A rehabilitation robot adaptive control method based on two-dimensional state evaluation

Abstract

This invention relates to the field of rehabilitation robot control technology and discloses an adaptive control method for rehabilitation robots based on dual-dimensional state assessment. The method includes: acquiring surface electromyography (EMG) signals, human-machine interaction force signals, kinematic signals, and physiological and psychological signals; generating a muscle fatigue index based on the EMG signals; generating a human-machine trust index based on the human-machine interaction force signals, kinematic signals, and physiological and psychological signals; generating a comprehensive human state index based on the muscle fatigue index and the human-machine trust index; and adjusting the impedance control parameters of the rehabilitation robot in response to the comprehensive human state index to generate torque control commands. This invention, by simultaneously assessing the dual-dimensional states of physiological fatigue and psychological trust and dynamically coupling them to drive fuzzy adaptive impedance control, achieves accurate perception and adaptive response of the rehabilitation robot to the patient's state, significantly improving the comfort, human-machine collaboration, and safety of rehabilitation training.

Description

Technical Field

[0001] This invention relates to the field of rehabilitation robot control technology, and specifically to an adaptive control method for rehabilitation robots based on two-dimensional state assessment. Background Technology

[0002] Rehabilitation robots, as wearable devices that assist patients in rehabilitation training, have been widely used in the field of clinical rehabilitation. Existing control methods for rehabilitation robots mainly include pre-programmed control, dynamic model-based control, sensitivity amplification control based on human-machine interaction forces, and motion intention recognition control based on surface electromyography signals. These methods, to a certain extent, realize the assistive function of exoskeletons for patient movement.

[0003] However, existing technologies still have the following shortcomings in practical applications: First, existing control methods mostly focus on single-dimensional physiological signals, such as judging muscle activation based solely on surface electromyography signals or identifying human movement intentions based solely on interaction force signals, failing to simultaneously consider the patient's physiological fatigue state and psychological trust state. During rehabilitation training, accumulated muscle fatigue affects the patient's motor ability and cooperation, while the patient's psychological trust in the exoskeleton directly affects training effectiveness and human-machine coordination. Single-dimensional assessments cannot comprehensively reflect the patient's true state. Second, existing control strategies are mostly based on fixed rule-based mode switching or parameter adjustment based on single indicators, lacking a decision-making mechanism that effectively integrates multi-dimensional human state information. Furthermore, control parameter adjustments are often discrete and abrupt, easily leading to a sense of jarring during human-machine interaction, and failing to achieve smooth adaptive control based on real-time changes in the patient's overall state.

[0004] To address the aforementioned issues, there is a need for a rehabilitation robot control method that can simultaneously assess a patient's physical fatigue and psychological trust status, and effectively integrate the two-dimensional information to achieve smooth adaptive control. Summary of the Invention

[0005] In view of this, the present invention provides an adaptive control method for rehabilitation robots based on dual-dimensional state assessment, in order to solve the technical problems of existing rehabilitation robots being unable to simultaneously perceive the patient's physical fatigue and psychological trust status, and having a single control strategy and uneven adjustment.

[0006] In a first aspect, the present invention provides an adaptive control method for a rehabilitation robot based on a two-dimensional state assessment, applied to a rehabilitation robot. The method includes the following steps: acquiring surface electromyography (EMG) signals, human-machine interaction force signals, kinematic signals, and physiological and psychological signals, wherein the physiological and psychological signals include heart rate variability indicators and skin conductance response indicators; generating a muscle fatigue index based on the EMG signals; generating a human-machine trust index based on the human-machine interaction force signals, kinematic signals, and physiological and psychological signals; generating a comprehensive human state index based on the muscle fatigue index and the human-machine trust index; and adjusting the impedance control parameters of the rehabilitation robot in response to the comprehensive human state index to generate a torque control command.

[0007] In one optional implementation, generating a muscle fatigue index based on the surface electromyography (EMG) signal specifically includes: performing singular spectrum analysis on the EMG signal, constructing a trajectory matrix and performing singular value decomposition to obtain multiple components arranged in descending order of singular values; calculating the proportion of the cumulative energy of the first k components to the total energy of all components, accumulating them sequentially according to the component order, selecting the first r principal components that make the cumulative energy proportion reach a preset threshold for the first time to reconstruct a trend component; extracting the root mean square value and average rectified value of the trend component, as well as the median frequency of the original signal, to form a fatigue feature vector; inputting the fatigue feature vector into the Mamba time series prediction model to output the muscle fatigue index.

[0008] In one optional implementation, generating a human-computer trust index based on the human-computer interaction force signal, kinematic signal, and physiological and psychological signal specifically includes: constructing a trust state vector based on the heart rate variability index, skin conductance response index, the human-computer interaction force signal, and kinematic signal. The trust state vector contains eight state components, specifically including: total heart rate variability (SDNN), short-term heart rate variability (RMSSD), peak skin conductance response count (f_SCR) per unit time, human-computer interaction force deviation (ε_F), trajectory following error (e_θ), trajectory following error change rate (Δe_θ), previous time trust (Trust(t-1), and current fatigue level (FI(t)); and inputting the trust state vector into an improved Dueling algorithm. The DoubleDQN network comprises an online network and a target network. The online network generates an estimated Q-value for each trust adjustment action based on the current trust state and selects an action to be executed. The target network generates a target Q-value for the action to be executed. The parameters of the online network are updated based on the error between the target Q-value and the actual reward obtained after executing the action. The Q-value is decomposed into a state value function and an action advantage function. The trust adjustment action is selected and executed based on the updated Q-value output by the online network. The human-machine trust index is updated and output based on the execution result.

[0009] In one optional implementation, generating a comprehensive human condition index based on the muscle fatigue index and the human-machine trust index specifically includes: calculating a coupled fatigue index and a coupled trust index after their mutual influence based on the muscle fatigue index and the human-machine trust index; then performing a weighted summation of the coupled fatigue index and the coupled trust index to obtain the comprehensive human condition index; wherein the weights of the weighted summation are adaptively adjusted according to the rehabilitation stage, and the weight of fatigue gradually increases while the weight of trust gradually decreases as the rehabilitation time progresses.

[0010] In an optional implementation, the step of calculating the coupled fatigue index and coupled trust index after their mutual influence based on the muscle fatigue index and the human-machine trust index is specifically achieved as follows: A dynamic coupling weight matrix is ​​constructed, the matrix containing four elements, each of which is a function of the muscle fatigue index FI and the human-machine trust index Trust; the coupled fatigue index FI_coupled = w11×FI + w12×Trust and the coupled trust index Trust_coupled = w21×FI + w22×Trust are calculated according to the formulas; where the first self-weight w11 = α + β×(1-Trust), the second self-weight w22 = α + β×FI, and the first cross weight w12 and the second cross weight w21 are both γ×(FI-Trust), and α, β, and γ are preset constants.

[0011] In one optional implementation, the step of adjusting the impedance control parameters of the rehabilitation robot in response to the Comprehensive Human State Index (CHSI) and generating a torque control command specifically includes: using the CHSI, its rate of change ΔCHSI, and the human-machine interaction force deviation ΔF as input variables of a fuzzy controller; performing inference based on a preset fuzzy rule base to obtain the stiffness adjustment amount ΔK and damping adjustment amount ΔB of the impedance parameters; updating the impedance control parameters based on the stiffness adjustment amount ΔK and damping adjustment amount ΔB, including updating the desired stiffness K_d and desired damping B_d, and synchronously updating the desired inertia M_d based on K_d and B_d according to the principle of maintaining a stable damping ratio; dynamically adjusting the maximum permissible torque τ_max and maximum permissible velocity θ̇_limit with the CHSI; calculating the desired acceleration θ̈_cmd based on the updated impedance parameters according to the target impedance equation, and then generating the torque control command τ_cmd in conjunction with the exoskeleton dynamics model.

[0012] Secondly, the present invention provides an adaptive control system for a rehabilitation robot based on dual-dimensional state assessment, comprising: a signal acquisition module for acquiring surface electromyography (EMG) signals, human-machine interaction force signals, kinematic signals, and physiological and psychological signals, wherein the physiological and psychological signals include heart rate variability indicators and skin conductance response indicators; a muscle fatigue assessment module connected to the signal acquisition module for generating a muscle fatigue index based on the EMG signals; a human-machine trust assessment module connected to the signal acquisition module for generating a human-machine trust index based on the human-machine interaction force signals, kinematic signals, and physiological and psychological signals; a fusion decision module connected to the muscle fatigue assessment module and the human-machine trust assessment module for generating a comprehensive human state index based on the muscle fatigue index and the human-machine trust index; and an adaptive control module connected to the fusion decision module and the signal acquisition module for adjusting the impedance control parameters of the rehabilitation robot in response to the comprehensive human state index and generating torque control commands.

[0013] Thirdly, the present invention provides a computer device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the adaptive control method for rehabilitation robots based on dual-dimensional state assessment described in the first aspect or any embodiment thereof.

[0014] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions, the computer instructions being used to cause a computer to execute the adaptive control method for a rehabilitation robot based on dual-dimensional state assessment as described in the first aspect or any embodiment thereof.

[0015] Fifthly, the present invention provides a computer program product, including computer instructions, which are used to cause a computer to execute the adaptive control method for a rehabilitation robot based on dual-dimensional state assessment as described in the first aspect or any embodiment thereof.

[0016] Beneficial Effects: This invention provides a comprehensive multimodal data foundation for subsequent state assessment by acquiring surface electromyography (EMG) signals, human-machine interaction force signals, kinematic signals, and physiological and psychological signals. A muscle fatigue index is generated based on EMG signals, and a human-machine trust index is generated simultaneously based on the human-machine interaction force signals, kinematic signals, and physiological and psychological signals. This enables parallel assessment of both the patient's physical fatigue and psychological trust, overcoming the limitations of existing technologies that focus on only a single dimension. A comprehensive human state index is generated based on the muscle fatigue index and the human-machine trust index, integrating the two-dimensional information into a unified control command, providing a comprehensive decision-making basis for subsequent control. The impedance control parameters of the rehabilitation robot are adjusted in response to the comprehensive human state index, and torque control commands are generated, enabling the control strategy to adapt to the patient's current state in real time. Through the synergistic effect of the above steps, this invention enables the rehabilitation robot to simultaneously perceive the patient's physical fatigue and psychological trust state, and adaptively adjust control parameters according to the comprehensive state, thereby improving the comfort, human-machine collaboration, and safety of rehabilitation training. Attached Figure Description

[0017] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is an overall flowchart of the adaptive control method for rehabilitation robots based on dual-dimensional state assessment according to an embodiment of the present invention; Figure 2 This is a flowchart of a muscle fatigue assessment method according to an embodiment of the present invention; Figure 3 This is a flowchart of a human-machine trust assessment method according to an embodiment of the present invention; Figure 4 This is a flowchart of the dynamic coupling and CHSI generation method according to an embodiment of the present invention; Figure 5 This is a flowchart of the CHSI-driven fuzzy adaptive control method according to an embodiment of the present invention; Figure 6 This is a structural block diagram of an adaptive control system for a rehabilitation robot based on dual-dimensional state assessment according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of 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, 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.

