Multi-source sensor fusion-based upper and lower limb collaborative rehabilitation training method

Abstract

The present application relates to the technical field of medical rehabilitation robot control, and discloses a multi-source sensing fusion upper and lower limb cooperative rehabilitation training method, comprising the following steps: synchronously collecting multi-source physiological and kinematic data and calibrating the system inherent total delay time; decoding the lower limb gait phase based on a hidden Markov model; calculating the real-time muscle fatigue index of the hand, dynamically adjusting the electromyography and human-machine interaction force weight, and extracting the force intention high-frequency mutation characteristic quantity; combining the gait phase, force intention characteristic and system delay time to perform feedforward trajectory advance prediction, calculating the lower limb target auxiliary torque; filtering and smoothing the target torque through estimating the flexible driver collective disturbance torque, and superimposing disturbance feedforward compensation to generate motor instructions. The present application constructs an upper and lower limb cooperative control closed loop, overcomes the intention recognition deviation caused by physiological fatigue, compensates the system time lag and driving nonlinear disturbance, and improves the control precision of the auxiliary torque.

Description

Technical Field

[0001] This invention relates to the field of medical rehabilitation robot control technology, specifically a multi-source sensor fusion-based upper and lower limb coordinated rehabilitation training method. Background Technology

[0002] With the development of rehabilitation robot technology, exoskeleton robots have been widely used in the rehabilitation training of patients with impaired neuromotor function. Current rehabilitation training methods mostly focus on the control of single limbs or independent joints, often leaving the upper and lower limbs in a separate training state, lacking a synergistic linkage mechanism between limbs, and making it difficult to effectively stimulate the neuromuscular synergistic effect of the human body.

[0003] To achieve coordinated control of the upper and lower limbs, existing technologies typically incorporate multimodal data, such as surface electromyography (EMG) signals and human-computer interaction forces, to identify human movement intentions. However, during prolonged continuous training, patients inevitably experience muscle fatigue, leading to frequency domain shifts and amplitude attenuation in EMG signals. Most existing intention recognition models employ fixed sensor feature weight allocation strategies, failing to consider the dynamic physiological state of muscle fatigue. When a patient experiences fatigue, if the system maintains the original EMG weights, signal distortion will cause deviations in the intention recognition results, making the output interaction torque unable to accurately respond to the patient's true force exertion intention.

[0004] Furthermore, the synchronous acquisition, feature extraction, and fusion calculation of multi-source sensor data require significant computing resources. Combined with the communication time of the underlying control bus, the entire system suffers from inherent processing and transmission delays that cannot be ignored. Existing control schemes typically issue calculated control commands directly, lacking advanced prediction and feedforward compensation mechanisms to address the aforementioned system time delays. This results in the physical output of the auxiliary torque lagging in phase with the patient's actual movement rhythm. Simultaneously, for human-computer interaction safety considerations, rehabilitation exoskeletons often employ flexible actuators as the actuating mechanism. In the underlying torque control, existing methods often neglect unmodeled dynamics such as nonlinear disturbances caused by elastic element deformation and frictional losses during mechanical transmission. Without disturbance observation compensation and anti-integral saturation mechanisms, the system's ability to cope with external load fluctuations is weak, easily leading to torque tracking errors and control overshoot, resulting in uneven auxiliary torque output and consequently reducing the compliance of rehabilitation training. Summary of the Invention

[0005] The technical problem solved by this invention is that existing upper and lower limb rehabilitation training methods usually lack a collaborative linkage mechanism between limbs, and fail to fully consider the impact of system communication delay, the patient's real-time muscle fatigue state, and the inherent disturbance of the flexible actuator on the accuracy of the auxiliary torque in human-computer interaction control, resulting in delayed recognition of movement intention, unsmooth auxiliary torque output, and poor human-computer compliance.

[0006] To address the above problems, the present invention provides the following technical solution: This invention provides a multi-source sensor fusion-based upper and lower limb coordinated rehabilitation training method, comprising the following steps: Simultaneously acquire multi-source physiological and kinematic data, perform system initialization, and calibrate the system's inherent total communication and processing latency. The system receives lower limb kinematic data based on a hidden Markov model, decodes it online, and outputs the lower limb gait phase. Acquire electromyographic signals on the surface of the hand and calculate the real-time muscle fatigue index. Adjust the weights dynamically based on the real-time muscle fatigue index and calculate the high-frequency abrupt change characteristic of the force exertion intention by combining the human-computer interaction force detected by the torque sensor. Based on the high-frequency abrupt change characteristics of lower limb gait phase and force exertion intention, and combined with the system's inherent communication and processing total delay time, feedforward trajectory prediction is performed, and lower limb target auxiliary torque is calculated. The lumped disturbance torque is estimated by obtaining the spring deformation displacement of the flexible actuator, the auxiliary torque of the lower limb target is filtered and smoothed, and the motor torque command is generated by combining the lumped disturbance torque and output to the flexible actuator for execution.

[0007] The innovative principle of this invention lies in establishing a coordinated control closed loop for the upper and lower limbs by synchronously processing multi-source physiological and kinematic data. During the control process, an online calibration mechanism for the total system delay is introduced, and dynamic time prediction step length calculation is performed by combining the high-frequency abrupt changes in lower limb gait phase and upper limb force exertion intention, achieving feedforward trajectory prediction to compensate for the time lag effect of system communication and hardware processing. Simultaneously, a real-time muscle fatigue index is introduced during the force exertion intention extraction stage, dynamically allocating the feature weights of electromyography and torque signals to overcome the problem of intention misjudgment easily caused by a single sensor source in a fatigued patient state. Furthermore, a discretized dynamics observer is used to estimate the lumped disturbance torque of the flexible actuator, and feedforward compensation is performed when generating the final motor command. Combined with filtering and smoothing processing targeting the target torque, the control error caused by mechanical structure deformation is reduced.

[0008] Furthermore, the method of receiving lower limb kinematic data based on a hidden Markov model, decoding and outputting the lower limb gait phase online includes: extracting the angular velocity and angular acceleration of both lower limbs and constructing an observation vector to characterize the lower limb motion state by combining it with joint angles; calculating the initial local maximum likelihood variable by combining the initial state probability vector in the optimized model parameters with the observation state transition probability matrix; performing logarithmic domain transformation on the local maximum likelihood variable, and calculating the maximum path probability of the current time step in each candidate state based on the local maximum likelihood variable retained in the previous time step and the state transition probability matrix; comparing the values ​​of the local maximum likelihood variables corresponding to all hidden states in the current time step, and extracting the state identifier corresponding to the maximum value of the local maximum likelihood variable as the lower limb gait phase of the current time step.

[0009] Furthermore, the acquisition of hand surface electromyography (EMG) signals and calculation of real-time muscle fatigue index includes: extracting hand surface EMG signals within a sliding time window, performing frequency conversion and calculating the power spectral density distribution; solving for the critical frequency point that precisely divides the total power spectral density area equally as the median frequency of the current time step; combining the baseline median frequency recorded in the initial stage with the median frequency of the current time step to obtain the frequency offset attenuation amount; calculating the normalized initial fatigue quantization value; using an exponential moving average algorithm to perform time-domain smoothing filtering on the initial fatigue quantization value to generate a smooth fatigue quantization value; and using saturation limiting logic to restrict the smooth fatigue quantization value to a real number boundary range of zero to one, outputting the final real-time muscle fatigue index.

[0010] Furthermore, the step of dynamically adjusting the weights based on the real-time muscle fatigue index and calculating the high-frequency abrupt change characteristic of the force exertion intention based on the human-computer interaction force detected by the torque sensor includes: normalizing the electromyography (EMG) signal on the hand surface and the human-computer interaction force to obtain the EMG activation degree and torque mapping value; subtracting the product of the real-time muscle fatigue index and the preset fatigue compensation coefficient from the preset initial EMG feature weight to obtain the dynamic EMG weight; adding the product of the real-time muscle fatigue index and the fatigue compensation coefficient to the preset initial torque feature weight to obtain the dynamic torque weight; calculating the comprehensive force exertion intention feature by combining the dynamic EMG weight, EMG activation degree, dynamic torque weight, and torque mapping value; performing first-order difference calculation on the comprehensive force exertion intention feature of adjacent time steps to obtain the first-order difference component; introducing a nonlinear dead zone activation function to truncate and filter the first-order difference component; and outputting the high-frequency abrupt change characteristic of the force exertion intention after filtering out static noise.

[0011] Furthermore, the step of predicting the target auxiliary torque of the lower limb based on the high-frequency abrupt change features of the lower limb gait phase and force intention, combined with the system's inherent communication and processing total delay time, includes: matching the corresponding discrete kinematic node data in the standardized healthy gait database according to the lower limb gait phase, performing smooth interpolation calculation using a cubic spline interpolation algorithm to generate the basic expected joint angle, basic expected joint angular velocity, and basic expected joint angular acceleration; introducing a nonlinear smoothing mapping mechanism based on a logic function to convert the high-frequency abrupt change features of force intention into normalized smooth triggering coefficients; wherein, 50% of the historical peak value of the high-frequency abrupt change features of force intention extracted within a preset sliding time window is dynamically set as the abrupt change activation center threshold.

[0012] Furthermore, after generating the basic expected joint angle, basic expected joint angular velocity, basic expected joint angular acceleration, and smooth trigger coefficient, the process also includes: adding the product of the preset maximum feedforward time compensation gain and the smooth trigger coefficient to the total system inherent communication and processing delay time to obtain the dynamic time prediction step size; based on the second-order Taylor series expansion principle, combining the basic expected joint angle, basic expected joint angular velocity, basic expected joint angular acceleration, and dynamic time prediction step size, calculating the advanced expected joint angle and advanced expected joint angular velocity; limiting the advanced expected joint angle to the upper and lower limits of the human physiological limit collision avoidance angle of the corresponding joint, and proportionally truncating the amplitude of the advanced expected joint angular velocity to the maximum safe angular velocity threshold of the motor.

[0013] Furthermore, after calculating the expected joint angle and expected joint angular velocity, the process also includes: calculating the root mean square value of the electromyography signal on the lower limb surface to obtain the degree of lower limb muscle activation; generating basic stiffness and basic damping based on the degree of lower limb muscle activation and a preset proportional coefficient; multiplying the smoothing trigger coefficient by the preset peak values ​​of additional stiffness and additional damping, and adding them to the basic stiffness and basic damping to synthesize the comprehensive stiffness and comprehensive damping at the current time step; wherein the peak value of additional damping and the peak value of additional stiffness are in a critical damping matching state; calculating the tracking deviation based on the expected joint angle and expected joint angular velocity, combined with the current actual joint angle and current actual joint angular velocity; calculating the target auxiliary torque of the lower limb based on the dynamic closed-loop equation combined with the comprehensive stiffness and comprehensive damping; and performing maximum auxiliary torque limiting processing on it.

[0014] Furthermore, the step of obtaining the lumped disturbance torque by estimating the spring deformation displacement of the flexible actuator includes: calculating the elastic transmission torque by combining the spring deformation displacement with a preset spring stiffness coefficient; performing discrete difference calculation after filtering the actual motor angular velocity collected by the rotary encoder using a low-pass filter to obtain the smooth actual motor angular acceleration; constructing a discretized dynamic observer to calculate the dynamic model residual by combining the elastic transmission torque, the smooth actual motor angular acceleration, the actual motor angular velocity, and the motor torque command output in the previous time step; and deducing the lumped disturbance torque of the current time step by combining the set observer filtering coefficients and the lumped disturbance torque retained in the previous time step.

[0015] Furthermore, the step of filtering and smoothing the lower limb target auxiliary torque, and generating and outputting the motor torque command in combination with the lumped disturbance torque, includes: introducing a nonlinear tracking differentiator based on kinematic limit constraints to perform first-order low-pass filtering and amplitude limiting smoothing on the lower limb target auxiliary torque, and extracting the lower limb smoothed auxiliary torque; subtracting the elastic transmission torque from the lower limb smoothed auxiliary torque to form the torque tracking error; combining the torque tracking error with the proportional-integral-derivative (PID) control law to generate a basic feedback adjustment command; wherein, in the calculation of the integral term of the PID control law, a dynamic anti-integral saturation limiting algorithm is introduced, and when the absolute value of the result of the cumulative sum of the torque tracking error over the historical time step multiplied by the integral control gain exceeds the preset integral saturation torque threshold, the accumulation calculation of the integral term is stopped; and the basic feedback adjustment command and the lumped disturbance torque are fed forward and superimposed for compensation to synthesize the final motor torque command.

[0016] Furthermore, the multi-source physiological and kinematic data are collected by sensors configured on the lower limb rehabilitation exoskeleton robot and the hand function rehabilitation exoskeleton robot; the calibration system's inherent total communication and processing delay time specifically includes: generating a benchmark test pulse and injecting it into the system control bus during the system startup phase, recording the time span from the issuance of the benchmark test pulse to the receipt of status feedback, and defining this time span as the system's inherent total communication and processing delay time.

