Knee exoskeleton collaborative control method and system fusing surface electromyography feedforward

Abstract

This application relates to a collaborative control method and system for a knee exoskeleton that integrates surface electromyography (SEM) feedforward, and pertains to the field of rehabilitation robot control. First, individualized baseline calibration is performed on dual-channel SEM signals acquired from the user in both resting and maximally contracted states. Then, based on the resting root mean square baseline and resting standard deviation, time-frequency situational analysis and feedforward intent trigger determination are performed on the real-time SEM signals. Next, using the SEM feedforward trigger signal as an access enable, kinematic parameters such as knee joint angle, angular velocity, and thigh tilt angle are integrated and a priority logic decision tree is used to determine the motion pattern. Based on the pattern identifier, a collaborative torque strategy is selected to allocate the stress torque branch to obtain the baseline auxiliary torque. Finally, fatigue detection is performed based on a dual-index joint criterion of root mean square descent rate and median frequency left shift rate, and adaptive safety adjustment of the baseline auxiliary torque is performed in conjunction with co-contraction anomaly monitoring to output the final driving torque command.

Description

Technical Field

[0001] This application relates to the field of rehabilitation robot control, specifically to a method and system for collaborative control of a knee exoskeleton that integrates surface electromyography feedforward. Background Technology

[0002] Knee exoskeleton robots have become an important assistive device for post-orthopedic rehabilitation and for patients with degenerative joint diseases to restore motor function. However, existing control schemes still have significant limitations in intent recognition and torque output. Traditional control strategies mostly adopt preset trajectory following or fixed torque output modes. The former forces the affected limb to move passively along a predetermined angle curve, making it difficult to stimulate the participation of active neuromuscular muscles; the latter, although allowing autonomous movement, is prone to overcompensation or insufficient assistance because the auxiliary torque cannot be dynamically adjusted according to the actual activation level of the muscles. In terms of intent recognition, existing schemes mostly rely on kinematic signals from inertial sensors for posterior judgment, that is, the exoskeleton response is triggered only after the joint movement has occurred, resulting in inherent control lag. Although surface electromyography (EMG) signals can be detected tens of milliseconds before muscle exertion, giving them a feedforward prediction advantage, existing schemes utilizing EMG signals mostly remain at simple threshold switching control, failing to establish a continuous proportional mapping between EMG amplitude and output torque, and also lacking differentiated torque strategies for multiple movement modes.

[0003] In terms of safety protection, existing knee exoskeletons mainly rely on passive protection measures such as angle limiting and torque limiting, lacking the ability to actively detect and adaptively respond to the user's muscle fatigue state. If muscles become fatigued during training and the system fails to adjust its output strategy in time, it can lead to decreased training effectiveness or even secondary injury. Furthermore, when agonist and antagonist muscles exhibit abnormal co-contraction patterns, the joint experiences bidirectional strain without warning or protective mechanisms, posing a potential safety hazard. Summary of the Invention

[0004] This application provides a method and system for coordinated control of the knee exoskeleton that integrates surface electromyography feedforward, in order to at least solve the above-mentioned technical problems existing in the prior art.

[0005] According to a first aspect of this application, a method for coordinated control of a knee exoskeleton incorporating surface electromyography (EMG) feedforward is provided, comprising: S1, performing individualized baseline calibration on dual-channel EMG signals acquired by the user in resting and maximally contracted states to obtain a resting root mean square baseline, resting standard deviation, and maximum root mean square reference value; S2, based on the resting root mean square baseline and resting standard deviation, performing time-frequency state analysis and feedforward intent trigger determination on real-time dual-channel EMG signals to obtain instantaneous root mean square values ​​of the rectus femoris, biceps femoris, median frequency of the rectus femoris, and an EMG feedforward trigger signal; S3, using the EMG feedforward trigger signal as an access enable, fusing knee joint angle, knee joint angular velocity, and thigh... The inclination angle and the growth rate of the instantaneous root mean square value of the rectus femoris relative to the maximum root mean square reference value are used to determine the motion pattern through a priority logic decision tree to obtain the motion pattern identifier; S4, based on the motion pattern identifier, the stress torque branch is selected, and a coordinated torque strategy is applied to the instantaneous root mean square value of the rectus femoris, the maximum root mean square reference value, the knee joint angle, and the knee joint angular velocity to obtain the baseline auxiliary torque; S5, fatigue detection is performed based on the dual-indicator joint criterion of the sliding window descent rate of the instantaneous root mean square value of the rectus femoris and the leftward shift rate of the median frequency value of the rectus femoris, and co-contraction abnormality monitoring is performed based on the instantaneous root mean square values ​​of the rectus femoris and the instantaneous root mean square values ​​of the biceps femoris, and adaptive safety adjustment is performed on the baseline auxiliary torque to obtain the adaptive driving torque command.

