Self-help stair descending intelligent fall prevention rehabilitation evaluation and training system and method for hemiplegic patients

Abstract

The application discloses a kind of hemiplegic patient autonomous stair rehabilitation evaluation and training system and method of falling down.The system uses cognitive-motor double-layer control architecture: upper VLM rehabilitation decision engine predicts the compensation intention of ill leg retraction and ankle inversion, abnormal movement mode of pelvis and fear psychological state through cross-modal semantic understanding, and dynamically adjusts the penalty weight in lower model predictive controller cost function by Sigmoid mapping;Lower MPC real-time stable predictive controller optimizes mass center stability in millisecond level cycle rolling, actively cooperates control pelvis electric push rod, ill leg pneumatic exoskeleton and ankle joint electric stimulator and pneumatic bionic constraint-guide mechanism, and implements prospective fall prevention intervention;Pneumatic bionic constraint-guide mechanism realizes the direction selectivity constraint that only generates constraint in ill leg retraction direction.The application synchronously solves four obstacles in hemiplegic patient stair, including visual fear, ill leg retraction crossing, foot drop inversion and pelvis instability, and improves rehabilitation safety and efficiency.

Description

Technical Field

[0001] This invention relates to the fields of cognitive neurorehabilitation engineering, intelligent robot control and artificial intelligence, specifically to an intelligent fall prevention rehabilitation assessment and training system and method for hemiplegic patients to descend stairs independently, integrating visual language large model, emotion computing, model predictive control, pneumatic soft robot, active pelvic stabilization and functional electrical stimulation. Background Technology

[0002] Descending stairs is far more biomechanically complex and requires far greater neural control than ascending stairs, making it one of the final functional movements to overcome in the recovery of daily living abilities for hemiplegic patients. Studies have shown that the failure rate of hemiplegic patients descending stairs is approximately 43% higher than ascending stairs, and the risk of falling is 2.1 times higher. The core mechanisms underlying this impairment involve four intertwined dimensions: First, the neuropsychological fear mechanism. Because patients have long experienced failure in motor control, the visual system triggers a strong fear response in the amygdala when receiving information about stair steps, leading to abnormally high muscle tone and impaired motor initiation, forming a vicious cognitive cycle of "fear-compensation-failure." Existing training devices almost entirely neglect intervention at the psychological and neuropsychological levels.

[0003] Second, the compensatory mechanism of the affected leg when stepping. Spasm of the hip adductor muscle on the affected side is activated during the stepping phase when going downstairs, causing the affected knee to turn inward and cross over the healthy leg to form a "scissor cross". At the same time, there is a plantar flexion-inversion pattern of the ankle joint (equinus inversion foot). This compensatory behavior has a very strong neurological habit and is difficult to correct continuously by manual guidance alone.

[0004] Third, the lack of pelvic stability mechanism. Insufficient trunk support on the affected side leads to uncontrolled forward swaying of the healthy pelvis during descent, disrupting the coordination of both lower limbs and further exacerbating compensatory behavior in the affected leg.

[0005] Fourth, there is a lack of real-time stability control. Existing electric assistive devices lack the ability to predict the patient's overall stability in advance, and often only react after instability has occurred, which can easily lead to secondary falls.

[0006] In existing techniques, the main manual training methods used by therapists to help hemiplegic patients descend stairs include: taking right-sided hemiplegia as an example, the therapist uses an arc-shaped structure to surround the patient's sacrum to shift the center of gravity towards the healthy side, placing one hand on the patient's healthy pelvis to prevent tipping, and the other hand on the affected knee to pull the knee forward and fully flex it, guiding the affected leg to step outward and avoiding adduction and crossing of the affected leg. However, this manual method is highly dependent on the therapist's experience and physical strength, making it difficult to standardize, quantify, and implement consistently over the long term.

[0007] A few existing devices employ simple feedback control strategies, intervening only after an abnormality is detected. They lack the ability to predict the patient's overall stability and cannot understand the patient's motor intentions and psychological state. In summary, current technologies lack a systematic solution to address the four-fold impairment mechanism of neurotic fear, compensatory patterns, pelvic instability, and lack of real-time control. Summary of the Invention

[0008] The purpose of this invention is to propose an intelligent fall prevention rehabilitation assessment and training system and method for hemiplegic patients to descend stairs independently, so as to realize the simultaneous intervention of psychological and motor obstacles in the training of hemiplegic patients descending stairs, the prospective prevention and control of fall risk, and the precise correction of compensatory patterns, thereby significantly improving the safety and efficiency of stair descent rehabilitation training.

[0009] To achieve the above objectives, in a first aspect, the present invention provides an intelligent fall prevention rehabilitation assessment and training system for hemiplegic patients to descend stairs independently, based on a visual language model and model predictive control, comprising: The neuro-fear desensitization module is used to identify and intervene in patients' fear-anxiety neurological responses when going downstairs; A multimodal three-dimensional motion capture subsystem is used to acquire the patient's kinematic, dynamic, and physiological electrical signals; The VLM rehabilitation decision engine is used to perform semantic understanding of the action of going down stairs and to predict the compensatory intention of the affected lower limb and the pelvic movement pattern. The MPC real-time stability predictive controller is used to predict the patient's center of mass trajectory and joint abnormality trends in millisecond cycles, and output multi-actuator collaborative control commands. An active pelvic stabilization module, worn on the patient's pelvis, is used to actively apply horizontal thrust according to the instructions of the controller to adjust the center of gravity shift. The affected leg assist exoskeleton covers at least the abduction degree of freedom of the hip joint and the flexion and extension degree of freedom of the knee joint on the affected side, and is used to output abduction torque and knee flexion assist torque according to the instructions of the controller; An ankle joint intelligent electrical stimulator is used to output electrical stimulation according to the instructions of the controller to inhibit foot drop and inversion; A pneumatic bionic restraint-guidance mechanism, including at least a hip-knee directional restraint soft exoskeleton on the affected side, is used to apply directional restraint force to the affected lower limb; And a context-aware voice interaction platform to provide voice guidance and feedback.

[0010] The system adopts a "cognitive-motor" two-layer control architecture: the upper layer is the VLM rehabilitation decision engine, which is used to classify and output the probability of adduction of the affected leg, the probability of ankle inversion, and the pelvic movement pattern; the lower layer is the MPC real-time stability prediction controller.

[0011] The VLM rehabilitation decision engine outputs the probabilities of the affected leg adduction and ankle inversion, which are then dynamically adjusted via a nonlinear sigmoid mapping function to control the hip adduction angle penalty weight and ankle inversion angle penalty weight in the cost function of the MPC real-time stability predictive controller. This enables cross-level propagation of semantically understood vectorized constraints. The MPC real-time stability predictive controller performs rolling optimization of the cost function within the prediction time domain and outputs the optimal control values ​​to each actuator.

[0012] Secondly, the present invention provides a method for intelligent fall prevention rehabilitation training for hemiplegic patients to descend stairs independently based on the aforementioned system, comprising the following steps: Step 1: Wearing calibration and fear baseline assessment, the patient wears each module of the system and completes resting baseline acquisition; Step 2: Establishing a motor baseline and systematically recording the patient's initial motor ability parameters; Step 3: Progressive closed-loop training. The patient begins training with lower steps and gradually increases the steps. In each downhill gait cycle: the VLM rehabilitation decision engine analyzes the semantics of the movement in real time and outputs the probability of compensatory intent prediction and pelvic movement pattern; the MPC real-time stability prediction controller predicts the centroid stability within the next 0.5 seconds with a period of 20ms and solves the cost function to obtain the optimal control quantity; according to the optimal control quantity, each actuator outputs the corresponding thrust, torque, electrical stimulation or constraint force; the context-aware voice interaction platform synchronously plays guiding voice and safety margin prompts. Step 4: Accurate recording of compensation events and stability margins. The system records each compensation event and the dynamic stability margin calculated in real time by the model predictive control. Step 5: Multi-dimensional training report generation. The system generates a rehabilitation assessment report that includes model prediction and control performance indicators.

[0013] The beneficial effects of this invention are as follows: (1) Psychological-motor dual-dimensional intervention: This invention introduces affective computing and neuro-fear desensitization technology into the training of hemiplegic stair climbing for the first time, simultaneously solving psychological and motor obstacles. Simulation results show that the training interruption rate is reduced by about 60% compared with traditional methods.

[0014] (2) Proactive fall prevention: The MPC real-time stability predictive controller predicts stability within the next 0.5 seconds and actively applies pelvic thrust, abduction torque and electrical stimulation, rather than post-event compensation. Theoretical analysis shows that the risk of falls is reduced by about 45% compared with traditional feedback control.

[0015] (3) Early warning and blocking: The VLM rehabilitation decision engine's compensation intention prediction mechanism completes the intervention decision on average 120ms before the compensation action occurs. Combined with the real-time correction of the MPC real-time stable prediction controller, passive correction is transformed into active prevention, effectively preventing the neural solidification of compensation habits.

[0016] (4) Precise correction of abnormal mode: pneumatic abduction + active push rod + electrical stimulation work together to effectively inhibit the adduction and crossing of the affected leg, foot inversion and pelvic instability. Simulation and preliminary clinical observation show that the correction success rate is about 70% higher than manual method.

[0017] (5) Directional selective constraint: The directional constraint design of the pneumatic soft exoskeleton only blocks the harmful compensatory direction and retains the normal degree of freedom of movement, realizing a new paradigm of rehabilitation engineering of "precise blocking rather than comprehensive restriction".

[0018] (6) Semantic-level motion understanding: The VLM engine's semantic understanding of the motion of going down stairs far exceeds that of traditional rule-driven systems. It can understand subtle compensatory behaviors in complex situations, thus improving the applicability and robustness of intelligent intervention.

[0019] (7) Quantitative safety assessment: Real-time calculation and recording of objective indicators such as stability margin and intervention success rate to provide data support for rehabilitation assessment and reduce subjective judgment errors of therapists.