[0020] Example 1: According to an embodiment of the present invention, an adaptive control method for a rehabilitation robot based on two-dimensional state assessment is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0021] This embodiment provides an adaptive control method for rehabilitation robots based on two-dimensional state assessment, which can be used in the controller of rehabilitation robots. Figure 1 This is a flowchart of a method according to an embodiment of the present invention, such as... Figure 1 As shown, the process includes the following steps: Step S101: Acquire surface electromyography signals, human-computer interaction force signals, kinematic signals, and physiological and psychological signals, wherein the physiological and psychological signals include heart rate variability indicators and skin conductance response indicators.

[0022] Specifically, surface electromyography (sEMG) signals are acquired by attaching Ag / AgCl surface electrodes to the target muscle group. The electrodes are attached along the muscle fiber direction with a spacing of 20 mm. Differential amplification is used, and the sampling frequency is set to 1000 Hz to meet the Nyquist sampling requirements of the main frequency components of sEMG (20-500 Hz). The raw signal is then subjected to a 20-450 Hz bandpass filter to remove low-frequency baseline drift and high-frequency noise, followed by a 50 Hz notch filter to eliminate electric field interference, resulting in a preprocessed sEMG signal. Human-computer interaction force signals are acquired by installing a six-dimensional force / torque sensor at the interface between the exoskeleton and the human body. This sensor collects triaxial forces and torques in real time during human-computer interaction at a sampling frequency of 500 Hz. The raw signal is then subjected to a low-pass filter to remove high-frequency vibration noise, and static calibration is used to eliminate sensor zero-point drift, resulting in a corrected interaction force signal. Kinematic signals are directly read from joint angles by installing absolute photoelectric encoders at the rotation axes of each active joint of the exoskeleton. The encoder resolution is 4096 pulses / revolution, and the sampling frequency is 500Hz. Angular velocity is obtained by numerically differentiating and filtering the angle signal. Physiological and psychological signals include heart rate variability (HRV) and skin conductance response (SCR) indices. HRV is acquired by collecting ECG signals using ECG electrodes attached to the chest at a sampling frequency of 250Hz. After R-wave peak detection, the RR interval is calculated to obtain the SDNN and RMSSD time-domain indices. SCR is acquired by collecting skin conductance sensors attached to the fingers. A constant microcurrent of 0.5μA is applied to measure skin conductance values. The baseline of skin conductance is extracted by minimum value filtering through a sliding window, and the rapidly changing skin conductance response (SCR) is separated. Since the sampling frequencies of each sensor are different (sEMG@1kHz, force / kinematics / skin conductance@500Hz, heart rate variability@250Hz), in order to eliminate the phase deviation caused by different sampling rates and ensure the time consistency of multimodal data, a unified time reference is established with a step size of 1ms corresponding to the 1kHz sampling of the aforementioned sEMG signal. By using linear interpolation, each signal is aligned to a unified time axis of 1kHz, and a time-synchronized multimodal data packet is constructed and transmitted to the subsequent human condition assessment stage.

[0023] Step S102: Based on the surface electromyography signal, generate a muscle fatigue index.

[0024] Specifically, singular spectrum analysis is performed on the preprocessed surface electromyography (sEMG) signals to extract fatigue features. First, a trajectory matrix is ​​constructed by embedding a one-dimensional sEMG time series into the trajectory matrix. The analysis window length is set to 2000 samples corresponding to a 2-second duration, the embedding dimension is 500, and the trajectory matrix dimension is 500×1501. Singular value decomposition (SVD) is performed on the trajectory matrix to obtain multiple components arranged in descending order of singular values ​​and their corresponding singular values ​​σ1, σ2, ..., σ_L. The proportion of the cumulative energy of the first k components to the total energy of all components is calculated. This cumulative energy proportion is the ratio of the sum of the singular values ​​of the currently selected first k components to the sum of the singular values ​​of all components, where k ranges from 1 to the total number of components. The components are accumulated sequentially. The first r principal components that cause the cumulative energy proportion to first reach a preset threshold are selected to reconstruct the trend components. In this embodiment, the preset threshold is 0.95, meaning that the first r principal components that retain more than 95% of the energy are selected. Statistical analysis of multiple sEMG samples determines that r is typically 5. During reconstruction, a matrix retaining the first r principal components is first constructed, and then the trend component is obtained by diagonal averaging. Next, fatigue features are extracted. Power spectrum estimation is performed on the preprocessed original signal, using the Welch method for segmented calculation (8 segments, 256 samples per segment, Hamming window). After obtaining the power spectral density, the frequency corresponding to the median of the power spectral area is calculated as the median frequency (MDF). Within a sliding window of 500 samples, the root mean square (RMS) and average rectified value (ARV) of the trend component are calculated. These three features are normalized to the 0-1 interval to form the fatigue feature vector. The fatigue feature vector is then input into the Mamba time-series prediction model. The Mamba model employs a selective state-space equation, with discrete forms h(t) = Ā(t)h(t-1) + B(t)z(t) and y(t) = C(t)h(t) + Dz(t). The state transition matrix, input matrix, and output matrix all rely on the current input to achieve a selective memory mechanism. The model is pre-trained and validated offline using sEMG time-series samples from multiple groups of rehabilitation patients. During the inference phase, it can directly output the muscle fatigue index. This embodiment uses a 4-layer stacked Mamba block, with each layer having a hidden state dimension of 64. After the model outputs the muscle fatigue index, it uses an exponential moving average for smoothing with a smoothing factor of 0.1. The final output muscle fatigue index FI ranges from 0 to 1, with a higher value indicating a higher degree of fatigue.

[0025] Step S103: Based on the human-computer interaction force signal, kinematic signal and physiological and psychological signal, generate the human-computer trust index.

[0026] Specifically, a trust state vector is constructed based on heart rate variability (HRV), skin conductance (SDR), human-computer interaction (HCI) force signals, and kinematic signals. This vector contains eight state components: SDNN (reflecting overall HRV), RMSSD (reflecting short-term HRV), peak SCR count f_SCR (reflecting emotional fluctuation frequency), HCI force deviation ε_F (the difference between actual interaction force and expected assistance force), trajectory following error e_θ (the difference between expected trajectory and actual angle), trajectory following error rate of change Δe_θ, the previous trust level Trust(t-1), and the current fatigue level FI(t). A larger HRV indicates higher parasympathetic nerve activity and higher trust; a larger SDR indicates greater emotional fluctuation and lower trust; and larger HCI force deviation and trajectory following error indicate stronger human-computer interaction resistance and lower trust.

[0027] The trust state vector is input into an improved Dueling Double DQN network, which outputs trust adjustment actions. These actions include five discrete levels: -2 corresponds to a significant decrease in stiffness, -1 to a slight decrease in stiffness, 0 to maintain stiffness, +1 to a slight increase in stiffness, and +2 to a significant increase in stiffness. The stiffness adjustments for each level are -40, -20, 0, +20, and +40 N·m / rad, respectively, and the corresponding damping adjustments are -16, -8, 0, +8, and +16 N·m / s / rad. The trust adjustment actions can simultaneously include damping adjustments, with the damping and stiffness adjustments matched in a fixed ratio to maintain the system's optimal damping ratio. The improved Dueling Double DQN network consists of an online network and a target network. The online network generates a predicted Q-value for each trust adjustment action based on the current trust state and selects the action with the highest predicted Q-value as the action to be executed. The target network generates a target Q-value for the action selected by the online network based on the current trust state.

[0028] After executing the pending action, the actual reward is obtained. This embodiment uses a multi-objective weighted reward function, where I(・) is an indicator function. The function value is 1 when the condition is met and 0 when the condition is not met. The specific calculation formula is: R = w_trust × R_trust + w_safety × R_safety + w_efficiency × R_efficiency. in: R_trust=[Trust(t)-Trust(t-1)-0.02] represents the reward for increased trust, with the portion exceeding the threshold of 0.02 counted as a reward; R_safety=-2.0×I(‖F_ext‖>50N)-1.5×I(|θ̇|>1.5rad / s) is penalized when the interaction force exceeds 50N or the joint speed exceeds 1.5rad / s; R_efficiency=0.5×exp(-|e_θ| / 0.1) is the rehabilitation efficiency reward calculated based on the trajectory tracking error e_θ, with a larger reward for a smaller error; w_trust=0.4, w_safety=0.3, and w_efficiency=0.3 are the weight coefficients for each item.

[0029] The parameters of the online network are updated based on the error between the target Q-value and the actual reward. Simultaneously, the online network decomposes the Q-value into a state value function and an action advantage function, i.e., Q(s,a) = V(s) + A(s,a) - mean(A), reflecting the characteristics of the Dueling architecture. This embodiment also employs a priority experience replay mechanism, prioritizing experience samples based on the absolute value of the temporal difference error δ, with priority p_t = |δ_t| + ε, where ε is set to 0.01 to prevent zero probability. The sampling probability P(i) = p_i^α / Σp_j^α, where α is set to 0.6. Importance sampling weights are used in the weighted loss function to eliminate sampling bias. Finally, a trust adjustment action is selected and executed based on the updated Q-value output by the online network. An ε-greedy strategy is used to balance exploration and exploitation. During the inference phase, ε is set to 0.01. The human-machine trust index Trust is updated and output based on the execution results, with a value ranging from 0 to 1; a larger value indicates higher trust. To avoid instantaneous fluctuations, this embodiment uses an exponential moving average to smooth the trust level, with a smoothing factor of 0.15, i.e., Trust_smooth(t) = 0.15 × Trust(t) + 0.85 × Trust_smooth(t-1).

[0030] Step S104: Generate a comprehensive human condition index based on the muscle fatigue index and the human-machine trust index.