[0006] According to a second aspect of this application, a knee exoskeleton collaborative control system integrating surface electromyography (EMG) feedforward is provided for executing the aforementioned knee exoskeleton collaborative control method integrating EMG feedforward, comprising: a baseline calibration module for performing individualized baseline calibration on dual-channel EMG signals acquired by the user in a resting state and a maximal contraction state to obtain a resting root mean square baseline, a resting standard deviation, and a maximum root mean square reference value; an intent trigger determination module for performing EMG signal time-frequency state analysis and feedforward intent trigger determination on real-time dual-channel EMG signals based on the resting root mean square baseline and resting standard deviation to obtain the instantaneous root mean square value of the rectus femoris, the instantaneous root mean square value of the biceps femoris, the median frequency value of the rectus femoris, and an EMG feedforward trigger signal; and a pattern discrimination module for using the EMG feedforward trigger signal as an access enable. The system integrates knee joint angle, knee joint angular velocity, thigh tilt angle, and the growth rate of the instantaneous root mean square value of the rectus femoris relative to the maximum root mean square reference value. A priority logic decision tree is used to determine the motion pattern to obtain a motion pattern identifier. A torque allocation module selects the corresponding torque branch based on the motion pattern identifier and performs a coordinated torque strategy allocation on the instantaneous root mean square value of the rectus femoris, the maximum root mean square reference value, the knee joint angle, and the knee joint angular velocity to obtain a baseline auxiliary torque. A safety adjustment module performs fatigue detection based on a dual-indicator joint criterion of the sliding window descent rate of the instantaneous root mean square value of the rectus femoris and the leftward shift rate of the median frequency value of the rectus femoris. It also monitors co-contraction anomalies based on the instantaneous root mean square values ​​of the rectus femoris and the biceps femoris, and performs adaptive safety adjustment on the baseline auxiliary torque to obtain an adaptive driving torque command.

[0007] Compared with existing technologies, the knee exoskeleton collaborative control method and system integrating surface electromyography (SEM) feedforward provided in this application has the following technical effects: It achieves a feedforward response lead of tens of milliseconds, matching the timing of exoskeleton torque establishment with the initial stage of human active force exertion, eliminating the inherent response lag of traditional feedback control, and significantly improving the time synchronization of human-computer interaction and the wearer's subjective sense of movement. Furthermore, the continuous proportional mapping between normalized electromyographic intensity and auxiliary torque allows the assistance amplitude to accurately follow the dynamic changes in the actual muscle activation level, avoiding the problems of over-compensation inhibiting active force exertion or insufficient assistance preventing the completion of functional movements in a fixed torque mode. Simultaneously, the inverse damping strategy achieves an adaptive safety characteristic in downward scenarios, where weaker muscles provide stronger buffering protection. Finally, the time-frequency dual-domain fatigue criterion has higher detection specificity than a single indicator, and can effectively distinguish between real muscle fatigue and interference factors such as changes in movement rhythm or electrode impedance drift. Meanwhile, the monitoring of antagonistic muscle co-contraction ratio fills the gap in the existing system's lack of ability to perceive abnormal muscle coordination patterns. The synergistic effect of the two enables the exoskeleton to have the ability to actively detect and intervene in changes in the wearer's physiological state, and overall reduces the risk of secondary injury during training. Attached Figure Description

[0008] The above and other objects, features, and advantages of exemplary embodiments of this application will become readily understood by reading the following detailed description with reference to the accompanying drawings. Several embodiments of this application are illustrated in the drawings by way of example and not limitation, wherein: in the drawings, the same or corresponding reference numerals denote the same or corresponding parts.

[0009] Figure 1 This is an overall flowchart of the knee exoskeleton collaborative control method incorporating surface electromyography feedforward according to an embodiment of this application; Figure 2 This is a schematic diagram of data flow in the knee exoskeleton collaborative control method incorporating surface electromyography feedforward according to an embodiment of this application. Figure 3 This paper illustrates a sub-process diagram of step S2 in the knee exoskeleton collaborative control method integrating surface electromyography feedforward according to an embodiment of this application. Figure 4 This paper illustrates a sub-process diagram of step S3 in the knee exoskeleton collaborative control method integrating surface electromyography feedforward according to an embodiment of this application. Figure 5 This paper illustrates a sub-process diagram of step S4 in the knee exoskeleton collaborative control method integrating surface electromyography feedforward according to an embodiment of this application. Figure 6 This is a block diagram of a knee exoskeleton collaborative control system with fused surface electromyography feedforward according to an embodiment of this application. Detailed Implementation

[0010] To further illustrate the technical means and effects adopted by this application in order to achieve the intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of this application is provided in conjunction with the accompanying drawings and preferred embodiments.