[0017] This invention provides a multi-source sensor fusion-based upper and lower limb coordinated rehabilitation training method. It has the following beneficial effects:

[0018] 1. This invention calculates a real-time muscle fatigue index by acquiring surface electromyography (EMG) signals and dynamically adjusts the weighting of EMG activation and human-computer interaction force in calculating force intention based on this index. This design solves the problem of intention recognition deviation caused by EMG signal attenuation due to patient muscle fatigue during long-term rehabilitation training. When fatigue occurs, the system automatically reduces the weighting of EMG signals and increases the weighting of physical torque, ensuring the accuracy of movement intention extraction and maintaining the stability of human-computer interaction control.

[0019] 2. This invention combines the inherent total communication and processing delay time of the system's initial calibration with the dynamic calculation of time prediction step length based on the high-frequency abrupt changes in the upper limb force exertion intention, and performs feedforward trajectory prediction for the lower limb target auxiliary torque. This mechanism directly compensates for the physical time delay caused by data transmission and algorithm calculation through time advance calculation of kinematic state, eliminates phase lag in the underlying hardware execution process, and improves the synchronization of upper and lower limb coordinated movements.

[0020] 3. This invention utilizes the spring deformation displacement of the flexible actuator and the actual motion state of the motor to construct a discretized dynamic observer, estimate the lumped disturbance torque, and perform feedforward superposition compensation. Combined with the filtering and smoothing processing of the nonlinear tracking differentiator and the anti-integral saturation mechanism, the interference of nonlinear deformation of the flexible transmission mechanism and external load fluctuations on torque execution is eliminated, reducing torque tracking error and avoiding control overshoot, thus achieving high-precision smooth output of auxiliary torque. Attached Figure Description

[0021] Figure 1 This is a system architecture block diagram of the present invention; Figure 2 This is a schematic diagram of the method flow of the present invention; Figure 3 This is a schematic diagram of the control flow of the present invention; Figure 4 This is a schematic diagram of the cross-limb collaborative control algorithm and torque command synthesis flowchart of the present invention.

[0022] Among them, 100 is the data synchronization module; 200 is the gait decoding module; 300 is the intent extraction module; 400 is the feedforward control module; and 500 is the underlying execution module. Detailed Implementation

[0023] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] See attached document Figure 1 The present invention provides a multi-source sensor fusion upper and lower limb coordinated rehabilitation training system, which may include a lower limb rehabilitation exoskeleton robot, a hand function rehabilitation exoskeleton robot and a central real-time control unit.

[0025] The lower limb rehabilitation exoskeleton robot is equipped with flexible actuators. The hand function rehabilitation exoskeleton robot is equipped with an arc-groove transmission mechanism. The central real-time control unit communicates with both the lower limb rehabilitation exoskeleton robot and the hand function rehabilitation exoskeleton robot.

[0026] The lower limb rehabilitation exoskeleton robot is equipped with an inertial measurement unit, rotary encoder, plantar pressure sensor, lower limb electromyography (EMG) sensor, and displacement sensor. The displacement sensor is housed within the flexible actuator. The hand function rehabilitation exoskeleton robot is equipped with a torque sensor and a hand EMG sensor.

[0027] The central real-time control unit contains a data synchronization module 100, a gait decoding module 200, an intent extraction module 300, a feedforward control module 400, and a low-level execution module 500.

[0028] The data synchronization module 100 is connected to the signals of each sensor, synchronizes the sensor sampling data and assigns timestamps.

[0029] The gait decoding module 200 receives lower limb kinematic data from the inertial measurement unit and the rotary encoder, calculates and outputs the lower limb gait phase.

[0030] The intent extraction module 300 receives data from the torque sensor and the hand electromyography sensor, calculates and outputs feature quantities.

[0031] The feedforward control module 400 receives the lower limb gait phase output by the gait decoding module 200 and the feature quantity output by the intent extraction module 300, and calculates the lower limb target auxiliary torque.

[0032] The underlying execution module 500 receives the lower limb target auxiliary torque output by the feedforward control module 400, combines it with the spring deformation data from the displacement sensor for filtering, and generates a motor torque command to be output to the flexible actuator of the lower limb rehabilitation exoskeleton robot.

[0033] See attached document Figure 2 This invention provides a multi-source sensor fusion-based upper and lower limb coordinated rehabilitation training method, comprising the following steps: S100, synchronously acquiring multi-source physiological and kinematic data, performing system initialization and calibrating communication and processing delay times; S200, receiving lower limb kinematic data based on a hidden Markov model, decoding online and outputting lower limb gait phase; S300, acquiring hand electromyographic signals and calculating muscle fatigue index, dynamically adjusting weights based on muscle fatigue index, and calculating high-frequency abrupt change features of force exertion intention based on hand torque data; S400, performing feedforward trajectory prediction based on lower limb gait phase and high-frequency abrupt change features of force exertion intention, and calculating lower limb target auxiliary torque based on communication and processing delay times; S500, acquiring spring deformation data of flexible actuator to estimate lumped disturbance torque, filtering and smoothing lower limb target auxiliary torque, generating motor torque command based on lumped disturbance torque and outputting it to flexible actuator for execution.

[0034] The specific implementation of the technical solution of the present invention will be described in detail below with reference to each step.

[0035] See attached document Figure 3 In specific implementation, the synchronous acquisition of multi-source physiological and kinematic data and the system initialization process can be further included in sub-steps S101 to S103. Sub-step S101 is that the data synchronization module 100 establishes a communication connection with each sensor at the bottom layer through the control bus and sets a unified discrete time step for the whole system. For the timestamp alignment mechanism of multi-source heterogeneous data, it can be executed according to the existing high-precision clock synchronization protocol. Under the premise of establishing a unified time step, the data synchronization module 100 reads the feedback signals of the inertial measurement unit and rotary encoder configured in the lower limb rehabilitation exoskeleton robot in real time.

[0036] The data synchronization module 100 extracts the angular velocity and angular acceleration of both lower limbs and combines them with joint angles to construct an observation vector characterizing the lower limb motion state. This observation vector is then input into the gait decoding module 200 for state estimation. The specific mathematical expression of the observation vector is as follows:

[0037] In the formula, For observations in discrete time steps, and These represent the velocities of the right and left lower limbs at the corresponding time steps. and These represent the angular accelerations of the right and left lower limbs at the corresponding time steps. This represents the angle of the right knee joint relative to the global coordinate system. The angle of the right lower leg relative to the global coordinate system.

[0038] While acquiring lower limb kinematic data, the data synchronization module 100 simultaneously extracts the lower limb surface electromyography signals collected by the lower limb electromyography sensor, the spring deformation displacement measured by the displacement sensor deployed inside the flexible actuator, and the plantar pressure signal collected by the plantar pressure sensor. In addition, the system synchronously reads the human-computer interaction force detected by the torque sensor configured on the hand function rehabilitation exoskeleton robot and the hand surface electromyography signals collected by the hand electromyography sensor via the communication network. The lower limb surface electromyography signals, spring deformation displacement, plantar pressure signals, human-computer interaction forces, and hand surface electromyography signals are all assigned corresponding timestamps at the same time step and packaged and sent to the shared memory of the central real-time control unit for downstream modules to call.

[0039] Sub-step S102 involves the data synchronization module 100 performing communication and processing delay time calibration. During the system startup phase, the data synchronization module 100 generates a benchmark test pulse and injects it into the system control bus. The system records the time span from the issuance of the benchmark test pulse to the feedback of the receiving status by the underlying execution module 500 through an internal hardware timer, and defines the time span as the system's inherent total communication and processing delay time. The system's inherent total communication and processing delay time will be stored and provided as a constant parameter to the feedforward control module 400 for phase correction during subsequent trajectory prediction.

[0040] Sub-step S103 initializes the hidden Markov model parameters for the gait decoding module 200. The gait decoding module 200 constructs a probability state space in memory to map the lower limb movement process, divides a complete gait cycle into seven independent discrete states, and constructs a set of hidden states. The hidden state set is structured as follows:

[0041] In the formula, For the set of hidden states, to These correspond to the seven standard gait phases from the initial landing phase of the lower limbs to the end of the swing phase.

[0042] The gait decoding module 200 constructs a probability matrix system for state calculation based on preset benchmark parameters. It defines an initial state probability vector to characterize the prior probability of the system being in each state at startup, and a state transition probability matrix to describe the statistical regularity of the temporal transitions between gait phases. Furthermore, it defines an observation state transition probability matrix to establish a mapping relationship between hidden states and observation vectors. The initialization expressions for the initial state probability vector, state transition probability matrix, and observation state transition probability matrix are as follows:

[0043] In the formula, Let be the initial state probability vector. This represents the prior probability of being in a specified state at the initial moment. Here is the state transition probability matrix. The transition probability from the current state to the next target state. Let be the observation state transition probability matrix. This represents the conditional probability for generating the corresponding observation vector under a specific state. and All are positive integers.

[0044] The process of receiving lower limb kinematic data based on a hidden Markov model, decoding it online, and outputting the lower limb gait phase can be further implemented by including sub-steps S201 to S202.

[0045] Sub-step S201 involves the gait decoding module 200 performing offline training of the Hidden Markov Model (HMM) parameters. The preprocessing and optimization mechanism for the HMM parameters can be implemented using the existing Baum-Welch algorithm. Specifically, by inputting a pre-collected historical observation vector sequence of healthy subjects under standard rehabilitation training conditions, the initial state probability vector, state transition probability matrix, and observed state transition probability matrix are iteratively updated. During the iteration process, human physiological gait constraints are introduced, restricting state transitions to only occur in subsequent adjacent states or while maintaining the current state. State transition probabilities that do not conform to normal gait timing are forcibly set to zero until the initial state probability vector, state transition probability matrix, and observed state transition probability matrix converge to a set error threshold. In some examples, the set error threshold can be set as the increment of the log-likelihood function between two consecutive iterations being less than a first preset minimum value. The range of the first preset minimum value can be... This allows us to obtain optimized model parameters, which are then used to support the efficient operation of subsequent real-time sequence decoding.

[0046] Sub-step S202 involves the gait decoding module 200 performing real-time sequence decoding based on the Viterbi algorithm. During this process, the Hidden Markov Model does not rely on cross-sectional data at a single moment for judgment, but rather seeks a gait state transition trajectory that best conforms to human physiological laws by evaluating the correlation between historical time sequences. This avoids abrupt jumps in gait phase when transient environmental noise exists in the sensors. After obtaining the observation vector at the current time step, the gait decoding module 200 initiates recursive calculations to find the hidden state path with the highest probability. When the time step is at the initial startup moment of the multi-source sensor fusion upper and lower limb coordinated rehabilitation training system, the gait decoding module 200 performs state variable initialization. It calculates the initial local maximum likelihood variable by combining the initial state probability vector in the optimized model parameters and the observation state transition probability matrix. The initial local maximum likelihood variable represents the joint probability that the multi-source sensor fusion upper and lower limb coordinated rehabilitation training system is in each independent state and generates the initial observation vector at the initial moment. The specific calculation expression for the initial local maximum likelihood variable is as follows:

[0047] In the formula, Let be the initial local maximum likelihood variable that is initially in a specified state with a state label of a positive integer. This represents the prior probability of being in a specified state at the initial moment. This represents the conditional probability for generating the initial observation vector under a specified state. The observation vector acquired at the initial time. It is a positive integer and its value ranges from one to seven.

[0048] After the multi-source sensor fusion upper and lower limb coordinated rehabilitation training system passes the initial moment and enters the subsequent discrete time steps, the gait decoding module 200 performs recursive derivation calculations. Based on the local maximum likelihood variables and state transition probability matrix retained from the previous time step, the gait decoding module 200 calculates the maximum path probability in each candidate state for the current time step. It also calculates the local maximum likelihood variables for the current time step by combining the observation vector and the observation state transition probability matrix obtained at the current time step. To avoid computational dead zones caused by multiplying consecutively low probability values, resulting in underflow of floating-point numbers, the recursive derivation process undergoes a logarithmic domain transformation. The specific derivation expression for the local maximum likelihood variables after the logarithmic domain transformation is as follows:

[0049] In the formula, This represents the natural logarithm of the local maximum likelihood variable for the target state at the current time step. This is the natural logarithm of the local maximum likelihood variable in the source state at the previous time step. Let be the transition probability from the source state to the target state. The conditional probability for generating the observation vector at the current time step under the target state. The time step number is greater than one. It is a positive integer identifier for the target state, and its value ranges from one to seven.

[0050] After completing the calculation of the local maximum likelihood variables for the current time step, the gait decoding module 200 performs optimal path backtracking. The gait decoding module 200 compares the local maximum likelihood variable values ​​corresponding to all seven hidden states at the current time step and extracts the state identifier corresponding to the maximum local maximum likelihood variable value. The extracted state identifier is used as the lower limb gait phase for the current time step. The backtracking solution expression for the lower limb gait phase is as follows:

[0051] In the formula, The lower limb gait phase is the decoded output of the current time step. The natural logarithmic value of the local maximum likelihood variable in each independent state at the current time step is used as the natural logarithmic value of the lower limb gait phase. The lower limb gait phase is output to the feedforward control module 400 in real time as the rhythm reference for generating the basic reference trajectory.

[0052] Acquire hand electromyography signals and calculate muscle fatigue index. Adjust weights dynamically based on muscle fatigue index. Combine hand torque data to calculate high-frequency abrupt change characteristics of force exertion intention. In specific implementation, this may further include sub-steps S301 to S302.