[0031] Specifically, the coupled fatigue index and coupled trust index are first calculated based on the muscle fatigue index and the human-machine trust index, respectively, after their mutual influence. Then, the two coupled indices are weighted and summed to obtain the comprehensive human state index. When calculating the coupled indices, a dynamic coupling weight matrix W_couple is constructed. This matrix contains four elements: a first self-weight w11, a second self-weight w22, a first cross-weight w12, and a second cross-weight w21. All four elements are functions of the muscle fatigue index FI and the human-machine trust index Trust. The coupled fatigue index FI_coupled = w11 × FI + w12 × Trust and the coupled trust index Trust_coupled = w21 × FI + w22 × Trust are calculated using matrix multiplication. Here, the first self-weight w11 = α + β × (1 - Trust), the second self-weight w22 = α + β × FI, and the first cross-weight w12 and the second cross-weight w21 are both equal to γ ​​× (FI - Trust). α, β, and γ are preset constants. In this embodiment, the optimal values ​​for α (0.5), β (0.3), and γ (-0.1) were determined through experimental optimization. This set of parameters effectively increases the focus on fatigue under low trust conditions and the focus on trust under high fatigue conditions. Furthermore, a penalty is applied through cross-terms when there is a significant difference between fatigue and trust. The weights of the weighted sum are adaptively adjusted according to the rehabilitation stage, with the weight of fatigue gradually increasing and the weight of trust gradually decreasing as rehabilitation time progresses. This embodiment employs a linear adjustment rule: fatigue weight ω_F(t) = ω_F0 + Δω × (t / T_total), where ω_F0 is the initial fatigue weight (0.4), Δω is the weight change magnitude (0.2), t is the current rehabilitation duration, T_total is the total duration of a single rehabilitation session (30 minutes), and trust weight ω_T(t) = 1 - ω_F(t). The final output comprehensive human condition index CHSI = ω_F × FI_coupled + ω_T × (1-Trust_coupled) takes a value from 0 to 1. The larger the value, the worse the overall human condition (where 1-Trust_coupled is used to convert the trust index into a state deterioration degree that is semantically consistent with the fatigue index, ensuring the logical consistency of the weighted summation).

[0032] In this embodiment, CHSI is divided into four semantic ranges: less than 0.25 is the excellent state, corresponding to standard impedance and control strategies that can increase active training components; 0.25 to 0.5 is the normal state, corresponding to control strategies that moderately reduce impedance stiffness; 0.5 to 0.75 is the poor state, corresponding to control strategies that significantly reduce impedance and increase auxiliary force; and greater than 0.75 is the extremely poor state, corresponding to the maximum auxiliary / protection mode.

[0033] Step S105: In response to the comprehensive human condition index, adjust the impedance control parameters of the rehabilitation robot and generate torque control commands.

[0034] Specifically, the impedance parameters are first initialized. Based on the rehabilitation task type and the patient's initial state, the expected inertia M_d0 is set to 2.0 kg·m², the expected stiffness K_d0 to 80 N·m / rad, and the expected damping B_d0 is calculated using a damping ratio ζ=0.7 as B_d0=2ζ√(M_d0K_d0)=2×0.7×√(2.0×80)=17.7 N·m·s / rad. Then, the Comprehensive Human State Index (CHSI), its rate of change ΔCHSI, and the human-machine interaction force deviation ΔF are used as input variables for the fuzzy controller, where the human-machine interaction force deviation ΔF is the difference between the actual interaction force and the expected auxiliary force. Based on a preset fuzzy rule base, the stiffness adjustment ΔK and damping adjustment ΔB of the impedance parameters are obtained through inference. The weighted average method is used to resolve fuzziness, yielding the stiffness adjustment ΔK and damping adjustment ΔB. ΔK has an output range of -40 to 40 N·m / rad, and ΔB has an output range of -16 to 16 N·m·s / rad. The impedance control parameters are updated based on these stiffness and damping adjustments using the formulas K_d(t) = K_d(t-1) + λ_K × ΔK and B_d(t) = B_d(t-1) + λ_B × ΔB, where λ_K = 0.8 and λ_B = 0.6 are the update step sizes. The desired inertia M_d is then updated synchronously based on K_d and B_d, maintaining a stable damping ratio, using the formula M_d = B_d² / (4ζ²K_d), with the damping ratio ζ set to 0.7. The maximum permissible torque τ_max and maximum permissible velocity θ̇_limit are dynamically adjusted according to the Comprehensive Human State Index (CHSI), where τ_max = τ_nominal × (1 - CHSI) and θ̇_limit = θ̇_nominal × (1 - CHSI), τ_nominal is 30 N·m and θ̇_nominal is 2.0 rad / s. A multi-layered safety constraint mechanism is established: when the actual angular velocity θ̇_actual exceeds θ̇_limit, a braking torque τ_brake = -50 × θ̇_actual is applied; when the joint angle θ approaches the mechanical limit, virtual wall protection is activated (θ_min and θ_max are the intersection of the physiological activity safety range of the corresponding rehabilitation joint and the mechanical limit of the exoskeleton, preset according to the target rehabilitation joint); an emergency stop is triggered when ΔF / Δt > 3000 N / s or ΔCHSI / Δt > 0.2 / s.Based on the updated impedance parameters, the desired acceleration θ̈_cmd is calculated according to the target impedance equation M_dë+B_dė+K_de=F_ext. Then, the torque control command τ_cmd is generated by combining the exoskeleton dynamics model τ_cmd=M(θ)θ̈_cmd+C(θ,θ̇)θ̇+G(θ)+τ_comp (the above parameters M(θ), C(θ,θ̇), and G(θ) are obtained in advance through exoskeleton CAD modeling and system identification experiments). The friction compensation term τ_comp=F_c×sign(θ̇)+F_v×θ̇, where F_c is 0.5N・m and F_v is 0.1N・m・s / rad.

[0035] The method provided in this embodiment achieves precise perception and adaptive response of the rehabilitation robot to the patient's state by simultaneously assessing two dimensions: physiological fatigue and psychological trust, and dynamically coupling them for impedance control. In the early stages of rehabilitation when trust is low, the system appropriately increases its focus on fatigue to prevent muscle tension caused by distrust from accelerating fatigue. In the later stages of rehabilitation when fatigue is high, the system increases its focus on maintaining trust to avoid negative emotions caused by excessive fatigue. By smoothly mapping the comprehensive human state index to impedance parameter adjustment through a fuzzy controller, a transition from discrete mode to continuous adaptive control is achieved, significantly improving the comfort and human-machine synergy of rehabilitation training. Simultaneously, the mechanism of dynamically adjusting the safety threshold according to the comprehensive human state index automatically provides stricter safety protection for the patient when their condition is poor.

[0036] Example 2: The present invention will be described in detail below with reference to a specific embodiment. This embodiment provides a specific implementation of an adaptive control method for a rehabilitation robot based on two-dimensional state assessment.

[0037] The adaptive control method for rehabilitation robots disclosed in this embodiment first performs signal acquisition and preprocessing. During the signal acquisition stage, surface electromyography (EMG) signals, human-machine interaction force signals, kinematic signals, and physiological and psychological signals are acquired in real time. This constructs a multimodal synchronous perception system covering human physiological state, human-machine interaction state, joint movement state, and psychological and emotional state, providing high-quality, time-aligned input data for subsequent muscle fatigue assessment, human-machine trust modeling, and adaptive impedance control. Surface EMG signals are acquired by attaching Ag / AgCl surface electrodes to the target muscle group surface. The electrodes are attached along the muscle fiber direction with a 20mm electrode spacing, using differential amplification. The sampling frequency is set to 1000Hz to meet the Nyquist sampling requirements of 20-500Hz for the main frequency components of sEMG. All the above acquisition parameters strictly adhere to the internationally recognized gold standard for surface EMG signal acquisition, SENIAM (European Initiative for Standardization of Non-invasive Assessment of Surface Electromyography), and are combined with the core requirements of accurate real-time assessment of muscle fatigue and synchronous fusion of multimodal signals in the rehabilitation exoskeleton scenario of this embodiment. Targeted optimization, the selection criteria for each parameter, the technical problems solved, and the specific technical effects are as follows: Regarding the selection of Ag / AgCl surface electrodes, sEMG is an ultra-weak bioelectrical signal with an amplitude of only 10~500μV, which is easily overwhelmed by electrochemical noise, skin-electrode interface polarization interference, and limb motion artifacts in rehabilitation training. Ag / AgCl electrodes are chosen because they have extremely low polarization voltage and stable half-cell potential. Combined with conductive gel, they can significantly reduce skin-electrode interface impedance, greatly suppressing baseline drift and motion artifacts from the hardware source, stably capturing weak potential changes related to muscle fatigue, and solving the problem of limb motion artifacts in rehabilitation training. The technical challenge of poor signal fidelity caused by continuous patient movement during training is addressed by providing high-fidelity raw signals for subsequent fatigue feature extraction. Regarding the electrode attachment along the muscle fiber direction, sEMG essentially involves the superposition of potential signals on the skin surface as the action potentials of the target muscle group's motor units propagate along the longitudinal axis of the muscle fiber. Attaching two electrodes along the muscle fiber direction ensures that the two electrodes synchronously capture the temporal potential difference of the action potentials of the same muscle fiber bundle, maximizing the amplitude of the effective signal from the target muscle group while minimizing the acquisition range. This reduces crosstalk from nearby non-target muscle groups caused by multi-muscle group synergistic movements during rehabilitation training, ensuring that the acquired signal comes only from... The target muscle group is crucial for ensuring the specificity and accuracy of subsequent fatigue assessments. Regarding the selection of a 20mm electrode spacing, this directly determines the signal-to-noise ratio and crosstalk level of sEMG signal acquisition. Too small a spacing leads to insufficient potential difference between the two electrodes and severe attenuation of the effective signal amplitude. Too large a spacing expands the acquisition range, exacerbates crosstalk between adjacent muscle groups, and causes action potential phase cancellation, resulting in frequency domain distortions related to muscle fatigue. 20mm is a universally optimal spacing explicitly recommended by the SENIAM standard and adapted to the physiological characteristics of human skeletal muscle. This is particularly relevant for the biceps brachii, quadriceps femoris, and other core rehabilitation muscle groups involved in this embodiment.This method maximizes the effective signal amplitude while achieving an optimal balance between crosstalk suppression and signal-to-noise ratio. Simultaneously, the fixed quantization interval eliminates human variables in the acquisition process, ensuring the consistency and repeatability of data collected from different rehabilitation batches and subjects, providing a stable and unified input data source for subsequent fatigue assessment models. Regarding the use of differential amplification, the clinical acquisition environment for rehabilitation training contains strong common-mode interference such as 50Hz mains power frequency interference and spatial electromagnetic radiation, whose intensity can reach hundreds of times that of the effective sEMG signal. Single-ended amplification cannot effectively filter out this type of interference, directly leading to… The effective signal is completely submerged. The differential amplification method only linearly amplifies the potential difference at the dual-electrode input, i.e., the target electromyographic signal. It has a high suppression capability of over 100dB for common-mode interference received by both electrodes. It significantly filters out environmental interference and common-mode baseline drift caused by motion at the hardware level, linearly amplifying the weak electromyographic signal at the microvolt level to a range that the acquisition device can accurately identify. At the same time, it completely preserves the time-domain and frequency-domain characteristics related to muscle fatigue, solving the core technical problem that weak electromyographic signals are easily submerged by environmental interference. Regarding the setting of the 1000Hz sampling frequency, according to the Nyquist (N... According to the yquist sampling theorem, the sampling frequency for distortion-free reconstruction of the original continuous signal must be greater than twice the highest effective frequency of the signal; otherwise, frequency aliasing will occur, resulting in complete distortion of the signal's time and frequency domain characteristics, rendering all subsequent fatigue analyses meaningless. In this embodiment, over 95% of the effective energy of sEMG related to muscle fatigue assessment is concentrated in the 20-500Hz frequency band, with the highest effective frequency being 500Hz. Theoretically, the minimum compliant sampling frequency should be greater than 1000Hz. Setting the sampling frequency to 1000Hz can completely avoid the risk of frequency aliasing and fully preserve the core of muscle fatigue assessment. The sampling frequency exhibits characteristics such as a decrease in median frequency (MDF) and an increase in root mean square frequency (RMS) in the time domain. This avoids the massive data storage pressure and real-time processing delays associated with excessively high sampling frequencies, perfectly matching the real-time and accuracy requirements of the rehabilitation exoskeleton system in this embodiment for fatigue assessment. Furthermore, the 1ms time step corresponding to 1000Hz sampling provides a unified time reference for aligning the sEMG signal with the multi-source timestamps of human-computer interaction forces, kinematics, and physiological and psychological signals in this embodiment, ensuring the temporal consistency of multimodal data and laying the foundation for subsequent multi-dimensional data fusion and comprehensive human condition assessment. The above parameter settings form a standardized and highly robust sEMG acquisition scheme, ensuring the authenticity, accuracy, and consistency of the data from the signal acquisition source. This provides reliable underlying data support for subsequent precise muscle fatigue assessment, human-computer trust modeling, and adaptive impedance control in this embodiment. Those skilled in the art can repeatedly implement the sEMG signal acquisition function of this embodiment according to the above parameter settings and achieve the expected signal acquisition effect. In the above original sEMG signal,Low-frequency baseline drift, high-frequency environmental noise, and residual interference from 50Hz power line frequencies, which cannot be completely eliminated by the hardware acquisition process, still exist and will directly lead to distortion of subsequent fatigue feature extraction results. Therefore, the original sEMG signal is preprocessed in two steps: First, a fourth-order Butterworth zero-phase 20-450Hz bandpass filter is used. This frequency band is compatible with the main effective frequency components of sEMG, 20-500Hz, and a filter roll-off transition band is reserved to avoid high-frequency aliasing. This can effectively remove low-frequency baseline drift and high-frequency noise that are beyond the physiological effective range. Then, a second-order Butterworth 50Hz notch filter is applied to the bandpass-filtered signal to specifically eliminate power line frequency interference that cannot be completely suppressed by differential amplification. Finally, a preprocessed high signal-to-noise ratio sEMG signal is obtained, which provides accurate and effective input for subsequent singular spectrum analysis feature extraction and muscle fatigue assessment. The simultaneous acquisition and preprocessing of human-computer interaction force signals aims to fully characterize the multi-dimensional mechanical interaction state at the interface between the human body and the exoskeleton. To avoid measurement distortion caused by the inability of single-dimensional force sensors to capture eccentric torques, a six-dimensional force / torque sensor is employed and installed at the direct contact interface between the exoskeleton and the human limbs, such as the forearm support or thigh support. This directly acquires the triaxial interaction force and torque applied by the human body to the exoskeleton, avoiding measurement delays and errors caused by transmission clearances and mechanical friction when the sensor is installed at the joint pivot. The sampling frequency is set to 500Hz, which matches the standard control cycle of the rehabilitation exoskeleton system and meets the Nyquist sampling requirements for the effective dynamic frequency band of human-computer interaction force (0-200Hz). This allows for distortion-free reproduction of real-time dynamic changes in interaction force and output of the original six-dimensional force / torque signal. To address two inherent errors that the hardware acquisition process cannot completely suppress in the original signal acquisition stage—high-frequency vibration noise from the exoskeleton motor operation and mechanical vibration, and sensor temperature drift and static stress during assembly—a six-dimensional force / torque sensor is used. Zero-point offset errors caused by interference and errors directly distort the force feedback signal, reducing the compliance and stability of adaptive impedance control. They also lead to errors in human-computer interaction behavior feature extraction and affect the accuracy of human-computer trust assessment. To address this, a fourth-order Butterworth low-pass filter is first used to filter the original signal, setting the cutoff frequency to 200Hz. This effectively filters out high-frequency noise while preserving the effective mechanical features of the human-computer interaction. Before rehabilitation training begins, the exoskeleton and the human body are kept still, and sensor data is continuously collected for 2 seconds. The average value is used as the zero-point offset, which is then removed from the subsequently collected real-time signals to eliminate systematic zero-point errors. Finally, after calibration, a high-fidelity human-computer interaction force signal is output. This signal is transmitted in two paths: one to the impedance control module as the force feedback closed-loop input, and the other to the human-computer trust assessment module as the behavior feature input. Simultaneously, it enters the multi-source signal timestamp alignment stage synchronously with the pre-processed surface electromyography signal, forming the core input of the multimodal synchronous sensing system in this embodiment.