[0011] Figure 1 The illustration shows a schematic flowchart of a knee exoskeleton collaborative control method incorporating surface electromyography feedforward according to an embodiment of this application. Figure 2 This is a schematic diagram of data flow in a knee exoskeleton collaborative control method incorporating surface electromyography feedforward according to an embodiment of this application. Figure 1 and Figure 2 As shown, this application provides a method for coordinated control of a knee exoskeleton that integrates surface electromyography feedforward, comprising: S1 involves individualized baseline calibration of the dual-channel surface electromyography (SEM) signals acquired from the user in both resting and maximally contracted states to obtain the resting root mean square baseline, resting standard deviation, and maximum root mean square reference value. It is understandable that due to significant individual differences in muscle volume, subcutaneous fat thickness, electrode attachment impedance, and neuromuscular recruitment ability among different users, the same amplitude SEM signal represents completely different levels of muscle activation in different individuals. If the system directly uses fixed thresholds or general reference values ​​for intent determination and torque calculation without individualized calibration, it will lead to oversensitivity and frequent false triggers for some users, while remaining unresponsive for others. Therefore, step S1 performs a structured calibration process before the control system officially runs, acquiring dual-channel SEM signals from the user in both fully relaxed and maximally contracted states. Statistical features characterizing the resting noise level and extreme value features characterizing the maximum activation potential are extracted, providing individualized numerical benchmarks for subsequent feedforward trigger determination, normalized EMG intensity calculation, and fatigue detection.

[0012] Specifically, the dual-channel surface electromyography (EMG) signal includes EMG signals from the rectus femoris and biceps femoris muscles. The rectus femoris is a superficial muscle of the quadriceps femoris group and is the primary executor of knee extension movements; its surface EMG signal amplitude directly reflects the active activation level of the knee extension torque. The biceps femoris belongs to the hamstring muscle group and is the primary executor of knee flexion movements, while also acting as an antagonist during knee extension. The selection of dual channels covers the pairing relationship between the agonist and antagonist muscles of the knee joint, enabling both the determination of movement intent and proportional torque mapping through the amplitude characteristics of the agonist channel, and the monitoring of co-contraction abnormalities through the signal ratio relationship between the antagonist and agonist channels.

[0013] In one possible implementation of this application, step S1 includes: S11, performing dual digital filtering on the dual-channel surface electromyography signal to remove motion artifacts and grid interference, obtaining a resting-state rectus femoris filtered signal and a resting-state biceps femoris filtered signal; S12, performing non-overlapping sliding window statistical calculation on the resting-state rectus femoris filtered signal and the resting-state biceps femoris filtered signal to obtain the resting root mean square baseline and resting standard deviation; S13, during multiple maximum isometric knee extension contractions performed by the user, using a peak sliding window to perform root mean square extreme value retrieval on the contraction segment of the maximum contraction-state rectus femoris filtered signal, and taking the absolute maximum value among multiple contractions as the maximum root mean square reference value.

[0014] First, step S11 is executed, performing analog-to-digital conversion on the acquired resting-state rectus femoris and biceps femoris raw analog signals. The analog-to-digital converter performs equally spaced digital sampling of the continuous analog voltage signals output by the dual-channel surface electromyography electrodes at a sampling rate of not less than 1000Hz, converting them into discrete digital signal sequences. Subsequently, a fourth-order Butterworth bandpass filter is applied to this discrete digital signal sequence, with the passband set to 20Hz to 450Hz, to filter out low-frequency motion artifacts caused by relative sliding between the skin and electrodes, as well as high-frequency electronic thermal noise exceeding the effective spectrum range of surface electromyography. The bandpass-filtered signal is further fed into a second-order double-pole-double-zero IIR notch filter, with the notch center frequency set to 50Hz and the quality factor Q set to 30, to steeply attenuate and suppress power grid frequency interference and its harmonic components. After the above dual filtering process, the resting-state rectus femoris filtered signal and the resting-state biceps femoris filtered signal are obtained, respectively. The same filtering process was also applied to the raw rectus femoris signal collected when the user performed the maximum voluntary contraction test to obtain the maximum contraction state rectus femoris filtered signal.

[0015] In the resting-state statistical feature solving stage, the system performs non-overlapping time window division on the resting-state rectus femoris filtered signal and the resting-state biceps femoris filtered signal, respectively. The window length is set to 200ms, and each window contains 200 sampling points at a sampling rate of 1000Hz. Based on a 60-second resting acquisition duration, a total of 300 independent analysis time windows are obtained. The local root mean square value is calculated for the filtered sample values ​​within each window, using the following formula:

[0016] in, This represents the number of sampling points within a single window, with a value of 200. This represents the discrete voltage amplitude of the electromyographic signal at the i-th sampling point within the current window after double filtering. The local root mean square values ​​of 300 windows each for the rectus femoris and biceps femoris channels are sorted in ascending order, and their medians are extracted. The medians of the two channels are then combined to form the resting root mean square baseline. The median is used instead of the arithmetic mean because it is more robust to occasional motion artifact spikes and will not be biased by individual abnormal windows. The standard deviations of the 300 local root mean square values ​​for each of the two channels are calculated separately, and the standard deviations of the two channels are combined to form the resting standard deviation.

