelligent walking assistance robot system based on human-machine force perception interaction.

[0014] The system comprises a perception layer, an estimation layer, a decision-making layer, a control layer, and an execution layer.

[0015] The perception layer is used to collect multi-dimensional human-computer interaction force signals, user motion posture signals and environmental information in real time, extract real-time motion state parameters based on the user motion posture signals, and generate user historical motion data based on historically collected user motion posture signals.

[0016] The estimation layer is used to estimate human-computer interaction force in real time based on the full exoskeleton dynamics model, the multidimensional human-computer interaction force signal and the user's motion posture signal, and to identify environmental features based on the environmental information.

[0017] The decision layer is used to identify the user's movement intention based on the human-computer interaction force, the user's movement posture signal and the environmental features, generate motion control commands, and output the motion control commands to the control layer.

[0018] The control layer is used to receive the motion control command and determine the desired interaction force according to the motion control command, and adaptively adjust the impedance parameter based on the error between the desired interaction force and the human-machine interaction force, the user's historical motion data and the real-time motion state parameters to generate joint control torque.

[0019] The execution layer is used to drive joint movement according to the joint control torque, providing assist output.

[0020] Secondly, based on the same inventive concept, this application provides a control method for an intelligent walking robot based on human-machine force perception interaction.

[0021] The method includes: S1: Real-time acquisition of multi-dimensional human-computer interaction force signals, user motion posture signals and environmental information, extraction of real-time motion state parameters based on the user motion posture signals, and generation of user historical motion data based on historically acquired user motion posture signals; S2: Based on the full exoskeleton dynamics model, the human-computer interaction force is estimated in real time according to the multi-dimensional human-computer interaction force signal and the user's motion posture signal, and environmental features are identified according to the environmental information; S3: Identify the user's motion intention based on the human-computer interaction force, the user's motion posture signal, and the environmental features, and generate motion control commands; S4: Receive the motion control command and determine the desired interaction force according to the motion control command. Based on the error between the desired interaction force and the human-machine interaction force, the user's historical motion data and the real-time motion state parameters, adaptively adjust the impedance parameters to generate joint control torque. S5: Drive joint movement according to the joint control torque to provide power assist output.

[0022] Furthermore, as an optimization or supplement to the first aspect of the technical solution described above, this application also provides the following preferred technical solution: In a preferred embodiment, the sensing layer further includes joint torque sensors and kinematic sensors. The estimation layer includes a floating base full exoskeleton dynamics model, used to estimate the gait cycle during the single-leg support phase and the double-leg support phase using a single-rigid-body dynamics model and a double-rigid-body dynamics model, respectively, based on the signals from the joint torque sensors and kinematic sensors. This model calculates human-machine interaction forces in real time and performs exoskeleton gravity compensation and dynamic force compensation, which are used to pre-counteract exoskeleton propriodynamic disturbances when generating control commands.

[0023] In another preferred embodiment, the control layer includes a virtual mass controller. The virtual mass controller receives the human-machine interaction force and the desired interaction force, calculates the interaction force error, and adaptively adjusts stiffness, damping, and inertia parameters based on the error to generate joint control torque; simultaneously, it updates the impedance parameters online based on the user's historical motion data using a policy gradient reinforcement learning algorithm.

[0024] In another preferred embodiment, the decision layer includes a multimodal intent recognition model. The model receives the human-computer interaction force, user motion posture signals, and environmental features, and outputs a recognition result for the user's motion intent category through fusion processing based on a long short-term memory network. The categories include starting, accelerating, decelerating, turning, stopping, going up stairs, and going down stairs. The decision layer is also used to generate motion control commands based on the recognition results.

[0025] In another preferred embodiment, the decision layer further includes a safety policy manager. The manager receives the user's motion posture signal, calculates a posture stability index, and uses the Sequential Probability Ratio Test (SPRT) algorithm to calculate a log-likelihood ratio of fall risk based on this index. When the log-likelihood ratio exceeds a preset threshold, a fall risk is determined, and a corrective torque is calculated and output to the control layer.

[0026] In another preferred embodiment, the perception layer further includes a lidar and a depth camera for collecting environmental information. The estimation layer also includes a terrain feature analysis module for extracting and classifying terrain features (such as slope, step height, etc.) based on the environmental information. The control layer further includes a control strategy library for progressively switching control strategies within a preset time period based on the terrain classification results and terrain feature parameters.

[0027] In another preferred embodiment, the system further includes a multi-layered safety redundancy mechanism (sensing, control, and communication redundancy). When the primary system fails, the backup system automatically takes over; when an extremely high risk is detected (such as a log-likelihood ratio exceeding an emergency threshold and abnormal attitude indicators), the execution layer locks the joints and issues an alarm.

[0028] In another preferred embodiment, the multidimensional force sensor array of the sensing layer is specifically arranged at the hip joint, knee joint, ankle joint and sole of the foot, for measuring three-dimensional force, three-dimensional torque and sole pressure distribution.

[0029] In another preferred embodiment, the perception layer further includes a user input module for receiving information such as the rehabilitation stage. The control layer also includes a mode switcher for automatically switching between various control modes such as zero torque, assist, impedance, and error correction based on the rehabilitation stage information or environmental characteristics.