[0020] (8) Full-process emotional support: The context-aware voice interaction platform combines emotional computing with a large language model to build the first AI training companion in the history of rehabilitation that truly has the ability to empathize, significantly improving patients' motivation for rehabilitation and training compliance.

[0021] The present invention has other features and advantages, which will be apparent from or will be set forth in detail in the accompanying drawings and the following detailed description, which together serve to explain the particular principles of the invention. Attached Figure Description

[0022] The above and other objects, features and advantages of the present invention will become more apparent from the accompanying drawings, in which like reference numerals generally denote like parts.

[0023] Figure 1 This is a schematic diagram of the mechanical structure of an intelligent rehabilitation system for hemiplegic patients to descend stairs independently, according to the present invention.

[0024] Figure 2 This is a block diagram illustrating the control principle of the MPC real-time stability predictive controller in an intelligent rehabilitation system for hemiplegic patients to descend stairs independently, as described in this invention.

[0025] Figure 3 This is a timing diagram of the VLM+MPC system in an intelligent rehabilitation system for hemiplegic patients to descend stairs independently, as described in this invention. Detailed Implementation

[0026] The invention will now be described in more detail with reference to the accompanying drawings. While preferred embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0027] For ease of understanding, the following explanations are provided for the English abbreviations used in this application: VLM stands for Vision-Language Model, MPC for Model Predictive Control, ToF for Time of Flight, IMU for Inertial Measurement Unit, sEMG for Surface Electromyography, TENS for Transcutaneous Electrical Nerve Stimulation (primarily used for pain relief), FES for Functional Electrical Stimulation (used to induce muscle contraction to restore motor function), PPG for Photoplethysmography, EDA for Electrodermal Activity, GSR for Galvanic Skin Response, HRV for Heart Rate Variability, RMSSD for Root Mean Square of Successive Differences, and VQA for Visual Question Answering. Answering), FMA is Fugl-Meyer Assessment, COP is Center of Pressure, SVM is Support Vector Machine, EKF is Extended Kalman Filter, OSQP is OperatorSplitting Quadratic Program, CAN is Controller Area Network, SPI is Serial Peripheral Interface, BLE is Bluetooth Low Energy, and GAD-7 is Generalized Anxiety Disorder 7-item scale.

[0028] I. System Overall Architecture

[0029] like Figure 1As shown, the hardware of the system of the present invention includes: an adjustable stair platform (including a six-dimensional force / torque sensor, lidar, and depth camera), a neuro-fear desensitization module (transparent screen and physiological sensor), a multimodal three-dimensional motion capture subsystem (active infrared markers, ToF depth sensor, and pressure-sensing carpet), an active pelvic stabilization module (independent left and right electric push rods and IMU), an auxiliary exoskeleton for the affected leg (pneumatic muscle abduction + motor knee flexion), an intelligent functional electrical stimulator for the ankle joint (sEMG-triggered FES), a pneumatic bionic restraint-guidance mechanism (soft exoskeleton, pelvic stabilization vest, and FES insole), and a context-aware voice interaction platform.

[0030] The adjustable stair platform consists of 2 to 4 intelligent stair platforms with adjustable height, each step height ranging from 12cm to 20cm (adjustable step length by 1cm), a step width of 30cm, and a step depth of 120cm. Handrails (90cm high) are provided on both sides. Each step comprises an aluminum alloy frame and engineering plastic treads. A sensor mounting cavity is located beneath each tread, embedding a high-density thin-film pressure sensor matrix (16×16 array, sampling frequency ≥100Hz) for real-time acquisition of foot pressure distribution and center of gravity trajectory.

[0031] The system adopts a two-layer control architecture: Upper layer (cognitive decision layer): A rehabilitation decision engine based on VLM, which integrates multimodal motion capture, emotion computing and multimodal perception data to complete action semantic decomposition, compensation intention prediction, pelvic movement pattern classification and personalized rehabilitation report generation.

[0032] Lower layer (real-time control layer): Based on MPC, the real-time stable predictive controller performs rolling optimization of state variables such as center of mass trajectory, center of pressure, and joint angles in a 20ms cycle, and outputs pelvic thrust, exoskeleton torque and electrical stimulation intensity to achieve proactive fall prevention control.

[0033] II. Desensitization Module for Nervous Fear

[0034] The neuro-fear desensitization module includes a progressive visual exposure subsystem, a context-aware voice companion subsystem, and an affective computing subsystem.

[0035] A transparent projection screen is installed 1.2m directly in front of the stair landing, with its center at a height of 1.2m and tilted downwards at a 15° angle. The field of view covers the entire stair landing and the buffer zones before and after it. When the patient stands in front of the first step, the screen is within their direct line of sight, used to present a progressive visual exposure scenario. A dual-microphone array and speaker are integrated behind the screen for voice interaction and guidance.

[0036] 2.1 Affective Computing Subsystem

[0037] The affective computing subsystem integrates multimodal physiological signals and outputs a comprehensive anxiety index (...). AI score (0-100, with higher values ​​indicating stronger anxiety). The calculation process consists of three layers: signal acquisition → feature extraction → fusion scoring.

[0038] (A) Facial Expression Recognition: An expression classifier based on a finely tuned ResNet-50 backbone network is used. The input is a 112×112 pixel face ROI image captured in real time by a depth camera, and the output is the probability of fear. P fear ∈[0,1]. The training dataset consists of AffectNet (containing 450,000 annotated facial expression images) plus 2,800 images of fear expressions from hemiplegic patients collected in our lab. The inference frame rate is 25fps, and the single inference latency is <12ms (GPU: NVIDIA Jetson Orin).

[0039] (B) Physiological signal feature extraction: Heart rate variability (HRV): RR interval sequences were extracted from chest-band photoplethysmography (PPG) waves, and RMSSD values ​​were calculated within a 30-second window. Electrodermal conductance (EDA / GSR): 4Hz sampling, sliding mean filtering (2-second window) followed by extraction of baseline conductance level (SCL) and fast response component (SCR peak count / minute). Respiratory rate (RR): Extracted from the acceleration signal of the IMU chest strap. Too fast (>22 breaths / minute) or irregular (coefficient of variation>0.3) both indicate anxiety.

[0040] (C) Anxiety Index Calculation: The affective computing subsystem calculates the anxiety composite index using the following weighted linear fusion model:

[0041] in, S 1 is the facial expression fear score ( S 1 = P fear (× 100) S 2 represents the heart rate variability score ( S 2 = max(0, (50-RMSSD) / 50) × 100), S 3 represents the skin conductance score ( S 3= min(1, GSR / 10) × 100), S 4 represents the respiratory rate score ( S 4= min(1, (RR-12) / 15) × 100), w 1.w 2. w 3. w 4 represents the corresponding weighting coefficient; The weighting coefficients are as follows: w 1 = 0.30 w 2 = 0.25, w 3 = 0.25, w 4 = 0.20. The weight values ​​can be recalibrated using multiple regression based on clinically collected data.

[0042] when AI score When the time is >60, the MPC controller automatically extends the prediction time domain from 25 steps (0.5s) to 40 steps (0.8s), while the limit on the rate of change of pelvic thrust is reduced from 20N / s to 8N / s, thus achieving a "flexible" buffer for motion intervention.

[0043] 2.2 Progressive Visual Exposure Subsystem

[0044] The progressive visual exposure subsystem adopts a five-level scenario grading scheme (L0 to L4). Each level corresponds to a different amount of visual information and psychological stimulation intensity. Patients must reach the "stable pass" standard in the current level before they can advance to the next level. When the anxiety index exceeds the safety threshold, the level will be automatically downgraded.

[0045] The scene settings for each level are as follows: L0 (Baseline Level): No staircase background, solid color base, no steps displayed, only ground markings, uniform white light 300 lux, 90° overhead view, no sense of depth; L1 (Static Beginner): Static staircase image, close-up overhead view, 2 steps, low contrast, natural light simulation 200 lux, 75° overhead view, 35mm focal length. L2 (Static Standard): Static staircase frontal view, high contrast static full-frame image, 3 steps, normal indoor light 350 lux, eye level 45°, focal length 28mm. L3 (Dynamic Slow Speed): Slow-motion dynamic staircase video (0.5× speed playback), including people walking, 4 steps, dynamic light and shadow 200-400 lux, slightly downward 30°, focal length 24mm; L4 (Real Scene Level): Real-world scene rendering at full speed in real time, depth camera rendering at original speed, 4-level steps (consistent with the training platform), completely consistent with the training environment, first-person perspective, 18mm focal length.

[0046] Upgrade Decision Logic: The system will automatically trigger an upgrade and present the next level scenario when all of the following conditions are met: ① The current level is completed consecutively. N pass Training sessions ( N pass Default values: L0→L1 is 3 times, L1→L2 is 5 times, L2→L3 is 8 times, L3→L4 is 10 times); ② The most recent training... AI score ≤40; ③RMSSD≥35ms (reflecting normal parasympathetic nerve function); ④GSR<4μS (no significant stress response). After the upgrade is triggered, the system plays positive voice feedback and records the upgrade timestamp and the baseline values ​​of physiological parameters at that time.

[0047] Degradation Decision Logic: The system will immediately trigger a degradation and revert to the previous scenario (L0 will no longer be downgraded) when any of the following conditions are met: ① Real-time AI score >70; ② Heart rate (HR) >110 bpm and lasts for more than 10 seconds; ③ GSR increases sharply for more than 6 μS within 5 seconds; ④ The patient actively presses the pause button or verbally says the preset stop word. After downgrading, enter a 2-minute visual-auditory relaxation period (blue-green gradient background + breathing guidance audio), wait... AI score Training can only continue after the value drops below 35.