[0038] The acquisition and preprocessing of kinematic signals and physiological and psychological signals are completed simultaneously. Together with the aforementioned surface electromyography signals and human-computer interaction force signals, they constitute the multi-modal synchronous perception system of this invention, which covers human physiology, human-computer interaction, joint movement, and psychological emotions. This provides a complete input data source for subsequent muscle fatigue assessment, human-computer trust modeling, and adaptive impedance control.

[0039] Kinematic signals are directly read from the joint angles by installing absolute photoelectric encoders at the pivots of each active joint of the exoskeleton (such as the elbow and knee joints). The encoder resolution is set to 4096 pulses / revolution, and the sampling frequency is 500Hz. Absolute photoelectric encoders are chosen because they can directly output the absolute joint angles without cumulative error, and do not require re-zeroing after power failure, avoiding the risk of positional deviation during rehabilitation training startup, thus meeting the core safety requirements of the rehabilitation exoskeleton. Direct installation at the pivots eliminates angle measurement errors caused by transmission backlash and mechanism deformation, ensuring the accuracy of the angle data. The 12-bit encoder with 4096 pulses / revolution has an angle resolution of approximately 0.088°, fully meeting the ±0.1° control accuracy requirement of the rehabilitation exoskeleton, while balancing hardware cost and real-time processing efficiency. The 500Hz sampling frequency meets the Nyquist sampling requirements of the effective frequency band of joint movement and is consistent with the aforementioned human-machine interaction force signal sampling frequency, matching the exoskeleton's 2ms standard control cycle and reducing the computational load of subsequent multi-source synchronization. Angular velocity is obtained by numerically differentiating and filtering the angle signal, eliminating the need for an additional angular velocity sensor, thus reducing hardware costs and assembly complexity. Filtering also eliminates high-frequency counting noise from differential amplification, resulting in a smooth angular velocity signal that provides accurate position and velocity feedback for subsequent adaptive impedance control.

[0040] Physiological and psychological signals include heart rate variability (HRV) indicators and skin conductance response indicators, both of which provide objective physiological basis for the human-machine trust modeling of this invention. HRV is measured by collecting ECG signals via ECG electrodes attached to the chest at a sampling frequency of 250Hz. Chest attachment is the international standard position for ECG acquisition, ensuring a high signal-to-noise ratio. 250Hz is a common standard sampling rate for clinical ECG acquisition, satisfying the sampling requirements for accurate R-wave identification while balancing data processing volume and real-time performance. The collected ECG signals are analyzed by detecting the R-wave peak value, and the RR interval is calculated to obtain two time-domain indicators: SDNN and RMSSD. These are the international gold standard for HRV analysis, objectively quantifying the patient's autonomic nervous activity and anxiety / tension state. This solves the technical problem of not being able to quantify the patient's psychological trust state in real time during rehabilitation scenarios, and the computational load is extremely small, making it suitable for the real-time processing needs of embedded exoskeleton systems. The skin conductance response (SCR) is collected using a skin conductance sensor attached to the finger. A constant microcurrent of 0.5 μA is applied to measure the skin conductance value. The fingertip is the most sensitive part of the skin conductance response, which can accurately capture the conductance changes brought about by emotional arousal. 0.5 μA is the international safety standard for skin conductance acquisition, which is non-irritating and harmless to patients, while ensuring measurement accuracy. The skin conductance baseline is extracted by minimum value filtering through a sliding window, and the rapidly changing skin conductance response (SCR) is separated. This can characterize the patient's long-term tension level and instantaneous emotional fluctuations, providing a real-time indicator of emotional changes for trust assessment.

[0041] Since the sampling frequencies of the various sensors are different (sEMG@1kHz, force / kinematics / electrodermal transference@500Hz, heart rate variability@250Hz), in order to eliminate the phase deviation caused by different sampling rates and ensure the time consistency of multimodal data, a unified time reference is established with a step size of 1ms corresponding to the 1kHz sampling of the aforementioned sEMG signal. The signals are aligned to a unified time axis of 1kHz through linear interpolation. Linear interpolation has low computational cost, good real-time performance, and no additional signal distortion, which can perfectly adapt to the real-time computing requirements of the exoskeleton embedded system. Finally, a time-synchronized multimodal data packet is constructed and transmitted to the subsequent two-dimensional human state assessment stage, providing a unified high-quality input for multi-source data fusion decision-making.

[0042] Based on the preprocessed high signal-to-noise ratio sEMG signal, this embodiment uses the Singular Spectrum Analysis-Mamba (SSA-Mamba) time-series prediction model to achieve accurate real-time assessment of muscle fatigue. This provides core decision-making basis from a physiological dimension for subsequent comprehensive human condition assessment, human-machine trust modeling, and adaptive impedance control, as detailed below: First, singular value decomposition (SVD) was performed on the preprocessed surface electromyography (sEMG) signals to extract robust features related to muscle fatigue. The first step involved constructing a trajectory matrix by embedding a one-dimensional sEMG time series into the matrix. The analysis window length was set to 2000 samples, corresponding to a 2-second duration at a 1000Hz sampling frequency. This window length ensures signal decomposition accuracy while meeting the latency requirements of real-time assessment for rehabilitation exoskeletons. The embedding dimension was set to 500, conforming to the industry standard of using a window length of 1 / 4 to 1 / 2 for SVD embedding dimensions, resulting in a 500×1501-dimensional trajectory matrix. The second step involved performing singular value decomposition (SVD) on the trajectory matrix, obtaining multiple components arranged in descending order of singular values ​​and their corresponding singular values ​​σ1, σ2, ..., σ_L. The magnitude of the singular value corresponds to the energy proportion of each component. The third step involves principal component screening and signal reconstruction. The cumulative energy of the first k components is calculated as the ratio of the total energy of all components to the sum of the singular values ​​of the currently selected first k components to the sum of the singular values ​​of all components. k ranges from 1 to the total number of components. The components are sequentially added, and the first r principal components that cause the cumulative energy ratio to reach a preset threshold for the first time are selected to reconstruct the trend components. In this embodiment, the preset threshold is 0.95, which can completely preserve the effective components of the signal while filtering noise and motion artifacts. Statistical analysis of sEMG samples from multiple rehabilitation scenarios shows that r typically meets this threshold (usually 5). During reconstruction, a matrix retaining the first r principal components is first constructed, and then the standard diagonal mean method of singular spectrum analysis is used to restore it to a one-dimensional time series, obtaining the low-frequency trend components related to muscle fatigue.