[0030] The technical solution provided in this application, compared with the closest prior art, can produce the following significant beneficial effects: By constructing a five-layer collaborative closed-loop system and introducing an online learning mechanism based on user data, high-precision perception, intelligent decision-making, and adaptive control are systematically integrated. This provides a unified framework for simultaneously solving multiple technical challenges such as natural interaction, personalized adaptation, active safety, and environmental adaptation, achieving synergistic gains in technical effects.

[0031] A floating base full exoskeleton dynamic model is adopted, and estimation strategies are switched for different gait stages. Gravity and dynamic force compensation are performed simultaneously, which can largely offset the interference of exoskeleton propriodynamics. This allows the system to reduce the average absolute interaction torque of the support phase to an extremely low level (e.g., below 0.05 Nm / kg), directly and effectively solving the core problems of "unnatural interaction and poor user experience," making human-machine collaborative movement as smooth as a "second skin."

[0032] By employing a virtual quality controller to adjust parameters based on real-time interactive force errors and integrating policy gradient reinforcement learning for online iterative optimization, impedance parameters can dynamically adapt to the user's individual characteristics and state evolution. This breaks the limitations of fixed parameters, substantially solves the problem of "lack of personalized adaptation," and achieves precise assistance from "one-size-fits-all" to "one-person-one-policy."

[0033] By employing a multimodal fusion model to extract and fuse deep features from force perception, kinematics, and environmental information, the accuracy and breadth of understanding user intentions in complex scenarios (such as climbing stairs) are significantly improved. This provides key input for generating precise and forward-looking control commands, enhancing the system's intelligence and practical application.

[0034] By performing sequential probability analysis on posture stability indicators to determine risk, early and high-confidence detection of subtle imbalance trends is achieved, and active corrective torque can be output. This revolutionizes safety protection from "post-incident braking" to "in-process intervention," effectively solving the problem of "passive and lagging safety protection." Experiments show that it can significantly reduce the incidence of falls.

[0035] By using LiDAR, depth cameras, and terrain analysis modules to perceive and classify terrain in real time, and driving the control strategy library to smoothly switch strategies, the system is given the ability to "understand" the environment and "adjust strategies." This fundamentally improves the deficiency of "poor environmental adaptability" and ensures the assist performance and safety boundaries in various real terrains such as slopes and steps.

[0036] A multi-layered safety redundancy mechanism provides hardware and logic backup for the system; a finely arranged multi-dimensional force sensor array provides more comprehensive and reliable sensing data; and switchable multi-control modes enable the system to flexibly adapt to different rehabilitation stages and scenario requirements. These improvements collectively enhance the overall reliability, accuracy, and applicability of the system, providing users with a safer and more comfortable comprehensive experience. Attached Figure Description

[0037] The specific embodiments of this application will be further described in detail below with reference to the accompanying drawings. It should be understood that the described drawings are only for explaining this application and do not constitute any undue limitation on the scope of protection of this application.

[0038] Figure 1 This is an overall architecture block diagram of an intelligent walking assistance robot system provided in an embodiment of this application.

[0039] Figure 2 This is a flowchart of an intelligent walking robot control method provided in an embodiment of this application.

[0040] Figure 3 This is a logical schematic diagram of a security decision-making mechanism provided in an embodiment of this application.

[0041] Figure 4 This is a schematic diagram of a multimodal intent recognition model provided in an embodiment of this application. Detailed Implementation

[0042] The specific embodiments of this application will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustrative purposes only and do not constitute any undue limitation on the scope of protection of this application. Those skilled in the art, after reading the above-described summary of this application, will be fully capable of implementing the technical solutions of this application.

[0043] In the following description, for clarity and brevity, detailed descriptions of known functions and constructions have been omitted. The same reference numerals in the accompanying drawings consistently denote the same elements.

[0044] Example 1 This embodiment provides an intelligent walking assistance robot system and its control method based on human-machine force perception interaction, aiming to achieve natural, safe and personalized walking assistance through high-precision force perception, intelligent decision-making and adaptive control.

[0045] Please see Figure 1 The diagram illustrates the overall architecture of the intelligent mobility assistance robot system 100 of this embodiment. The system 100 includes a perception layer 110, an estimation layer 120, a decision layer 130, a control layer 140, and an execution layer 150 connected in sequence, forming a complete closed loop from information perception to assistance execution.

[0046] 1. Perception Layer 110 The perception layer 110 is the physical interface for information interaction between the system and users and the external environment, and is responsible for collecting raw data in real time and in multiple dimensions.

[0047] The perception layer 110 includes a multi-dimensional force sensor array 111, a motion posture sensor 112, a joint torque sensor 113, a kinematic sensor 114, an environmental perception module 115, and a user input module 116.

[0048] The multi-dimensional force sensor array 111 is arranged at key interaction points between the user and the robot. In a specific example, such as... Figure 1 As shown, the array 111 includes: a first multidimensional force sensor 111a disposed at the hip joint exoskeleton connection for measuring the three-dimensional forces (Fx, Fy, Fz) and three-dimensional torques (Mx, My, Mz) of human-computer interaction at the hip joint; a second multidimensional force sensor 111b disposed at the knee joint exoskeleton connection; a third multidimensional force sensor 111c disposed at the ankle joint exoskeleton connection; and a fourth multidimensional force sensor 111d integrated into the plantar pressure insole for measuring the pressure distribution in various areas of the sole. The signals collected by the multidimensional force sensors 111a, 111b, 111c, and 111d together constitute the multidimensional human-computer interaction force signal. The sensors can be selected from six-dimensional force / torque sensors (such as the ATI Mini series) and flexible piezoresistive pressure sensor arrays.