[0048] III. Multimodal 3D Motion Capture Subsystem

[0049] The multimodal three-dimensional motion capture subsystem includes a pressure sensor matrix (16×16 array, sampling frequency ≥100Hz) embedded in the step pedal, a lidar array (horizontal scanning frequency 25Hz, accuracy ±3mm) installed on the front and sides of the step, a ToF depth camera array (frame rate 30-50fps, resolution 640×480), a six-axis inertial measurement unit (sampling frequency 200Hz) fixed to the patient's waist, and 17 active infrared markers.

[0050] The 17 active infrared markers correspond to the main joints of the lower limbs and trunk (both hips, knees, ankles, shoulders, elbows, wrists, and head). The markers are made of rechargeable LED infrared light-emitting tubes, emitting wavelengths of 850nm, which are synchronously captured by the infrared receiver built into the depth camera. Each marker is fixed to the patient's body surface by an elastic strap, and the three-dimensional coordinates of the joints are output at a rate of 50 frames per second.

[0051] A pressure sensor matrix is ​​used to collect real-time data on plantar pressure distribution and center of gravity trajectory. A lidar sensor is used to detect the position of the step edge, with the step edge corners calibrated in real-time to sub-centimeter accuracy (±3mm). A depth camera is used to extract the 3D coordinates of the toe (average values ​​of the front end of key foot points) and calculate the horizontal distance d from the front of the foot to the edge of the step. edge .

[0052] A six-axis inertial measurement unit (IMU) is installed in a rigid plastic box behind the lumbar support belt, corresponding to the L3-L4 spinous processes. It is integrated and secured to the flexible belt of the lumbar stabilization device via elastic straps. The IMU and the ToF depth camera array undergo data fusion using an extended Kalman filter (EKF): the angular velocity integral from the IMU is used as the prediction model, and the joint positions identified by the depth camera are used as observations to estimate the optimal joint angles and center-of-mass state. Process noise covariance... Q com_pos ≈1e -4 m², observation noise covariance R cam ≈5e -4 m². When the exoskeleton's built-in angle sensor is available, the exoskeleton sensor data is used preferentially, with the camera and IMU serving as redundant calibration sources.

[0053] The pelvic angle status is directly converted from the quaternion of the IMU and then output through a low-pass filter (cutoff frequency 10Hz, 2nd order Butterworth), with a delay of about 3ms.

[0054] All sensors aggregate data via the ROS2 communication framework and achieve time synchronization using the IEEE 1588 protocol, with a synchronization error of less than 1ms. The communication methods for each sensor are as follows: the pressure sensor matrix and actuator built-in sensors connect to the main control computer via a CAN bus (1Mbps baud rate); the depth camera and LiDAR connect via USB 3.0; the IMU and surface electromyography sensor transmit wirelessly via Bluetooth 5.0 BLE; and the tension sensor connects to an independent safety monitoring module (using an STM32F407 chip) via an SPI interface, forming a dual-channel redundant design with the main control computer.

[0055] IV. VLM Rehabilitation Decision Engine

[0056] 4.1 Basic Architecture of Pre-trained Models

[0057] The VLM rehabilitation decision engine of this system uses LLaVA-Med-1.5 (based on LLaVA-1.5, with CLIP ViT-L / 14@336px as the backbone visual encoder and Vicuna-13B-v1.5 as the language model) as the basic pre-training model, and performs two-stage fine-tuning for hemiplegic motor rehabilitation tasks.

[0058] Model architecture overview: Visual encoder: CLIP ViT-L / 14@336px, output dimension 1024, extracting 576 visual tokens per frame (24×24 patch grid). Modal projection layer (MLP Connector): Two MLP layers (1024→4096) map visual features to the language model input space; Language Reasoning Module (LLM): Vicuna-13B (32-layer Transformer), based on a multi-turn dialogue paradigm, accepts visual and linguistic input and outputs decision results in structured JSON format; Temporal fusion module: Performs temporal self-attention aggregation on visual tokens of the current frame and the previous 4 frames (5 frames in total, time window = 200ms) to capture joint motion trends.

[0059] The fine-tuning dataset consists of: Hemiplegic gait video library (self-built): 3,200 videos totaling 48 hours, labeled with 17 joint trajectories and compensatory behavior tags (adduction / inversion / pelvic tilt), used for stage 1 fine-tuning of skeleton-semantic alignment; PD-Motor-QA (modification): 28,000 VQA question-answer pairs for stage 1 fine-tuning of motion semantic understanding question answering; Clinical rehabilitation report matching (self-built): 856 patients × an average of 12 training sessions, including quantitative assessment reports + video frames (FMA score, Berg balance scale), used for stage 2 fine-tuning of personalized report generation; Fear expression-gait joint data (self-built): 420 training videos, including fear expression tags and corresponding gait compensatory behavior timestamps, used to improve the cross-modal joint reasoning ability of the VLM rehabilitation decision engine.

[0060] Input (multimodal): 5 frames of RGB images (336×336px each) + synchronous skeleton projection map (17 key points, OpenPose format); structured system prompts (including basic patient parameters: affected side, FMA grade, historical compensation frequency) + current frame action description; current frame COP coordinates, IMU pelvic angle (Pitch / Roll / Yaw), sEMG activation ratio (affected / healthy side ratio).

[0061] The output (JSON format) includes:P adduction (Probability of adduction of the affected leg) P inversion (Probability of ankle inversion), Pelvis pattern (Pelvic movement pattern classification tags: forward thrust / stable / lateral tilt) risk level (Comprehensive risk level), specific intervention parameters, including: recommended hip abduction assist torque value. abduction_torque Recommended value of knee flexion assist torque (knee) flex_moment Recommended FES current intensity value fes intensity Recommended pelvic thrust values ​​(pelvis) push_force wait.

[0062] 4.2 Calculation Model for the Probability of Adduction of the Affected Leg

[0063] probability of adduction of affected leg P adduction This method is used to quantify the risk of hip adduction compensation (scissor gait) in the swing phase of the affected lower limb, with a value ranging from 0 to 1. This probability is calculated by fusing two parallel models: geometric judgment and temporal semantic understanding, to achieve dual prediction based on the current joint state and movement trend.

[0064] Path ① Geometric Determination of Path P geo : Based on the skeleton data of the current frame, extract the hip adduction angle θ. hip (Transverse distance of hip on the healthy side / pelvic width) and the medial deviation angle of the knee joint θ knee (The coronal plane offset angle of the tibia relative to the femur), the probability of adduction is calculated using a logistic regression model:

[0065] in, P geo is the inward probability output by the geometric path, with a value of 0 to 1; 4.5, 3.2, and -3.8 are the logistic regression model coefficients, obtained by fitting the data to the hemiplegic gait training dataset using maximum likelihood estimation.

[0066] The physical significance of this path lies in the hip adduction angle θ. hip and the inner deviation angle θ of the knee joint knee The larger the value, the higher the linear weighted sum. P geo The closer a value is to 1, the higher the risk of adduction based on the current joint posture. When both joint angles are simultaneously close to the abnormal threshold, P geo A value close to 0.98 indicates an extremely high probability of internal compensation.

[0067] Path ② Timing VLM Path P vlm : The temporal VLM path is implemented by a visual language model: taking the RGB image sequence of the current frame and the previous 4 frames (5 frames in total, with a time window of 200ms) and the synchronous skeleton projection map as input, the spatial features are extracted by the visual encoder, the joint motion trend is captured by the temporal fusion module, and the compensatory trend probability is output by the language inference module. P vlm , with a value ranging from 0 to 1.

[0068] P vlm This reflects the confidence level of VLM's prediction of future compensatory behavior based on temporal features such as increased hip adduction angular velocity and intensified knee inward deviation. Unlike the "current state judgment" of geometric paths, temporal paths capture the "change trend judgment" and can issue an early warning 120ms before the compensatory action is fully formed.

[0069] Fusion strategy and threshold response: The two probabilities are adaptively weighted and fused to obtain the final ingress probability: P adduction =0.55× P geo +0.45× P vlm The geometric path weight is 0.55 and the temporal VLM path weight is 0.45. This weight ratio reflects the principle of prioritizing current state judgment and supplementing it with trend prediction. It was determined through grid search optimization on the hemiplegic gait dataset.

[0070] After fusion P adduction As one of the core outputs of the VLM engine, the affected hip adduction angle penalty weight in the MPC cost function is dynamically adjusted via the Sigmoid mapping function. Q hip (See Section 4.5 for details on the mapping mechanism), and simultaneously trigger tiered intervention responses based on thresholds: when P adduction When the hip adduction angle is ≥0.65 or >2° on the affected side, a preparatory hip abduction torque (10~15 N·m) is applied to the auxiliary exoskeleton of the affected leg; when P adduction When the value is ≥0.80, the auxiliary exoskeleton of the affected leg outputs the maximum hip abduction torque and the pneumatic bionic constraint-guide mechanism outputs directional constraint force synchronously.

[0071] 4.3 Model for calculating the probability of ankle inversion

[0072] probability of ankle inversion (foot drop inversion pattern) P inversion Based on trimodal fusion computing: The first mode is the surface electromyography ratio. R emg = EMG PL / EMG TA ,in EMG PL The root mean square value of surface electromyography of the peroneus longus muscle on the affected side. EMG TA The root mean square value of surface electromyography of the tibialis anterior muscle on the affected side is the normal stepping phase when going downstairs. R emg ∈[0.8,1.2]; The second modality is the coronal inversion angle θ of the ankle joint. ankle (Positive values ​​indicate inversion), θ ankle A value >5° is abnormal; The third modality is the ratio of medial to lateral plantar pressure detected by FES insoles. R press = P medial / P lateral >1.5 indicates inversion.

[0073] Fusion formula: P inversion = sigmoid (2.8×max(0, R emg -1.2)+1.5×( θ ankle / 10)+2.0×max(0, R press -1.5)-1.2) Judgment threshold: P inversion ≥0.60 triggers FES electrical stimulation of the peroneus longus muscle (intensity 30mA, frequency 50Hz); P inversion When the amplitude is ≥0.75, the tibialis anterior muscle is activated simultaneously (40mA), and the output is coordinated with the functional electrical stimulation array in the intelligent orthotic insole for inhibiting plantar flexion-inversion of the affected foot ankle.