[0053] Sub-step S301 involves the intent extraction module 300 acquiring the electromyographic (EMG) signal from the hand surface and calculating the real-time muscle fatigue index. The intent extraction module 300 retrieves the hand surface EMG signal within a sliding time window from the shared memory of the central real-time control unit. In some examples, the time-frequency domain conversion of the hand surface EMG signal can be performed using existing fast Fourier transform algorithms. From a physical perspective, as the contraction time of the hand muscles increases, lactic acid accumulation in the muscle fibers leads to a gradual decrease in the action potential conduction velocity. This is macroscopically manifested as a gradual decrease in the high-frequency components and an increase in the low-frequency components of the hand surface EMG signal. The intent extraction module 300 calculates the power spectral density distribution of the hand surface EMG signal after frequency conversion and solves for the median frequency corresponding to the area division points. The median frequency characterizes the leftward shift and attenuation of the spectrum of hand muscle fibers during continuous contraction. The intent extraction module 300 combines the baseline median frequency, recorded in the initial stage of the multi-source sensor fusion upper and lower limb coordinated rehabilitation training system and whose value is strictly greater than zero, to calculate the real-time muscle fatigue index. The specific calculation expression for the real-time muscle fatigue index is as follows:

[0054] In the formula, The real-time muscle fatigue index is calculated at the current time step. The baseline median frequency recorded when the multi-source sensor fusion upper and lower limb coordinated rehabilitation training system is in its initial, fatigue-free state. The median frequency is obtained by solving the short-time power spectral density at the current time step. In specific ecological records, the real-time muscle fatigue index is set to maintain a limit range within a continuous real number interval between zero and one to prevent extreme dilution anomalies from causing the feedforward control module 400 to misconfigure.

[0055] Sub-step S302 involves the intention extraction module 300 dynamically adjusting weights based on the real-time muscle fatigue index and calculating high-frequency abrupt change features of the force exertion intention in conjunction with hand torque data. Simultaneously, the intention extraction module 300 reads the human-computer interaction force from the shared memory of the central real-time control unit. It then calculates the root mean square value of the electromyography (EMG) signal on the hand surface and normalizes it to obtain the EMG activation level. Simultaneously, it normalizes the absolute value of the human-computer interaction force to obtain the torque mapping value. The intention extraction module 300 establishes a dynamic weight allocation mechanism based on the real-time muscle fatigue index. Since the amplitude of the EMG signal on the hand surface usually increases non-linearly when hand muscles are fatigued, continuing to rely on a fixed proportion of EMG features would inevitably cause excessive output from the feedforward control module 400. Therefore, as the degree of hand muscle fatigue deepens, the weight of the EMG features is reduced linearly, and the weight of the torque features is simultaneously increased to compensate, thereby constructing a comprehensive force exertion intention feature that can resist physiological fatigue interference. The specific calculation expression for the comprehensive force exertion intention feature is as follows:

[0056] In the formula, This represents the overall intent to exert force as fused and output at the current time step. The initial electromyographic feature weights are preset. The initial torque feature weights are preset. The sum of the initial electromyography feature weights and the initial torque feature weights is equal to one, and the values ​​of both the initial electromyography feature weights and the initial torque feature weights are between zero and one. A preset fatigue compensation coefficient, ranging from 0.1 to 0.5, is used to maintain the stability of the multi-source sensor fusion upper and lower limb coordinated rehabilitation training system. Simultaneously, the fatigue compensation coefficient is limited to be less than or equal to the initial electromyographic feature weights to avoid negative values ​​in weight allocation. This is a real-time muscle fatigue index. The electromyographic activation level is calculated at the current time step. This is the torque mapping value calculated at the current time step.

[0057] After completing the dynamic adaptive fusion of the comprehensive force intention features, the intention extraction module 300 performs first-order difference calculation on the comprehensive force intention features of adjacent time steps to separate the transient high-frequency components characterizing the start and stop of movement. The intention extraction module 300 introduces a nonlinear dead-zone activation function to truncate and filter the transient high-frequency components, filtering out low-amplitude high-frequency noise caused by hand muscle tremors and retaining the high-frequency abrupt components of the true hand force intention. The calculation expression of the high-frequency abrupt feature of force intention is as follows:

[0058] In the formula, This refers to the high-frequency abrupt change characteristic of the exertion intention calculated at the current time step. This represents the overall intention and characteristics of exerting force at the current time step. This retains the characteristics of the comprehensive effort intention from the previous time step. The high-frequency mutation determination dead zone threshold is preset. The value of the high-frequency mutation determination dead zone threshold is determined by the pre-calibration of the peak variance of the signal noise floor of a healthy subject in a resting state. In some specific embodiments, the value range of the high-frequency mutation determination dead zone threshold is usually set to three to ten percent of the maximum amplitude of the comprehensive force intention feature. The high-frequency mutation feature of the force intention is packaged in real time and output to the feedforward control module 400 as a cross-limb feedforward control source signal to trigger the generation of the lower limb target axis auxiliary moment.

[0059] Based on the high-frequency abrupt change characteristics of lower limb gait phase and force exertion intention, and combined with communication and processing delay time, feedforward trajectory prediction and calculation of lower limb target auxiliary torque can be further included in the specific implementation as sub-steps S401 to S403.

[0060] Sub-step S401 generates a basic reference trajectory and constructs smooth trigger coefficients for the feedforward control module 400. The feedforward control module 400 receives the lower limb gait phase output by the gait decoding module 200. Based on the lower limb gait phase, the feedforward control module 400 matches the expected joint angle and expected joint angular velocity corresponding to the lower limb gait phase in a pre-stored standardized healthy gait database. Considering the possible inaccurate alignment between discrete sampling points and gait phase nodes in the standardized healthy gait database, the feedforward control module 400 uses a linear interpolation algorithm to perform smooth interpolation calculations on the data nodes of adjacent gait phases, and uses the expected joint angle and expected joint angular velocity corresponding to the lower limb gait phase as the basic reference trajectory for the current time step. At the same time, the feedforward control module 400 receives the intent prompt. To prevent the high-frequency abrupt change in force application intention from being directly input into the underlying execution module 500, which could cause a step-impact overload in the flexible actuators of the lower limb rehabilitation exoskeleton robot, the feedforward control module 400 introduces a smooth mapping mechanism based on logic functions to convert the high-frequency abrupt change in force application intention into normalized smooth triggering coefficients. From a system dynamics perspective, the smooth mapping mechanism acts as a virtual buffer, transforming the discrete torque intention jump characteristics into a continuously differentiable, gently rising curve, thereby avoiding destructive spikes in the underlying motor current. The smooth triggering coefficients are used to dynamically adjust the time-domain step size of subsequent feedforward predictions and the target gain of impedance control. The specific calculation expression for the smooth triggering coefficients is as follows:

[0061] In the formula, The smooth triggering coefficient is calculated based on the current time synchronization. The preset smooth transition slope constant is set, and the value range of the smooth transition slope constant is set to two to five in some embodiments to balance the response agility and mechanical stability of the multi-source sensor fusion upper and lower limb coordinated rehabilitation training system. This represents a high-frequency mutation characteristic of the current time synchronization effort. The preset mutation activation center threshold is used to avoid the control threshold dead zone caused by feature baseline drift during long-term operation. The mutation activation center threshold is set to 50% of the historical peak value of the high-frequency mutation feature of the force exertion intention extracted within the preset sliding time window. The length of the preset sliding time window can be set to two to five seconds.

[0062] Sub-step S402 involves the feedforward control module 400 performing advanced trajectory prediction calculations to compensate for communication delays. The underlying hardware transmission and command processing of the multi-source sensor fusion upper and lower limb coordinated rehabilitation training system typically exhibits time lag. Directly controlling the system based on the current time step's reference trajectory can easily cause the lower limb rehabilitation exoskeleton robot's movements to lag behind the subject's actual movement intentions. The feedforward control module 400 retrieves the system's inherent communication and processing total delay time stored in the data synchronization module 100 during the initialization phase of the multi-source sensor fusion upper and lower limb coordinated rehabilitation training system, and dynamically adjusts the time prediction based on the smooth trigger coefficient. The step length is increased when the subject's hand exertion intention is strong to provide stronger cross-limb advance assistance. Based on the principle of cross-limb collaborative control, the future movement state of the lower limb is predicted in advance by utilizing the hand exertion intention, so that the lower limb rehabilitation exoskeleton robot enters the force preparation state in advance, forming a rehabilitation traction effect of the upper limb leading the lower limb. The feedforward control module 400 performs kinematic advance calculation on the basic reference trajectory based on the Taylor series expansion principle, thereby obtaining the advance expected joint angle and advance expected joint angular velocity after compensating for the time lag phenomenon. The specific calculation expression of the advance trajectory prediction after compensating for the time lag phenomenon is as follows:

[0063]

[0064]

[0065] In the formula, The time prediction step size is dynamically adjusted for the current time step. This is the total inherent communication and processing delay of the system. The maximum feedforward time compensation gain is preset. This maximum feedforward time compensation gain is used to limit the risk of kinematic divergence caused by excessively large prediction step size. The value range of the maximum feedforward time compensation gain can be determined experimentally to be 0.05 seconds to 0.2 seconds. The smoothing trigger coefficient for the current time step. To predict the expected joint angle, The base expected joint angle for the current time step. The base expected joint angular velocity for the current time step. The base expected joint angle acceleration for the current time step. To predict the expected joint angular velocity.

[0066] Sub-step S403 involves the feedforward control module 400 performing dynamic adaptive impedance control target torque calculation. The feedforward control module 400 retrieves lower limb surface electromyography (EMG) signals from the shared memory of the central real-time control unit and calculates the root mean square (RMS) value of the EMG signals to reflect the degree of lower limb muscle activation. Based on the degree of lower limb muscle activation, the feedforward control module 400 generates basic stiffness and basic damping by linearly mapping the values. Specifically, the mapping mechanism involves multiplying the preset stiffness and damping ratios by the lower limb muscle activation degree to calculate the basic stiffness and basic damping. The basic stiffness and basic damping are used to maintain the basic compliance support required for rehabilitation training. Simultaneously, the feedforward control module 400 uses a smoothing trigger coefficient as an additional control gain to induce a strong force exertion in the subject's hand. The system dynamically superimposes additional stiffness and damping to achieve a collaborative rehabilitation logic where the subject's hand exerts force to drive lower limb movements. Based on the dynamic principle of impedance control, the dynamically superimposed additional stiffness makes the lower limb rehabilitation exoskeleton robot exhibit stronger rigidity in tracking the target trajectory direction. This actively assists subjects with weak lower limb muscle strength to overcome the resistance dead point during gait transitions. The feedforward control module 400 calculates the actual lower limb joint angle and angular velocity deviation of the lower limb rehabilitation exoskeleton robot based on the expected joint angle and expected joint angular velocity. Combined with the comprehensive stiffness and comprehensive damping dynamically synthesized by the feedforward control module 400, the impedance equation is solved in a closed loop to calculate the target auxiliary torque of the lower limb. The specific calculation expression for the target auxiliary torque of the lower limb is as follows:

[0067]

[0068]

[0069] In the formula, The overall stiffness synthesized in the current time synchronization. The basic stiffness is generated based on the degree of lower limb muscle activation. The preset additional stiffness peak value is determined offline by the rehabilitation physician based on the subject's degree of motor impairment. The additional stiffness coefficient for the current time. The overall damping is synthesized synchronously at the current time. The base damping is generated based on the degree of lower limb muscle activation. The preset additional damping peak value is set to a value that is in a critical damping matching state with the additional stiffness peak value in order to suppress high-frequency joint oscillations of the lower limb rehabilitation exoskeleton robot. The lower limb target auxiliary torque is calculated and output at the current time step. To predict the expected joint angle, This represents the actual joint angles of the lower limb rehabilitation exoskeleton robot at the current time step. To predict the expected joint angular velocity, The actual joint angular velocity of the lower limb rehabilitation exoskeleton robot is obtained at the current time step, and the lower limb target auxiliary torque is output to the bottom execution module 500 in real time to drive the flexible actuator configured on the lower limb rehabilitation exoskeleton robot to perform the corresponding physical movement.

[0070] See attached document Figure 4 The process involves obtaining spring deformation data from the flexible actuator to estimate the lumped disturbance torque, filtering and smoothing the auxiliary torque of the lower limb target, and generating a motor torque command by combining the lumped disturbance torque and outputting it to the flexible actuator for execution. In specific implementations, this process may further include sub-steps S501 to S502.