[0049] The motion attitude sensor 112 is used to collect the user's overall kinematic information. In a preferred embodiment, the motion attitude sensor 112 includes an inertial measurement unit (IMU) 112a worn on the user's torso (e.g., waist), and IMUs 112b and 112c arranged on the thigh and calf struts. The IMU includes a three-axis accelerometer, a three-axis gyroscope, and a three-axis magnetometer, used to output the user's attitude angles (roll angle, pitch angle, yaw angle), angular velocity, and acceleration information in real time, constituting the user's motion attitude signal.

[0050] The joint torque sensor 113 and kinematic sensor 114 are unique sensors used in the estimation layer for high-precision dynamic calculations. The joint torque sensor 113 (such as a strain gauge torque sensor) is directly mounted on the output shaft of the hip, knee, and ankle joint actuators to directly measure the joint output torque, thus forming a joint torque signal. The kinematic sensor 114 includes a high-precision encoder (such as a photoelectric absolute encoder), mounted on the joint motor, to measure the precise angles and angular velocities of the joint, thus forming a kinematic signal.

[0051] The environment perception module 115 is used to acquire information about the robot's environment. In one specific embodiment, the environment perception module 115 includes a 16-line LiDAR 115a mounted in front of the robot's torso and an RGB-D depth camera 115b. The LiDAR 115a is used to scan the environment in front and generate three-dimensional point cloud data; the depth camera 115b is used to capture color images and corresponding depth images. These data together constitute the environmental information.

[0052] The user input module 116 is an optional module used to receive direct commands or information from the user. For example, the module 116 can be a touch screen or physical button integrated into the handrail, allowing the user or therapist to select the current rehabilitation stage (such as "bedridden stage", "standing stage", "early walking stage", "community walking stage"), or manually switch control modes.

[0053] The core function of the perception layer 110 is not only to acquire raw signals, but also to perform preliminary signal processing and feature extraction. Specifically, the perception layer 110 (or its associated data processing unit) is configured as follows: The user's motion posture signal (from IMU 112a-c) is filtered (e.g., Kalman filtering), posture is calculated, and the real-time motion state parameters are extracted in real time. These parameters include, but are not limited to: gait cycle phase (e.g., heel strike, full foot strike, toe lift-off events), cadence (steps / minute), stride length, walking speed, trunk tilt angle, and angular velocity.

[0054] The system continuously stores user motion posture signals and extracted state parameters from multiple historical frames to form the user's historical motion data. This data can be used to create independent profiles based on user IDs for long-term personalized learning.

[0055] 2. Estimated layer 120 The estimation layer 120 receives multi-dimensional human-computer interaction force signals, user motion posture signals, joint torque signals, kinematic signals and environmental information from the perception layer 110. Its core task is to "understand" the current physical interaction state and environmental conditions.

[0056] The estimation layer 120 includes a floating base full exoskeleton dynamic model 121 and a terrain feature analysis module 122.

[0057] (1) Dynamic model of floating base full exoskeleton 121 This model 121 is used to solve the problem of inaccurate interaction force estimation caused by model simplification in the prior art. Unlike the traditional model that regards the exoskeleton as fixed to the ground, the "floating base" model of this application regards the exoskeleton and the wearer's body as a whole system that moves freely in space, and the contact between its feet and the ground is dynamic and intermittent.

[0058] The model 121 is based on the joint torque signal collected by the joint torque sensor 113 ( ) and the kinematic signals (joint angles) collected by the kinematic sensor 114 angular velocity Real-time calculation of human-computer interaction force ( The specific implementation process is as follows: First, establish the Lagrange dynamic equations of the system:

[0059] in, It is the system's inertia matrix. It is the matrix of Coriolis force and centripetal force. It is a gravity term. It is the output torque of the joint motor (which can be obtained from...) approximate), It is the Jacobian matrix of the foot end relative to the joint space.

[0060] Because of direct solution Acceleration required However, acceleration estimation is noisy. This embodiment adopts a two-step estimation strategy to make full use of gait phase information: During the single-leg support phase: the system can be approximated as a single rigid body chain structure with the supporting leg as its base. Model 121 uses a single rigid body dynamics model for estimation. In this phase, the dynamic influence of the non-supporting leg is relatively small; by using the aforementioned dynamic equations and utilizing known... , , and estimated It can accurately solve the human-machine interaction force between the supporting leg and the ground (mainly manifested as the reaction force of the foot).

[0061] During the bipedal support phase: At this point, the system is a closed-loop dynamic structure with complex forces. Model 121 is switched to a dual-rigid-body dynamic model for estimation. This model treats the legs as two interacting rigid-body chains, establishes coordinated dynamic equations for the two chains, and introduces plantar pressure distribution information (from sensor 111d) as constraints to jointly solve the human-machine interaction forces between the feet and the ground, as well as between the legs.