[0074] 4.4 Pelvic Movement Pattern Classification Algorithm

[0075] Extracting 6-dimensional feature vectors from the multimodal motion capture subsystem F pelvis : F pelvis=[Δ x pelvis ,Δ y pelvis , ω pitch , ω roll , ω yaw , COP shift ] Where Δ x pelvis The displacement velocity (m / s) of the pelvic center of mass in the coronal plane (left and right), corrected by IMU integration and ToF; Δ y pelvis The displacement velocity (m / s) of the pelvic center of mass in the sagittal plane (anterior-posterior). ω pitch , ω roll , ω yaw The pelvic triaxial angular velocity (deg / s) is directly output by a 6-DOF IMU. COP shift It is the ratio of the anterior-posterior displacement of the pressure center to the anterior-posterior displacement of the pelvis (dimension 1), which measures the degree of active forward thrust of the pelvis.

[0076] Support Vector Machine (SVM, kernel function: Radial Basis Function (RBF), γ=0.08, C=15) was used for tri-class classification: anterior, stable, and lateral. The training samples consisted of manually annotated pelvic movement segments from 3,200 hemiplegic patients descending stairs, approximately 1,067 cases per class. Cross-validation accuracy was 87.3% (5-fold cross-validation). Real-time inference latency was <5ms, meeting the 20ms MPC control cycle requirement. The output consisted of the probability distributions of the three classes (SoftMax normalized) plus the label of the class with the highest confidence.

[0077] 4.5 Dynamic Weight Mapping Relationship (VLM→MPC Linkage Core)

[0078] like Figure 3 As shown, the probability value output by VLM dynamically adjusts the weights of each penalty term in the MPC cost function through a nonlinear mapping function, thereby achieving cross-level linkage between "semantic understanding" and "quantization constraints".

[0079] There are three penalty weights related to compensation in the MPC cost function: Q hip (Hip adduction angle penalty weight) Q ankle (Ankle inversion angle penalty weight) Qpelvis (Pelvic lateral displacement penalty weight). The mapping rules are as follows: Inward weight mapping: Q hip ( k )= Q hip_base +Δ Q hip_max × f sig ( P adduction ( k )) in Q hip_base =5, Δ Q hip_max =45, f sig This is a general sigmoid mapping function.

[0080] Ankle inversion weight mapping: Q ankle (k)=Q ankle_base +ΔQ ankle_max ×f sig (P inversion (k)) in Q ankle_base =3, Δ Q ankle_max =30.

[0081] The general sigmoid mapping function is:

[0082] Where the steepness coefficient k a =8 (adduction) or 10 (ankle inversion), bias threshold θ a =0.50 (adduction) or 0.45 (ankle inversion). This function achieves a smooth non-linear transition, avoiding abrupt weight changes.

[0083] Physical meaning: P adduction When the value is 0.5, the weight increment is 50% of the maximum value. Q hip =27.5), P adduction When the weight increment is 0.8, it is close to the maximum value. Q hip ≈48), MPC will strongly constrain the deviation of the hip adduction angle.

[0084] Pelvic pattern mapping (three-category → weight selection table): Anterior type Q pelvis =40, stable Q pelvis =20, lateral type Q pelvis =50.

[0085] V. MPC Real-time Stable Predictive Controller

[0086] 5.1 Prediction Model Types and Dynamic Equations

[0087] like Figure 2 As shown, the MPC prediction model of this system adopts a three-degree-of-freedom linearized gait dynamics model, with the center of mass (CoM) as the core of the state variables. It integrates the pelvic angle and key joint angles to construct a continuous state equation, which is then discretized into a difference equation for rolling optimization.

[0088] The equations of motion for the center of mass (an extended version of the simplified inverted pendulum model). In the sagittal plane (the plane of motion of the lower part of the pendulum), the state of the center of mass is described by a second-order differential equation:

[0089] in , This represents the displacement of the center of mass in the horizontal / vertical direction. The equivalent pendulum length (distance from the center of mass to the support point, approximately 0.85m for hemiplegic gait); g = 9.81m / s²; m is the patient's body weight (kg). The horizontal thrust applied to the pelvic push rod; The ground reaction force of the supporting leg; , This is an estimate of the external disturbance.

[0090] Pelvic angle dynamics equation: pelvis =(τ external -τ passive ) / I pelvis

[0091] Where θ pelvis This refers to the pelvic pitch angle (sagittal anterior tilt angle). τ external External torque applied to the exoskeleton and push rods; τ passiveFor passive muscle elastic torque (linearization: τ passive = k p × θ pelvis + b p × pelvis ); I pelvis ≈0.12 kg·m² (pelvic rotational inertia).

[0092] With sampling period T s =0.02s (50Hz), using Euler forward difference discretization, the standard state-space form is obtained: x ( k +1)= A d × x ( k )+ B d × u ( k )+ E d × d ( k ) y ( k )= C d × x ( k ) State vector x = [x com , com , y com , com , θ pelvis , pelvis , θ hip , hip ] T (8-dimensional). Control input vector u = [ Fpelvis_L , F pelvis_R, τ hip_abd , τ knee_flex , I FES , F constraint ] T (6 dimensions, representing left and right thrust, hip abduction torque, knee flexion torque, FES current intensity, and aerodynamic constraint force). The output vector y represents observables (measurable by sensors): COP position, pelvic angle, and hip joint angle.

[0093] 5.2 Complete Expression of Cost Function

[0094] MPC in the prediction time domain N Step (default) N =25, T s Within a 0.02s (i.e., a 0.5-second prediction window) period, the optimal control sequence is solved by minimizing the following cost function:

[0095] The parameters are defined as follows: (1) Optimize variables and states J The cost function value, a scalar, represents the total cost over the entire prediction time domain. MPC minimizes this cost. J To solve for the optimal control sequence.

[0096] k: Discrete time step index, corresponding to the sampling time of the MPC controller, k=0,1,2,… N -1.

[0097] N : Number of time-domain steps for prediction, default N =25, corresponding to 0.5 seconds (sampling period) T s =0.02s); when the anxiety index exceeds the standard or triggers a level 2 warning, N Extended to 40 steps, corresponding to 0.8 seconds.

[0098] (2) Output quantity and reference trajectory

[0099] y(k): The system output vector at step k, which is an observable (directly measurable by the sensor), including: the coordinates of the center of plantar pressure (COP). x COP y The output vector y has a dimension of 5, and includes the pelvic pitch and roll angles, as well as the angle of the affected hip joint.

[0100] y ref The reference trajectory for output quantities, i.e., the expected values ​​of each output quantity under normal gait, is generated by a clinical gait standard template. At each time step k in the prediction time domain, yref provides the corresponding reference value. MPC guides the patient to move according to the healthy gait pattern by minimizing the deviation between the actual output and the reference trajectory.

[0101] (3) Control input quantity

[0102] u(k): The control input vector (6-dimensional) at step k, including: F pelvis_L(Push force of the electric push rod on the left side of the pelvis) F pelvis_R (Pelvic right electric push rod thrust), τ hip_abd (Assistive torque for hip abduction), τ knee_flex (Assisted torque for knee flexion) I FES (Ankle joint electrical stimulation intensity), F constraint (Pneumatic constraint). Each control variable is subject to hard constraints (see constraints below).

[0103] (4) Terminal status

[0104] x(N): Predicted system state vector (8-dimensional) at the end of the time domain, including centroid position and velocity, pelvic tilt angle and angular velocity, and hip adduction angle and angular velocity on the affected side. Terminal cost term. This is used to ensure the stability of the prediction terminal and to ensure that the system can still converge stably after the prediction time domain ends.

[0105] (5) Weight matrix

[0106] Q y Output tracking weight matrix (diagonal matrix), used to penalize the output quantity y(k) against the reference trajectory y. ref Tracking deviation. Specifically, this includes: Q com (Center of mass tracking weight): diag(30, 15, 20, 10), penalizes deviation of the center of mass trajectory from the reference gait, determined by clinical gait standards. The larger the value, the stricter the requirements for center of mass trajectory tracking. Q pelvis (Pelvic stability weight): The value is 20 (stable type), 40 (forward type) or 50 (lateral type), which is dynamically output by the VLM pelvic motion pattern classifier. The larger the value, the stronger the suppression of pelvic angle / displacement deviation. Q hip (Hip adduction penalty weight): dynamically varies within the range of 5 to 50, from P adduction As determined by the Sigmoid mapping, a larger value results in a stronger penalty for the deviation of the adduction angle of the affected hip, which is used to prevent scissor gait compensation. Q ankle (Ankle inversion penalty weight): dynamically changes within the range of 3 to 33, from P inversion Determined by the Sigmoid mapping, a larger value results in a stronger penalty for ankle inversion deviation, which is used to prevent clubfoot.

[0107] RThe control input penalty weight matrix (diagonal matrix), diag(0.1, 0.1, 0.5, 0.3, 0.01, 0.2), is used to penalize the magnitude of the control input to prevent the actuator output from becoming too large or overloaded. Each element corresponds to... F pelvis_L , F pelvis_R τ hip_abd , I FES The penalty coefficient is a factor, and a larger value indicates a stricter restriction on the output of the actuator.

[0108] Δ Q comp The control increment penalty term, diag(5, 5, 2, 1, 0.1, 0.5), is used to penalize the change in control input between adjacent time steps (i.e., Δu(k) = u(k) - u(k-1)), ensuring the smoothness of control commands and avoiding sudden changes in actuator output that could cause discomfort or harm to the patient. Each element corresponds to a penalty for the rate of change of the four control variables.

[0109] Q N: Terminal cost weight matrix Q N = 5 × Q y This is used to penalize terminal state deviations at the end of the prediction time domain, ensuring the system's stability in the Lyapunov sense. The terminal cost weight coefficient of 5 has been verified through simulation to ensure stability without significantly increasing the computational burden.