[0043] Next, multi-dimensional complementary fatigue features are extracted to construct a fatigue feature vector: The first type is frequency domain features. Power spectrum estimation is performed on the preprocessed raw sEMG signal. The Welch method is used to calculate the power spectrum in segments. The number of segments is set to 8, with 256 samples per segment, and a Hamming window is added. This parameter setting can reduce the variance of the spectrum estimation and reduce spectral leakage while taking into account the real-time performance of the calculation. After obtaining the power spectral density, the frequency corresponding to the median of the power spectral area is calculated as the median frequency (MDF). The MDF is the gold standard frequency domain index for muscle fatigue assessment and can stably reflect the physiological changes of decreased muscle fiber conduction velocity during muscle fatigue. The second type is time domain features. Within a sliding window of 500 samples (corresponding to a duration of 0.5 seconds), the root mean square (RMS) value and average rectified value (ARV) of the trend component are calculated respectively. These two can reflect the changes in muscle activation and complement the MDF feature, solving the problem of poor robustness of single features. Finally, the three features MDF, RMS, and ARV were normalized to the 0 to 1 range to eliminate individual differences among different subjects and different muscle groups, thus forming a standardized fatigue feature vector.

[0044] The fatigue feature vector is then input into the Mamba time-series prediction model to achieve accurate prediction and quantitative output of fatigue development trends. The Mamba model is constructed using selective state-space equations, with discrete forms h(t) = Ā(t)h(t-1) + B(t)z(t) and y(t) = C(t)h(t) + Dz(t). The state transition matrix, input matrix, and output matrix all depend on the current input to achieve a selective memory mechanism, which can adaptively focus on slow-changing features related to fatigue and filter out invalid interference information. Compared with traditional LSTM and Transformer models, it has better long-range memory capabilities and lower computational complexity, perfectly adapting to the long-term real-time scenarios of rehabilitation exoskeletons. This embodiment uses a 4-layer stack of Mamba blocks, with each layer having a hidden state dimension of 64, which can meet the real-time inference requirements of embedded devices while ensuring the model's fitting ability. After the model outputs the muscle fatigue prediction results, it is smoothed by exponential moving average with a smoothing factor of 0.1 to avoid frequent jumps in fatigue values ​​that could cause fluctuations in subsequent control parameters. Finally, a standardized muscle fatigue index FI is output, with a value ranging from 0 to 1. The larger the value, the higher the degree of fatigue.

[0045] This embodiment combines clinical guidelines in rehabilitation medicine with the control requirements of exoskeletons, dividing the muscle fatigue index into four semantic intervals: FI less than 0.3 indicates mild or no fatigue, corresponding to normal rehabilitation training mode; 0.3 to 0.6 indicates moderate fatigue, corresponding to appropriately reducing resistance and increasing assistive force; 0.6 to 0.8 indicates severe fatigue, corresponding to significantly increasing assistive force and reducing training intensity; and greater than 0.8 indicates extreme fatigue, corresponding to recommending rest and activating the maximum assistive protection mode. This fatigue index will be synchronously transmitted to the subsequent fatigue-trust coupling weight calculation stage, working together with the human-machine trust index to complete a comprehensive assessment of the human body's condition.

[0046] Based on the heart rate variability index, skin conductance index, human-computer interaction force signal, and kinematic signal obtained from the aforementioned multimodal signal acquisition and preprocessing, as well as the muscle fatigue index FI output earlier, this embodiment constructs an 8-dimensional trust state vector and uses an improved Dueling Double DQN (ID3QN) reinforcement learning network to achieve real-time quantitative evaluation of human-computer trust, providing core decision-making basis in the psychological dimension for subsequent comprehensive human state evaluation and adaptive impedance control, as detailed below: First, a full-dimensional trust state vector is constructed, covering four dimensions: physiological and psychological, interactive behavior, temporal continuity, and fatigue coupling, with a total of eight state components. All inputs come from the standardized outputs of the aforementioned multimodal perception system, specifically including: The first category is physiological and psychological indicators, namely SDNN reflecting overall heart rate variability, RMSSD reflecting short-term parasympathetic activity, and peak SCR count f_SCR per unit time (reflecting the frequency of emotional fluctuations). Among them, the larger the heart rate variability index, the more relaxed the patient is, the higher the parasympathetic activity, and the higher the corresponding trust level; the larger the skin conductance index, the stronger the patient's emotional fluctuations and tension. The first category consists of three indicators: the human-computer interaction behavior indicators, namely the human-computer interaction force deviation ε_F (the difference between the actual interaction force and the expected assistance force), the trajectory following error e_θ (the difference between the expected trajectory and the actual angle), and the trajectory following error change rate Δe_θ. The larger the values ​​of these three indicators, the stronger the human-computer confrontation and the lower the patient's willingness to follow, which corresponds to a lower level of trust. The second category consists of temporal and coupling indicators, namely the trust level Trust(t-1) at the previous moment and the muscle fatigue index FI(t) at the current moment. These indicators ensure the temporal continuity of trust assessment and model the coupling relationship between fatigue and trust, forming a strong binding with the muscle fatigue assessment module mentioned above.

[0047] The aforementioned trust state vector is input into the improved Dueling Double DQN network. This network is designed for the sparse reward, long time series, and high safety requirements of rehabilitation exoskeleton scenarios. It integrates three major improvements: Dueling architecture, DoubleQ-learning, and priority experience replay, which solve the shortcomings of traditional DQN networks such as Q-value overestimation, slow convergence, and poor generalization ability. The network output is a confidence level adjustment action, with five discrete levels: -2 corresponds to a significant decrease in stiffness, -1 corresponds to a slight decrease in stiffness, 0 corresponds to maintaining stiffness, +1 corresponds to a slight increase in stiffness, and +2 corresponds to a significant increase in stiffness. The stiffness adjustment amounts corresponding to each level are -40, -20, 0, +20, and +40 N·m / rad, respectively, and the corresponding damping adjustment amounts are -16, -8, 0, +8, and +16 N·m·s / rad, respectively. This adjustment amount is fully compatible with the previously defined impedance stiffness safety range of 5.0~200.0 N·m / rad. The discrete level design takes into account both the fineness of control and the real-time performance of model inference. The confidence level adjustment action can simultaneously include the synchronous adjustment of damping. The damping adjustment amount is matched with the stiffness adjustment amount in a fixed ratio to maintain the stability of the optimal damping ratio of the system.

[0048] The improved Dueling Double DQN network comprises a dual-branch structure: an online network and a target network. The online network generates a predicted Q-value for each trust-adjusting action based on the current trust state and selects the trust-adjusting action with the highest predicted Q-value as the action to be executed. The target network generates a target Q-value for the action selected by the online network based on the current trust state. This dual-network structure eliminates the Q-value overestimation bias of traditional DQN. Simultaneously, the online network employs a Dueling architecture, decomposing the Q-value into a state value function and an action advantage function, i.e., Q(s,a)=V(s)+A(s,a)-mean(A), separating the value of the state itself from the relative merits of the action, thus improving the model's robustness to decisions under similar trust states.

[0049] After performing the actions to be performed, actual rewards are obtained. This embodiment adopts a multi-objective weighted reward function that aligns with the clinical priorities of rehabilitation, balancing the three core objectives of trust building, safety constraints, and rehabilitation efficiency. The specific calculation formula is R = w_trust × R_trust + w_safety × R_safety + w_efficiency × R_efficiency, where: the trust building reward R_trust = [Trust(t) - Trust(t-1) - 0.02], guiding the model to optimize towards increasing patient trust; the safety constraint reward R_safety = -2.0 × I(‖F_ext The force threshold is 50N - 1.5 × I(|θ̇|>1.5 rad / s), where 50N force and 1.5 rad / s speed are the general clinical safety standards for rehabilitation exoskeletons. The penalty is used to absolutely prevent patients from being injured. The rehabilitation efficiency reward is R_efficiency=0.5×exp(-|e_θ| / 0.1), which ensures the accuracy of trajectory following during rehabilitation training and avoids sacrificing rehabilitation effect for the sake of increasing trust. The weight coefficients are set to w_trust=0.4, w_safety=0.3, and w_efficiency=0.3, which are in line with the core requirements of trust priority, safety backup, and efficiency guarantee in rehabilitation scenarios.

[0050] This embodiment also employs a priority experience replay mechanism to improve model convergence speed and address the issue of sparse trust change events in rehabilitation scenarios. Experience samples are prioritized based on the absolute value of the temporal difference error δ, calculated using the formula p_t = |δ_t| + ε, where ε is set to 0.01 to prevent zero-error samples from being completely ignored and to ensure sampling diversity. The sampling probability is calculated using the formula P(i) = p_i^α / Σp_j^α, where α is set to 0.6 to balance the focused sampling of high-value samples with the uniform sampling of the entire sample set. The parameters of the online network are updated based on the temporal difference error between the target Q-value and the actual return. The target network is synchronized with the online network using a soft update method, with a soft update coefficient τ set to 0.001. The target network parameters are synchronized every 10 training steps to ensure training stability.

[0051] During the model inference phase, an ε-greedy strategy is employed, with ε set to 0.01. This ensures a 99% probability of executing the optimal action to maintain control stability while reserving 1% exploration space to address sudden changes in the patient's state. Trust-adjusting actions are selected and executed based on the updated Q-value of the online network output. The human-machine trust index, Trust, is updated and output based on the execution results, with a value ranging from 0 to 1. To avoid trust jumps caused by instantaneous emotional fluctuations, an exponential moving average is used to smooth the trust index, with a smoothing factor of 0.15, adapting to the dynamic changes in psychological state and balancing smoothness and real-time performance.

[0052] This embodiment combines clinical rehabilitation needs with resistance control strategies, dividing the human-machine trust index into four semantic ranges: Trust less than 0.3 indicates low trust or resistance, corresponding to a control mode of maximum compliance and minimum coercion; 0.3 to 0.6 indicates moderate trust or hesitation, corresponding to a control mode of moderate assistance and gradual adaptation; 0.6 to 0.8 indicates high trust or cooperation, corresponding to a normal rehabilitation training mode; and greater than 0.8 indicates high trust or initiative, corresponding to a mode where more challenging training can be added. The smoothed human-machine trust index is then simultaneously transmitted to the subsequent fatigue-trust coupling weight calculation stage, working together with the muscle fatigue index FI to complete a comprehensive assessment of the human body's state.