[0071] Sub-step S501 involves the bottom-level execution module 500 acquiring the spring deformation displacement of the flexible actuator and constructing a discretized dynamics observer to dynamically estimate the lumped disturbance torque. In the specific implementation, the bottom-level execution module 500 extracts the spring deformation displacement measured by the displacement sensor from the shared memory of the central real-time control unit. Considering the physical misalignment and frictional wear phenomena in the mechanical transmission mechanism of the lower limb rehabilitation exoskeleton robot during long-term operation, the physical degradation of the mechanical transmission mechanism combined with external environmental resistance can easily generate interference torque that hinders the precise movement of the bottom-level motor. To eliminate the influence of interference torque on the cross-limb feedforward control accuracy, the bottom-level execution module 500 calculates the elastic transmission torque based on the spring deformation displacement and the pre-calibrated spring stiffness coefficient. The bottom-level execution module 500 then combines the elastic transmission torque with the actual motor angular velocity collected by the rotary encoder configured on the lower limb rehabilitation exoskeleton robot and the actual... A discretized dynamics observer is constructed for the motor angular acceleration. To prevent high-frequency noise amplification caused by direct differential of the rotary encoder signal from leading to divergence of the discretized dynamics observer, the underlying execution module 500 uses a Butterworth low-pass filter to filter the actual motor angular velocity and then differentiates it to obtain a smooth actual motor angular acceleration. The discretized dynamics observer derives the lumped disturbance torque by calculating the residual between the output of the ideal dynamics model and the actual physical state of the underlying motor. The lumped disturbance torque includes nonlinear friction components and parameter perturbation components that are not accurately modeled. From a physical principle perspective, the introduction of the discretized dynamics observer equates the unknown external environmental resistance and internal friction to a lumped disturbance term, enabling the subsequent control law to counteract the lumped disturbance term in reverse, thereby maintaining the dynamic compliance of the flexible actuator output. The specific calculation expressions for the lumped disturbance torque and the elastic transmission torque are as follows:

[0072] In the formula, The elastic transmission torque is calculated synchronously at the current time. The preset spring stiffness coefficient for the built-in spring of the flexible actuator. The spring deformation displacement is measured by the displacement sensor at the current time step. Estimate the lumped disturbance torque output step by step for the current time. The preset observer filter coefficients are set, and the range of the observer filter coefficients is set to 0.8 to 0.9 to ensure the convergence stability of the discretized dynamics observer. This is the lumped disturbance torque retained from the previous time step. The equivalent moment of inertia of the motor is obtained through offline frequency sweep experiment parameters. The actual motor angular acceleration obtained at the current time synchronization. The equivalent viscous friction coefficient of the motor is obtained through offline frequency sweep experiment parameters. The actual motor angular velocity obtained at the current time synchronization. This is the motor torque command output by the underlying execution module 500 in the previous time step.

[0073] Sub-step S502 involves the bottom-level execution module 500 performing differential smoothing processing on the lower limb target auxiliary torque and combining it with the lumped disturbance torque to synthesize the final motor torque command. The bottom-level execution module 500 receives the lower limb target auxiliary torque output by the feedforward control module 400. Addressing the technical issue that multi-source heterogeneous signals often carry high-frequency noise during processing, this high-frequency noise easily causes high-frequency jitter in the lower limb target auxiliary torque between adjacent time steps. This high-frequency jitter directly affects the hardware of the lower limb rehabilitation exoskeleton robot, easily triggering mechanical resonance in the flexible actuator and shortening the lower limb rehabilitation timeline. To determine the physical lifespan of the skeletal robot, the underlying execution module 500 introduces a nonlinear tracking differentiator based on kinematic constraints to perform first-order low-pass filtering and amplitude-limiting smoothing on the lower limb target auxiliary torque. The nonlinear tracking differentiator can extract the smoothed lower limb auxiliary torque after removing jitter components under low-latency constraints. The underlying execution module 500 uses the smoothed lower limb auxiliary torque as the ideal output torque of the flexible actuator. The underlying execution module 500 subtracts the elastic transmission torque from the smoothed lower limb auxiliary torque to constitute the torque tracking error. The underlying execution module 500 combines the torque tracking error with proportional-integral-differential (PID)... The control law generates basic feedback adjustment commands. Simultaneously, the underlying execution module 500 introduces a dynamic anti-integral saturation limiting algorithm in the integral term calculation of the proportional-integral-derivative control law, setting an integral saturation torque threshold. When the cumulative sum of torque tracking errors over historical time steps multiplied by the integral control gain exceeds the integral saturation torque threshold, the accumulation calculation of the integral term is immediately stopped to prevent the risk of uncontrolled overload of the underlying motor output commands when the patient's lower limbs experience spasticity and resistance. The underlying execution module 500 performs feedforward superposition compensation of the basic feedback adjustment commands and the lumped disturbance torque to synthesize the final motor torque command. At the control execution principle level, the proportional-integral-derivative control law provides basic error feedback tracking capability, while the feedforward superposition of the lumped disturbance torque achieves proactive cancellation of external physical environment and internal transmission mechanism disturbances. The feedforward superposition mechanism of the basic feedback adjustment commands and the lumped disturbance torque effectively balances the dynamic response speed and disturbance rejection steady-state accuracy of the system control commands. The motor torque command is output in real time to the flexible actuator of the lower limb rehabilitation exoskeleton robot to drive the flexible actuator to perform corresponding rehabilitation assistive actions. The specific calculation expression for the motor torque command synthesis process is as follows:

[0074]

[0075]

[0076] In the formula, The lower limb smoothing auxiliary torque is the output after filtering by the nonlinear tracking differentiator at the current time step. The lower limb smoothing assist torque retained from the previous time step, The preset discrete time step size for the multi-source sensor fusion upper and lower limb coordinated rehabilitation training system is set to a value between one millisecond and ten milliseconds. The optimal control synthesis function is preset, and its specific implementation can be based on existing active disturbance rejection control theory. The lower limb target auxiliary torque is output at 400 in the feedforward control module at the current time step. The speed tracking factor is set to a value between ten and fifty for the nonlinear tracking differentiator to balance the smoothness of the filter and the speed of signal following. The acceleration limiting threshold is set for the nonlinear tracking differentiator, and the value of the acceleration limiting threshold is determined by the maximum physical acceleration calibration parameters of the motor configured with the flexible actuator. This represents the torque tracking error calculated at the current time step. The elastic transmission torque calculated at the current time step. The motor torque command synthesized at the current time step is output to the flexible driver. To control the gain according to the pre-defined proportional control method of Ziegler-Nichols's law, To control the gain based on the integral control preset according to the Ziegler-Nichols rule, This represents the cumulative sum of torque tracking errors over historical time steps after processing by a dynamic anti-integral saturation limiting algorithm. The differential control gain is preset according to the Ziegler-Nichols rule. This represents the torque tracking error from the previous time step. The lumped disturbance torque is estimated by the discretized dynamics observer at the current time step.

[0077] In practice, acquiring electromyographic signals from the surface of the hand and calculating the real-time muscle fatigue index may further include sub-steps S311 to S313.

[0078] Sub-step S311 involves the intent extraction module 300 acquiring electromyographic (EMG) signals from the hand surface and constructing a dynamically overlapping sliding time window. In practice, the intent extraction module 300 retrieves the real-time EMG signals from the hand surface sensors configured on the hand function rehabilitation exoskeleton robot from the shared memory of the central real-time control unit. Addressing the technical issue of hardware thermal noise and power grid interference easily contaminating the raw physiological signals, the intent extraction module 300 uses a bandpass filter combined with a notch filter to clean the hand surface EMG signals. The effective cutoff frequency range of the bandpass filter is set to 10 Hz to 500 Hz to extract the core energy frequency band of the EMG signal. The notch filter... The heart frequency is set to 50 Hz or 60 Hz to eliminate power frequency interference generated by the local power grid. For bandpass filtering and power frequency notch processing of the electromyography (EMG) signal on the hand surface, it can be performed according to existing digital signal processing theory. The intent extraction module 300 truncates the cleaned EMG signal on the hand surface into multiple continuous data sequences at a fixed time step. The intent extraction module 300 introduces a sliding time window mechanism to encapsulate the continuous data sequences. In order to reduce system processing delay while ensuring frequency resolution, the time span of the sliding time window is set to 200 milliseconds to 500 milliseconds, and a 50% data overlap rate is maintained between adjacent sliding time windows to ensure the continuity of spectral feature extraction.

[0079] Sub-step S312 involves the intent extraction module 300 performing frequency domain transformation on the electromyography (EMG) signals of the hand surface within the sliding time window and calculating the median frequency of the current time step. The intent extraction module 300 extracts the data sequence sequentially according to the sliding time window and performs a Fast Fourier Transform (FFT). For the time-domain to frequency-domain transformation calculation of the hand surface EMG signals, existing FFT algorithms can be used. The intent extraction module 300 calculates the power spectral density distribution of the hand surface EMG signals after frequency transformation. Based on the principles of human muscle electrophysiology, with the increase in rehabilitation training time and the intensification of hand muscle contraction, metabolic products such as lactic acid within the hand muscle fibers increase. The accumulation of substances leads to a decrease in the speed of action potential conduction along muscle fibers. Macroscopically, this change in the electrophysiological state of muscle fibers manifests as a gradual weakening of high-frequency components and a gradual increase in low-frequency components in the electromyographic signal spectrum distribution on the hand surface. To quantify the leftward shift and attenuation of the spectrum caused by this change in the electrophysiological state of muscle fibers, the intention extraction module 300 performs numerical integration along the frequency coordinate axis on the power spectral density distribution. The intention extraction module 300 searches for the critical frequency point that can precisely divide the total power spectral density area into two equal halves, and defines the critical frequency point as the median frequency of the current time step. The specific expression for solving the median frequency equation is as follows:

[0080] In the formula, The median frequency is the value obtained at the current time step. The effective frequency band upper limit is set for the intention extraction module 300, and the value of the effective frequency band upper limit is usually set to 10 Hz to filter out baseline drift interference. The effective frequency band upper limit is set for the intent extraction module 300, and the value of the effective frequency band upper limit is set to half of the sampling frequency according to the Nyquist sampling theorem. This represents the power spectral density of the electromyography (EMG) signal on the hand surface at a specified frequency within the sliding time window at the current time step. Let be the frequency variable of the continuous integral.

[0081] Sub-step S313 involves the intent extraction module 300 calculating the initial fatigue quantization value based on the benchmark median frequency and performing smoothing and amplitude limiting processing to output the final real-time muscle fatigue index. The intent extraction module 300 retrieves the benchmark median frequency recorded by the multi-source sensor fusion upper and lower limb coordinated rehabilitation training system during the initial unforced initiation phase. The benchmark median frequency represents the healthy spectral characteristics of the subject's hand muscles in a completely relaxed state without fatigue accumulation. The intent extraction module 300 subtracts the median frequency of the current time step from the benchmark median frequency to obtain the frequency offset attenuation. The intent extraction module 300 divides the frequency offset attenuation by the benchmark median frequency to obtain the normalized initial fatigue quantization value. To address the interference problem caused by poor electrode contact or sudden arm tremors in the subject, the intent extraction module 300 uses an exponential moving average algorithm to perform time-domain smoothing filtering on the initial fatigue quantization value to filter out high-frequency non-physiological jump noise, generating a smooth fatigue quantization value. To prevent… To prevent the initial muscle relaxation of the subject from causing a negative dead zone in the subsequently calculated smoothed fatigue quantization value, or to prevent the smoothed fatigue quantization value from exceeding one due to sensor data distortion, thus triggering divergence in the downstream control module, the intent extraction module 300 introduces saturation limiting logic to restrict the smoothed fatigue quantization value to the real number boundary range of zero to one. Specifically, when the smoothed fatigue quantization value is less than zero, the intent extraction module 300 forces the final real-time muscle fatigue index to zero; when the smoothed fatigue quantization value is greater than one, the intent extraction module 300 forces the final real-time muscle fatigue index to one; and when the smoothed fatigue quantization value is between zero and one, the intent extraction module 300 assigns the final real-time muscle fatigue index to the smoothed fatigue quantization value. The real-time muscle fatigue index generated after smoothing and limiting is written into the shared memory of the central real-time control unit for subsequent dynamic weight adjustment. The specific calculation expression for the real-time muscle fatigue index is as follows:

[0082]

[0083] In the formula, The initial fatigue quantization value calculated at the current time step. The baseline median frequency for the initial recording of a multi-source sensor fusion upper and lower limb coordinated rehabilitation training system. The median frequency obtained at the current time step. The smoothed fatigue quantization value generated for the current time step and used for amplitude limiting processing. The preset smoothing weight coefficient is set for the exponential moving average algorithm, and the value range of the preset smoothing weight coefficient is set to 0.05 to 0.2 to balance smoothness and dynamic tracking response speed. The real-time muscle fatigue index is retained from the previous time step.

[0084] The calculation of high-frequency abrupt change characteristics of force exertion intention based on dynamic adjustment of weights according to muscle fatigue index and combined with hand torque data can be further included in sub-steps S321 to S323 during specific implementation.