[0017] In the maximum voluntary contraction peak extraction sub-stage, the system guides the user to perform three maximum isometric knee extension contraction tests, maintaining maximum force for 3 seconds each time, with a 30-second interval between tests to ensure sufficient muscle recovery. Based on the test timing markers, three active force intervals are located in the filtered signal of the rectus femoris muscle at maximum contraction. Within each force interval, a sliding time window of 500ms is set, and the window slides with a step size of 1 sampling point. The local root mean square value at each window position is calculated, and the maximum local root mean square peak value of that contraction is extracted. The absolute maximum value of the maximum local root mean square values ​​obtained from the three tests is taken as the reference value of the maximum root mean square value of the rectus femoris muscle for the user. The calculation formula is as follows:

[0018] in, This is the test sequence number for the contraction test, with a value range of 1, 2, and 3; This represents the number of sampling points within the peak retrieval window, and is set to 500 at a window width of 500ms and a sampling rate of 1000Hz. This is the index of the starting sampling point of the window when the root mean square value reaches its maximum within a 500ms window during the kth contraction test. This represents the discrete voltage amplitude of the rectus femoris channel at the i-th sampling point after double filtering. The reason for using a 500ms window width instead of the entire 3-second contraction segment to calculate the root mean square value is that the muscle activation level is not constant during maximal voluntary contraction, typically peaking in the middle of the contraction and being lower at the beginning and end. The peak window can more accurately capture the individual's true maximum activation potential.

[0019] In one specific embodiment, the user sits on a standard-height chair with the knees naturally flexed at 90 degrees. Dual-channel surface electromyography (SMMG) electrodes are attached to the center of the rectus femoris muscle belly and the center of the biceps femoris muscle belly, respectively. The system acquires 60 seconds of resting dual-channel signals at a sampling rate of 1000Hz and 24-bit analog-to-digital conversion accuracy. After processing with a 20-450Hz fourth-order Butterworth bandpass filter and a 50Hz second-order IIR notch filter, 300 local root mean square (RMS) values ​​are calculated using a 200ms non-overlapping window. The median yields a resting RMS baseline of 5.2 μV for the rectus femoris channel and 4.8 μV for the biceps femoris channel. The standard deviations yield a resting standard deviation of 1.1 μV for the rectus femoris channel and 0.9 μV for the biceps femoris channel. The user then performed three maximum isometric knee extensions and retractions. The system extracted the peak root mean square (RMS) values ​​for each repetition using a 500ms sliding window, which were 186 μV, 192 μV, and 189 μV, respectively. The absolute maximum value of 192 μV was taken as the maximum RMS reference value. The above-mentioned resting RMS baseline, resting standard deviation, and maximum RMS reference value were solidified as system parameters for the current control session, and used in subsequent steps S2 to S5 for threshold construction, normalization calculation, and fatigue baseline establishment.

[0020] S2, based on the resting root mean square baseline and resting standard deviation, performs time-frequency situational analysis and feedforward intent trigger determination on real-time dual-channel surface electromyography (EMG) signals to obtain the instantaneous root mean square values ​​of the rectus femoris, biceps femoris, median frequency of the rectus femoris, and the EMG feedforward trigger signal. It can be understood that surface EMG signals, as a direct electrophysiological representation of muscle contraction, can be detected on the body surface tens of milliseconds before the actual observable joint movement. This time-advanced characteristic provides the physical basis for the intention-feedforward control of exoskeletons. However, the raw EMG signal is a highly random, amplitude-fluctuating interference pattern signal that cannot be directly used for continuous control. Therefore, it is necessary to obtain smooth instantaneous activation amplitude features through time-domain sliding window envelope extraction, obtain spectral features reflecting changes in muscle fiber conduction velocity through frequency-domain analysis, and convert the continuous amplitude signal into a discrete trigger enable signal through statistical threshold criteria. The output of step S2 includes the instantaneous root mean square value of the rectus femoris muscle, the instantaneous root mean square value of the biceps femoris muscle, the median frequency value of the rectus femoris muscle, and the electromyographic feedforward trigger signal. These four outputs serve the proportional torque mapping calculation, cocontraction ratio monitoring, fatigue frequency domain index tracking, and access enable control for motion pattern recognition in subsequent steps, respectively.

[0021] In one of the possible implementation methods of this application, such as Figure 3 As shown, step S2 includes: S21, after bandpass filtering and power frequency notch preprocessing of the real-time dual-channel surface electromyography (EMG) signal, the instantaneous root mean square (RMS) values ​​of the rectus femoris channel and the biceps femoris channel are calculated using a short-time sliding window to obtain the instantaneous RMS values ​​of the rectus femoris and biceps femoris; S22, the resting RMS baseline is added to a preset multiple of the resting standard deviation to construct a dynamic trigger threshold. When the instantaneous RMS value of the rectus femoris exceeds the dynamic trigger threshold and the continuous holding time reaches the confirmation window duration, it is determined as an active movement intention trigger and an EMG feedforward trigger signal is generated; S23, the filtered signal of the rectus femoris channel is subjected to a fast Fourier transform using a long-time sliding window, and the median frequency of the power spectral density is extracted to obtain the median frequency value of the rectus femoris.

