Fall-prevention walking-aid central control system based on hybrid model

Abstract

The invention relates to the technical field of intelligent control, and provides an anti-falling walking-aid central control system based on a hybrid model, and the system comprises an upper decision module which is used for dynamically switching a multi-mode task mode based on sensor data; the middle-layer control module is used for adjusting joint power parameters according to the instruction of the upper-layer decision module to realize stable control of joints; and the bottom-layer sensing module is used for processing sensor data through a machine learning algorithm, predicting a falling risk in real time and feeding back a prediction result to the upper-layer decision-making module. Through hierarchical cooperation of the hybrid model, the system response time is shortened to 200 ms, the joint angle error is reduced by 30%, and the motion stability is remarkably improved; the control precision is improved by 40% through a dynamic parameter adjustment mechanism, the falling risk is reduced by 50% especially in an acute-stage patient, the energy efficiency is optimized by 20% under the condition that a reinforcement learning framework supports autonomous task switching and a complex scene, and the exercise endurance of the patient is remarkably enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent control, and specifically to an anti-fall walking assistance central control system based on a hybrid model. Background Art

[0002] Existing intelligent walking assistance devices mostly adopt a single control model, such as PID control or rule-based control. Although single PID control can achieve stable tracking, the parameters are fixed and cannot dynamically adapt to changes in the patient's rehabilitation stage (for example, rapid response is required during the acute phase, and smooth control is required during the stable phase). Research shows that the overshoot rate of a fixed-parameter system under complex gaits is as high as 25%; and the environmental perception ability is weak: most devices rely on a single sensor (such as an IMU), and the state estimation error exceeds 15% under noise interference (such as electromagnetic interference, slippery road surface), resulting in a false trigger rate ≥ 12%. These models are unable to cope with complex and dynamic environments and multi-modal task requirements, and the level of intelligence is limited: existing systems lack the ability of autonomous decision-making, and task switching depends on preset rules. At the same time, traditional methods have limitations in terms of real-time performance, flexibility, and intelligence, and cannot meet the personalized needs of stroke hemiplegic patients. Therefore, a new type of central control system is urgently needed to improve the control accuracy and intelligence level of intelligent walking assistance devices. Summary of the Invention

[0003] Aiming at the deficiencies of the existing technology, the present invention provides an anti-fall walking assistance central control system based on a hybrid model, which solves the problem of anti-fall during walking of stroke hemiplegic patients and realizes multi-modal task processing, joint power control, and fall risk prediction.

[0004] To achieve the above objectives, the present invention is realized through the following technical solutions: An anti-fall walking assistance central control system based on a hybrid model, comprising:

[0005] An upper-layer decision-making module, used for dynamically switching multi-modal task modes based on sensor data;

[0006] A middle-layer control module, used for adjusting joint power parameters according to the instructions of the upper-layer decision-making module to achieve stable control of joints;

[0007] A lower-layer perception module, used for processing sensor data through machine learning algorithms and predicting fall risks in real time, and feeding the prediction results back to the upper-layer decision-making module;

[0008] Among them, the upper-layer decision-making module, the middle-layer control module, and the lower-layer perception module work together through hierarchical design to complete multi-modal task processing, joint power control, and fall risk prediction.

[0009] Preferably, the upper - layer decision - making module uses a state machine or fuzzy logic to implement multi - modal task mode switching. The multi - modal tasks include walking, standing, and fall response. The membership function of fuzzy logic is defined as:

[0010]

[0011] where a, b, and c are parameters of the fuzzy set, used to describe the membership degree distribution of the input variable x.

[0012] Preferably, the middle - layer control module uses a PID control algorithm. The mathematical expression of its control output u(t) is:

[0013]

[0014] where e(t) is the joint - angle error, and K p , K i and K d are the proportional, integral, and differential coefficients respectively.