[0062] More importantly, the model 121 performs dynamic compensation in real time. It does so based on the kinematic signals. Real-time calculation of the gravity term generated by the exoskeleton mechanical body in the current configuration and inertial force term Then, in generating the final estimate... At the same time, through gravity compensation and dynamic force compensation, these forces are pre-countered in the control commands. This means that the torque commands issued by the controller already include the part of "lifting the robot's own weight and overcoming its inertia," thus significantly reducing the interactive torque that the user needs to actively provide to overcome the body dynamics. Experiments show that through this precise compensation, the average absolute interactive torque at the hip and knee joints of the support phase can be reduced to below 0.05 Nm / kg, achieving a nearly imperceptible compliant interaction.

[0063] (2) Terrain Feature Analysis Module 122 The module 122 receives environmental point cloud data and depth images from the environmental perception module 115 to address the robot's adaptability in complex terrain.

[0064] Its processing flow includes: By integrating and fusing the LiDAR point cloud with the depth image using a coordinate system, a dense 3D environment map of the area in front of the robot (e.g., within a 3-meter radius) can be constructed.

[0065] Key terrain features are extracted from the 3D map using a plane fitting algorithm and an edge detection algorithm based on Random Sample Consensus (RANSAC). These features include: slope (the angle between the fitted ground plane normal vector and the gravity direction), step height (the detected vertical height difference), and ground roughness (the standard deviation of the point cloud relative to the fitted plane).

[0066] Based on the extracted features, a pre-trained lightweight classifier (such as a Support Vector Machine, SVM) is used to classify the terrain ahead. The classification categories include: "flat ground", "uphill", "downhill", "ascending steps", "descending steps", and "uneven road surface".

[0067] Module 122 outputs terrain classification results and key feature parameters (such as slope value and step height) to decision layer 130 and control layer 140.

[0068] 3. Decision-making level 130 The decision layer 130 is the system's "intelligent brain," responsible for integrating multi-source information, understanding user intent, and assessing security status. The decision layer 130 includes a multimodal intent recognition model 131 and a security policy manager 132.

[0069] (1) Multimodal intent recognition model 131 This model 131 is used to accurately identify a user's movement intentions in complex environments, such as Figure 4 As shown Model 131 receives three inputs: 1) Human-computer interaction force from estimation layer 120 ( 1) The user's intention to exert force; 2) User motion posture signals from the perception layer 110 (especially gait phase and torso tilt angle), representing the user's motion state; 3) Environmental features (terrain classification results) from the estimation layer 120, representing external environmental constraints.

[0070] The core of Model 131 is a multimodal fusion model based on a Long Short-Term Memory (LSTM) network. The specific implementation is as follows: First, the original input is transformed into a high-level feature vector through three independent feature encoding sub-networks. The force-sensing encoding sub-network (a fully connected network) will... The gait phase and angle are converted into kinematic modal feature vectors; the kinematic coding sub-network (another fully connected network) converts gait phase, angle, etc. into kinematic modal feature vectors; the environment coding sub-network (a lightweight convolutional network) converts the one-hot encoding and parameters of terrain classification into environment modal feature vectors.

[0071] The feature vectors of the three modalities, including the current moment and several past frames (e.g., data from the past 0.5 seconds), are input into an LSTM network in chronological order. The LSTM network can learn the temporal dependencies of multimodal information. For example, it can identify the temporal coordination pattern of a user's "forward leaning of the torso" and "increased knee joint force" in an "uphill" environment.

[0072] The hidden state of the last time step of the LSTM is input into a fully connected classification layer, and the probability distribution is output through the Softmax function.

[0073] Model 131 outputs the recognition result of the user's motion intention category, and the decision layer 130 is further used to generate motion control commands based on the recognition result. The category is predefined and closely related to the control system, including: "start", "accelerate", "constant speed", "decelerate", "stop", "turn left", "turn right", "go up stairs", and "go down stairs". The recognition result is given in the form of a probability vector, and the system selects the category with the highest probability as the current decision.

[0074] (2) Security Policy Manager 132 The manager 132 is used to implement proactive and forward-looking safety protection, the core of which is to use a sequential probability ratio test algorithm to make fall risk decisions.

[0075] The manager 132 continuously receives user motion posture signals from the perception layer 110, especially the tilt angle (θ) and tilt angular velocity (ω) of the torso.

[0076] Its workflow is a clear decision-making chain: A comprehensive attitude stability index (SI) is calculated based on θ and ω. For example, ,in The steady-state upright angle is represented by α and β, which are weighting coefficients. The larger the value of this index, the more unstable the attitude.

[0077] A continuous SI sequence is input into the SPRT algorithm. SPRT is a statistical sequence analysis algorithm that defines two hypotheses: H0 (no risk of falling) and H1 (risk of falling). The algorithm updates the log-likelihood ratio (LLR) based on the new SI value at each time step.

[0078] in, and The probability distributions (such as normal distributions) fitted to historical fall data and normal walking data are obtained respectively.

[0079] Set two thresholds: risk threshold A and emergency threshold B (B>A).

[0080] like If the value is >A, it is determined that there is a risk of falling. At this time, the manager 132 does not simply issue an alarm, but calculates a corrective torque based on the current user's motion posture signals (such as center of gravity position and velocity). This torque is designed to generate an increment of joint torque that returns the torso to a stable position, and is output to the control layer 140 via a "feedforward" method, where it is superimposed on the assist torque generated by the control layer. This active intervention can make fine adjustments in the early stages of imbalance.