[0022] First, step S21 is executed, where the raw signals from the rectus femoris channel and the raw signals from the biceps femoris channel, captured in real time by the surface electromyography acquisition module during exoskeleton operation, undergo dual digital filtering processing with the same configuration as in step S1. Specifically, a fourth-order Butterworth bandpass filter is first executed, with a passband of 20Hz to 450Hz, to filter out low-frequency motion artifacts and high-frequency electronic noise; then a second-order IIR notch filter is executed, with a center frequency of 50Hz and a quality factor Q of 30, to suppress power grid frequency interference and its harmonics. After dual filtering, real-time rectus femoris filtered signals and real-time biceps femoris filtered signals are obtained, respectively.

[0023] After obtaining the filtered signal, the system establishes an extremely short sliding time window of 50ms. At a sampling rate of 1000Hz, this window contains 50 sampling points, and the window slides point by point with a step size of 1 sampling point. The instantaneous root mean square (RMS) value of the real-time rectus femoris filtered signal is calculated for the 50 sampling points within the current window, yielding the instantaneous RMS value of the rectus femoris muscle. The calculation formula is as follows:

[0024] in, The instantaneous root mean square of the rectus femoris muscle. This represents the number of sampling points within the instantaneous analysis window, with a value of 50. This represents the discrete voltage amplitude of the rectus femoris channel at the i-th sampling point within the current window after double filtering. The same calculation method is used to perform window root mean square (RMS) calculation on the real-time biceps femoris filtered signal to obtain the instantaneous RMS value of the biceps femoris. The reason for choosing 50ms as the window length is that: if the window is too short, the RMS value fluctuates wildly and is difficult to use for continuous control; if the window is too long, the temporal resolution decreases and it cannot capture the rapid onset characteristics of muscle activation. A 50ms window width achieves a suitable trade-off between smoothness and temporal responsiveness.

[0025] Then, step S22 is executed, using the rectus femoris channel component in the resting root mean square baseline and the rectus femoris channel component in the resting standard deviation obtained in step S21 to construct a dynamic trigger threshold. This threshold is constructed by adding the resting root mean square baseline to a preset multiple of the resting standard deviation; the determination formula is as follows:

[0026] in, The rectus femoris channel component in the resting root mean square baseline calibrated in step S1; The rectus femoris channel component in the resting standard deviation calibrated in step S1; This is the preset standard deviation factor. The value of determines the balance between trigger sensitivity and resistance to false triggering. Under the assumption that the resting electromyographic signal approximately follows a Gaussian distribution, A value of 3 corresponds to a statistical significance level of 0.13% for one-sided transcendence, effectively distinguishing active muscle activation from background noise fluctuations. In practical applications... The typical value range is 2 to 4, and the value in this embodiment is 3.

[0027] When the aforementioned inequality is true for the first time at a certain moment, the system does not immediately determine it as a valid trigger, but instead starts a confirmation time window timer. Only when the duration of the consecutively true inequality reaches the preset confirmation window duration does the system determine that the user has generated an active movement intention and set the electromyographic feedforward trigger signal to valid (logic high level). The purpose of setting the confirmation window duration is to exclude single, occasional over-threshold events caused by transient electrode contact jitter or external electromagnetic pulse interference. The typical value range of the confirmation window duration is 15ms to 30ms, and in this embodiment, it is set to 20ms. If the consecutively true inequality is interrupted within the confirmation window duration, the timer is reset to zero and waits for the next over-threshold event.

[0028] When the EMG feedforward trigger signal is active, the system continuously monitors whether the instantaneous root mean square value of the rectus femoris muscle falls below the reset threshold. The reset threshold is set as the sum of the resting root mean square baseline and the resting standard deviation of a preset reset factor. A typical reset factor is 1.5, lower than the standard deviation factor used during triggering to create hysteresis and prevent high-frequency trigger-reset oscillations when the signal fluctuates near the threshold. When the instantaneous root mean square value of the rectus femoris muscle remains below the reset threshold for a duration equal to the reset confirmation time, the system restores the EMG feedforward trigger signal to inactive (logic low). The typical reset confirmation time is 50ms, slightly longer than the trigger confirmation window duration, to ensure that the reset decision is only executed after the signal has indeed returned to the resting level. The amplitude difference between the trigger threshold and the reset threshold constitutes a hysteresis dead zone. This hysteresis dead zone makes the switching between active and inactive states of the trigger signal directionally dependent, thereby suppressing repeated jumps caused by natural fluctuations of the EMG signal near the threshold boundary.

[0029] The feedforward time advantage of this triggering mechanism stems from the inherent electromechanical delay characteristics of the human neuromuscular system. In the human rectus femoris muscle, from the detection of activation initiation via surface electromyography (EMG) signals to the completion of excitation-contraction coupling within the muscle fibers and the generation of sufficient joint torque to induce an observable angular change in the knee joint, there is an electromechanical delay of approximately 50 ms to 80 ms. This scheme triggers the exoskeleton into a torque preparation state during the initial stage of EMG signal activation, matching the timing of exoskeleton torque establishment with the timing of active muscle force generation in the human body. The time from the generation of the EMG trigger signal to the completion of exoskeleton torque establishment does not exceed 30 ms, thus achieving feedforward control relative to the occurrence of joint movement.