[0110] (6) Hard constraints

[0111] Control input upper and lower limit constraints: u min ≤ u(k) ≤ u max ; Pelvic thrust: F pelvis ∈ [-60N, 60N] (negative values ​​indicate that a thrust is applied to the opposite side); Hip abduction moment: τ hip_abd ∈ [0, 25N·m]; FES current intensity: I FES ∈ [0, 60mA]; Knee flexion moment: τ knee_flex ∈ [0, 12 N·m]; Aerodynamic constraint force: F constraint ∈ [0, 25N]; Joint angle hard constraint: θ hip ≥ -8° (inward angle not exceeding 8°), θ ankle ∈ [-5°, 15°] (to prevent excessive inversion or eversion).

[0112] (7) Solver

[0113] Solution method: The OSQP (Operator Splitting Quadratic Program) solver is used to solve the above quadratic programming problem based on the augmented Lagrange method. The single-step solution time is <8ms (on the ARM Cortex-A72 processor), which meets the real-time requirement of 20ms control cycle.

[0114] 5.3 State Estimation Stage

[0115] The 8-dimensional state vector x required for MPC cannot be measured directly by a single sensor and must be obtained through multi-sensor data fusion (state estimation).

[0116] Sensor configuration and observable measurements: ToF depth camera array (4 units, 50Hz, output key point 3D coordinates); pelvic IMU (6-DOF, 200Hz, output acceleration / angular velocity); six-dimensional force sensor (step embedded × 4, 1000Hz, output ground reaction force); pressure sensing carpet (100Hz, output foot pressure distribution and COP coordinates); exoskeleton encoder (200Hz, output hip / knee joint angle and angular velocity).

[0117] The state estimation of the centroid position / velocity employs an extended Kalman filter (EKF), fusing depth camera skeleton coordinates and IMU integral data: ① Prediction Steps (High Frequency, IMU-based, 200Hz):

[0118] ②Update steps (based on depth camera, 50Hz):

[0119] The parameters in the formula are defined as follows: This is the posterior state estimate at step k (the optimal estimate that incorporates the observations at the current time). This is the prior prediction of the state at step (k+1) based on the information from step k. This is the prior prediction of the state at step k based on the information at step k-1; Let $\mathbf{k}$ be the posterior estimation error covariance matrix. Let the prior prediction error covariance matrix be the (k+1)th step. Let be the prior prediction error covariance matrix at step k; K k is the Kalman gain matrix at step k, used to balance the weights of predicted and observed values ​​in the update; Ad is the discretized state transition matrix, which describes the evolution of the state variables from one step to the next without control. It is obtained by Euler discretization of the continuous state equations in Section 5.1. Bd is the discretized control input matrix, which describes the influence of the control quantity on the state quantity. It is obtained by Euler discretization of the continuous state equation in Section 5.1. u ( k ) represents the control input vector at step k, which is the optimal control quantity obtained by MPC solution; H The observation matrix maps the state vector x to the observation space (i.e., the observable y), and its specific values ​​are determined by the correspondence between the depth camera output and the state variables. z k The actual observations at step k are obtained by sensors such as depth cameras and exoskeleton encoders. Q noise The process noise covariance matrix characterizes the uncertainty of the prediction model and is obtained from measured sensor noise calibration. R noise The noise covariance matrix was obtained from the measured sensor noise calibration to characterize the uncertainty of sensor observations. I It is an identity matrix with the same dimensions as the state vector (8×8).

[0120] 5.4 Visual Risk Warning Judgment Logic

[0121] Visual risk warning is based on LiDAR point cloud and depth camera data to assess the spatial relationship between the foot and the edge of the step in real time and trigger hierarchical control logic.

[0122] Foot-step edge distance detection: LiDAR detects the position of the step edge, and the corner points of the step edge are calibrated in real time with sub-centimeter accuracy (±3mm); a depth camera extracts the 3D coordinates of the toe and calculates the horizontal distance d from the front of the foot to the edge of the step. edge (Unit: cm); swing speed of the affected leg, v swing The horizontal velocity of the foot marker point (cm / s) is calculated from the temporal difference of the skeleton key points.

[0123] Risk warnings are divided into two levels: Level 1 Warning (Yellow Alert): Triggering condition is d edge ≤5cm and v swing<25cm / s (Logical AND, both conditions must be met simultaneously). Control response: ① Voice prompt "Please lift your feet"; ② MPC weights the step avoidance term. Q edge Increase to 25; ③ FES tibialis anterior pre-activation (10mA). Upper limit of response delay ≤40ms.

[0124] Level 2 Warning (Red Emergency): Triggering condition is d edge ≤3cm and v swing <15cm / s (logical AND, both conditions must be met simultaneously). Control response: ① Exoskeleton knee joint assisted flexion torque activation (12N·m, i.e., maximum flexion assisted torque); ② Pelvic pusher active support (40N forward force); ③ FES tibialis anterior muscle full activation (40mA); ④ Controller prediction time domain extended to N =40 steps (0.8s). Maximum response latency ≤20ms.

[0125] Control logic timing: After the level 2 warning is triggered, the following timing operations shall be completed within 20ms: T =0ms (Trigger condition detected) → T =3ms (State estimation update) → T =8ms (MPC re-optimization completed) → T =12ms (control command sent to actuator driver) → T =20ms (exoskeleton / push rod torque output in place).

[0126] VI. Active Pelvic Stabilization Module

[0127] The active pelvic stabilization module is a soft-rigid hybrid structure worn on the patient's pelvis. It integrates a linear electric actuator on each side (stroke ±5cm, thrust ≤200N), capable of independently applying horizontal thrust to either the healthy or affected side. The actuator ends in an arc-shaped silicone pad (120° radius, 14cm width), with a 4×6 thin-film pressure sensor array (range 0~100N, accuracy 0.5N) integrated on its inner surface for detecting contact pressure distribution. The actuators are driven by a brushless DC motor and ball screw (peak thrust 200N, response time ≤50ms). The module incorporates a triaxial IMU (six-axis, acceleration ±8g, angular velocity ±1000° / s, sampling frequency 200Hz) to detect pelvic tilt angle and angular velocity. The arc-shaped silicone pad is 5mm thick, with a Shore hardness of A30, matching the anatomical shape of the patient's iliac crest.

[0128] This module is directly linked to the MPC controller: when the MPC predicts insufficient center of gravity transfer or a tendency for the healthy side pelvis to tilt forward, it applies a horizontal thrust of 50–150 N to the healthy side to guide the center of gravity to shift towards the healthy side; when the center of gravity shifts backward after the affected foot lands, the thrust is slowly withdrawn to assist the patient in tilting forward. In addition, this module works in conjunction with the "healthy side pelvic triaxial stabilization vest" in the pneumatic bionic constraint-guidance mechanism—the vest provides cushioning resistance, and the electric push rod provides active thrust, achieving composite pelvic stabilization.

[0129] VII. Auxiliary exoskeleton for the affected leg

[0130] The auxiliary exoskeleton for the affected leg only covers the abduction / adduction range of motion of the hip joint and the flexion / extension range of motion of the knee joint on the affected side.

[0131] Hip abduction joint: Utilizing a pneumatic muscle antagonistic actuator, it generates a maximum abduction torque of 15 N·m to counteract adductor overactivation. The pneumatic muscle is controlled by a miniature air pump (maximum pressure 0.6 MPa) in conjunction with a high-speed proportional valve (response time ≤ 10 ms). This actuator is cascaded with VLM prediction and MPC output: when the VLM predicts adduction intention (… P adduction When the angle is ≥0.65° or the inward angle is >2°, or when the MPC detects an inward angle >2°, the outward torque is output immediately. The joint housing integrates a magnetic encoder angle sensor (resolution 0.01°) and a six-dimensional force / torque sensor (range ±50N·m, accuracy 0.1N·m) for real-time detection of joint angle and human-machine interaction torque.

[0132] Knee flexion and extension: Motor-driven (brushless DC motor + harmonic reducer, reduction ratio 100:1) to assist knee flexion (0~12N·m) and avoid toe dragging.

[0133] This exoskeleton complements the "affected hip-knee directional restraint soft exoskeleton" in the pneumatic bionic restraint-guidance mechanism: the former provides active torque, and the latter provides directional restraint force; the two can work independently or collaboratively. The system automatically selects the mode (active mode / restraint mode / hybrid mode) based on the patient's ability level. The exoskeleton is fixed to one of the uprights of the two handrails via a three-degree-of-freedom adjustable bracket, and the end of the shell has a quick-release interface for easy adjustment of position and angle, as well as replacement and maintenance.

[0134] The weight of the entire assisted exoskeleton for the affected leg does not exceed 3kg (excluding the pneumatic muscle supply pump and tubing). The hip joint actuator unit weighs approximately 1.2kg, the knee joint actuator unit weighs approximately 0.9kg, and the support and quick-release interface weigh approximately 0.6kg. The exoskeleton transfers most of the weight to the handrail column through a three-degree-of-freedom adjustable support, so the patient only bears a dynamic load of approximately 0.8kg, avoiding additional burden on the affected lower limb.

[0135] 8. Intelligent electrical stimulator for ankle joint

[0136] The ankle joint intelligent electrical stimulator uses a transcutaneous functional electrical stimulation (TENS) patch triggered by sEMG, applied to the motor points of the tibialis anterior and peroneus longus muscles on the affected side. The TENS patch is a disposable Ag-AgCl electrode (30mm in diameter, 2mm thick), employing a bipolar differential configuration. The stimulating electrode and reference electrode for the same muscle are spaced 20mm apart, with the reference electrode placed at a bony prominence (tibial tuberosity or fibular head). The electrode patch is connected to the main control unit via a flexible 60cm wire, laid along the lateral side of the lower leg and secured with medical tape.