[0053] Based on the muscle fatigue index FI and human-machine trust index Trust obtained through real-time calculation, this embodiment generates a comprehensive human condition index CHSI through a fatigue-trust dynamic coupling mechanism and adaptive weighting throughout the rehabilitation cycle. This provides a unique core decision basis for subsequent adaptive impedance control, completely solving the problem of human condition assessment distortion caused by independent assessment of fatigue and trust and neglect of their inherent coupling relationship in existing technologies. Specifically, as follows: First, the coupling index resulting from the interaction between muscle fatigue and human-machine trust is calculated. A dynamic coupling weight matrix W_couple is constructed to model the inherent negative coupling relationship between the two. This matrix contains four adaptively adjusted weight elements: a first self-weight w11, a second self-weight w22, a first cross-weight w12, and a second cross-weight w21. All four elements are functions of the current muscle fatigue index FI and the human-machine trust index Trust, and can be dynamically updated according to the patient's real-time status. The standardized indices after coupling are calculated using matrix multiplication: Coupled fatigue index FI_coupled = w11 × FI + w12 × Trust, Coupled trust index Trust_coupled = w21 × FI + w22 × Trust.

[0054] In this embodiment, the first self-weight w11 = α + β × (1 - Trust) is the self-weight of the fatigue index, which adaptively increases as the patient's trust decreases, matching the clinical rule that "the lower the trust and the higher the muscle tension, the greater the reference value of the fatigue assessment." The second self-weight w22 = α + β × FI is the self-weight of the trust index, which adaptively increases as the patient's fatigue increases, matching the physiological-psychological coupling characteristic that "the higher the fatigue, the stronger the negative impact on the patient's trust." The first cross-weight w12 and the second cross-weight w21 are both equal to γ ​​× (FI - Trust), reflecting the cross-influence of the two and capturing the imbalance between fatigue and trust. In this embodiment, through controlled experiments with multiple groups of rehabilitation clinical samples, the preset constants α = 0.5, β = 0.3, and γ = -0.1 were determined. This set of parameters can fully reflect the coupling effect of the two and ensure that the coupled index remains stable within the standardized range of 0-1, avoiding jumps and distortions in the assessment results.

[0055] The two coupled indices are then weighted and summed to obtain the final Comprehensive Human State Index (CHSI). The weights are adaptively adjusted according to the rehabilitation training stages, with the weight of fatigue gradually increasing and the weight of trust gradually decreasing as rehabilitation time progresses. This aligns with the clinical rehabilitation cycle principle of "prioritizing the establishment of human-machine trust in the early stages of rehabilitation and focusing on preventing and controlling muscle fatigue in the middle and later stages." This embodiment adopts a linear adjustment rule that is easy to implement in engineering and has smooth changes. The fatigue weight ω_F(t) = ω_F0 + Δω × (t / T_total) is used, where the initial fatigue weight ω_F0 is 0.4, the weight adjustment range Δω throughout the cycle is 0.2, the total duration of a single rehabilitation training session T_total is taken as the clinical standard of 30 minutes, and the trust weight ω_T(t) = 1 - ω_F(t) ensures that the sum of the weights is always 1, eliminating assessment bias.

[0056] The final output of the Comprehensive Human State Index (CHSI) is calculated using the formula: CHSI = ω_F × FI_coupled + ω_T × (1 - Trust_coupled), with a value ranging from 0 to 1. A higher value indicates a worse overall patient condition, consistent with the semantic definitions of fatigue and trust levels mentioned earlier. This embodiment, combining clinical rehabilitation needs with impedance control strategies, divides CHSI into four semantic intervals: less than 0.25 represents an excellent state, corresponding to standard impedance and a control strategy that can increase active training components; 0.25 to 0.5 represents a normal state, corresponding to a control strategy that moderately reduces impedance stiffness; 0.5 to 0.75 represents a poor state, corresponding to a control strategy that significantly reduces impedance and increases assistance; and greater than 0.75 represents an extremely poor state, corresponding to the maximum assistance / protection mode. The final output CHSI is transmitted in real-time to the subsequent fuzzy adaptive impedance control module as the core decision input for impedance parameter adjustment.

[0057] Based on the Comprehensive Human State Index (CHSI) output in real time by the aforementioned fusion decision module, this embodiment constructs a CHSI-driven three-input fuzzy adaptive impedance controller to achieve a real-time compliant response of the rehabilitation robot to the patient's dual-dimensional states of physiological fatigue and psychological trust. This controller serves as the execution layer of the technical solution of this invention, as detailed below: First, impedance parameters are initialized. Based on the type of rehabilitation task and the patient's initial overall condition, the expected nominal inertia value M_d0 is set to 2.0 kg·m², which is the industry-standard nominal inertia for upper limb elbow joint and lower limb knee joint rehabilitation exoskeletons, closely matching the actual inertia range of human limbs. The expected nominal stiffness value K_d0 is set to 80 N·m / rad, which is a compromise initial value between passive assistance and active following modes in rehabilitation training, taking into account both trajectory accuracy and patient comfort. The expected damping B_d0 is calculated based on the optimal damping ratio ζ=0.7 of the second-order system, with the formula B_d0=2ζ√(M_d0K_d0), and the final value is 17.7 N·m·s / rad, ensuring that the system has no overshoot, no oscillation, and good compliance and response speed. At the same time, a safe adjustment range constraint for the impedance parameters is set to avoid patient injury or failure of rehabilitation effect due to parameter exceeding the limit. Specifically, the expected inertia M_d∈[0.5,5.0]kg・m², the expected stiffness K_d∈[5,200]N・m / rad, and the expected damping B_d∈[1,50]N・m・s / rad are completely matched with the stiffness adjustment level of the trust adjustment module mentioned above, ensuring the consistency of the control logic.

[0058] Then, in response to the comprehensive human condition index (CHSI), the impedance control parameters of the rehabilitation robot are adjusted in real time and torque control commands are generated. Specifically, the comprehensive human condition index (CHSI), the rate of change of CHSI (ΔCHSI), and the human-computer interaction force deviation (ΔF) are used as input variables of the fuzzy controller. ΔCHSI is the change in CHSI within a 1-second time window, and ΔF is the difference between the actual interaction force and the expected assistive force. These three inputs correspond to the patient's overall condition, the trend of condition change, and the real-time human-computer interaction, respectively, achieving full-dimensional control of "condition perception - trend prediction - real-time correction." This solves the core defect of traditional impedance control, which cannot simultaneously consider the patient's physiological fatigue and psychological trust.

[0059] Fuzzy inference is performed based on a pre-defined fuzzy rule base constructed from the knowledge of rehabilitation clinical experts to obtain the stiffness adjustment amount ΔK and damping adjustment amount ΔB of the impedance parameters. The fuzzy rule base used in this embodiment includes a main rule table and a correction rule table. The main rule table determines the basic adjustment amount of the impedance parameters based on the patient's overall condition and changing trends. The correction rule table compensates and corrects the adjustment amount based on real-time human-computer interaction forces, as detailed below: Main rule table (determining ΔK and ΔB based on CHSI and ΔCHSI): rule CHSI ΔCHSI ΔK ΔB R1 VL N PS ZE R2 VL Z ZE ZE R3 VL P NS PS R4 L N PS NS R5 L Z ZE ZE R6 L P NS PS R7 M N ZE NS R8 M Z ZE ZE R9 M P NS PS R10 H N NS ZE R11 H Z NB PS R12 H P NB PB R13 VH * NB PB Correction rule table (correcting ΔK and ΔB based on ΔF): rule ΔF ΔK correction ΔB correction R14 NB +PS -NS R15 ZE 0 0 R16 PB -NB +PS The aforementioned fuzzy rules are constructed based on the experience of rehabilitation clinical experts. The core principle of rehabilitation medicine is that "the better the patient's overall condition, the higher the impedance stiffness and the stronger the training challenge; the worse the patient's overall condition, the lower the impedance stiffness and the higher the damping, and the higher the priority of compliance and safety." This ensures that the control strategy is fully adapted to the clinical rehabilitation needs.

[0060] The fuzzy set of the input variables is defined as follows: CHSI is divided into five levels: VL (very low), L (low), M (medium), H (high), and VH (very high). VL corresponds to CHSI < 0.25, L corresponds to 0.25 ≤ CHSI < 0.5, M corresponds to 0.5 ≤ CHSI < 0.75, H corresponds to 0.75 ≤ CHSI < 0.9, and VH corresponds to CHSI ≥ 0.9, which completely corresponds to the semantic range of CHSI mentioned above; ΔCHSI is divided into three levels: N (negative), Z (zero), and P (positive); ΔF is divided into three levels: NB (large negative), ZE (zero), and PB (large positive). The fuzzy sets of output variables ΔK and ΔB are divided into five levels: NB (negative large), NS (negative small), ZE (zero), PS (positive small), and PB (positive large). The corresponding output center values ​​are: ΔK corresponds to -40, -20, 0, +20, and +40 N·m / rad, which perfectly matches the stiffness level of the confidence adjustment mentioned above; ΔB corresponds to -16, -8, 0, +8, and +16 N·m·s / rad, which maintains a fixed ratio with ΔK to maintain the optimal damping ratio of the system.

[0061] In the fuzzy inference process, the activation degree of each rule is calculated using a product operator to achieve AND logic operations on multiple input conditions. Finally, a weighted average method is used for defuzzification to obtain precise stiffness adjustment ΔK and damping adjustment ΔB. This method has extremely low computational cost and can perfectly adapt to the real-time control requirements of the rehabilitation exoskeleton embedded system. The impedance parameters are updated based on the adjustment values ​​obtained from defuzzification. Combined with the exoskeleton dynamics model and real-time interactive forces and kinematic signals, joint torque control commands are finally generated and sent to the exoskeleton actuators for execution, forming a complete technical closed loop of "multimodal perception → dual-dimensional evaluation → fusion decision → adaptive control → execution feedback".

[0062] Based on the stiffness adjustment ΔK and damping adjustment ΔB output from the defuzzification output of the aforementioned fuzzy controller, this embodiment completes the impedance parameter update, torque control command generation, and constructs a multi-layered safety constraint mechanism driven by CHSI. This mechanism serves as the execution and safety assurance layer of the technical solution of this invention, as detailed below: First, impedance parameters are smoothly updated and inertia is adjusted synchronously to ensure the system is always in the optimal damping state. The update formulas are K_d(t)=K_d(t-1)+0.8×ΔK and B_d(t)=B_d(t-1)+0.6×ΔB. Partial updates are used instead of full updates to avoid the adjustment amount of the fuzzy controller output directly causing impedance parameter jumps, greatly improving the smoothness of control and avoiding discomfort or resistance from patients due to parameter mutations. The coefficients 0.8 / 0.6 were optimized and determined through controlled experiments on multiple groups of rehabilitation patients. The stiffness update coefficient is slightly larger (0.8) to ensure control responsiveness, and the damping update coefficient is slightly smaller (0.6) to prioritize system stability, perfectly matching the core requirement of "balancing responsiveness and stability" of rehabilitation exoskeletons. The updated parameters are strictly limited to the safe adjustment range set above (K_d∈[5,200]N・m / rad, B_d∈[1,50]N・m・s / rad), with no risk of exceeding the limit. Simultaneously, the expected inertia M_d is updated synchronously based on K_d and B_d. The calculation formula is M_d=B_d² / (4×0.7²×K_d), which is the standard derivation formula for maintaining a fixed optimal damping ratio ζ=0.7 in a second-order impedance system. This ensures that no matter how K_d and B_d are adjusted, the system is always in the optimal state near the critical damping, without overshoot or oscillation, thus avoiding system instability that could lead to patient injury. ζ=0.7 directly uses the optimal damping ratio initialized by impedance mentioned earlier, ensuring complete consistency of the control logic before and after.