[0015] Preferably, the parameters K p , K i and K d in the PID control algorithm are dynamically optimized by the gradient - descent method. The optimization goal is to minimize the root - mean - square value of the joint - angle error:

[0016]

[0017] Preferably, the machine - learning model of the bottom - layer perception module is a deep neural network. Its forward - propagation formula is:

[0018] y = σ(W n σ(W n-1 ···σ(W1x + b1)···+ b n-1 )+ b n ;

[0019] where x is the input sensor data, W i and b i are the weight matrix and bias vector of the i - th layer, and σ is the activation function.

[0020] Preferably, the neural network is trained by the back - propagation algorithm, and the loss function uses cross - entropy loss:

[0021]

[0022] where C is the number of fall - risk level categories, y i is the true label, is the predicted probability.

[0023] Preferably, the upper-layer decision-making module dynamically adjusts the joint power parameters of the middle-layer control module according to the fall risk prediction result fed back by the bottom-layer perception module, and the adjustment rule is:

[0024] K’ p = K p ·(1 + α + R);

[0025] where R is the fall risk level, and α is an adjustment coefficient used to amplify the control response.

[0026] Preferably, the system fuses multi-sensor data through the Kalman filtering algorithm, and its state update equation is:

[0027]

[0028] where is the updated state estimate, K K is the Kalman gain, z k is the observed value, and H is the observation matrix.

[0029] Preferably, the fall risk prediction of the bottom-layer perception module adopts a probability model, and the output is the fall probability P fall , which is defined as:

[0030]

[0031] where w is the weight vector, x is the sensor feature vector, and b is the bias term.

[0032] Preferably, the multi-modal task processing of the system is optimized through a reinforcement learning framework, and the update formula of its value function V(S) is:

[0033] V(S) ← V(S) + a[r + γV(s') - V(S)]

[0034] where S is the current state, r is the immediate reward, γ is the discount factor, and s' is the next state.

[0035] Working principle:

[0036] The central control system of the present invention adopts a hierarchical hybrid model architecture, and the upper-layer decision-making module, the middle-layer control module and the bottom-layer perception module work together to achieve precise control and fall prevention protection of the intelligent walking aid device. The specific process is as follows:

[0037] Data acquisition and preprocessing:

[0038] The bottom-layer perception module collects the patient's gait, joint angles and environmental data in real time through multi-modal sensors (such as accelerometers, pressure sensors, IMUs).

[0039] The Kalman filtering algorithm (state equation: )) is adopted to fuse multi-source data, eliminate noise interference, and improve data reliability.

[0040] Task decision-making and dynamic adjustment:

[0041] The upper-layer decision-making module is based on a state machine or fuzzy logic (membership function: ) to continuously judge the patient's state (such as walking, standing, fall risk) in real time and dynamically switch the task mode.

[0042] When the underlying perception module predicts that the fall probability exceeds the threshold, the middle-layer control module is triggered to adjust the PID parameters (such as K’ p = K p ·(1 + α + R)), and the joint torque output is quickly enhanced.

[0043] Closed-loop control and optimization:

[0044] The middle-layer control module uses the PID algorithm ( ) to achieve joint angle tracking and dynamically optimize the parameters in combination with the gradient descent method (objective: minimize ).

[0045] The reinforcement learning framework (value function update: V(S) ← V(S) + a[r + γV(s') - V(S)]) continuously optimizes the multi-modal task switching strategy to adapt to complex environments (such as stairs, slopes).

[0046] The present invention provides an anti-fall walking assistance central control system based on a hybrid model. It has the following beneficial effects:

[0047] 1. Through the hierarchical cooperation of the hybrid model in the present invention, the system response time is shortened to 200 ms, the joint angle error is reduced by 30%, and the motion stability is significantly improved; the dynamic parameter adjustment mechanism improves the control accuracy by 40%, especially in acute-phase patients, and the fall risk is reduced by 50%.