[0085] Sub-step S321 involves the intent extraction module 300 acquiring the electromyographic (EMG) signal of the hand surface and the human-computer interaction force, and performing feature normalization processing. The intent extraction module 300 extracts the EMG signal of the hand surface from the shared memory of the central real-time control unit. The calculation of the root mean square (RMS) value of the EMG signal of the hand surface can be performed using existing time-series statistical algorithms. The intent extraction module 300 retrieves the maximum RMS value generated by the subject's hand muscles under maximum voluntary contraction state, recorded during the offline calibration phase of the multi-source sensor fusion upper and lower limb coordinated rehabilitation training system. The maximum RMS value generated by the subject's hand muscles under maximum voluntary contraction state is obtained by measuring the peak RMS value of the hand surface EMG signal by instructing the subject to clench their fist and contract their hand muscles as much as possible before the rehabilitation training begins, holding the contraction for three to five seconds. The intent extraction module 300 divides the RMS value of the hand surface EMG signal calculated at the current time step by the maximum RMS value generated by the subject's hand muscles under maximum voluntary contraction state, recorded during the offline calibration phase of the multi-source sensor fusion upper and lower limb coordinated rehabilitation training system, to obtain the dimensionless value. The intent extraction module 300 simultaneously extracts the human-computer interaction force measured by the torque sensor configured on the hand function rehabilitation exoskeleton robot from the shared memory of the central real-time control unit. The intent extraction module 300 extracts the absolute value of the human-computer interaction force and divides it by the maximum torque threshold recorded in the offline calibration phase to obtain a dimensionless torque mapping value. The maximum torque threshold recorded in the offline calibration phase is obtained by instructing the subject to apply the maximum equal-length resistance to the arc-groove transmission mechanism of the hand function rehabilitation exoskeleton robot. At the same time, the intent extraction module 300 sets a minimum effective value safety bottom line to ensure that the maximum root mean square value generated by the upper and lower limb coordinated rehabilitation training system with multi-source sensor fusion under the maximum voluntary contraction state of the subject's hand muscles recorded in the offline calibration phase and the maximum torque threshold recorded in the offline calibration phase are both strictly greater than zero, thereby avoiding the low-level control dead zone of computer division-by-zero operation overflow. The intent extraction module 300 limits the value range of both the electromyographic activation and the torque mapping value to between zero and one through a boundary limiting function to unify the numerical scale of subsequent dynamic feature fusion.

[0086] Sub-step S322 involves the intention extraction module 300 constructing a dynamic weight allocation mechanism based on the real-time muscle fatigue index and calculating comprehensive force exertion intention features. Based on the principles of human muscle electrophysiology, when hand muscles enter a fatigued state, the central nervous system recruits more motor units to maintain the same external output torque. This is macroscopically manifested as a compensatory increase in the amplitude of the electromyographic signal on the hand surface. If a fixed proportion of electromyographic features is relied upon in cross-limb control, it is easy to cause the feedforward control module 400 to over-output, thereby increasing the risk of excessive traction on the subject's lower limbs. The intention extraction module 300 extracts the calculated real-time muscle fatigue index. The intention extraction module 300 establishes dynamic electromyographic weights that decrease linearly with the real-time muscle fatigue index and correspondingly linearly increasing dynamic torque weights. The intention extraction module 300 sets initial electromyographic feature weights and initial torque feature weights and initially... The initial electromyography (EMG) feature weight plus the initial torque feature weight equals one. The intent extraction module 300 sets a fatigue compensation coefficient, and the value of the fatigue compensation coefficient is less than or equal to the initial EMG feature weight to prevent the dynamic EMG weight from becoming negative and causing a dead zone in the control logic. The intent extraction module 300 subtracts the product of the real-time muscle fatigue index and the fatigue compensation coefficient from the initial EMG feature weight to obtain the dynamic EMG weight of the current time step. The intent extraction module 300 adds the product of the real-time muscle fatigue index and the fatigue compensation coefficient to the initial torque feature weight to obtain the dynamic torque weight of the current time step. The intent extraction module 300 multiplies the dynamic EMG weight of the current time step by the dimensionless EMG activation degree and adds the product of the dynamic torque weight of the current time step and the dimensionless torque mapping value, and then fuses and calculates to obtain the comprehensive force exertion intent feature. The specific calculation expression of the comprehensive force exertion intent feature is as follows:

[0087] In the formula, This represents the overall intent and characteristics of the current time synchronization and fusion output. The initial electromyographic feature weights are preset for the intention extraction module 300. The initial torque feature weights are preset for the intention extraction module 300. The fatigue compensation coefficient is preset in the intention extraction module 300, and the value range of the fatigue compensation coefficient is set to 0.1 to 0.5 to avoid control system divergence caused by excessive fatigue compensation amplitude in a single operation. This represents the real-time muscle fatigue index received at the current time. The electromyographic activation level is calculated synchronously at the current time. The torque mapping value is calculated synchronously at the current time.

[0088] Sub-step S323 involves the intention extraction module 300 performing first-order difference and dead-zone filtering on the comprehensive force exertion intention features and outputting high-frequency abrupt change features of the force exertion intention. The intention extraction module 300 extracts the comprehensive force exertion intention features of the current time step and the comprehensive force exertion intention features retained from the previous time step, and performs a difference calculation to obtain a first-order difference component used to characterize the transient rate of change of the subject's force exertion movement. Addressing the issue of low-amplitude, high-frequency physiological tremors in the subject's hand muscles due to insufficient strength during rehabilitation, which easily lead to high-frequency noise unrelated to the true force exertion intention mixed into the first-order difference component, the intention extraction module 300 introduces a nonlinear dead-zone activation function to truncate and filter the first-order difference component. When the absolute value of the first-order difference component is less than or equal to a predetermined value, the filter is applied. When the high-frequency mutation determination dead zone threshold is set, the intent extraction module 300 forces the high-frequency mutation feature quantity of the force intention to be zero to filter out static background noise. When the first-order difference component is greater than the high-frequency mutation determination dead zone threshold, the intent extraction module 300 uses the difference between the first-order difference component and the high-frequency mutation determination dead zone threshold as the high-frequency mutation feature quantity of the force intention to extract the true mutation intent. When the first-order difference component is less than the negative value of the high-frequency mutation determination dead zone threshold, it indicates that the subject is in a state of rapid force withdrawal and does not generate positive stimulation for cross-limb control. In the state where the subject is in a state of rapid force withdrawal and does not generate positive stimulation for cross-limb control, the intent extraction module 300 also forces the high-frequency mutation feature quantity of the force intention to be zero. The specific calculation expression of the high-frequency mutation feature quantity of the force intention is as follows:

[0089] In the formula, This refers to the high-frequency abrupt change characteristic of the exertion intention calculated at the current time step. The overall force exertion intention characteristics are calculated at the current time step. This retains the characteristics of the comprehensive effort intention from the previous time step. The high-frequency mutation determination dead zone threshold is preset, and the value of the high-frequency mutation determination dead zone threshold is determined by the historical cortical peak signal-to-noise ratio measured by healthy subjects in a completely resting state. After the high-frequency mutation characteristic quantity is calculated, it is packaged and written into the shared memory of the central real-time control unit for use by the feedforward control module 400.

[0090] The feedforward control module 400 generates a basic reference trajectory and constructs a smooth triggering coefficient. In specific implementation, this may be further included in sub-steps S411 to S412.

[0091] Sub-step S411 involves the feedforward control module 400 receiving the lower limb gait phase and generating a basic reference trajectory based on a standardized healthy gait database. The feedforward control module 400 retrieves the lower limb gait phase output in real-time by the gait decoding module 200 from the shared memory of the central real-time control unit at the current time step. The feedforward control module 400 inputs the lower limb gait phase as a query index parameter into the pre-built standardized healthy gait database within itself. In some specific embodiments, the standardized healthy gait database is generated by a multi-source sensor fusion upper and lower limb coordinated rehabilitation training system during the offline configuration phase, collecting multi-joint three-dimensional kinematic data from multiple healthy subjects walking at a constant speed on a standard rehabilitation training device, and then performing a five-fold polynomial smoothing fitting process. The feedforward control module 400 retrieves the corresponding discrete kinematic node data from the standardized healthy gait database based on the lower limb gait phase and then... From a principle perspective, since the sampling frequency of actual control systems is usually greater than the partition density of discrete kinematic node data in a standardized healthy gait database, directly using discrete kinematic node data has a high probability of causing numerical jumps in the generated control reference trajectory between adjacent nodes, which in turn leads to mechanical vibration of the lower limb rehabilitation exoskeleton robot. The feedforward control module 400 uses a cubic spline interpolation algorithm to perform smooth transition calculations on adjacent discrete kinematic node data. The cubic spline interpolation algorithm can ensure that the piecewise polynomial has the same first and second derivatives at the connection points of adjacent intervals, thereby generating a basic expected joint angle, basic expected joint angular velocity, and basic expected joint angular acceleration with second-order continuous differentiability at the current time step. The basic expected joint angle, basic expected joint angular velocity, and basic expected joint angular acceleration together constitute the basic reference trajectory used to guide the movement of the lower limb rehabilitation exoskeleton robot.

[0092] Sub-step S412 involves the feedforward control module 400 receiving the high-frequency abrupt change characteristic of the force application intention and constructing a smooth trigger coefficient based on a nonlinear logic function. The feedforward control module 400 simultaneously extracts the high-frequency abrupt change characteristic of the force application intention calculated and output by the intention extraction module 300 at the current time step. If the high-frequency abrupt change characteristic of the force application intention is directly used as the hardware trigger signal for cross-limb traction and input to the underlying execution module 500, the discrete intention step jump can easily cause spike currents in the internal motors of the flexible actuators configured in the lower limb rehabilitation exoskeleton robot and increase the risk of impact damage to the mechanical transmission mechanism. Therefore, the feedforward control module 400 introduces a nonlinear smooth mapping mechanism based on logic functions to establish a transition mapping boundary in the control command dimension. Those skilled in the art can understand that the nonlinear smooth mapping mechanism utilizes the asymptotic saturation characteristics of logic functions to transform abrupt step signals into gently rising continuous variables. The feedforward control module 400 converts the high-frequency abrupt change characteristic of the force application intention into continuously changing smooth trigger coefficients. The feedforward control module 400 sets a smooth transition slope constant to adjust the geometric slope of the smooth mapping curve. Simultaneously, the feedforward control module 400 constructs a sliding time window with a preset length of two to five seconds and records the historical peak value of the high-frequency mutation characteristic of the force intention within the sliding time window in real time. The feedforward control module 400 dynamically sets the mutation activation center threshold to 50% of the historical peak value of the high-frequency mutation characteristic of the force intention to avoid trigger dead zones caused by baseline drift of the subject's force application during long-term rehabilitation training. At the same time, the feedforward control module 400 sets a minimum anti-false trigger boundary limit for the mutation activation center threshold. The specific value of the minimum anti-false trigger boundary limit is determined by the maximum amplitude of the background noise measured and calibrated by the multi-source sensor fusion upper and lower limb coordinated rehabilitation training system in the initial resting state. When 50% of the historical peak value of the high-frequency mutation characteristic of the force intention is less than the minimum anti-false trigger boundary limit, the feedforward control module 400 forcibly assigns the mutation activation center threshold to the minimum anti-false trigger boundary limit to suppress control drift caused by random background noise in the environment. The feedforward control module 400 calculates the smooth trigger coefficient of the current time step based on the nonlinear smooth mapping mechanism. The specific calculation expression of the smooth trigger coefficient is as follows:

[0093] In the formula, The smoothing trigger coefficient is calculated for the current time step, and its value is restricted to between zero and one by the boundary function. It is a natural constant. The feedforward control module 400 is preset with a smooth transition slope constant, and the value range of the smooth transition slope constant is set to two to five in some embodiments to balance the response agility of cross-limb triggering and the mechanical stability of motor execution. This represents the high-frequency mutation characteristic of the exertion intent received at the current time step. The mutation activation center threshold is dynamically updated by the feedforward control module 400 based on historical data. The smoothing trigger coefficient is retained by the feedforward control module 400 after the calculation is completed for use in subsequent control algorithms to dynamically adjust the time prediction step size and the target gain of impedance control.

[0094] Combining the delay time and the smoothing trigger coefficient, the kinematic expansion formula is used to calculate the future advanced trajectory position and velocity. In specific implementation, this can be further included in sub-steps S421 to S422.

[0095] Sub-step S421 involves the feedforward control module 400 acquiring the smooth trigger coefficient and the system's inherent communication and processing total delay time, and calculating the dynamic time prediction step size. The feedforward control module 400 retrieves the system's inherent communication and processing total delay time, calibrated and stored by the data synchronization module 100 during the initialization phase, from the shared memory of the central real-time control unit. This inherent communication and processing total delay time represents the physical time required for the sensor-acquired signals to be transmitted via the hardware bus to the central real-time control unit for data processing and ultimately for issuing control commands to the lower-level execution module 500. Without time feedforward compensation for the system's inherent communication and processing total delay time, the mechanically assisted movements of the lower limb rehabilitation exoskeleton robot may lag behind the subject's actual human effort intentions, leading to human-machine motion interference resistance. The feedforward control module 400 synchronously extracts the smooth trigger coefficient generated at the current time step. Based on the system's inherent communication and processing total delay time and the smooth trigger coefficient, the feedforward control module 400 constructs a dynamic feedforward time compensation equation. In specific implementation scenarios of cross-limb collaborative control, the dynamic feedforward time compensation equation is used to provide two-layer time advance compensation. The mechanism utilizes the inherent communication and processing latency of the system to offset the inherent physical lag in data transmission and instruction calculation in the hardware. Simultaneously, it uses a smooth trigger coefficient corresponding to the subject's hand exertion intention, superimposed with the cross-limb pre-traction time, allowing the lower limb rehabilitation exoskeleton robot to enter an assisted following state before the subject's actual lower limb exertion. The feedforward control module 400 sets a maximum feedforward time compensation gain to limit the maximum time boundary of the pre-calculation. The specific value of the maximum feedforward time compensation gain is determined offline based on the natural switching frequency of the human gait cycle, and the range of the maximum feedforward time compensation gain is set to 0.05 seconds to 0.2 seconds to prevent trajectory divergence risk caused by excessively large time calculation spans. The feedforward control module 400 adds the product of the maximum feedforward time compensation gain and the smooth trigger coefficient to obtain the dynamic time prediction step size. The dynamic calculation mechanism of the dynamic time prediction step size can increase the pre-calculation span when the subject's hand exertion intention is strong, thus providing a more agile cross-limb initiation traction assistance function. The specific calculation expression of the dynamic time prediction step size is as follows:

[0096] In the formula, The dynamic time prediction step size is dynamically calculated based on the current time synchronization. The system's inherent total communication and processing latency is pre-calibrated and stored for the data synchronization module 100. The maximum feedforward time compensation gain is preset for the feedforward control module 400. The smooth trigger coefficients are received synchronously at the current time and used for calculation.