[0063] Then, a CHSI-driven dynamic safety threshold and multi-layered safety constraint mechanism are constructed to achieve an adaptive safety strategy of "increasing training challenge when the patient is in good condition and tightening safety protection when the patient is in poor condition." This is the core innovation of this solution, which distinguishes it from existing technologies with fixed safety thresholds. The dynamic safety threshold is adjusted in real time according to CHSI: the maximum allowable torque τ_max = 30 × (1-CHSI) N·m and the maximum allowable speed θ̇_limit = 2.0 × (1-CHSI) rad / s, where 30 N·m and 2.0 rad / s are the general clinical maximum safety thresholds for rehabilitation exoskeletons, corresponding to the case of CHSI = 0 (optimal patient condition); when CHSI = 1 (extremely poor patient condition), the safety threshold automatically drops to 0, and the maximum protection mode is activated, which completely corresponds to the semantic range of CHSI mentioned above. The multi-layered safety constraints form a complete safety chain of "torque limiting → speed monitoring → angle boundary → mutation detection," protecting patient safety from the source to the execution end in all aspects. Torque limiting: Limits the generated torque control command τ_cmd to the range of [-τ_max, τ_max] to prevent excessive torque from harming the patient; Speed ​​monitoring: When the actual angular velocity θ̇_actual exceeds θ̇_limit, a braking torque τ_brake=-50×θ̇_actual is applied. 50N・m・s / rad is the universal braking damping coefficient of the rehabilitation exoskeleton, which can quickly and smoothly reduce the angular velocity to a safe range without the impact caused by hard braking. Virtual wall protection: When the joint angle θ approaches the mechanical limit, an elastic virtual wall is activated. If θ < θ_min + 0.1 rad, apply τ_boundary = 500 × (θ_min + 0.1 - θ). If θ > θ_max - 0.1 rad, apply τ_boundary = 500 × (θ_max - 0.1 - θ). 0.1 rad is the boundary margin to prevent the joint from directly impacting the mechanical limit and causing damage. 500 N·m / rad is the virtual wall stiffness, providing a flexible transition and improving the patient experience. Emergency stop for mutation detection: When ΔF / Δt>3000N / s (corresponding to the patient's sudden violent struggle, such as pain or fear) or ΔCHSI / Δt>0.2 / s (corresponding to the patient's condition deteriorating rapidly), an emergency stop is triggered to cut off the exoskeleton's power output. This is the last resort safety measure and complies with clinical emergency care guidelines.

[0064] Finally, based on the updated impedance parameters, joint torque control commands are generated. First, the desired acceleration θ̈_cmd is calculated according to the target impedance equation M_dë + B_dė + K_de = F_ext, where the position error e = θ_ref - θ_actual, the velocity error ė = θ̇_ref - θ̇_actual, θ_ref and θ̇_ref are the desired rehabilitation trajectory and velocity, θ_actual and θ̇_actual are the outputs of the kinematics acquisition module mentioned earlier, and F_ext is the output of the human-computer interaction force acquisition module mentioned earlier, completely eliminating logical gaps. Then, combined with the exoskeleton dynamics model τ_cmd = M (θ)θ̈_cmd + C (θ,θ̇)θ̇ + G (θ) + τ_comp, the final torque control command is generated, where M (θ), C (θ,θ̇), and G (θ) are the exoskeleton's inertia matrix, centrifugal force and Coriolis force matrix, and gravity vector, respectively (these parameters are obtained through exoskeleton CAD). Model modeling and system identification experiments (obtained in advance) can eliminate the influence of the exoskeleton's own dynamics, making the actual system behavior approximate the desired impedance characteristics and improving control accuracy; τ_comp=0.5×sign(θ̇)+0.1×θ̇ is the friction compensation term, 0.5×sign(θ̇) is Coulomb friction compensation, and 0.1×θ̇ is viscous friction compensation, which can eliminate the frictional nonlinearity of the exoskeleton joints, improve the control accuracy during low-speed movement, and avoid patient discomfort caused by low-speed crawling.

[0065] The final generated torque control command is limited and corrected by a multi-layer safety constraint mechanism before being sent to the exoskeleton actuator for execution. The interactive force and kinematic signals collected in real time are fed back to the perception and evaluation module mentioned above, forming a complete technical closed loop of multimodal perception → dual-dimensional evaluation → fusion decision → adaptive control → safety assurance → execution feedback.

[0066] This embodiment uses a standard 30-minute single rehabilitation training cycle as an example to detail the collaborative working logic of each module of the present invention and fully reproduce the entire process of the invention: In the initial rehabilitation stage at t=0s, the patient's muscles are not fatigued, and the muscle fatigue index FI=0.2; however, the patient is nervous when using the exoskeleton for the first time, with low heart rate variability and high skin conductance, resulting in a human-machine trust index Trust=0.3. Based on the dynamic coupling weight formula mentioned above, the coupling fatigue index FI_coupled=0.145 and the coupling trust index Trust_coupled=0.17 are calculated; combined with the initial rehabilitation weights ω_F=0.4 and ω_T=0.6, the comprehensive human state index CHSI=0.16 is calculated, indicating an excellent state. The fuzzy controller matches rule R2, outputting ΔK=0 and ΔB=0, maintaining the initial normal impedance parameters of the exoskeleton, while the dynamic safety threshold is adjusted to τ_max=25.2N・m and θ̇_limit=1.68rad / s. At t=15min during the mid-rehabilitation period, the patient's muscles began to accumulate fatigue as training progressed, with FI=0.5. After 15 minutes of adaptation, the patient had established basic trust in the exoskeleton, with Trust=0.7. Based on the dynamic coupling weight formula mentioned earlier, FI_coupled=0.309 and Trust_coupled=0.465 were calculated. Combining the adaptively adjusted weights ω_F=0.5 and ω_T=0.5 during the mid-rehabilitation period, CHSI=0.387 was calculated, which is in a normal state, and CHSI shows a positive upward trend with fatigue accumulation. Based on the input of moderate CHSI and a positive rate of change, the fuzzy controller matched the rule base R9 mentioned earlier, outputting ΔK=-20 and ΔB=+8. The robot reduced stiffness and increased damping, making the operation mode more compliant, while the dynamic safety threshold was tightened to τ_max=18.4N・m and θ̇_limit=1.23rad / s. At the end of rehabilitation (t=30min), the patient's muscles showed significant fatigue, with FI=0.8. After full-course training and adaptation, the patient had established a stable high level of trust in the exoskeleton, with Trust maintained at 0.8. Based on the dynamic coupling weight formula mentioned earlier, FI_coupled=0.448 and Trust_coupled=0.592 were calculated. Combining the adaptively adjusted weights ω_F=0.6 and ω_T=0.4 at the end of rehabilitation, CHSI=0.506 was calculated, indicating a poor state, and CHSI showed a positive upward trend with continuous fatigue accumulation. Based on the high CHSI and positive rate of change input, the fuzzy controller matched the aforementioned rule base R12, outputting ΔK=-40 and ΔB=+16. The robot entered a compliant protection mode, and the dynamic safety threshold was further tightened to τ_max=14.8N・m and θ̇_limit=0.99rad / s.

[0067] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can be adapted and modified in various ways to suit different rehabilitation scenarios and individual patient differences, adjusting the parameters and rules of each module accordingly. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0068] Example 3: This example also provides an adaptive control system for a rehabilitation robot based on two-dimensional state assessment. This device is used to implement the above embodiments and preferred embodiments, and details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that implements a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0069] This embodiment provides an adaptive control system for a rehabilitation robot based on dual-dimensional state assessment, including: The signal acquisition module 201 is used to acquire surface electromyography signals, human-computer interaction force signals, kinematic signals and physiological and psychological signals, wherein the physiological and psychological signals include heart rate variability indicators and skin conductance response indicators. The muscle fatigue assessment module 202 is connected to the signal acquisition module 201 and is used to generate a muscle fatigue index based on the surface electromyography signal. The human-machine trust assessment module 203 is connected to the signal acquisition module 201 and is used to generate a human-machine trust index based on the human-machine interaction force signal, kinematic signal and physiological and psychological signal. The fusion decision module 204 connects the muscle fatigue assessment module 202 and the human-machine trust assessment module 203, and is used to generate a comprehensive human state index based on the muscle fatigue index and the human-machine trust index. The adaptive control module 205, connected to the fusion decision module 204 and the signal acquisition module 201, is used to adjust the impedance control parameters of the rehabilitation robot in response to the comprehensive human state index and generate torque control commands.

[0070] In some optional implementations, the muscle fatigue assessment module 202 includes: The singular spectrum analysis unit 2021 is used to perform singular spectrum analysis on the surface electromyography signal, construct a trajectory matrix and perform singular value decomposition, and select the first r principal components that make the cumulative energy ratio reach a preset threshold for the first time to reconstruct the trend component. The feature extraction unit 2022 is used to extract the root mean square value and average rectified value of the trend component, as well as the median frequency of the original signal, to form a fatigue feature vector. The Mamba time series prediction unit 2023 is used to input the fatigue feature vector into the Mamba time series prediction model and output the muscle fatigue index.

[0071] In some optional implementations, the human-machine trust assessment module 203 includes: The trust state vector construction unit 2031 is used to construct a trust state vector based on the heart rate variability index, skin conductance index, human-computer interaction force signal, and kinematic signal. The trust state vector contains eight-dimensional state components, specifically including: total heart rate variability SDNN, short-term heart rate variability RMSSD, peak skin conductance response f_SCR per unit time, human-computer interaction force deviation ε_F, trajectory following error e_θ, trajectory following error change rate Δe_θ, previous trust level Trust(t-1), and current fatigue level FI(t). The improved Dueling Double DQN network unit 2032 includes an online network and a target network. The online network is used to generate an estimated Q value for each trust level and select an action to be executed based on the current trust state. The target network is used to generate a target Q value for the action to be executed and update the parameters of the online network based on the error between the target Q value and the actual reward. At the same time, the Q value is decomposed into a state value function and an action advantage function. The trust update unit 2033 is used to select and execute a trust adjustment action based on the updated Q value of the online network output, and update and output the human-machine trust index based on the execution result.