[0137] When foot drop or inversion tendency is detected, an electrical stimulus (intensity ≤60mA) with a frequency of 50Hz and a pulse width of 200μs is automatically output to induce ankle dorsiflexion and eversion.

[0138] The stimulator shares a control strategy with the FES array (16 channels) in the intelligent orthotic insole for plantar flexion-inversion inhibition of the affected foot ankle; the two can be used in tandem or switched. The VLM engine automatically selects the optimal stimulation mode based on the real-time ankle angle. P inversion ≥0.60 triggers FES electrical stimulation of the peroneus longus muscle (intensity 30mA, frequency 50Hz); when P inversion At ≥0.75, the tibialis anterior muscle is activated synchronously (40mA) and outputs in synergy with the FES array in the orthotic insole.

[0139] IX. Aerodynamic Bionic Constraint-Guidance Mechanism

[0140] The pneumatic bionic restraint-guidance mechanism includes a hip-knee directional restraint soft exoskeleton on the affected side, a triaxial stabilizing vest for the healthy side pelvis, and an intelligent orthotic insole that inhibits plantar flexion-inversion of the affected ankle.

[0141] The miniature air pump (maximum pressure 0.6MPa, flow rate ≥5L / min) is integrated into a rigid shell at the back of the patient's waist vest, weighing no more than 0.6kg. The air pump is connected via 4mm inner diameter polyurethane hoses to the affected hip-knee directional restraint soft exoskeleton, the inflatable chambers of the healthy side pelvic triaxial stabilization vest, and the pneumatic muscles of the affected leg auxiliary exoskeleton. The air lines are laid along the waist belt and exoskeleton frame, protected by spiral sheaths to prevent tangling and excessive bending.

[0142] The hip-knee directional restraint soft exoskeleton uses an isotropic fiber-reinforced pneumatic actuator to generate a directional restraint force of 0–25 N. The isotropic fiber-reinforced pneumatic actuator is a cylindrical braided tube structure composed of three layers: an inner layer of elastic airtight tubing (natural latex, 0.5 mm wall thickness), a middle layer of an isotropic fiber braided mesh (aramid fiber, braiding angle 15°–25°, fiber diameter 0.3 mm), and an outer layer of flexible restraint sheath (polyurethane, 0.3 mm wall thickness). During inflation, the fiber braided mesh creates circumferential restraint and shortens axially, thus generating a contractile restraint force in a specific direction (adduction of the affected leg); in non-restraint directions (flexion / extension), due to the asymmetric design of the fiber braiding angle, the actuator does not generate significant resistance, achieving "directional selective restraint." The actuator has a maximum working pressure of 0.4 MPa, a response time ≤50 ms, and a mass of approximately 0.15 kg / m.

[0143] The healthy side pelvic triaxial stabilization vest has three independent inflation chambers with a working air pressure of 50-120 kPa, corresponding to the stabilization control in three directions: sagittal forward / backward tilt, coronal tilt, and horizontal rotation.

[0144] The intelligent orthotic insole for inhibiting plantar flexion-inversion of the affected foot and ankle has a built-in 16-channel functional electrical stimulation array, which can independently control the stimulation intensity and duration of each channel to achieve precise foot movement control.

[0145] All control commands for this device can be issued by the VLM engine or the MPC controller (priority is configurable), achieving dual protection of "advanced intent prediction" and "low-level real-time correction". This device works in conjunction with an active pelvic stabilization module, an auxiliary exoskeleton for the affected leg, and an intelligent electrical stimulator for the ankle joint.

[0146] 10. Context-Aware Voice Interaction Platform

[0147] The context-aware voice interaction platform is built on a large language model and has the ability to empathize. The voice platform can announce the "current safety margin" calculated by MPC in real time (such as "the current stable margin is 2.3 cm, continue to maintain it"), conveying a quantitative sense of security to patients and reducing uncertainty and anxiety.

[0148] The platform also supports the following functions: psychological reassurance and guidance before training, real-time movement guidance during training (such as "pay attention to controlling the affected leg" and "please raise your foot"), and positive feedback and encouragement after training. When the emotion computing module detects that the anxiety index exceeds the threshold, the voice platform automatically switches to cognitive reframing mode, using breathing guidance and positive cues to help patients alleviate fear.

[0149] XI. Waist stabilization device and safety suspension system

[0150] The lumbar stabilization device includes a vertical support rod (70cm long, 25mm outer diameter, made of aluminum alloy), an arc-shaped support pad (lined with high-density sponge, 3cm thick, located in the patient's T10-L2 thoracolumbar region) fixed to the upper end of the support rod, and a flexible strap (8cm wide, nylon webbing) connected to the middle of the support rod. The bottom of the support rod is free and not fixed to the ground, remaining 3-5cm above the ground or step surface. The support rod is fixed to the patient's waist by the flexible strap and can move with the patient. When the patient's torso leans back beyond a safe angle, the bottom of the support rod touches the ground to provide rigid support and prevent falls; during normal walking, the support rod does not contact the ground and does not interfere with the patient's natural movement.

[0151] The safety suspension system includes a ceiling-embedded track, an electric winch, a steel wire rope, a chest- and abdomen-mounted soft vest worn by the patient, and a tension sensor connected between the steel wire rope and the vest. The ceiling-embedded track is arranged longitudinally along the stair platform (length ≥ 150cm), with movable hanging points within the track. The electric winch is fixed to the track via these hanging points and can be adjusted forward and backward along the track according to the patient's standing position. The steel wire rope is connected to a D-ring on the back of the vest via a guide ring fixed to the top of the waist stabilization device support rod, ensuring that the suspension force is always along the longitudinal axis of the patient's torso. The electric winch has a rated tension of ≥ 200N and a lifting speed of 0.1m / s; the tension sensor has a range of 0~500N and an accuracy of 1N.

[0152] The suspension system's tension automatically adjusts based on the patient's torso tilt angle and center of gravity position: when the torso tilt angle is >5°, the tension increases by 30N; when it is >10°, it increases to the maximum safe value (150-200N) and triggers emergency stop locking and audible and visual alarms. The suspension system is linked to the MPC controller: when the MPC predicts a fall risk (stability margin <1cm), the suspension system automatically increases the tension to the safe value, achieving predictive protection.

[0153] 12. System Working Sequence

[0154] Step 1: Baseline Fear Assessment

[0155] After the patient dons all the sensors, the system first performs a fear baseline assessment. The neuro-fear desensitization module collects the patient's resting physiological parameters (heart rate, HRV, GSR, respiratory rate) in a Level 0 scenario to establish an individualized physiological baseline. Simultaneously, facial expression images are acquired, and an expression classifier is trained to adapt to the patient's current facial features. The baseline data is stored in the patient's file as a reference standard for subsequent assessments.

[0156] Step 2: Establishing the motion baseline

[0157] The multimodal 3D motion capture subsystem collects kinematic data (joint range of motion, electromyographic activation level, and center of gravity swing range) of patients during standing on flat ground and simple lower limb movements, establishing an individualized baseline of motor ability. The VLM engine records the frequency and type of the patient's initial compensatory behavior patterns.

[0158] Step 3: Progressive Closed-Loop Training

[0159] Training begins with low steps (10cm) and gradually increases in size as the patient's ability improves. The entire process of descending the stairs step by step is as follows: (1) Intention to go downstairs recognition: The pressure sensor detects that the body is leaning forward and the vertical force on the healthy foot is reduced, and the system enters the preparation mode for going downstairs.

[0160] (2) Healthy leg initiation assistance: When the MPC predicts that the stability margin is low when the affected side is supported, it sends a command to the pelvic module, and the electric push rod applies a horizontal thrust of 50-150N to the healthy side to guide the center of gravity to the healthy side.

[0161] (3) Correction of flexion and abduction of the affected leg: When the affected leg begins to swing, VLM predicts or MPC detects an adduction tendency ( P adduction (≥0.65 or adduction angle >2°), immediately output pneumatic muscle abduction torque (10~15N·m), while the knee joint outputs flexion auxiliary torque, and the pneumatic bionic constraint-guide mechanism synchronously outputs directional constraint force.

[0162] (4) Foot placement and positioning: The depth camera provides real-time feedback on the relative position of the foot to the next step, and the MPC adjusts the knee flexion torque according to the distance; the electrical stimulator responds to sEMG or VLM commands ( P inversion (≥0.60) Triggers stimulation of the tibialis anterior muscle to prevent foot drop and inversion. At the moment of landing, a six-dimensional force sensor monitors the shear force; if abnormal, it is recorded and the abduction assist torque is increased for the next step.

[0163] (5) The center of gravity shifts again: After the affected foot lands, the MPC recalculates the position of the center of gravity. If it is biased to the rear, the pelvic module slowly withdraws the thrust and prompts the patient to lean forward.

[0164] (6) Voice synchronous guidance: The context-aware voice interaction platform plays guidance voice and real-time safety margin prompts throughout the process.

[0165] Step 4: Accurate Recording of Compensation Events and Stability Margin

[0166] The system records the compensatory events (adduction, inversion, pelvic tilt) and timestamps in each training session, and also records the dynamic stability margin (minimum distance from the centroid to the boundary of the supporting polygon) calculated in real time by MPC, generating a stability margin curve.

[0167] Step 5: Generation of Multi-Dimensional Training Report

[0168] After training, the system generates a rehabilitation assessment report containing the following: a movement quality radar chart (covering dimensions such as hip abduction control, ankle dorsiflexion control, pelvic stability, center of gravity transfer, and fear index), a compensatory behavior frequency trend chart, MPC control performance indicators (average stability margin, intervention success rate, and actuator output statistics), a stability margin curve, and a movement quality score assessed by VLM.

[0169] Thirteen, Iterative Learning Mechanism

[0170] After each downstairs training session, the system uses the "minimum fall margin", "number of abnormal patterns", and the action quality score evaluated by VLM as performance indicators. It then uses Bayesian optimization to adjust the MPC weight parameters (such as the adduction penalty coefficient and pelvic thrust smoothing coefficient) online, while updating the probabilistic graphical model of VLM to achieve two-level personalized adaptation.