[0081] like If the value is B, it is determined that there is an extremely high and irreversible risk of fall. At this time, the manager 132 sends an emergency command to the execution layer 150 and the user interface.

[0082] For details, please refer to Figure 3 The figure illustrates a logical diagram of the safety decision-making mechanism in this embodiment. As shown, it employs the Sequential Probability Ratio Test (SPRT) algorithm for fall risk decision-making. The decision logic architecture of the safety policy manager 132 is as follows: The manager 132 continuously receives user motion posture signals from the perception layer 110.

[0083] The manager 132 calculates the attitude stability index based on the attitude signal and calculates the fall risk log-likelihood ratio (LLR) in real time based on the SPRT algorithm.

[0084] Risk assessment and corrective intervention: Determine whether the LLR is greater than the risk threshold A.

[0085] If LLR > A, a risk of falling is determined. Manager 132 calculates a corrective torque based on the current motion state and outputs it to control layer 140 to achieve active intervention.

[0086] If LLR ≤ A, the process ends without intervention.

[0087] Emergency braking determination: After determining that there is a risk (LLR>A) and outputting a corrective torque, further determine whether LLR is greater than the higher emergency threshold B.

[0088] If LLR > B, an extremely high and irreversible risk of fall is identified. Manager 132 sends an emergency command to execution layer 150, triggering joint locking and an alarm.

[0089] If LLR ≤ B, the process ends.

[0090] In addition, the system also includes a multi-layered security redundancy mechanism (such as sensing redundancy, control redundancy, and communication redundancy). When the main system fails, the backup system automatically takes over, forming another line of defense.

[0091] 4. Control Layer 140 The control layer 140 is responsible for receiving motion control commands output by the decision layer 130, determining the desired interaction force based on the motion control commands, and generating executable, personalized joint control commands based on the error between the desired interaction force and the human-machine interaction force. The control layer 140 includes a virtual mass controller 141, a control strategy library 142, and a mode switcher 143.

[0092] (1) Virtual Quality Controller 141 The controller 141 is the core of personalized adaptive control, which is based on a virtual mass (or admittance / impedance) control framework, but the parameters can be learned and updated online.

[0093] Specifically, it receives motion control commands from the decision layer 130 to determine the desired interaction force, human-computer interaction force from the estimation layer 120, and user historical motion data and real-time motion state parameters from the perception layer 110.

[0094] Calculate the interaction force error. The desired interaction force is determined based on the motion control command; for example, an "accelerate" command corresponds to a forward desired interaction force.

[0095] The internal model of controller 141 is a second-order virtual system: ,in , , These are virtual inertia, damping, and stiffness parameters. By adjusting these three parameters, the robot's dynamic response characteristics to external forces can be changed (such as "heavy / light" or "sticky / compliant"). In this embodiment, the controller 141... The size and direction of the data, as well as the user's real-time motion state (such as cadence), are adaptively fine-tuned using a fuzzy rule table or a shallow neural network. and To achieve a more comfortable interaction.

[0096] Based on the adjusted impedance model and the desired motion (determined by motion control commands), the desired joint acceleration, velocity, and position are calculated. Then, combined with the robot dynamics model, the joint control torque is finally generated.

[0097] The controller 141 integrates an online learning module for a policy gradient reinforcement learning algorithm. This module uses the user's historical motion data (long-term) and current motion performance (short-term) as evaluation criteria. Its learning objective is to maximize a "comfort" reward function, which may integrate metrics such as interaction force smoothness, gait symmetry, and user energy consumption estimation. The learning algorithm periodically (e.g., every 10 gait cycles) adjusts the impedance parameters. , , Fine-tuning is performed to gradually adapt to the user's personalized exercise patterns. This means that the control parameters are no longer fixed, but are continuously optimized as the user interacts with them.

[0098] (2) Control strategy library 142 and mode switcher 143 These two modules work together to enable multi-environment adaptation and multi-mode switching.

[0099] Control strategy library 142: Stores multiple sets of pre-optimized control parameters or control laws for different terrain types (from module 122). For example, an "uphill" strategy may include a larger base assist gain and a more forward body center of gravity setting; a "downhill" strategy includes a more cautious foot trajectory and higher joint stiffness.

[0100] Mode switcher 143: Automatically or suggest switching the global control mode based on recovery stage information from user input module 116 in perception layer 110, or environmental features (terrain classification) from estimation layer 120. The control modes include: Zero torque control mode: suitable for the early stage of rehabilitation, the robot completely follows the user's passive movement, only compensating for gravity, without providing active assistance.

[0101] Assisted control mode: Suitable for the walking training period, the robot provides partial assistance based on intent recognition to help users complete walking actions.

[0102] Impedance control mode: Suitable for strength training, the robot simulates different impedance environments, allowing users to actively overcome resistance during training.

[0103] Error correction control mode: When the security policy manager 132 is activated, an active correction torque is superimposed on this mode.

[0104] The mode switcher 143 ensures that the system behavior matches the user's capabilities and environmental requirements.

[0105] 5. Execution layer 150 The execution layer 150 is the "limb" of the system, responsible for precisely executing control commands. It includes joint actuators 151 (such as frameless motors or harmonic reducers), brakes 152, and audible and visual alarms 153. The execution layer 150 receives joint control torque from the control layer 140. The actuator 151 converts the motion into actual joint movement, providing assist output.