[0097] Sub-step S422 involves the feedforward control module 400 performing kinematic advance calculations based on the dynamically predicted step size and outputting the expected joint angles and angular velocities. The feedforward control module 400 acquires the base expected joint angles, angular velocities, and angular accelerations generated at the current time step. Since the human lower limb gait trajectory can be approximated as a continuous and smooth physical motion process within a millisecond-level extremely short prediction time window, by introducing first-order velocity integrals and second-order acceleration integrals, the future kinematic node positions can be estimated with low latency without relying on complex dynamic model calculations. For approximate extrapolation calculations of nonlinear trajectories within a short time window, existing Taylor series expansion algorithms can be used. Module 400 uses the second-order Taylor series expansion principle to perform a physical time projection mapping of the lower limb movement state at the current time step to the predicted future time. The feedforward control module 400 calculates the advanced expected joint angle by adding the product of the basic expected joint angular velocity and the dynamic time prediction step length to the basic expected joint angle, and then adding half of the product of the basic expected joint angular acceleration and the square of the dynamic time prediction step length. Simultaneously, the feedforward control module 400 calculates the advanced expected joint angular velocity by adding the product of the basic expected joint angular acceleration and the dynamic time prediction step length to the basic expected joint angular velocity. To prevent the mathematically theoretical trajectory generated by the time projection calculation from exceeding the actual physiological activity limits of the subject's joints, the feedforward control module 400 pre-calculates the lower limb... The hip, knee, and ankle joints of the rehabilitation exoskeleton robot are each configured with upper and lower limits for human physiological limit collision avoidance angles. The specific values ​​of these upper and lower limits are entered and set by the rehabilitation physician during the offline configuration phase of the multi-source sensor fusion upper and lower limb coordinated rehabilitation training system based on the subject's passive joint range of motion measurements. This constitutes a safety envelope constraint for hardware execution. The feedforward control module 400 sets the expected joint angle and the corresponding upper and lower limb physiological limit collision avoidance angles for specific joints of the lower limb rehabilitation exoskeleton robot. The lower limit of the collision avoidance angle is calculated using boundary comparison. When the expected joint angle is greater than the upper limit of the human physiological limit collision avoidance angle, the feedforward control module 400 forcibly truncates the expected joint angle and assigns it the value of the upper limit of the human physiological limit collision avoidance angle. When the expected joint angle is less than the lower limit of the human physiological limit collision avoidance angle, the feedforward control module 400 forcibly limits the expected joint angle and assigns it the value of the lower limit of the human physiological limit collision avoidance angle. This is to ensure that the subject can carry out rehabilitation training within a controllable range of activity as much as possible. At the same time, to avoid the transient power overload dead zone of the motor configured in the underlying execution module 500 caused by excessive predicted angular velocity, the feedforward control module 400 simultaneously extracts the maximum safe angular velocity threshold of the motor of the lower limb rehabilitation exoskeleton robot.When the absolute value of the expected joint angular velocity exceeds the maximum safe angular velocity threshold of the motor, the feedforward control module 400 proportionally truncates the amplitude of the expected joint angular velocity to the maximum safe angular velocity threshold of the motor to maintain the sign of the motion direction. After the anti-exceeding limit amplitude processing and speed limiting processing, the expected joint angle and expected joint angular velocity are stored in real time in the shared memory of the central real-time control unit by the feedforward control module 400 for subsequent error closed-loop calculation of the target torque for dynamic adaptive impedance control. The specific calculation expression for the kinematic advance calculation is as follows:

[0098] In the formula, This is the expected joint angle derived from the current time step and after physiological limiting processing. The base expected joint angle is input at the current time step. The base desired joint angular velocity is input at the current time step. The dynamic time prediction step size calculated by the feedforward control module 400. The base desired joint angular acceleration is input at the current time step. This is the expected joint angular velocity derived from the current time step and after amplitude limiting.

[0099] The implementation of dynamic adaptive impedance control target torque calculation includes sub-steps S431 to S433.

[0100] Sub-step S431 involves the feedforward control module 400 retrieving real-time lower limb surface electromyography (EMG) signals from the shared memory of the lower limb rehabilitation exoskeleton robot configured with the central real-time control unit, extracting the lower limb muscle activation level, and performing RMS value calculation and normalization processing on the lower limb surface EMG signals to obtain the lower limb muscle activation level. Specifically, this involves calculating the RMS value of the lower limb surface EMG signals by setting a time sliding window, dividing the calculated RMS value by the calibrated RMS value corresponding to the maximum voluntary contractile force of the lower limb muscles to obtain the lower limb muscle activation level, which ranges from zero to one. The feedforward control module 400 constructs a linear mapping relationship based on the lower limb muscle activation level to generate basic stiffness and basic damping. Specifically, the feedforward control module 400 presets stiffness and damping proportional coefficients; the specific values ​​of these coefficients are determined by the technical personnel. The staff pre-calibrates the lower limb rehabilitation exoskeleton robot based on its hardware mass and the rated parameters of the joint motors. The feedforward control module 400 multiplies the stiffness proportional coefficient with the degree of lower limb muscle activation to obtain the basic stiffness. At the same time, the feedforward control module 400 multiplies the damping proportional coefficient with the degree of lower limb muscle activation to obtain the basic damping. The basic stiffness and basic damping constitute the underlying compliant support parameters of the multi-source sensor fusion upper and lower limb collaborative rehabilitation training system in steady-state following state. The above dynamic impedance setting method is physically equivalent to connecting a virtual spring damper with stiffness and damping that changes in real time with the muscle exertion state at the joint of the lower limb rehabilitation exoskeleton robot. This allows the lower limb rehabilitation exoskeleton robot to provide only the minimum safe impedance required to maintain basic standing or basic gait walking when the subject has no obvious intention to move. This reduces the joint rigidity of the lower limb rehabilitation exoskeleton robot and avoids disuse atrophy of the lower limb muscle groups due to over-assistance.

[0101] Sub-step S432 involves the feedforward control module 400 acquiring smooth triggering coefficients and dynamically synthesizing integrated stiffness and integrated damping. The feedforward control module 400 synchronously retrieves the smooth triggering coefficients calculated and retained in the current time step. The range of the smooth triggering coefficients is strictly constrained to between zero and one to ensure the continuity and boundedness of the impedance control algorithm output and prevent the algorithm from falling into a logic dead zone. The feedforward control module 400 pre-sets the additional stiffness peak value and the additional damping peak value. The additional stiffness peak value is configured and entered offline by the rehabilitation physician based on the subject's specific movement disorder rating. The additional damping peak value is set to a specific value that is in a critical damping matching state with the additional stiffness peak value. The specific calculation logic for the critical damping matching state is that the additional damping peak value equals the square root of the product of twice the additional stiffness peak value and the equivalent rotational inertia of the corresponding joint of the lower limb rehabilitation exoskeleton robot. The setting of the critical damping matching state can suppress the lower limb rehabilitation exoskeleton robot... When the robot experiences a sudden change in the expected joint trajectory, it triggers high-frequency mechanical oscillations in the joints. The feedforward control module 400 uses the smooth trigger coefficient as a continuously changing weight to adjust the gain. The feedforward control module 400 multiplies the smooth trigger coefficient by the peak value of the additional stiffness to obtain the dynamic additional stiffness, and adds the dynamic additional stiffness to the basic stiffness to synthesize the comprehensive stiffness of the current time step. Similarly, the feedforward control module 400 multiplies the smooth trigger coefficient by the peak value of the additional damping to obtain the dynamic additional damping, and adds the dynamic additional damping to the basic damping to synthesize the comprehensive damping of the current time step. The dynamic impedance adjustment mechanism established based on the dynamic principle of impedance control enables the lower limb rehabilitation exoskeleton robot to exhibit stronger rigidity characteristics in the direction of the expected joint trajectory when the smooth trigger coefficient increases, thereby actively providing greater cross-limb traction force to assist subjects with weak lower limb muscle strength to smoothly cross the mechanical resistance dead point during gait transition.

[0102] Sub-step S433 involves the feedforward control module 400 combining the expected joint trajectory with the actual lower limb motion state to perform closed-loop solution of the impedance equation and output the target auxiliary torque for the lower limb. The feedforward control module 400 retrieves the current actual joint angle and current actual joint angular velocity from the shared memory of the central real-time control unit, fed back in real-time by the rotary encoder configured on the lower limb rehabilitation exoskeleton robot. Simultaneously, the feedforward control module 400 extracts the expected joint angle and expected joint angular velocity generated at the current time step. The feedforward control module 400 subtracts the current actual joint angle from the expected joint angle to obtain the cross-limb position tracking deviation, and subtracts the current actual joint angular velocity from the expected joint angular velocity to obtain the cross-limb velocity tracking deviation. The feedforward control module 400 multiplies the comprehensive stiffness by the cross-limb position tracking deviation and adds the comprehensive damping multiplied by the cross-limb velocity tracking deviation. The product is then used to calculate the target auxiliary torque for the lower limb through a dynamic closed-loop equation. After the calculation is completed, the feedforward control module 400 sets a maximum auxiliary torque limiting threshold. The maximum auxiliary torque limiting threshold is preset based on the smaller of the rated peak torque of the joint drive motor of the lower limb rehabilitation exoskeleton robot and the maximum safe torque that the subject's joint can withstand. When the absolute value of the target auxiliary torque for the lower limb is greater than the maximum auxiliary torque limiting threshold, the feedforward control module 400 forcibly cuts the amplitude of the target auxiliary torque for the lower limb to the maximum auxiliary torque limiting threshold to build an overload protection safety barrier for the underlying hardware. The target auxiliary torque for the lower limb after safety limiting is sent to the underlying execution module 500 in real time to drive the flexible actuator of the lower limb rehabilitation exoskeleton robot to perform the rehabilitation physical traction action of the lower limb rehabilitation exoskeleton robot. The closed-loop calculation expression of the target auxiliary torque for the lower limb is as follows:

[0103]

[0104]

[0105] In the formula, The overall stiffness is the synthesized output at the current time synchronization. The basic stiffness is generated based on the degree of lower limb muscle activation. The preset additional stiffness resistance value, The smooth trigger coefficient received at the current time step and used for control adjustment. The overall damping of the synthesized output at the current time step. The base damping is generated based on the degree of lower limb muscle activation. The preset additional damping peak value, The lower limb target auxiliary torque is calculated and output at the current time step. The expected joint angle is derived from the feedforward control module 400 and used for closed-loop calculation. This represents the actual joint angle obtained at the current time step. The expected joint angular velocity is derived from the feedforward control module 400 and used for closed-loop calculation. This represents the actual joint angular velocity obtained at the current time step.

[0106] In practice, obtaining spring deformation data of the flexible actuator to estimate the lumped disturbance torque may further include sub-steps S511 to S513.

[0107] Sub-step S511 involves the bottom-level execution module 500 extracting the spring deformation displacement measured by the displacement sensor and calculating the elastic transmission torque. The bottom-level execution module 500 extracts the spring deformation displacement measured by the displacement sensor inside the flexible actuator configured in the lower limb rehabilitation exoskeleton robot from the shared memory of the central real-time control unit. During long-term operation, the mechanical transmission mechanism of the lower limb rehabilitation exoskeleton robot inevitably experiences physical misalignment and frictional wear. The physical degradation of the mechanical transmission mechanism, combined with external environmental resistance, forms an interference torque that hinders the precise movement of the bottom-level motor. To actively compensate for the impact of this interference torque on the accuracy of cross-limb feedforward control, the bottom-level execution module 500 needs to accurately quantify the actual torque transmitted to the outside by the flexible actuator. The bottom-level execution module 500 calculates the elastic transmission torque by linearly multiplying the spring deformation displacement with the preset spring stiffness coefficient of the spring built into the flexible actuator. The specific value of the preset spring stiffness coefficient is determined by technicians through offline static loading experiments before the multi-source sensor fusion upper and lower limb collaborative rehabilitation training system leaves the factory to ensure the accuracy of physical quantification. The elastic transmission torque characterizes the effective physical torque output by the flexible actuator to the external load. The specific calculation expression for the elastic transmission torque is as follows:

[0108] In the formula, The elastic transmission torque calculated at the current time step. The preset spring stiffness coefficient for the built-in spring of the flexible actuator. This represents the spring deformation displacement measured by the displacement sensor at the current time step.