[0171] The objective function of Bayesian optimization is to maximize the action quality score and minimize the number of abnormal patterns, with the constraint that the minimum fall margin must be greater than the 1cm safety threshold. The optimization variable is MPC. Q hip_base Δ Q hip_max , Q pelvis The system assigns values ​​to key weight parameters. After each training iteration, the system updates the Gaussian process surrogate model based on the performance metrics and uses the Expectation Boosting (EI) acquisition function to determine the next set of weight parameter combinations to be tested. Specific Implementation

[0173] Example 1: Initial training for fear-induced hemiplegic patients to descend stairs

[0174] Patient information: Male, 68 years old, left hemiplegia (6 months after cerebral infarction), Fugl-Meyer lower limb score 18, Berg balance scale score 28, GAD-7 score 14 (moderate anxiety), previous stair climbing training was interrupted multiple times due to fear.

[0175] System configuration: The initial scene complexity of the neuropsychiatric fear desensitization module is set to the lowest level (L0, single-step 5cm virtual step), and the anxiety alarm threshold for emotion calculation is set to 60. The MPC controller prediction time domain is 0.5 seconds, the initial maximum thrust of the pelvic push rod is 80N, the upper limit of hip abduction torque is 12N·m, and the upper limit of knee flexion torque is 8N·m. The initial output force of the pneumatic bionic mechanism is set to 12N, the anterior chamber air pressure of the pelvic stabilizing vest is 70kPa, and the initial stimulation pulse width of the FES is 200μs, with a frequency of 30Hz.

[0176] Training record: Two minutes after the start of training, the emotion computing module detected the patient's comprehensive anxiety index. AI score The reading reached 72, exceeding the threshold. The voice platform immediately switched to cognitive reframing mode, and MPC automatically extended the prediction time domain to 0.8 seconds, reducing the pelvic pusher change rate by 50%. Three minutes later... AI score The temperature dropped to 48, returning to normal. During the first attempt to go downstairs, the VLM engine predicted the intention to compensate for adduction 95ms after the patient's affected leg began to sway (hip adduction angular velocity 23° / s). P adduction =0.78), the MPC controller simultaneously detected a rapid increase in the adduction angle and immediately output a hip abduction torque of 12 N·m. Simultaneously, the pneumatic restraint mechanism inflated to 14 N, successfully blocking the compensatory movement; the pelvic pusher, under the MPC command, applied a 60 N thrust to the healthy side, guiding the transfer of the center of gravity. At the moment the foot left the ground, the electrical stimulator detected... P adduction =0.68, outputting a 35mA stimulus, induced ankle dorsiflexion. A total of 15 step-down exercises were completed in this training, with a compensatory occlusion success rate of 87%. The system recorded a minimum stability margin of 2.1cm (safe threshold >1cm), which is significantly better than the historical average of manual training for similar patients (61%).

[0177] Example 2: Enhanced training for hemiplegic patients to descend stairs independently in the later stages of rehabilitation

[0178] Patient information: Female, 52 years old, right hemiplegia (10 months after cerebral hemorrhage), Fugl-Meyer lower limb score 35, Berg balance scale score 46, GAD-7 score 5 (mild anxiety), basically independent in daily walking, the goal is to restore the ability to go down stairs independently without assistance.

[0179] System configuration: The neurodesensitization module uses a real staircase visual scene (3 steps, 20cm step height) as input, and the fear desensitization function is reduced to auxiliary mode. The MPC controller predicts time in the 0.5-second range, but the adduction penalty coefficient for the affected hip is adjusted. Q hip_base Reduce from the default value of 5 to 3 (Bayesian optimization history recommendation). Q hip During dynamic adjustment, the peak value does not exceed 30. The pelvic push rod operates in "touch-assisted" mode (maximum thrust 40N, only intervening when the stability margin is <1.5cm); the pneumatic bionic mechanism operates in "prediction-touch" mode (maximum intervention force 6N, only intervening when the VLM predicts the probability). P adduction Intervention occurs when the frequency is >0.75; FES activates low-frequency sensory enhancement mode.

[0180] Training focus: VLM engine analysis shows that the patient's main residual impairment is compensatory anteversion of the contralateral pelvis (frequency approximately 35%). Q pelvis The ankle joint compensatory inversion has been basically corrected (set to 40). P inversion (Mean 0.25). Historical MPC data showed that the patient maintained good voluntary stability, with stability margin only briefly falling below the threshold during rapid steps. This training specifically strengthened pelvic stability, with the MPC controller providing only minimal thrust when needed, and monitoring the entire time. After 8 training sessions (4 weeks), the frequency of the patient's compensatory pelvic forward thrust decreased to 8%. VLM assessed that the patient had reached the functional standard of independently and safely descending stairs. The system generated a phased rehabilitation assessment report and recommended a gradual transition to the unassisted training phase.

[0181] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. A smart fall prevention rehabilitation assessment and training system for hemiplegic patients to descend stairs independently, characterized in that, include: The neuro-fear desensitization module is used to identify and intervene in patients' fear-anxiety neurological responses when going downstairs; A multimodal three-dimensional motion capture subsystem is used to acquire the patient's kinematic, dynamic, and physiological electrical signals; The VLM rehabilitation decision engine is used to perform semantic understanding of the action of going down stairs and to predict the compensatory intention of the affected lower limb and the pelvic movement pattern. The MPC real-time stability predictive controller is used to predict the patient's center of mass trajectory and joint abnormality trends in millisecond cycles, and output multi-actuator collaborative control commands. An active pelvic stabilization module, worn on the patient's pelvis, is used to actively apply horizontal thrust to adjust the center of gravity shift according to the instructions of the MPC real-time stability prediction controller. The affected leg assist exoskeleton covers at least the abduction degree of freedom of the hip joint and the flexion and extension degree of freedom of the knee joint on the affected side, and is used to output abduction torque and knee flexion assist torque according to the instructions of the MPC real-time stability prediction controller; An ankle joint intelligent electrical stimulator is used to output electrical stimulation according to the instructions of the MPC real-time stability prediction controller to inhibit foot drop and inversion; A pneumatic bionic restraint-guidance mechanism, including at least a hip-knee directional restraint soft exoskeleton on the affected side, is used to apply directional restraint force to the affected lower limb; And a context-aware voice interaction platform to provide voice guidance and feedback; The system adopts a cognitive-motor dual-layer control architecture: the upper layer is the VLM rehabilitation decision engine, which is used to classify and output the probabilities of adduction of the affected leg, inversion of the ankle, and pelvic movement patterns; the lower layer is the MPC real-time stability prediction controller. The adduction probability of the affected leg and the inversion probability of the ankle output by the VLM rehabilitation decision engine are dynamically adjusted by the nonlinear Sigmoid mapping function to adjust the adduction angle penalty weight of the affected hip and the inversion angle penalty weight of the cost function of the MPC real-time stable prediction controller, thereby realizing the cross-level transmission of semantic understanding vectorized constraints. The MPC real-time stability predictive controller performs rolling optimization to solve the cost function in the prediction time domain, and outputs the optimal control quantity obtained to the active pelvic stabilization module, the affected leg auxiliary exoskeleton, the ankle joint intelligent electrical stimulator, and the pneumatic bionic constraint-guidance mechanism, respectively.

2. The system according to claim 1, characterized in that, The state variables of the MPC real-time stability prediction controller include at least the position and velocity of the center of mass, the pelvic tilt angle and angular velocity, and the adduction angle and angular velocity of the affected hip joint; the control output includes at least the thrust of the left and right electric push rods of the active pelvic stabilization module, the hip abduction torque and knee flexion assist torque of the affected leg auxiliary exoskeleton, the electrical stimulation intensity of the ankle joint intelligent electrical stimulator, and the constraint force of the pneumatic bionic constraint-guidance mechanism. The cost function includes at least a real-time penalty term for the deviation of the hip adduction angle and the deviation of the foot inversion angle, and a control output smoothing term; The nonlinear Sigmoid mapping function is: Where k is the discrete time step index, corresponding to the sampling time of the MPC real-time stable predictive controller; The probability of adduction of the affected leg output by the VLM rehabilitation decision engine at the k-th sampling time; The penalty weight for the hip adduction angle at the kth sampling time; Basic weights; Adjust the amplitude for weighting; This represents the steepness coefficient of the Sigmoid curve. This is the bias threshold; The VLM rehabilitation decision engine employs a geometric-semantic parallel fusion strategy to calculate the adduction probability of the affected leg: the geometric judgment path calculates the geometric probability based on the adduction angle of the affected hip joint and the medial deviation angle of the knee joint; the temporal visual language model path outputs the compensatory trend probability based on multi-frame temporal visual features from the visual language model; the adduction probability of the affected leg is a weighted fusion of the geometric probability and the temporal visual language model probability; when the adduction probability of the affected leg is greater than or equal to 0.65 or the adduction angle of the affected hip joint exceeds 2°, the auxiliary exoskeleton for the affected leg applies a preparatory hip abduction torque; when the adduction probability of the affected leg is greater than or equal to 0.80, the auxiliary exoskeleton for the affected leg outputs the maximum hip abduction torque and the pneumatic bionic constraint-guidance mechanism simultaneously outputs directional constraint force.