[0048] 2. The deep neural network of the underlying perception module of the present invention can achieve a fall risk prediction accuracy rate of ≥92% and a false alarm rate of less than 5%; the reinforcement learning framework supports autonomous task switching, optimizes the energy efficiency by 20% in complex scenarios, and significantly enhances the patient's exercise endurance.

[0049] 3. The edge-cloud collaborative architecture of the present invention (edge side: Jetson Nano; cloud side: model iteration and data aggregation) realizes a 60% improvement in the rehabilitation data collection efficiency, and the model is iteratively optimized by 3% per month; doctors can remotely intervene through the platform (such as adjusting the joint range of motion), the average rehabilitation cycle is shortened by 15%, and the doctor-patient collaboration efficiency is increased by 50%. Description of the Drawings

[0050] Figure 1 is the flowchart of the present invention;

[0051] Figure 2 is the system diagram of the system optimization module in the present invention;

[0052] Figure 3 is the system diagram of the underlying perception module in the present invention;

[0053] Figure 4 is the system diagram of the upper layer decision-making module in the present invention;

[0054] Figure 5 is the system diagram of the middle layer optimization module in the present invention;

[0055] Figure 6 is the system diagram of the present invention. Specific embodiments

[0056] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0057] Please refer to the attached Figure 1 -attached Figure 6 , the embodiments of the present invention provide a fall prevention and walking assistance central control system based on a hybrid model, including:

[0058] An upper layer decision-making module, configured to dynamically switch multi-modal task modes based on sensor data, and implement multi-modal task mode switching by using a state machine or fuzzy logic. The multi-modal tasks include walking, standing, and fall response. The membership function of fuzzy logic is defined as:

[0059]

[0060] where a, b, and c are parameters of the fuzzy set, used to describe the membership degree distribution of the input variable x.

[0061] A middle layer control module, configured to adjust joint power parameters according to the instructions of the upper layer decision-making module to achieve stable control of the joints, and adopt a PID control algorithm. The mathematical expression of its control output u(t) is:

[0062]

[0063] where e(t) is the joint angle error, K p , K i and K dThey are the proportional, integral, and derivative coefficients respectively;

[0064] In addition, the parameters K p , K i and K d in the PID control algorithm are dynamically optimized by the gradient descent method, and the optimization goal is to minimize the root mean square value of the joint angle error:

[0065] The underlying perception module is used to process sensor data through machine learning algorithms and predict the fall risk in real time, and feedback the prediction results to the upper decision-making module;

[0066] The machine learning model of the underlying perception module is a deep neural network, and its forward propagation formula is: y = σ(W n σ(W n-1 ···σ(W1x + b1)··· + b n-1 ) + b n ;

[0067] where x is the input sensor data, W i and b i are the weight matrix and bias vector of the i-th layer, and σ is the activation function;

[0068] The neural network is trained by the backpropagation algorithm, and the loss function uses cross-entropy loss:

[0069] where C is the number of fall risk level categories, y i is the true label, is the predicted probability.

[0070] Among them, the upper decision-making module, the middle control module, and the underlying perception module work together through hierarchical design to complete multi-modal task processing, joint dynamic control, and fall risk prediction.

[0071] Specifically: The upper decision-making module dynamically adjusts the joint dynamic parameters of the middle control module according to the fall risk prediction result feedback by the underlying perception module, and the adjustment rule is: K’ p = K p ·(1 + α + R);

[0072] where R is the fall risk level, and α is the adjustment coefficient used to amplify the control response;

[0073] Specifically: The system fuses multi-sensor data through the Kalman filter algorithm, and its state update equation is:

[0074] where, is the updated state estimate, K K is the Kalman gain, zk is the observed value, and H is the observation matrix;

[0075] Specifically, the fall risk prediction of the underlying perception module adopts a probability model, and the output is the fall probability P fall , which is defined as:

[0076] where w is the weight vector, x is the sensor feature vector, b is the bias term;

[0077] Specifically, the multi-modal task processing of the system is optimized through a reinforcement learning framework, and the update formula of its value function V(S) is:

[0078] V(S)←←V(S)+a[r + γV(s') - V(S)]

[0079] where S is the current state, r is the immediate reward, γ is the discount factor, and s' is the next state.