[0030] Finally, step S23 is executed, opening a long-term sliding analysis window with a length of 5 seconds and a step size of 1 second. A Fast Fourier Transform is performed on the real-time rectus femoris filtered signal within this window to obtain its power spectral density function. After obtaining the power spectral density function, a frequency segmentation point is found within the range of the lower limit frequency (20Hz) to the upper limit frequency (450Hz) of the frequency axis, such that the power spectral area to the left of this segmentation point is equal to the power spectral area to the right of this segmentation point. This segmentation point frequency is the median frequency value of the rectus femoris. The physical meaning of the median frequency value lies in reflecting the average firing frequency characteristics of the current group of motor units participating in contraction. When the muscle is in a normal, non-fatigue state, the median frequency value remains relatively stable. When the muscle gradually enters a fatigue state, due to the withdrawal of fast motor units from recruitment and the decrease in muscle fiber conduction velocity, the power spectrum shifts towards lower frequencies, and the median frequency value decreases accordingly. This characteristic quantity does not directly participate in the torque calculation or trigger determination of the current control cycle, but is continuously updated as a frequency domain input indicator for the joint criterion of fatigue detection dual indicators in step S5.

[0031] In one specific embodiment, the system operates with a control cycle of 2ms. Within a certain control cycle, the double-filtered signal from the rectus femoris channel calculates an instantaneous root mean square (RMS) value of 18.6 μV for the rectus femoris and 6.2 μV for the biceps femoris channel within the current 50ms window. The system reads the baseline resting RMS value of 5.2 μV and the resting standard deviation of 1.1 μV for the rectus femoris channel calibrated in step S1, and constructs a dynamic trigger threshold of 5.2 + 3 multiplied by 1.1, which equals 8.5 μV. Since the current instantaneous RMS value of 18.6 μV for the rectus femoris is greater than the threshold of 8.5 μV, and this over-threshold state has been maintained for 22ms, exceeding the 20ms confirmation window duration, the system determines that the feedforward trigger is valid, and the electromyographic feedforward trigger signal output is a logic high level. Simultaneously, the system performs a Fast Fourier Transform based on the filtered rectus femoris signal within the most recent 5-second window, calculating the current median frequency of the rectus femoris to be 78Hz. The above four output quantities are immediately transmitted to the motion mode recognition module in step S3 and the fatigue monitoring module in step S5 for subsequent processing after the current control cycle ends.

[0032] S3 uses electromyography (EMG) feedforward trigger signals as the access enable, integrating knee joint angle, knee joint angular velocity, thigh tilt angle, and the growth rate of the instantaneous root mean square value of the rectus femoris muscle relative to the maximum root mean square reference value. A priority logic decision tree is then used to determine the movement mode to obtain the movement mode identifier. It should be understood that EMG signals alone cannot distinguish the specific type of movement performed by the user, because the rectus femoris muscle will actively contract and activate in different movement modes, and the amplitude and shape of its EMG signals overlap. The knee exoskeleton needs to implement drastically different torque control strategies in different movement modes. For example, zero-impedance follow-up compensation is needed for transparent following when walking on flat ground; positive proportional assistance is needed to help knee extension overcome gravity when climbing stairs; damping buffer is needed to provide eccentric braking protection when descending stairs; and phased decreasing assistance is needed to assist in standing up when transferring from a seated to a standing position. Therefore, before executing the torque control strategy, the system must accurately identify the current user's movement mode in order to select the correct torque calculation branch. Step S3 uses the electromyography feedforward trigger signal as the admission enable threshold for pattern recognition, integrates kinematic parameters such as knee joint angle, knee joint angular velocity, and thigh tilt angle output by the nine-axis inertial measurement unit, as well as the growth rate of the instantaneous root mean square value of the rectus femoris relative to the maximum root mean square reference value, and classifies and distinguishes the current motion state through a multi-condition logic decision tree, outputting a unique motion pattern identifier.

[0033] In one of the possible implementation methods of this application, such as Figure 4 As shown, the decision rules of the priority logic decision tree in step S3 are as follows: S31, when the electromyographic feedforward trigger signal is valid and the knee joint angle is greater than the preset sitting flexion angle and begins to extend, while the thigh tilt angle exceeds the forward tilt threshold and the growth rate of the instantaneous root mean square value of the rectus femoris relative to the maximum root mean square reference value within the preset time period exceeds the burst threshold, it is determined to be a sit-to-stand transition mode; S32, when the electromyographic feedforward trigger signal is valid and the knee joint angle is greater than the preset stair-climbing flexion angle and moves in the extension direction, while the knee joint angular velocity is negative and its absolute value exceeds the extension direction, it is determined to be a sit-to-stand transition mode; S33. When the knee joint angular velocity is above the flexion velocity threshold and the thigh tilt angle exceeds the forward tilt threshold, it is determined to be the stair-climbing mode; S34. When the electromyographic feedforward trigger signal is valid, the knee joint angular velocity is positive and its absolute value exceeds the flexion velocity threshold, and the knee joint angle shows a continuous increasing trend within the preset short-term observation window, it is determined to be the stair-descent mode; S35. When the electromyographic feedforward trigger signal is valid, the knee joint angle is within the normal gait range, and the knee joint angular velocity shows alternating positive and negative changes and the absolute value of the thigh tilt angle is lower than the forward tilt discrimination threshold, it is determined to be the flat ground walking mode.