[0072] In some optional implementations, the fusion decision module 204 includes: The dynamic coupling weight matrix calculation unit 2041 is used to construct a dynamic coupling weight matrix, wherein the elements of the matrix are functions of the muscle fatigue index and the human-machine trust index, and the coupled fatigue index and the coupled trust index are calculated according to matrix operations. CHSI generation unit 2042 is used to perform a weighted summation of the coupled fatigue index and the coupled trust index to obtain a comprehensive human condition index, wherein the weights of the weighted summation are adaptively adjusted according to the rehabilitation stage.

[0073] In some alternative implementations, the adaptive control module 205 includes: The fuzzy controller 2051 is used to take the comprehensive human body state index, its rate of change, and human-computer interaction force deviation as input variables, perform inference according to the preset fuzzy rule base, and output the stiffness adjustment amount and damping adjustment amount of the impedance parameter. The impedance parameter update unit 2052 is used to update the desired stiffness and desired damping according to the stiffness adjustment amount and the damping adjustment amount, and to update the desired inertia synchronously according to the principle of maintaining the damping ratio stable. The safety threshold linkage unit 2053 is used to dynamically adjust the maximum permissible torque and maximum permissible speed according to the comprehensive human body state index, and to establish a multi-layer safety constraint mechanism. The torque calculation unit 2054 is used to calculate the desired acceleration based on the updated impedance parameters and the target impedance equation, and then generate torque control commands by combining the exoskeleton dynamics model.

[0074] Further functional descriptions of the above modules and units are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0075] Example 4: In this example, the adaptive control device for a rehabilitation robot based on dual-dimensional state assessment is presented in the form of a functional unit. Here, a unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0076] This invention also provides a computer device having the above-described features. Figure 7 The diagram shows an adaptive control device for a rehabilitation robot based on two-dimensional state assessment.

[0077] Please see Figure 7 , Figure 7 This is a schematic diagram of the structure of a computer device provided in an optional embodiment of the present invention, such as... Figure 7 As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 7 Take a processor 10 as an example.

[0078] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0079] The memory 20 stores instructions executable by at least one processor 10 to cause the at least one processor 10 to perform the method shown in the above embodiments.

[0080] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0081] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0082] The computer device also includes an input device 30 and an output device 40. The processor 10, memory 20, input device 30, and output device 40 can be connected via a bus or other means. Figure 7 Taking the example of a connection between China and Israel via a bus.

[0083] Input device 30 can receive input numerical or character information, and generate key signal inputs related to user settings and function control of the computer device, such as a touchscreen, keypad, mouse, trackpad, touchpad, joystick, one or more mouse buttons, trackball, joystick, etc. Output device 40 may include display devices, auxiliary lighting devices (e.g., LEDs), and haptic feedback devices (e.g., vibration motors). The aforementioned display devices include, but are not limited to, liquid crystal displays, light-emitting diodes, displays, and plasma displays. In some alternative embodiments, the display device may be a touchscreen.

[0084] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0085] A portion of this application can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to this application through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0086] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. An adaptive control method for rehabilitation robots based on two-dimensional state assessment, applied to rehabilitation robots, characterized in that, Includes the following steps: Acquire surface electromyography signals, human-computer interaction force signals, kinematic signals, and physiological and psychological signals, wherein the physiological and psychological signals include heart rate variability indicators and skin conductance response indicators; Based on the surface electromyography signal, a muscle fatigue index is generated; Based on the human-computer interaction force signals, kinematic signals, and physiological and psychological signals, a human-computer trust index is generated. Based on the muscle fatigue index and the human-machine trust index, a comprehensive human condition index is generated. In response to the comprehensive human condition index, the impedance control parameters of the rehabilitation robot are adjusted to generate torque control commands.

2. The method according to claim 1, characterized in that, The generation of a muscle fatigue index based on the surface electromyography signal specifically includes: Singular spectrum analysis was performed on the surface electromyography signal to construct a trajectory matrix and perform singular value decomposition to obtain multiple components arranged in descending order of singular values. Calculate the proportion of the cumulative energy of the first k components to the total energy of all components. The cumulative energy proportion is the ratio of the sum of the singular values ​​of the currently selected first k components to the sum of the singular values ​​of all components, where k = 1, 2, ..., n, and n is the total number of components. The trend component is obtained by sequentially accumulating the components and selecting the first r principal components that cause the cumulative energy ratio to reach a preset threshold for the first time. The root mean square value and average rectified value of the trend component, along with the median frequency of the original signal, are extracted to form a fatigue feature vector. The fatigue feature vector is input into the Mamba time series prediction model, which outputs a muscle fatigue index.

3. The method according to claim 1, characterized in that, The generation of the human-computer trust index based on the human-computer interaction force signals, kinematic signals, and physiological and psychological signals specifically includes: Based on the heart rate variability index, skin conductance index, human-computer interaction force signal, and kinematic signal, a trust state vector is constructed. The trust state vector contains eight state components, specifically including: total heart rate variability SDNN, short-term heart rate variability RMSSD, peak skin conductance response f_SCR per unit time, human-computer interaction force deviation ε_F, trajectory following error e_θ, trajectory following error change rate Δe_θ, previous trust level Trust(t-1), and current fatigue level FI(t). The trust state vector is input into an improved Dueling Double DQN network, which is used to output a trust adjustment action, including at least one of increasing stiffness, decreasing stiffness, increasing damping, and decreasing damping. The improved Dueling Double DQN network comprises an online network and a target network, wherein: The online network is used to generate an estimated Q value for each trust adjustment action based on the current trust status, and select the trust adjustment action with the highest estimated Q value as the action to be executed. The target network is used to generate a target Q value for the action to be executed selected by the online network based on the current trust state; The parameters of the online network are updated based on the error between the target Q value and the actual reward obtained after performing the action to be performed. The actual reward is determined based on the changes in heart rate variability, skin conductance response, and human-computer interaction force deviation after performing the trust adjustment action. The online network decomposes the Q-value into a state value function and an action advantage function; Based on the updated Q value output from the online network, select and execute the trust adjustment action, and update and output the human-machine trust index based on the execution result.

4. The method according to claim 1, characterized in that, The generation of a comprehensive human condition index based on the muscle fatigue index and the human-machine trust index specifically includes: Based on the muscle fatigue index and the human-machine trust index, calculate the coupled fatigue index and coupled trust index after their mutual influence. The combined human body state index is obtained by weighted summation of the coupling fatigue index and the coupling trust index. The weights of the weighted sum are adaptively adjusted according to the rehabilitation stage, and as the rehabilitation time progresses, the weight of fatigue gradually increases and the weight of trust gradually decreases.

5. The method according to claim 4, characterized in that, The calculation of the coupled fatigue index and coupled trust index, based on the muscle fatigue index and the human-machine trust index, after their mutual influence, is specifically achieved through the following method: Construct a dynamic coupling weight matrix, which contains four elements: a first self-weight, a second self-weight, a first cross weight, and a second cross weight. All four elements are functions of the muscle fatigue index FI and the human-machine trust index Trust. The coupling fatigue index FI_coupled is calculated according to the formula: First self weight × FI + First cross weight × Trust; The coupling trust index is calculated according to the formula: Trust_coupled = Second Cross Weight × FI + Second Self Weight × Trust; Wherein, the first self-weight w11 = α + β×(1-Trust), the second self-weight w22 = α + β×FI, and the first cross weight w12 and the second cross weight w21 are both γ×(FI-Trust), where α, β and γ are preset constants.

6. The method according to claim 1, characterized in that, The process of adjusting the impedance control parameters of the rehabilitation robot in response to the comprehensive human condition index and generating torque control commands specifically includes: The Comprehensive Human State Index (CHSI), its rate of change ΔCHSI, and the human-computer interaction force deviation ΔF are used as input variables of the fuzzy controller, where the human-computer interaction force deviation ΔF is the difference between the actual interaction force and the expected auxiliary force. The stiffness adjustment amount ΔK and damping adjustment amount ΔB of the impedance parameters are obtained by reasoning based on the preset fuzzy rule base. The fuzzy rule base includes a main rule table that judges based on the comprehensive human body state index and its rate of change, and a correction rule table that corrects based on the human-computer interaction force deviation. The impedance control parameters are updated according to the stiffness adjustment amount ΔK and the damping adjustment amount ΔB, including updating the desired stiffness K_d and the desired damping B_d, and the desired inertia M_d is updated synchronously according to K_d and B_d in accordance with the principle of keeping the damping ratio stable. The maximum permissible torque τ_max and the maximum permissible speed θ̇_limit are dynamically adjusted according to the Comprehensive Human State Index (CHSI), where τ_max = τ_nominal × (1 - CHSI), θ̇_limit = θ̇_nominal × (1 - CHSI), τ_nominal is the nominal maximum torque preset according to the driving capability of the rehabilitation robot, and θ̇_nominal is the nominal maximum speed preset according to the range of motion of the rehabilitation robot; Based on the updated impedance parameters, the desired acceleration θ̈_cmd is calculated according to the target impedance equation M_d ë + B_d ė + K_d e = F_ext, where M_d is the desired inertia, B_d is the desired damping, K_d is the desired stiffness, e is the position tracking error, and F_ext is the external interaction force. Then, the torque control command τ_cmd is generated by combining the exoskeleton dynamic model τ_cmd = M(θ)θ̈_cmd + C(θ,θ̇)θ̇ + G(θ) + τ_comp, where M(θ), C(θ,θ̇), and G(θ) are dynamic parameters obtained in advance through system identification or CAD model, and τ_comp is the friction compensation term.

7. An adaptive control system for a rehabilitation robot based on dual-dimensional state assessment, characterized in that, include: The signal acquisition module is used to acquire surface electromyography signals, human-computer interaction force signals, kinematic signals, and physiological and psychological signals, including heart rate variability indicators and skin conductance response indicators. A muscle fatigue assessment module, connected to the signal acquisition module, is used to generate a muscle fatigue index based on the surface electromyography signal. The human-machine trust assessment module is connected to the signal acquisition module and is used to generate a human-machine trust index based on the human-machine interaction force signal, kinematic signal and physiological and psychological signal. The fusion decision module connects the muscle fatigue assessment module and the human-machine trust assessment module, and is used to generate a comprehensive human state index based on the muscle fatigue index and the human-machine trust index. An adaptive control module, connected to the fusion decision module and the signal acquisition module, is used to adjust the impedance control parameters of the rehabilitation robot in response to the comprehensive human state index and generate torque control commands.

8. A computer device, characterized in that, include: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the adaptive control method for a rehabilitation robot based on dual-dimensional state assessment as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the adaptive control method for a rehabilitation robot based on two-dimensional state assessment as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, Includes computer instructions for causing a computer to execute the adaptive control method for a rehabilitation robot based on two-dimensional state assessment as described in any one of claims 1 to 6.

Images