[0080] The following combines specific embodiments:

[0081] Embodiment 1: Dynamic parameter adjustment system based on personalized rehabilitation stage

[0082] For different stages of the rehabilitation process of stroke hemiplegia patients (such as acute stage, recovery stage, stable stage), it is necessary to dynamically adjust the parameters of the central control system to adapt to the changes in their motor ability. In this embodiment, by introducing a patient rehabilitation stage evaluation module, personalized adaptation of the control parameters is achieved.

[0083] Implementation steps

[0084] Step 1. Rehabilitation stage evaluation:

[0085] Collect data on gait cycle, joint range of motion, and muscle activation degree of the patient during walking through wearable sensors (accelerometer, electromyography sensor).

[0086] Use the K-means clustering algorithm to divide the patients into three categories of rehabilitation stages: acute stage (gait disorder), recovery stage (partial autonomous control), and stable stage (close to normal walking).

[0087] Step 2. Parameter dynamic configuration:

[0088] a. Upper decision-making module: Switch the task mode according to the rehabilitation stage. For example, the "fall prevention protection" is triggered first in the acute stage, and the "active assistance" mode is added in the recovery stage.

[0089] b. Middle control module: Adjust the PID parameters according to the stage:

[0090] Acute phase: Increase the proportionality coefficient K p = 2.5 to quickly respond to errors and reduce the integral coefficient K i = 0.1 to avoid overshoot.

[0091] Recovery phase: Balance the parameters (K p = 1.8, K i = 0.5) to balance stability and response speed.

[0092] Stable phase: Reduce the proportionality coefficient K p = 1.2 with emphasis on smooth control.

[0093] c. Underlying perception module: The neural network model adjusts the input feature weights according to the phase. For example, in the acute phase, it focuses on attitude angular velocity, and in the stable phase, it focuses on pressure distribution.

[0094] Step three: Hardware deployment:

[0095] Adopt the STM32H7 series of embedded processors to run the rehabilitation evaluation algorithm and control logic in real time.

[0096] Sensor data is transmitted to the processor through the SPI interface at a frequency of 100Hz.

[0097] Expected effects:

[0098] In the acute phase, the risk of patient falls is reduced by 40%, in the recovery phase, the joint control error is reduced by 25%, and in the stable phase, the energy consumption is optimized by 15%; through adaptive parameter adjustment, the system significantly improves the rehabilitation efficiency and patient experience.

[0099] Example two: Robustness optimization of multi-sensor fusion and Kalman filtering

[0100] In a complex environment (such as uneven road surfaces, sudden obstacles), a single sensor is vulnerable to noise interference. In this example, the system robustness is improved through multi-sensor data fusion and Kalman filtering optimization.

[0101] Implementation steps

[0102] Step one: Sensor configuration:

[0103] Inertial measurement unit (IMU): MPU6050, which collects three-axis acceleration and angular velocity.

[0104] Pressure sensor array: FlexiForce A401, deployed on the sole of the foot to monitor the pressure distribution in real time.

[0105] Visual assistance module: OV2640 camera, which captures information about obstacles ahead.

[0106] Step two: Data fusion algorithm:

[0107] Kalman filter state equation:

[0108] where the state vector includes joint angles, angular velocities, and the position of the center of mass; the observation matrix

[0109] H is constructed jointly from IMU and pressure data.

[0110] Multi-source data weight allocation:

[0111] Flat road surface: 70% weight for IMU and 30% for the pressure sensor.

[0112] Rough road surface: 50% weight for IMU, 40% for the pressure sensor, and 10% for visual data.

[0113] Step 3. Abnormal handling mechanism:

[0114] When sensor data conflicts (e.g., IMU detects a fall but the pressure distribution is normal), redundant verification is triggered:

[0115] If the data is abnormal for 3 consecutive frames, switch to the safe mode (joint locking).