[0106] When an emergency command is received from the decision-making layer 130 and the security policy manager 132 ( >B) When the brake 152 (such as an electromagnetic power failure brake) is activated, the relevant joint is locked, and at the same time the audible and visual alarm 153 is activated to prevent falls to the greatest extent and to alert the caregiver.

[0107] 6. Multi-layered security redundancy mechanism In addition to the aforementioned algorithm-level security strategies, this embodiment incorporates a multi-layered security redundancy mechanism in its system hardware and architecture.

[0108] Perception redundancy: Key sensors (such as IMU and joint encoders) are equipped with dual sets for cross-validation of data.

[0109] Control redundancy: In addition to the main control loop (based on the virtual mass controller 141), there is an independent backup control loop with lower computational load (such as position control based on PD). When the main system fails (such as software crash or communication timeout), the backup system automatically takes over within milliseconds to maintain the basic stability of the joint.

[0110] Communication redundancy: Critical data channels use dual-bus transmission (such as CAN bus and Ethernet in parallel).

[0111] Emergency braking: As mentioned above, when the following conditions are met... When B and SI > preset angle threshold, the brake 152 of the execution layer 150 is directly activated to achieve final hardware-level safety protection.

[0112] Example 2 This embodiment provides a control method for an intelligent mobility assistance robot based on human-machine force perception interaction. Please refer to [link / reference]. Figure 2 The flowchart of the method is shown. The method can be implemented using the system described in Example 1, and includes the following steps: S210: Signal acquisition and parameter extraction.

[0113] Multidimensional force sensor arrays and motion posture sensors are used to collect multidimensional human-computer interaction force signals, user motion posture signals, and environmental information in real time. Based on the user motion posture signals, real-time motion state parameters such as gait phase, stride frequency, stride length, and torso tilt angle are extracted through filtering and calculation. Simultaneously, the user database is accessed to obtain or generate historical user motion data based on historical data.

[0114] S220: Interaction force estimation and environment recognition.

[0115] Based on the full exoskeleton dynamics model, the system estimates accurate human-computer interaction forces in real time by collecting multidimensional human-computer interaction force signals and user motion posture signals (combining joint torque and kinematic signals). Specifically, a floating base model is used, and the system switches between single-leg and dual-leg rigid body models for estimation during the single-leg / dual-leg support phase. At the same time, the system calculates and compensates for the exoskeleton's gravity and dynamic forces.

[0116] In parallel, based on environmental information (LiDAR point cloud, depth image), features such as slope and step height are extracted through the terrain feature analysis module, and terrain classification is performed to identify environmental features.

[0117] S230: Motion Intent Recognition and Safety Decision Making.

[0118] Based on the human-computer interaction force estimated in step S220, the user motion posture signal collected in step S210, and the identified environmental features, a multimodal intent recognition model (based on LSTM) is used for fusion processing to identify the user's current motion intent categories such as starting, accelerating, turning, and going up and down stairs, and motion control commands are generated based on the recognition results.

[0119] Simultaneously, an attitude stability index is calculated based on the user's motion posture signal, and the sequential probability ratio test (SPRT) algorithm is used to calculate the log-likelihood ratio (LLR) of fall risk in real time. Risk is determined based on the LLR value.

[0120] S240: Adaptive control and torque generation.

[0121] The system receives the motion control command generated in step S230 and determines the desired interaction force based on the command. Based on the error between the desired interaction force and the human-computer interaction force estimated in step S220, the user's historical motion data obtained in step S210, and real-time motion state parameters, the impedance control parameters are adaptively adjusted. Specifically, the interaction force error is calculated using a virtual mass controller, and the virtual damping and stiffness are adjusted accordingly. Furthermore, the impedance parameters are updated online using the user's historical motion data through a policy gradient reinforcement learning algorithm, achieving personalized adaptation.

[0122] Simultaneously, the desired interaction force is determined based on the motion control commands, and combined with the adjusted impedance model, the final joint control torque is generated. If step S230 determines that there is a risk of fall (LLR>A), a corrective torque calculated by the safety strategy manager is superimposed on this torque. Furthermore, based on environmental characteristics or the user-defined rehabilitation stage, the control mode is automatically selected or switched via the control strategy library and mode switcher.

[0123] S250: Facilitating Execution.

[0124] The joint control torque generated in step S240 is sent to the joint driver in the execution layer to drive the joint movement and provide power assist output. If step S230 determines that there is an extremely high risk (LLR>B), the joint locking and alarm functions in the execution layer are triggered simultaneously.

[0125] Example 3 This embodiment is a further detailed description of the multimodal intent recognition model 131 in Embodiment 1, providing an optional model training and deployment method.

[0126] The multimodal intent recognition model 131 needs to be trained before deployment. The training process is as follows: Data Preparation: Collect a large-scale quadruple dataset {force sequence, kinematic sequence, environmental label, true intention label}. The data can come from real human-computer interaction experiments or high-quality simulation environments. Divide the dataset into training, validation, and test sets in an 8:1:1 ratio.

[0127] Model training: Using the training set data, the parameters of the LSTM network and feature encoding subnetwork are trained with the goal of minimizing the cross-entropy loss for intent classification, employing the backpropagation algorithm and the Adam optimizer. A validation set is used to monitor the training process and prevent overfitting.