[0034] First, the instantaneous root mean square (RMS) value of the rectus femoris muscle is normalized to eliminate the influence of individual differences on the electromyographic (EMG) related conditions in subsequent discrimination rules. The normalization calculation extracts the rectus femoris component from the instantaneous RMS value, the maximum RMS reference value calibrated in step S1, and the resting RMS baseline. Cross-temporal bias removal and extreme value compression operations are then performed to obtain the normalized EMG activation characteristics, calculated using the following formula:

[0035] in, This represents the instantaneous root mean square value of the rectus femoris muscle. The rectus femoris channel component in the resting root mean square baseline calibrated in step S1; This is the maximum root mean square reference value calibrated in step S1. The formula first subtracts the resting bias to remove background noise contribution, then divides by the maximum activation range for normalization, and finally constrains the result to the range of 0 to 1 through amplitude limiting. A normalized electromyographic activation characteristic value of 0 indicates that the muscle is currently at rest, and a value of 1 indicates that the maximum voluntary contraction level has been reached.

[0036] Simultaneously, the system reads the absolute knee angle, knee angular velocity, and thigh tilt angle output by the nine-axis inertial measurement unit at a sampling rate of 200Hz. In this application, the knee angle is defined as the degree of knee flexion, with 0 degrees at full extension and the maximum value at full flexion; a larger angle value indicates a deeper degree of flexion. The knee angular velocity is defined as the derivative of the knee angle with respect to time; a positive angular velocity indicates an increase in angle, i.e., joint movement in the flexion direction, while a negative angular velocity indicates a decrease in angle, i.e., joint movement in the extension direction. These three kinematic parameters undergo sampling time axis alignment processing to ensure consistency with the time reference of the electromyographic characteristics within the same control cycle. When calculating the growth rate of the instantaneous root mean square value of the rectus femoris relative to the maximum root mean square reference value, the system takes the increase in the instantaneous root mean square value of the rectus femoris within the most recent 200ms time period, divides this increase by the maximum root mean square reference value to obtain the percentage growth rate, which is used to characterize the explosive characteristics of rectus femoris activation in the sit-to-stand transition mode discrimination rule.

[0037] In the multi-condition logic decision tree branch discrimination sub-stage, the system uses the electromyography (EMG) feedforward trigger signal as the admission enable condition for the entire decision tree. When the EMG feedforward trigger signal is invalid (logic low), it indicates that the user has not generated an active movement intention, the system does not perform any pattern discrimination calculation, and the movement mode identifier remains at the locked value of the previous cycle or the default static standby state. Only when the EMG feedforward trigger signal is valid (logic high) does the system enter the rule-by-rule verification process of the decision tree.

[0038] The decision tree contains four discrimination rules, corresponding to the sitting / standing transfer mode, the upstairs mode, the downstairs mode, and the flat ground walking mode, respectively. The condition combinations of each rule are as follows.

[0039] The criteria for determining the sit-to-stand transition mode require the following conditions to be met simultaneously: the electromyographic feedforward trigger signal is valid, the knee angle is greater than the preset sitting flexion angle and begins to move in the extension direction, the thigh tilt angle exceeds the forward tilt threshold, and the rate of increase of the instantaneous root mean square value of the rectus femoris relative to the maximum root mean square reference value within a preset time period exceeds the burst threshold. The typical value for the preset sitting flexion angle is 90 degrees, indicating that the user is in a deeply flexed sitting position; the typical value for the forward tilt threshold is 30 degrees, indicating that the thigh has a significant forward tilt relative to the vertical direction, consistent with the characteristics of a forward shift of the center of gravity in a sitting position; the typical value for the burst threshold is 40%, meaning that the increase in the instantaneous root mean square value of the rectus femoris within 200ms exceeds 40% of the maximum root mean square reference value, characterizing the explosive recruitment characteristics of the quadriceps femoris in the initial stage of standing up.

[0040] The criteria for determining the stair-climbing mode require the following conditions to be met simultaneously: the electromyographic feedforward trigger signal is valid, the knee joint angle is greater than the preset stair-climbing flexion angle and moves in the extension direction, the knee joint angular velocity is negative and its absolute value exceeds the extension velocity threshold, and the thigh tilt angle exceeds the stair-climbing forward tilt threshold. The typical preset stair-climbing flexion angle is 60 degrees, indicating that the knee joint is in the flexion preparation state at the end of the stair-climbing swing; a negative knee joint angular velocity indicates that the joint angle is decreasing, i.e., the joint is moving in the extension direction; the extension velocity threshold is set to 30 degrees per second, meaning that the absolute value of the angular velocity must exceed 30 degrees per second to be considered an active knee extension movement, used to distinguish it from passive swaying; the typical stair-climbing forward tilt threshold is 20 degrees, indicating that the thigh has a forward tilt posture, but the degree is less than that of a seat-to-stand transition.