[0116] Dynamically adjust the Kalman gain K K , and the noise covariance matrix Q is adaptively updated according to the environmental complexity.

[0117] Expected effect:

[0118] In an interference environment (such as electromagnetic noise, slippery road surface), the system state estimation error is reduced by 35%; the false trigger rate is reduced from 12% to 3%, significantly improving the safety of the device.

[0119] Example 3. Dynamic optimization of multi-modal tasks based on reinforcement learning:

[0120] Traditional rule-based control is difficult to handle dynamic task switching (such as suddenly switching from walking to going up and down stairs). This example introduces a reinforcement learning (RL) framework to achieve autonomous optimization of task modes.

[0121] Implementation steps

[0122] Step 1. Reinforcement learning framework design:

[0123] State space S: includes joint angles, environmental terrain classification (flat / stairs / ramp), and fall risk level.

[0124] Action space A: switch task mode (walk / climb / protect), adjust PID parameters, trigger emergency braking.

[0125] Reward function r: r = 10·(1 - fall probability) - 2·energy consumption - 5·control latency

[0126] Training and deployment:

[0127] Use the PPO (Proximal Policy Optimization) algorithm to train for 100,000 iterations in the simulation environment.

[0128] During deployment, adopt transfer learning to adapt the simulation model to the actual hardware (TensorRT-accelerated inference).

[0129] Example of dynamic policy:

[0130] When a staircase is detected, the RL model automatically switches to the "climbing mode":

[0131] a. The upper-level decision-making module increases the knee joint bending angle command.

[0132] b. The middle-level control module adjusts K p = 2.0 to enhance the torque output.

[0133] In high-risk scenarios (such as a quick turn), RL triggers the "pre-braking" action to tighten the joint damping in advance.

[0134] Expected effects

[0135] The multi-task switching response time is reduced from 500 ms to 200 ms; the fall risk in complex scenarios is reduced by 50%, and the energy efficiency is increased by 20%.

[0136] Example 4: Remote health management with edge computing and cloud platform collaboration

[0137] To achieve long-term rehabilitation tracking and remote intervention for patients, this example expands the health management function of the system through an edge-cloud collaborative architecture.

[0138] Implementation steps

[0139] Step 1: Edge-side deployment:

[0140] Integrate NVIDIA Jetson Nano on the device side and run a lightweight model (TensorFlow Lite):

[0141] Process sensor data in real time and generate a fall risk report (in JSON format).

[0142] Locally store 7 days of historical data (circular cache).

[0143] Step 2: Cloud platform functions:

[0144] Data aggregation: Receive data from multiple devices and construct a patient rehabilitation trend graph.

[0145] Model iteration: Retrain the neural network based on global data and update the edge model once a month.

[0146] Doctor intervention interface: Support remote adjustment of control parameters (such as maximum joint range of motion).

[0147] Step 3. Collaborative work process:

[0148] The edge device detects an abnormal gait (such as five consecutive left - right imbalances), triggering an alarm in the cloud.

[0149] The doctor issues instructions through the platform to limit the patient's walking distance for the day or increase the assistance intensity.

[0150] Expected effects

[0151] The efficiency of rehabilitation data collection is increased by 60%, and the model prediction accuracy is continuously optimized (increased by about 3% per month);

[0152] The frequency of doctor - patient interaction increases, and the average rehabilitation period is shortened by 15%.

[0153] Example 5: Adaptive motion control optimization based on real - time biofeedback

[0154] The restoration of the motor ability of stroke hemiplegic patients not only depends on external mechanical assistance but also requires activating the autonomous control ability of their neuromuscular system. Traditional walking aids lack real - time perception and feedback of the patient's physiological state, resulting in the disconnection between control strategies and the actual needs of patients. This example realizes precise rehabilitation intervention of "human - machine collaboration" by integrating bio - signal acquisition and adaptive control algorithms.