[0109] Sub-step S512 involves the bottom-level execution module 500 acquiring the actual motor angular velocity and calculating the smoothed actual motor angular acceleration. The bottom-level execution module 500 retrieves the actual motor angular velocity collected by the rotary encoder configured on the lower limb rehabilitation exoskeleton robot from the shared memory of the central real-time control unit. Addressing the technical issue of increased high-frequency noise amplification risk due to direct differential processing of the rotary encoder signal, which causes numerical divergence in the discretized dynamics observer, the bottom-level execution module 500 uses a Butterworth low-pass filter to filter the actual motor angular velocity. The cutoff frequency range of the Butterworth low-pass filter is set to 10 Hz to 20 Hz to effectively filter out high-frequency mechanical vibration interference caused by the commutation process of the bottom-level motor while preserving the macroscopic motion trend of the bottom-level motor. The filtering of the actual motor angular velocity can be performed according to existing digital filter design theories. The bottom-level execution module 500 performs discrete differential calculation on the filtered actual motor angular velocity to obtain the smoothed actual motor angular acceleration. The smoothed actual motor angular acceleration is used to reflect the true transient dynamic characteristics of the bottom-level motor and serves as the core state input variable of the discretized dynamics observer.

[0110] Sub-step S513 involves the bottom-level execution module 500 constructing a discretized dynamics observer based on kinematic state variables and dynamically estimating the lumped disturbance torque. Those skilled in the art can understand that the physical essence of introducing the discretized dynamics observer is to compare the deviation between the theoretical torque calculated by the ideal dynamics model and the actual control output torque, and to extract the external disturbances and internal friction equivalents that were not accurately modeled, thereby achieving active feedforward cancellation of unknown disturbances. The bottom-level execution module 500 obtains the lumped disturbance torque retained from the previous time step and the motor torque command output by the bottom-level execution module 500 from the previous time step. The execution module 500 combines the elastic transmission torque with the smooth actual motor angular acceleration and actual motor angular velocity to construct an ideal underlying motor dynamics equation model. This underlying motor dynamics equation model unifies the inaccurately modeled nonlinear friction components and parameter perturbation components into a lumped disturbance torque. The execution module 500 sets the observer filter coefficients to adjust the convergence speed and noise immunity of the discretized dynamics observer. The range of the observer filter coefficients is set from 0.8 to 0.95 to achieve control stability between estimation accuracy and noise amplification. The execution module 500 then converts the previous time step data from the previous execution module 500. The dynamic model residual for the current time step is obtained by subtracting the elastic transmission torque from the output motor torque command, then subtracting the product of the motor's equivalent moment of inertia and the smooth actual motor angular acceleration, and finally subtracting the product of the motor's equivalent viscous friction coefficient and the actual motor angular velocity. To prevent sudden changes in the smooth actual motor angular acceleration caused by brief loss or frame dropping of the rotary encoder signal, which could trigger a calculation overload in the discretized dynamics observer, the underlying execution module 500 introduces a disturbance residual limiting threshold. When the absolute value of the dynamic model residual for the current time step exceeds the disturbance residual limiting threshold, the underlying execution module 500 will adjust the current time step... The residual of the dynamic model is forcibly limited to a disturbance residual limiting threshold. The value of the disturbance residual limiting threshold is determined by calibrating 30% of the rated peak torque of the underlying motor. The underlying execution module 500 multiplies the observer filter coefficient by the lumped disturbance torque retained in the previous time step, adds a difference between the observer filter coefficient and the observed filter coefficient, multiplies it by the product of the dynamic model residual of the current time step after the limiting process, and finally calculates the lumped disturbance torque of the current time step. The lumped disturbance torque is stored in real time in the shared memory of the central real-time control unit for subsequent compensation execution calls. The specific calculation expression of the lumped disturbance torque is as follows:

[0111] In the formula, The lumped disturbance moment of the output is dynamically estimated for the current time step. These are the preset observer filter coefficients. The lumped disturbance moment retained from the previous time step. The equivalent moment of inertia of the motor is obtained through offline frequency sweep experiment parameters. The smoothed actual motor angular acceleration obtained at the current time step. The equivalent viscous friction coefficient of the motor is obtained through offline frequency sweep experiment parameters. The actual motor angular velocity obtained at the current time step. This refers to the motor torque command output by the underlying execution module 500 in the previous time step. This is the elastic transmission torque calculated at the current time step.

[0112] The filtering and smoothing of the lower limb target auxiliary torque, combined with the lumped disturbance torque synthesis, and the final output to the motor execution command can be further included in the specific implementation as sub-steps S521 to S523.

[0113] Sub-step S521 involves the bottom-level execution module 500 receiving the target auxiliary torque of the lower limb and introducing a nonlinear tracking differentiator for first-order low-pass filtering and amplitude limiting smoothing to extract the smoothed auxiliary torque of the lower limb. The bottom-level execution module 500 receives the target auxiliary torque of the lower limb output by the feedforward control module 400. Addressing the technical problem of high-frequency noise carried by multi-source heterogeneous signals during processing, which can easily cause high-frequency jitter in the target auxiliary torque of the lower limb between adjacent time steps and directly induce mechanical resonance in the flexible actuators configured in the lower limb rehabilitation exoskeleton robot, the bottom-level execution module 500 introduces a nonlinear tracking differentiator based on kinematic constraints. The lower limb target auxiliary torque is filtered and smoothed. The nonlinear tracking differentiator can effectively filter out jitter components and extract a smooth and continuous lower limb smooth auxiliary torque under low time delay constraints. The specific discrete algorithm implementation of the optimal control synthesis function inside the nonlinear tracking differentiator can be performed according to the definition of the discrete-time optimal control synthesis function in the existing active disturbance rejection control theory. Those skilled in the art can understand that the optimal control synthesis function achieves the fastest tracking of the target control signal without static error without generating system response overshoot by constructing the fastest nonlinear feedback characteristic. The calculation expression of the lower limb smooth auxiliary torque is as follows:

[0114] In the formula, The lower limb smoothing auxiliary torque is the output after filtering by the nonlinear tracking differentiator at the current time step. The lower limb smoothing assist torque retained from the previous time step, A preset discrete time step is set, with the value of the preset discrete time step ranging from one millisecond to ten milliseconds to maintain the numerical stability of the discretization calculation. This is the preset optimal control synthesis function. The lower limb target auxiliary torque received at the current time step. The velocity tracking factor is set to a value between ten and fifty for the nonlinear tracking differentiator to balance the smoothness of the filter and the speed of the signal following response. The acceleration limiting threshold is set for the nonlinear tracking differentiator, and the value of the acceleration limiting threshold is determined by the maximum physical acceleration calibration parameters of the motor configured with the flexible actuator.

[0115] Sub-step S522 calculates the torque tracking error for the bottom-level execution module 500 and generates a basic feedback adjustment command by combining the proportional-integral-derivative control law and the dynamic anti-integral saturation limiting algorithm. The bottom-level execution module 500 obtains the lower limb smoothing auxiliary torque and uses it as the ideal output torque of the flexible actuator. The bottom-level execution module 500 subtracts the elastic transmission torque from the lower limb smoothing auxiliary torque to obtain the current torque tracking error. The bottom-level execution module 500 combines the torque tracking error with the preset proportional control gain, preset integral control gain, and preset derivative control gain to perform mathematical calculation of the basic feedback adjustment command. When the subject's lower limbs exhibit abnormal spastic resistance, the torque tracking error is prone to continuous unidirectional accumulation. Simple integral accumulation calculation can easily lead to a dangerous situation where the motor torque command output by the bottom-level execution module 500 becomes uncontrolled and overloaded. The underlying execution module 500 introduces a dynamic anti-integral saturation limiting algorithm in the integral term calculation of the proportional-integral-derivative control law. The underlying execution module 500 pre-sets an integral saturation torque threshold. The value of the integral saturation torque threshold is determined by offline calibration of 80% of the rated continuous output torque of the motor configured by the underlying execution module 500, leaving a 20% control margin for high-frequency mechanical disturbance compensation. When the absolute value of the product of the cumulative sum of torque tracking error in the historical time step and the preset integral control gain exceeds the integral saturation torque threshold, the underlying execution module 500 immediately stops the accumulation calculation of torque tracking error in the current time step and maintains the cumulative sum of torque tracking error in the historical time step at the value level of the previous time step, thereby strictly limiting the output value of the integral term of the proportional-integral-derivative control law within the safe tolerance range of the motor.

[0116] Sub-step S523 involves the bottom-level execution module 500 performing feedforward superposition compensation of the basic feedback adjustment command and the lumped disturbance torque, and synthesizing the final motor torque command. The bottom-level execution module 500 extracts the lumped disturbance torque estimated at the current time step from the shared memory of the central real-time control unit. The bottom-level execution module 500 multiplies the preset proportional control gain by the torque tracking error, adds the preset integral control gain multiplied by the cumulative sum of the torque tracking error over historical time steps, and adds the preset derivative control gain multiplied by the difference between the torque tracking error of the current time step and the torque tracking error of the previous time step. The specific values ​​of the preset proportional control gain, the preset integral control gain, and the preset derivative control gain are pre-determined offline according to the existing Ziegler-Nichols rule. Calibration is performed to ensure the response stability of the underlying proportional-integral-derivative (PID) control law, thereby calculating the basic feedback adjustment command. The underlying execution module 500 adds the basic feedback adjustment command to the lumped disturbance torque and finally synthesizes it into a motor torque command output to the flexible actuator. At the physical mechanism level of control execution, the PID control law provides the underlying motor with basic closed-loop error feedback tracking capability, while the feedforward superposition mechanism of the lumped disturbance torque realizes the proactive counter-cancellation of external complex physical environment and internal friction interference of mechanical transmission mechanism. The feedforward superposition mechanism of basic feedback adjustment command and lumped disturbance torque takes into account both the dynamic tracking response speed of system control command and the steady-state output accuracy against external disturbances. The specific calculation expression of the motor torque command synthesis process is as follows:

[0117] In the formula, This represents the torque tracking error calculated at the current time step. The lower limb smoothing assist torque extracted at the current time step. The elastic transmission torque calculated at the current time step. The motor torque command synthesized at the current time step is output to the flexible driver. The preset proportional control gain, The preset integral control gain, This is the cumulative sum of the torque tracking error over historical time steps after processing with dynamic anti-integral saturation and amplitude limiting algorithms. The preset differential control gain, This is the torque tracking error retained from the previous time step. The lumped disturbance torque extracted for the current time step and used for feedforward compensation.

[0118] To aid in understanding the technical solution of this invention, a specific application embodiment is provided below.

[0119] A patient with right-sided limb motor dysfunction was selected as the subject. The subject wore a lower limb rehabilitation exoskeleton robot and a hand function rehabilitation exoskeleton robot for cross-limb coordinated walking training. The discrete time step of the central real-time control unit was set to one millisecond. During the system startup initialization phase, the data synchronization module 100 injected a benchmark test pulse, calibrating the inherent total communication and processing latency of the multi-source sensor fusion upper and lower limb coordinated rehabilitation training system to be 21.4 milliseconds.

[0120] At the fifteenth minute of rehabilitation training, the subject attempted to step out with their right leg. The subject applied a gripping force to the arc-shaped groove transmission mechanism, and the torque sensor on the hand function rehabilitation exoskeleton robot and the hand electromyography (EMG) sensor simultaneously acquired signals. The intent extraction module 300, based on short-time power spectral density distribution calculations, found that the median frequency of the hand muscles decreased from 85.2 Hz recorded in the initial unforced initiation phase to 61.7 Hz. The intent extraction module 300 calculated a real-time muscle fatigue index of 0.275. Following a dynamic weight allocation mechanism, the intent extraction module 300 reduced the initial EMG feature weight from 0.6 to 0.517, simultaneously increasing the initial torque feature weight from 0.4 to 0.483, fusing them to obtain a comprehensive force exertion intent feature. This feature was then processed using first-order difference and dead-zone filtering to extract high-frequency abrupt changes in the force exertion intent.

[0121] The feedforward control module 400 receives the high-frequency abrupt change characteristic of the force application intention and maps it to a smooth triggering coefficient using a logic function. Combining the inherent communication and processing total delay time with the smooth triggering coefficient, the feedforward control module 400 calculates a dynamic time prediction step size of 48.6 milliseconds. Based on the second-order Taylor series expansion principle, the feedforward control module 400 uses the dynamic time prediction step size to deduce the expected joint angle. Simultaneously, the feedforward control module 400 uses the smooth triggering coefficient, superimposed with additional stiffness and additional damping, to calculate the lower limb target auxiliary torque through a closed-loop solution of the impedance equation.