[0128] Federated Learning Optimization: To protect user privacy and enable continuous model evolution, this embodiment preferably employs a federated learning framework. Multiple robot systems (clients) deployed on user terminals fine-tune the model locally using their newly generated data, and then upload only the updated model parameters (not the original data) to the cloud server. The server aggregates the updates from all clients, generates a globally improved model version, and then distributes it to each client. This allows the model to learn from all user interactions, continuously optimizing the accuracy of intent recognition while ensuring data privacy.

[0129] Example 4 This embodiment is a further detailed explanation of the online learning function of the virtual quality controller 141 in Embodiment 1, and provides specific implementation parameters for the reinforcement learning algorithm.

[0130] The policy gradient reinforcement learning algorithm employs the proximal policy optimization (PPO) algorithm. Its key settings are as follows: State: Defined as the characteristics of the most recent N gait cycles, including the average interaction force, gait cycle variation coefficient, stride symmetry index, and user-reported fatigue score (collected periodically through user input module 116).

[0131] Action: Defined as an action on the virtual impedance parameter: Small, continuous adjustments (within ±10%) are made near the baseline value.

[0132] Reward: The reward function R is designed as follows:

[0133] Among them, w1~w4 are weighting coefficients used to balance interaction accuracy, gait quality, user comfort and system stability.

[0134] Training: The learning process is performed offline and asynchronously on the robot's local embedded computing unit. Every K gait cycles, the algorithm updates the policy network (a small neural network, with state as input and action distribution as output) using the collected state-action-reward trajectories. The updated policy is then used for parameter tuning in the next time interval.

[0135] Example 5 This embodiment provides a specific implementation of a system hardware architecture to illustrate the feasibility of the technical solution of this application.

[0136] System 100's hardware can adopt a distributed computing architecture: Lower layer: Each joint module is an independent node, containing a motor driver, joint torque sensor 113, encoder 114 and local microcontroller (such as STM32 series), responsible for high-frequency (1kHz) current loop, position loop control and sensor data acquisition.

[0137] Middle layer: The robot's torso houses the main controller (such as one based on the NVIDIA Jetson AGX Orin platform), which runs a real-time Linux kernel and integrates the core algorithms of the perception layer 110, estimation layer 120, decision layer 130, and control layer 140, with an operation frequency of 100Hz.

[0138] Upper layer: The user interface (tablet computer) is connected to the main controller via Wi-Fi and is used to display status, set parameters (user input module 116), and receive alarms.

[0139] Sensors: Multi-dimensional force sensor 111 (ATI Mini45) communicates via CAN bus; IMU 112 (Xsens MTi series) communicates via UART; LiDAR 115a (Velodyne VLP-16) and depth camera 115b (Intel RealSense D435) communicate via Ethernet.

[0140] All critical communication links are designed with dual redundancy.

[0141] Example 6 This embodiment illustrates the specific technical effects achieved by the technical solution of the present invention in solving the problems in the background technology, and verifies them through experimental data.

[0142] In a preclinical study of 10 stroke patients with hemiplegia, the system described in this embodiment was compared with a mainstream commercially available walking exoskeleton (using fixed-parameter admittance control).

[0143] Interaction naturalness: The mean absolute interaction torque of the support phase was measured. The result for this system was (0.04 ± 0.01) Nm / kg, while that for the control system was (0.15 ± 0.03) Nm / kg. The result for this system was significantly lower (p<0.01), demonstrating that its dynamic compensation is effective and the interaction is smoother.

[0144] Personalized adaptation: Records the impedance parameters automatically adjusted by the system before and after one week of use for each patient. The results showed that the parameter changes ranged from 35% to 60% of the initial values, and the direction of change was consistent with the improvement trend of the patient's muscle strength assessment results, proving that the system can make personalized adjustments.

[0145] Active safety protection: In simulated sudden disturbance experiments, the SPRT-based early warning system provided by this system was on average 1.2 seconds earlier than the fixed threshold early warning system of the control system. The application of active corrective torque reduced the user's active force required to restore a stable attitude by approximately 40%.

[0146] Environmental adaptability: On test roads containing ramps and steps, the system automatically switched between different terrains according to the terrain strategy, and the coefficient of variation of the user's walking speed was reduced by 25%, with no gait interruption or staggering caused by terrain switching.

[0147] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, such as replacing the LSTM network with a Transformer network for intent recognition; replacing policy gradient reinforcement learning with a deep deterministic policy gradient (DDPG) algorithm; replacing LiDAR with a solid-state LiDAR or a binocular vision system, etc. These modifications or substitutions should all be considered to be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An intelligent walking assistance robot system based on human-machine force perception interaction, characterized in that, include: The perception layer includes a multi-dimensional force sensor array and a motion posture sensor, which are used to collect multi-dimensional human-computer interaction force signals, user motion posture signals and environmental information in real time, extract real-time motion state parameters based on the user motion posture signals, and generate user historical motion data based on historically collected user motion posture signals. The estimation layer is used to estimate the human-computer interaction force in real time based on the multidimensional human-computer interaction force signal and the user's motion posture signal, and to identify environmental features based on the environmental information, based on the full exoskeleton dynamic model. The decision layer is used to identify the user's movement intention based on the human-computer interaction force, the user's movement posture signal and the environmental features, generate motion control commands, and output the motion control commands to the control layer; The control layer is used to receive the motion control command and determine the desired interaction force according to the motion control command, and adaptively adjust the impedance parameter based on the error between the desired interaction force and the human-machine interaction force, the user's historical motion data and the real-time motion state parameters to generate joint control torque; An execution layer is used to drive joint movement according to the joint control torque, providing assist output.