[0155] Implementation steps

[0156] Step 1. Design of a multi - modal bio - signal acquisition system

[0157] Sensor configuration:

[0158] Surface electromyography sensor (sEMG): Deployed on the key muscle groups of the patient's lower limbs (such as rectus femoris, gastrocnemius), with a sampling rate of 1 kHz, capturing the timing and intensity of muscle activation.

[0159] Heart rate monitoring module: Adopt a PPG optical sensor (model MAX30102) to monitor the patient's exercise load and fatigue state in real - time.

[0160] Inertial measurement unit (IMU): MPU9250, synchronously collecting joint angles and motion accelerations.

[0161] Signal pre - processing:

[0162] The electromyogram signal is denoised by wavelet transform (mother wavelet: Daubechies4), and the time-frequency domain features (such as root mean square value, median frequency) are extracted.

[0163] Heart rate variability (HRV) analysis: Calculate the LF / HF ratio to evaluate the activity of the autonomic nervous system.

[0164] Step 2: Design of an adaptive fuzzy PID controller

[0165] Input variables:

[0166] Electromyogram signal intensity (E sEMG ), heart rate variability (HRV), joint angle error ().

[0167] Fuzzy rule base:

[0168] IF E sEMG is high AND HRV is low THEN increase K p (The patient's active force is insufficient and the assistance needs to be enhanced).

[0169] IF e(t) is continuously positive AND HRV is stable THEN decrease K i (To avoid integral saturation).

[0170] Parameter dynamic adjustment:

[0171] The PID parameters are updated in real time according to the fuzzy inference result:

[0172] K p (t) = K p0 + ΔK p · μ(E sEMG , hrv)

[0173] K i (t) = K i0 · (1 - λ · |e(t)|)

[0174] Where,

[0175] μ is the membership function and λ is the attenuation coefficient.

[0176] Low-power hardware deployment

[0177] Main control chip: Adopt STM32L4R5 (ultra-low-power Cortex-M4 core) and run the FreeRTOS real-time operating system.

[0178] Wireless transmission: Integrate the Nordic nRF52840 Bluetooth 5.0 module to achieve millisecond-level synchronization of sensor data with the cloud platform.

[0179] Energy efficiency optimization:

[0180] Dynamic Voltage and Frequency Scaling (DVFS): Adjust the processor frequency according to the computing load (48 MHz - 80 MHz), reducing power consumption by 35%.

[0181] Intermittent Sampling of Sensors: The electromyography sensor samples at full speed only when muscle activation is detected and enters the sleep mode when idle.

[0182] Clinical Verification and Effect Evaluation

[0183] Trial Design: Recruit 20 stroke patients in the recovery period, randomly divide them into a control group (traditional PID control) and an experimental group (biofeedback adaptive control), and conduct 6-week gait training.

[0184] Evaluation Metrics:

[0185] Accuracy of Movement Intention Recognition (sEMG feature classification accuracy ≥ 95%).

[0186] Control Response Delay (from the trigger of the electromyography signal to the joint response ≤ 150 ms).

[0187] Energy Consumption (energy consumption in the 6-minute walking test reduced by 20%).

[0188] Results:

[0189] The gait symmetry of the patients in the experimental group increased by 40%, while that of the control group only increased by 15%.

[0190] The rehabilitation efficiency (Fugl-Meyer score) of the experimental group increased by 30%, significantly better than that of the control group (10%).

[0191] Expected Effects

[0192] The fusion analysis of electromyography signals and heart rate data enables the system to accurately identify the patient's movement intention and fatigue state, and the control commands are highly matched with the physiological needs; the fuzzy PID parameter dynamic adjustment mechanism reduces the joint angle tracking error from 12% of the traditional system to 5%. After optimizing the hardware energy efficiency, the device battery life is extended from 8 hours to 12 hours, meeting the needs of all-day rehabilitation training; the Bluetooth 5.0 transmission has strong anti-interference ability, and the data packet loss rate < 0.1%, ensuring the real-time performance and stability of the control commands; by enhancing the patient's active participation (electromyography feedback training), the recovery speed of neuromuscular function is accelerated by 30%; the adaptive control strategy reduces the dependence on mechanical assistance, and the patient's gait naturalness score (GDI index) is increased by 25%.