3. The system according to claim 1, characterized in that, The active pelvic stabilization module is a soft-rigid hybrid structure worn on the patient's pelvis. It integrates a linear electric actuator on each side, with a stroke of ±5cm and a thrust of ≤200N, capable of independently applying horizontal thrust to either the healthy or affected side. Each linear electric actuator has an arc-shaped silicone pad at its end, and the inner surface of the arc-shaped silicone pad integrates a thin-film pressure sensor array. The active pelvic stabilization module embeds a six-axis inertial measurement unit for detecting pelvic tilt angle and angular velocity. The active pelvic stabilization module is linked with the MPC real-time stability prediction controller: when the MPC real-time stability prediction controller predicts insufficient center of gravity transfer or a tendency for the healthy side pelvis to tilt forward, the linear electric push rod applies a horizontal thrust of 50N to 150N to the healthy side to guide the center of gravity to shift to the healthy side; when the center of gravity shifts backward after the affected foot lands, the linear electric push rod slowly withdraws the thrust to assist the patient in tilting forward. The affected leg assistive exoskeleton only covers the abduction and adduction degrees of freedom of the hip joint and the flexion and extension degrees of freedom of the knee joint on the affected side; the hip abduction degree of freedom of the affected leg assistive exoskeleton is driven by pneumatic muscle antagonism, which can generate a maximum abduction torque of 15 N·m; the knee flexion and extension degree of freedom of the affected leg assistive exoskeleton is driven by a motor, which can generate a flexion assist torque of 0 to 12 N·m; the joint shell of the affected leg assistive exoskeleton integrates an angle sensor and a force / torque sensor for real-time detection of joint angle and human-machine interaction torque; The affected leg auxiliary exoskeleton works in conjunction with the affected hip-knee directional constraint soft exoskeleton in the pneumatic bionic constraint-guidance mechanism. The system automatically selects active mode, constraint mode, or hybrid mode for training based on the patient's ability level.

4. The system according to claim 1, characterized in that, The ankle joint intelligent electrical stimulator includes a transcutaneous functional electrical stimulation patch triggered by surface electromyography signals and a multi-channel functional electrical stimulation array within an intelligent orthotic insole for inhibiting plantar flexion-inversion of the affected ankle. The transcutaneous functional electrical stimulation patch is applied to the motor points of the tibialis anterior and peroneus longus muscles on the affected side. When foot drop or inversion tendency is detected, it automatically outputs electrical stimulation with a frequency of 50Hz, a pulse width of 200μs, and an intensity of ≤60mA to induce ankle dorsiflexion and eversion. The ankle joint intelligent electrical stimulator shares a control strategy with the multi-channel functional electrical stimulation array, and the stimulation mode is automatically selected by the VLM rehabilitation decision engine or the MPC real-time stability prediction controller based on the real-time ankle joint angle. The VLM rehabilitation decision engine calculates the ankle inversion probability based on a three-modal fusion: the first modality is the surface electromyography ratio of the peroneus longus and tibialis anterior muscles on the affected side; the second modality is the coronal plane inversion angle of the ankle joint; and the third modality is the ratio of medial to lateral plantar pressure. When the ankle inversion probability is greater than or equal to 0.60, the ankle joint intelligent electrical stimulator triggers percutaneous electrical stimulation of the peroneus longus functional muscle. When the ankle inversion probability is greater than or equal to 0.75, the ankle joint intelligent electrical stimulator simultaneously activates the tibialis anterior muscle and outputs in coordination with the multi-channel functional electrical stimulation array.

5. The system according to claim 1, characterized in that, The VLM rehabilitation decision engine is built on a cross-modal pre-trained model finely tuned on a hemiplegic rehabilitation training video database. The cross-modal pre-trained model takes RGB image sequences and synchronous skeleton projection maps as inputs and outputs the probability of adduction of the affected leg, the probability of ankle inversion, and the classification label of pelvic movement patterns. The VLM rehabilitation decision engine can provide an early warning on average 120ms before the compensated adduction of the affected leg occurs, and output the pelvic movement pattern in real time in three categories: forward thrust, stable and lateral tilt. The pelvic motion pattern classification label is used to determine the value of the pelvic stability penalty weight in the cost function.

6. The system according to claim 1, characterized in that, The pneumatic bionic restraint-guidance mechanism includes a directional restraint soft exoskeleton for the affected hip and knee, a triaxial stabilizing vest for the healthy pelvis, and an intelligent orthotic insole for inhibiting plantar flexion-inversion of the affected ankle. The directional restraint soft exoskeleton for the affected hip and knee uses an isotropic fiber-reinforced pneumatic actuator, which can generate a directional restraint force of 0–25 N. The triaxial stabilizing vest for the healthy pelvis has three independent inflation chambers with a working air pressure of 50 kPa–120 kPa. The intelligent orthotic insole for inhibiting plantar flexion-inversion of the affected ankle has a built-in 16-channel functional electrical stimulation array. The pneumatic bionic restraint-guidance mechanism works in conjunction with the active pelvic stabilization module, the affected leg auxiliary exoskeleton, and the ankle joint intelligent electrical stimulator. The control commands for the pneumatic bionic restraint-guidance mechanism are issued by the VLM rehabilitation decision engine or the MPC real-time stability prediction controller, and their priority is configurable.

7. The system according to claim 1, characterized in that, The neuro-fear desensitization module includes an affective computing subsystem and a progressive visual exposure subsystem; The emotion computing subsystem calculates the anxiety comprehensive index according to the following weighted linear fusion model: in, S 1 is a rating for fear of facial expressions. S 2 represents the heart rate variability score. S 3 represents the skin conductance score. S 4 represents the respiratory rate score. w 1. w 2. w 3. w 4 represents the corresponding weighting coefficient; When the anxiety index AI score When the value exceeds the first preset threshold, the MPC real-time stable prediction controller automatically extends the prediction time domain from 0.5 seconds to 0.8 seconds and reduces the pelvic thrust change rate limit from 20 N / s to 8 N / s. The progressive visual exposure subsystem adopts a five-level scene classification scheme from L0 to L4. When the current level has been continuously trained for a preset number of times and the anxiety index and physiological parameters meet preset conditions, it will automatically upgrade. When the anxiety index or physiological parameters exceed the preset safety range, it will automatically downgrade.

8. The system according to claim 1, characterized in that, The multimodal three-dimensional motion capture subsystem includes a pressure sensor matrix embedded in the step pedal, a lidar array and a ToF depth camera array installed in front of and on the sides of the step, and a six-axis inertial measurement unit fixed to the patient's waist. The six-axis inertial measurement unit and the ToF depth camera array perform data fusion through an extended Kalman filter: the angular velocity integral of the six-axis inertial measurement unit is used as the prediction model, and the joint point positions identified by the ToF depth camera array are used as the observation values ​​to estimate the joint angles and the centroid state. The lidar array and the ToF depth camera array detect the horizontal distance d from the front of the foot to the edge of the step in real time. edge and the swing speed v of the affected leg swing ; When d edge ≤5cm and v swing When the speed is <25cm / s, a level one warning is triggered. The MPC real-time stable predictive controller increases the weight of the step avoidance term and pre-activates electrical stimulation; when d edge ≤3cm and v swing When the speed is less than 15cm / s, a level 2 warning is triggered. The MPC real-time stability prediction controller completes state estimation update, MPC re-optimization and control command issuance within 20ms. At the same time, the affected leg auxiliary exoskeleton outputs the maximum flexion assist torque, the active pelvic stabilization module outputs forward support force, the ankle joint intelligent electrical stimulator fully activates the tibialis anterior muscle, and the MPC real-time stability prediction controller extends the prediction time domain to 0.8 seconds.

9. A method for intelligent fall prevention rehabilitation training for hemiplegic patients to descend stairs independently, based on the system described in any one of claims 1 to 8, characterized in that, Includes the following steps: Step 1: Wearing calibration and fear baseline assessment, the patient wears each module of the system and completes resting baseline acquisition; Step 2: Establishing a motor baseline and systematically recording the patient's initial motor ability parameters; Step 3: Progressive closed-loop training. The patient begins training with lower steps and gradually increases the number of steps. During each downhill gait cycle: The VLM rehabilitation decision engine analyzes action semantics in real time and outputs the probability of compensatory intent prediction and pelvic movement patterns. The MPC real-time stability predictive controller predicts the centroid stability within the next 0.5 seconds with a period of 20ms, and solves the cost function to obtain the optimal control quantity; According to the optimal control amount, the active pelvic stabilization module outputs pelvic thrust, the affected leg auxiliary exoskeleton outputs hip abduction torque and knee flexion torque, the ankle joint intelligent electrical stimulator outputs electrical stimulation, and the pneumatic bionic constraint-guidance mechanism outputs directional constraint force. The context-aware voice interaction platform simultaneously plays guiding voice messages and safety margin prompts; Step 4: Accurate recording of compensation events and stability margins. The system records each compensation event and the dynamic stability margin calculated in real time by the model predictive control. Step 5: Multi-dimensional training report generation. The system generates a rehabilitation assessment report that includes model prediction and control performance indicators.

10. The method according to claim 9, characterized in that, In the gait cycle of descending stairs in step three: Once the intention to go downstairs is recognized, if the MPC real-time stability prediction controller predicts that the stability margin is low when supporting the affected side, it sends a command to the active pelvic stabilization module to apply a horizontal thrust of 50N to 150N to the healthy side to guide the center of gravity to shift to the healthy side. When the affected leg begins to swing, if the VLM rehabilitation decision engine predicts an adduction trend, or if the MPC real-time stability prediction controller detects that the adduction angle exceeds 2°, the affected leg auxiliary exoskeleton immediately outputs a hip abduction torque of 10 N·m to 15 N·m. At the same time, the pneumatic bionic constraint-guidance mechanism outputs a directional constraint force, and the affected leg auxiliary exoskeleton outputs a flexion assist torque. As the affected foot descends and approaches the next step, the ToF depth camera provides real-time feedback on the relative distance between the foot and the edge of the step. The MPC real-time stability prediction controller adjusts the knee flexion torque based on the distance, while the ankle joint intelligent electrical stimulator triggers anterior tibial electromyography stimulation based on surface electromyography signals or instructions from the VLM rehabilitation decision engine. After the affected foot lands, the MPC real-time stability prediction controller recalculates the center of gravity position. If the center of gravity is biased to the rear, the active pelvic stabilization module slowly withdraws the thrust and provides a voice prompt to the patient to lean forward.

Images