2. The system according to claim 1, characterized in that, The sensing layer also includes joint torque sensors and kinematic sensors; The estimation layer includes a floating base full exoskeleton dynamics model, which is used to estimate the gait cycle in the single-leg support phase and the double-leg support phase using a single rigid body dynamics model and a double rigid body dynamics model, respectively, based on the joint torque signal collected by the joint torque sensor and the kinematic signal collected by the kinematic sensor. It calculates the human-machine interaction force in real time and calculates the exoskeleton gravity compensation force and dynamic force compensation force based on the kinematic signal. The gravity compensation force and the dynamic force compensation force are used to pre-counteract the exoskeleton propriodynamic interference when generating control commands.

3. The system according to claim 1, characterized in that, The control layer includes a virtual quality controller; The virtual quality controller is used for: Receive the human-computer interaction force and the desired interaction force as input; Calculate the difference between the human-computer interaction force and the expected interaction force, and use this difference as the interaction force error; Based on the interaction force error, the stiffness, damping, and inertia parameters are adaptively adjusted to generate joint control torque; The controller also updates the impedance parameters online based on the user's historical motion data using a policy gradient reinforcement learning algorithm.

4. The system according to claim 3, characterized in that, The decision layer includes a multimodal intent recognition model, which is used for: The system receives the human-computer interaction force, the user's motion posture signal, and the environmental features as inputs. By using fusion processing based on Long Short-Term Memory (LSTM) networks, the human-computer interaction force is converted into force-feeling modal features, the user's motion posture signal is converted into kinematic modal features, the environmental features are converted into environmental modal features, and the force-feeling modal features, the kinematic modal features, and the environmental modal features are fused together. The system outputs the recognition results of the user's motion intention categories, which include starting, accelerating, decelerating, turning, stopping, going up stairs, and going down stairs. The decision layer is also used to generate motion control commands based on the recognition results.

5. The system according to claim 1, characterized in that, The decision-making layer also includes a security policy manager; The security policy manager is used for: Receive the user's motion posture signal as input; Calculate the attitude stability index based on the user's motion posture signal; The sequential probability ratio test (SPRT) algorithm is used to calculate the log-likelihood ratio of fall risk based on the attitude stability index. When the log-likelihood ratio exceeds a preset threshold, it is determined that there is a risk of falling, and a corrective torque is calculated based on the user's motion posture signal; The corrective torque is output to the control layer.

6. The system according to claim 1, characterized in that, The perception layer also includes a lidar and a depth camera for collecting environmental point cloud data and depth images; The estimation layer also includes a terrain feature analysis module, used for: Receive the environmental point cloud data and the depth image as input; Extract and classify terrain features, including slope, step height, and ground roughness; Output the terrain classification results and the terrain feature parameters; The control layer also includes a control strategy library, which is used to gradually switch control strategies within a preset time period based on the terrain classification results and the terrain feature parameters.

7. The system according to claim 5, characterized in that, The system also includes a multi-layered security redundancy mechanism, which includes sensing redundancy, control redundancy, and communication redundancy. When the primary system fails, the backup system automatically takes over; When the fall risk log-likelihood ratio exceeds the emergency threshold and the posture stability index exceeds the preset threshold, the execution layer locks the joint and issues an alarm.

8. The system according to claim 1, characterized in that, The multidimensional force sensor array is arranged at the hip joint, knee joint, ankle joint and sole of the foot to measure three-dimensional force and three-dimensional torque and sole pressure distribution. The multidimensional human-machine interaction force signal includes the three-dimensional force, the three-dimensional torque and the sole pressure distribution.

9. The system according to claim 3, characterized in that, The perception layer also includes a user input module for receiving rehabilitation stage information input by the user; The control layer also includes a mode switcher for automatically switching between multiple control modes based on the rehabilitation stage information or the environmental characteristics. The control modes include zero torque control mode, assist control mode, impedance control mode, and error correction control mode.

10. A control method for an intelligent walking robot based on human-machine force perception interaction, characterized in that, Includes the following steps: S1: Real-time acquisition of multi-dimensional human-computer interaction force signals, user motion posture signals and environmental information through a multi-dimensional force sensor array and motion posture sensor, extraction of real-time motion state parameters based on the user motion posture signals, and generation of user historical motion data based on historically acquired user motion posture signals; S2: Based on the full exoskeleton dynamics model, the human-computer interaction force is estimated in real time according to the multi-dimensional human-computer interaction force signal and the user's motion posture signal, and environmental features are identified according to the environmental information; S3: Identify the user's motion intention based on the human-computer interaction force, the user's motion posture signal, and the environmental features, and generate motion control commands; S4: Receive the motion control command and determine the desired interaction force according to the motion control command. Based on the error between the desired interaction force and the human-machine interaction force, the user's historical motion data and the real-time motion state parameters, adaptively adjust the impedance parameters to generate joint control torque. S5: Drive joint movement according to the joint control torque to provide power assist output.

Images