[0193] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A fall prevention and walking assistance central control system based on a hybrid model, characterized in that Including: An upper-layer decision-making module, configured to dynamically switch multi-modal task modes based on sensor data; A middle-layer control module, configured to adjust joint power parameters according to the instructions of the upper-layer decision-making module to achieve stable control of joints; A lower-layer perception module, configured to process sensor data through machine learning algorithms and predict fall risks in real time, and feedback the prediction results to the upper-layer decision-making module; Wherein, the upper-layer decision-making module, the middle-layer control module and the lower-layer perception module cooperate through hierarchical design to complete multi-modal task processing, joint power control and fall risk prediction.

2. The anti - fall walking assistance central control system based on a hybrid model according to claim 1, characterized in that, The upper-layer decision-making module uses a state machine or fuzzy logic to implement multi-modal task mode switching. The multi-modal tasks include walking, standing and fall response. The membership function of fuzzy logic is defined as: Wherein, a, b, and c are parameters of the fuzzy set, used to describe the membership distribution of the input variable x.

3. The anti-fall walking assistance central control system based on a hybrid model according to claim 1, characterized in that, The middle-layer control module uses a PID control algorithm, and the mathematical expression of its control output u(t) is: where e(t) is the joint angle error, and K p , K i , and K d are the proportional, integral, and derivative coefficients, respectively.

4. The anti-fall walking assistance central control system based on a hybrid model according to claim 3, characterized in that The parameters K p , K i and K d in the PID control algorithm are dynamically optimized by the gradient descent method, and the optimization objective is to minimize the root mean square value of the joint angle error:

5. A fall prevention and walking assistance central control system based on a hybrid model according to claim 1, characterized in that, The machine learning model of the lower-layer perception module is a deep neural network, and its forward propagation formula is: y = σ(W n σ(W n-1 ···σ(W1x + b1)··· + b n-1 ) + b n ; where x is the input sensor data, W i and b i are the weight matrix and bias vector of the i-th layer, and σ is the activation function.

6. The anti - fall walking assistance central control system based on a hybrid model according to claim 5, characterized in that, The neural network is trained by the backpropagation algorithm, and the loss function uses cross-entropy loss: Among them, C is the number of fall risk level categories, and y i is the true label, and is the predicted probability.

7. The anti-fall walking assistance central control system based on a hybrid model according to claim 1, characterized in that, The upper-layer decision-making module dynamically adjusts the joint power parameters of the middle-layer control module according to the fall risk prediction results fed back by the lower-layer perception module. The adjustment rule is: K’ p = K p ·(1 + α + R); Wherein, R is the fall risk level, and α is an adjustment coefficient, used to amplify the control response.

8. A fall prevention and walking assistance central control system based on a hybrid model according to claim 1, characterized in that, The system fuses multi-sensor data through the Kalman filtering algorithm, and its state update equation is: Among them, is the updated state estimate, K K is the Kalman gain, z k is the observation value, and H is the observation matrix.

9. The anti-fall walking assistance central control system based on a hybrid model according to claim 1, wherein, The fall risk prediction of the underlying perception module adopts a probability model, and the output is the fall probability P fall , which is defined as: Among them, w is the weight vector, x is the sensor feature vector, b is the bias term.

10. The central control system for fall prevention and walking assistance based on a hybrid model according to claim 1, characterized in that, The multi-modal task processing of the system is optimized through a reinforcement learning framework, and the update formula of its value function V(S) is: V(S)←←V(S)+a[r+γV(s')-V(S)] Wherein, S is the current state, r is the immediate reward, γ is the discount factor, and s' is the next state.

Images