[0122] The bottom-level execution module 500 receives the target auxiliary torque for the lower limb and extracts the smoothed auxiliary torque for the lower limb using a nonlinear tracking differentiator. Based on the spring deformation displacement measured by displacement sensors deployed inside the flexible actuator, the bottom-level execution module 500 calculates the elastic transmission torque and estimates the lumped disturbance torque, which includes friction from the mechanical transmission mechanism and external resistance, to be 3.12 Nm, using a discretized dynamics observer. The bottom-level execution module 500 feeds forward the basic feedback adjustment command and the lumped disturbance torque to generate a motor torque command, which is then sent to the flexible actuator. The flexible actuator drives the lower limb rehabilitation exoskeleton robot to enter the assisted following state in advance.

[0123] Table 1. Performance Comparison Test Data Between Multi-Source Sensor Fusion Control Strategy and Traditional Single Control Strategy 1 Traditional single control strategy 76.42 4.81 9.14 2.37 1 System strategy of this invention 22.18 1.34 4.08 0.45 2 Traditional single control strategy 72.89 5.03 8.87 2.51 2 System strategy of this invention 24.31 1.48 4.22 0.52 3 Traditional single control strategy 78.15 4.96 9.53 2.68 3 System strategy of this invention 21.05 1.29 3.96 0.41 4 Traditional single control strategy 74.56 5.22 9.02 2.44 4 System strategy of this invention 23.47 1.55 4.31 0.49 5 Traditional single control strategy 79.03 4.78 9.25 2.59 According to the data in Table 1, the multi-source sensor fusion upper and lower limb coordinated rehabilitation training system exhibits a significantly lower intention response lag time than the traditional single control strategy. The intention response lag time of the traditional single control strategy ranges from 72.89 milliseconds to 79.03 milliseconds, while the intention response lag time of the system strategy of this invention remains between 21.05 milliseconds and 24.31 milliseconds. Test data shows that the feedforward control module, combining inherent communication delay calibration with kinematic advance calculation mechanism based on cross-limb force exertion intention characteristics, directly offsets the physical lag phenomenon of underlying hardware data transmission and instruction calculation, enabling the physical movements of the lower limb rehabilitation exoskeleton robot to be aligned with the subject's actual force exertion intention in the time domain.

[0124] According to the data in Table 1, in terms of the maximum tracking error of the knee joint, the system strategy of this invention limits the maximum tracking error of the knee joint from approximately 5 degrees in the traditional single control strategy to a range of 1.29 to 1.55 degrees. The improvement in tracking accuracy verifies the effectiveness of the discretized dynamic observer in the underlying execution module in estimating the lumped disturbance torque and performing feedforward superposition compensation. The underlying execution module actively cancels the interference torque caused by internal friction of the mechanical transmission mechanism and external physical environment resistance, preventing the interference torque from disrupting the position tracking closed loop. At the same time, the peak value of the transient current impact decreases from approximately 9 amperes in the traditional single control strategy to approximately 4 amperes in the system strategy of this invention. The decrease in the peak value of the transient current impact is attributed to the feedforward control module using a nonlinear logic function to construct a smooth trigger coefficient after the intention extraction module outputs the high-frequency abrupt change characteristic of the force application intention, and the underlying execution module introducing a nonlinear tracking differentiator to perform first-order low-pass filtering and amplitude limiting smoothing on the lower limb target auxiliary torque. The nonlinear tracking differentiator filters out high-frequency noise and abrupt step components carried during the fusion of multi-source heterogeneous signals, transforming discrete torque intentional jumps into continuously differentiable physical quantities. This avoids spikes in the underlying motor current and reduces the probability of mechanical resonance within the flexible actuator. The steady-state torque fluctuation variance is reduced from approximately 2.5 square Newton-meters (SMM) in traditional single control strategies to below 0.5 SMM. This demonstrates that the dynamic weight adaptive allocation mechanism combined with the muscle fatigue index can resist the control baseline drift caused by the compensatory increase in the amplitude of surface electromyography (EMG) signals when human muscles are fatigued, maintaining the steady-state torque output accuracy of the multi-source sensor fusion upper and lower limb coordinated rehabilitation training system during long-term operation.

[0125] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-source sensor fusion-based upper and lower limb coordinated rehabilitation training method, characterized in that, Includes the following steps: Simultaneously acquire multi-source physiological and kinematic data, perform system initialization, and calibrate the system's inherent total communication and processing latency. The system receives lower limb kinematic data based on a hidden Markov model, decodes it online, and outputs the lower limb gait phase. Acquire electromyographic signals on the surface of the hand and calculate the real-time muscle fatigue index. Adjust the weights dynamically based on the real-time muscle fatigue index and calculate the high-frequency abrupt change characteristic of the force exertion intention by combining the human-computer interaction force detected by the torque sensor. Based on the high-frequency abrupt change characteristics of the lower limb gait phase and the force exertion intention, and combined with the system's inherent total communication and processing delay time, feedforward trajectory prediction is performed to calculate the lower limb target auxiliary torque. The lumped disturbance torque is estimated by obtaining the spring deformation displacement of the flexible actuator, the auxiliary torque of the lower limb target is filtered and smoothed, and the motor torque command is generated by combining the lumped disturbance torque and output to the flexible actuator for execution.

2. The multi-source sensor fusion-based upper and lower limb coordinated rehabilitation training method according to claim 1, characterized in that, The method of receiving lower limb kinematic data based on a hidden Markov model, decoding it online, and outputting the lower limb gait phase includes: Angular velocity and angular acceleration of both lower limbs are extracted and combined with joint angles to construct an observation vector to characterize the motion state of the lower limbs; The initial local maximum likelihood variable is calculated by combining the initial state probability vector in the optimized model parameters with the observed state transition probability matrix. The local maximum likelihood variables are transformed into logarithmic domains. Based on the local maximum likelihood variables retained from the previous time step and the state transition probability matrix, the maximum path probability of the current time step under each candidate state is calculated, and the local maximum likelihood variables of the current time step are calculated. Compare the local maximum likelihood variables corresponding to all hidden states at the current time step, and extract the state identifier corresponding to the maximum local maximum likelihood variable value as the lower limb gait phase at the current time step.

3. The multi-source sensor fusion-based upper and lower limb coordinated rehabilitation training method according to claim 1, characterized in that, The acquisition of electromyographic signals on the surface of the hand and the calculation of the real-time muscle fatigue index include: The electromyographic signals of the hand surface within the sliding time window are extracted, frequency conversion is performed, and the power spectral density distribution is calculated. The critical frequency point that precisely divides the total power spectral density area is used as the median frequency of the current time step. The frequency offset attenuation is obtained by combining the reference median frequency recorded in the initial stage with the median frequency of the current time step, and the normalized initial fatigue quantization value is calculated. The initial fatigue quantization value is processed by time-domain smoothing filtering using the exponential moving average algorithm to generate a smooth fatigue quantization value. The smooth fatigue quantization value is then limited to the real number boundary range of zero to one by saturation limiting logic, and the final real-time muscle fatigue index is output.

4. The multi-source sensor fusion-based upper and lower limb coordinated rehabilitation training method according to claim 3, characterized in that, The step of dynamically adjusting weights based on the real-time muscle fatigue index and calculating high-frequency abrupt change characteristics of force exertion intention by combining the human-computer interaction force detected by the torque sensor includes: The electromyographic signals on the hand surface and the human-computer interaction force are normalized to obtain electromyographic activation degree and torque mapping value, respectively. The dynamic electromyography weight is obtained by subtracting the product of the real-time muscle fatigue index and the preset fatigue compensation coefficient from the preset initial electromyography feature weight, and the dynamic torque weight is obtained by adding the product of the real-time muscle fatigue index and the fatigue compensation coefficient to the preset initial torque feature weight. The comprehensive force application intention characteristics are calculated by combining the dynamic electromyography weights, the electromyography activation degree, the dynamic torque weights, and the torque mapping values. The first-order difference component is obtained by performing first-order difference calculation on the comprehensive force exertion intention feature of adjacent time steps. A nonlinear dead zone activation function is introduced to perform truncation filtering on the first-order difference component. After filtering out static noise, the high-frequency change feature of force exertion intention is output.

5. The multi-source sensor fusion-based upper and lower limb coordinated rehabilitation training method according to claim 1, characterized in that, The step of calculating the lower limb target auxiliary torque based on the high-frequency abrupt change characteristics of the lower limb gait phase and the force exertion intention, combined with the inherent communication and processing total delay time of the system, includes: Based on the lower limb gait phase, the corresponding discrete kinematic node data is matched in the standardized healthy gait database, and a cubic spline interpolation algorithm is used to perform smooth interpolation calculation to generate the basic expected joint angle, basic expected joint angular velocity, and basic expected joint angular acceleration. A nonlinear smoothing mapping mechanism based on logic functions is introduced to convert the high-frequency mutation feature of the force exertion intention into a normalized smooth triggering coefficient; wherein, 50% of the historical peak value of the high-frequency mutation feature of the force exertion intention extracted within a preset sliding time window is dynamically set as the mutation activation center threshold.

6. The multi-source sensor fusion-based upper and lower limb coordinated rehabilitation training method according to claim 5, characterized in that, After generating the basic expected joint angle, basic expected joint angular velocity, basic expected joint angular acceleration, and smooth trigger coefficient, the method further includes: adding the product of the preset maximum feedforward time compensation gain and the smooth trigger coefficient to the total communication and processing delay time inherent in the system to obtain the dynamic time prediction step size. Based on the principle of second-order Taylor series expansion, and combined with the basic expected joint angle, the basic expected joint angular velocity, the basic expected joint angular acceleration, and the dynamic time prediction step size, the advanced expected joint angle and the advanced expected joint angular velocity are calculated. The expected joint angle is limited to the upper and lower limits of the human physiological limit collision avoidance angle of the corresponding joint, and the amplitude of the expected joint angular velocity is proportionally truncated to the maximum safe angular velocity threshold of the motor.

7. The multi-source sensor fusion-based upper and lower limb coordinated rehabilitation training method according to claim 6, characterized in that, After calculating the expected joint angle and expected joint angular velocity, the following steps are also included: The root mean square value of the electromyography signal on the lower limb surface is calculated to obtain the degree of lower limb muscle activation. Based on the degree of lower limb muscle activation and a preset proportional coefficient, the basic stiffness and basic damping are generated. The smoothing trigger coefficient is multiplied by the preset additional stiffness peak value and additional damping peak value respectively, and then added to the basic stiffness and the basic damping to synthesize the comprehensive stiffness and comprehensive damping at the current time step; wherein the additional damping peak value and the additional stiffness peak value are in a critical damping matching state. Based on the expected joint angle and the expected joint angular velocity, the tracking deviation is calculated by combining the current actual joint angle and the current actual joint angular velocity. The target auxiliary torque of the lower limb is calculated based on the dynamic closed-loop equation, combined with the comprehensive stiffness and the comprehensive damping, and then the maximum auxiliary torque is limited.

8. The multi-source sensor fusion-based upper and lower limb coordinated rehabilitation training method according to claim 1, characterized in that, The method of obtaining the spring deformation displacement of the flexible actuator to estimate the lumped disturbance torque includes: calculating the elastic transmission torque by combining the spring deformation displacement with a preset spring stiffness coefficient; After filtering the actual motor angular velocity collected by the rotary encoder using a low-pass filter, discrete difference calculation is performed to obtain the smooth actual motor angular acceleration. By combining the elastic transmission torque, the smooth actual motor angular acceleration, the actual motor angular velocity, and the motor torque command output in the previous time step, a discretized dynamic observer is constructed to calculate the dynamic model residual. The lumped disturbance torque of the current time step is then calculated by combining the set observer filter coefficients and the lumped disturbance torque retained in the previous time step.

9. The multi-source sensor fusion-based upper and lower limb coordinated rehabilitation training method according to claim 8, characterized in that, The step of filtering and smoothing the auxiliary torque of the lower limb target, and combining it with the lumped disturbance torque to generate and output the motor torque command includes: A nonlinear tracking differentiator based on kinematic limit constraints is introduced to perform first-order low-pass filtering and amplitude limiting smoothing on the lower limb target auxiliary torque, and the smoothed lower limb auxiliary torque is extracted. The torque tracking error is calculated by subtracting the elastic transmission torque from the lower limb smoothing auxiliary torque. The torque tracking error is combined with the proportional-integral-derivative (PID) control law to generate a basic feedback adjustment command; wherein, a dynamic anti-integral saturation limiting algorithm is introduced in the integral term calculation of the PID control law. When the absolute value of the result of the cumulative sum of the torque tracking error over the historical time step multiplied by the integral control gain exceeds the preset integral saturation torque threshold, the accumulation calculation of the integral term is stopped. The basic feedback adjustment command and the lumped disturbance torque are fed forward and superimposed for compensation to synthesize the final motor torque command.

10. The multi-source sensor fusion-based upper and lower limb coordinated rehabilitation training method according to claim 1, characterized in that, The multi-source physiological and kinematic data were collected by sensors configured on the lower limb rehabilitation exoskeleton robot and the hand function rehabilitation exoskeleton robot; The calibration system's inherent total communication and processing delay time specifically includes: generating a benchmark test pulse and injecting it into the system control bus during the system startup phase, recording the time span from the issuance of the benchmark test pulse to the receipt of status feedback, and defining this time span as the system's inherent total communication and processing delay time.

Images