[0041] The criteria for determining the stair-descent pattern require the following conditions to be met simultaneously: a valid electromyographic feedforward trigger signal, a positive knee joint angular velocity with an absolute value exceeding the flexion velocity threshold, and a continuously increasing knee joint angle within a preset short-term observation window. A positive knee joint angular velocity indicates that the joint is moving in the flexion direction. The flexion velocity threshold is set at 20 degrees per second to exclude false positives caused by noise from the angular velocity sensor in static or micro-movement states. A continuously increasing knee joint angle within the preset short-term observation window means that, using the most recent 500ms as the observation window, a linear regression of the knee joint angle sequence within the window is performed. When the regression slope is positive and exceeds 5 degrees per second, it is considered an increasing trend. This condition complements the instantaneous angular velocity direction condition on a time scale: instantaneous angular velocity reflects the direction of movement in the current sampling period, and the increasing trend in the short-term observation window confirms that the knee joint has been in a process of deepening flexion over the past 500ms, rather than experiencing transient fluctuations. The two conditions together ensure that the system only identifies the stair-descent mode when the knee joint exhibits sustained eccentric flexion (rather than brief disturbances or momentary flexion during the gait swing phase). This combined feature reflects the kinematic characteristics of the knee joint bearing weight and undergoing eccentric braking during the stair-descent support phase.

[0042] The criteria for determining the flat-ground walking pattern require the following conditions to be met simultaneously: a valid electromyographic feedforward trigger signal, a knee joint angle within the normal gait range, alternating positive and negative knee joint angular velocities, and an absolute value of the thigh tilt angle below the forward tilt threshold. The typical range of normal gait is 10 to 65 degrees, covering the angle interval from the support phase extension to the swing phase flexion during normal walking; the frequency of alternating positive and negative angular velocities within the range of 0.8 Hz to 1.2 Hz corresponds to a normal gait frequency; and the typical forward tilt threshold is 15 degrees, indicating that the thigh is nearly vertical with no significant forward or backward tilt.

[0043] The four rules are executed in descending order of priority: sitting-to-standing transfer, climbing stairs, descending stairs, and walking on flat ground. This priority order is based on the following: the sitting-to-standing transfer mode has the most unique characteristics and the highest safety requirements, and should be responded to first once its condition combination is met; the knee extension assistance requirement is the most urgent among the four modes in the climbing stairs mode; the descending stairs mode requires timely damping protection; and the walking on flat ground mode, as the most common default movement state, has the lowest recognition priority. When all conditions of a rule are met simultaneously, the mode identifier corresponding to that rule becomes the candidate output for the current cycle, and lower-priority rules are no longer checked. If all four rules are not met, the movement mode identifier remains the locked value from the previous cycle.

[0044] During the timing-based debouncing locking sub-stage, the system performs timing-based debouncing processing on the candidate mode identifiers output by the decision tree to prevent high-frequency oscillations in the mode caused by transient fluctuations in sensor signals. The system establishes a debouncing hold window. Only when the candidate mode identifiers continuously output by the decision tree differ from the currently locked motion mode identifier, and the duration of this new mode identifier remaining unchanged reaches the debouncing hold window duration, does the system perform a mode switching update, updating the motion mode identifier to the new candidate mode. If the candidate mode identifier changes or regresses within the debouncing window duration, the debouncing timer is reset to zero, and the motion mode identifier retains its current locked value.

[0045] The following provides two specific implementations of priority logic decision tree discrimination rules.

[0046] In the first embodiment, each discrimination rule is arbitrated according to a fixed priority order of sitting-to-standing transition, climbing stairs, descending stairs, and walking on flat ground. There are no additional accessibility constraints or confidence scores between the rules. The system applies a uniformly long de-jitter hold window to lock the timing of all mode transition events. A typical value for the de-jitter hold window length is 200ms, meaning that the candidate mode must remain unchanged for 200ms to trigger a mode switch. This implementation is simple in structure and has low computational complexity, making it suitable for standardized rehabilitation training scenarios with low movement mode switching frequency and good sensor signal quality.

[0047] In the second embodiment, the system introduces a three-layer enhancement mechanism based on the fixed priority arbitration framework: directed transfer reachability constraint, transfer confidence score, and adaptive de-shake window, in order to address the problem of pattern misjudgment caused by sensor transient noise, electrode contact jitter, or limb posture changes in actual wearable rehabilitation scenarios.

[0048] In the second embodiment of this application, the arbitration of each discrimination rule and the temporal locking of the motion pattern identifier include: S35, based on the pre-constructed directed transition reachability adjacency matrix, extracting the reachability mask vector with the currently locked motion pattern identifier as the row index, performing element-wise logical AND masking on the candidate mode trigger flag vectors independently output by each discrimination rule, forcibly invalidating candidate modes that are biomechanically inaccessible in the current mode, and obtaining the reachability constraint candidate mode vector; S36, for each candidate mode that is still valid in the reachability constraint candidate mode vector, calculating the normalized margin distance of the knee joint angle, knee joint angular velocity, thigh tilt angle, and instantaneous root mean square value of the rectus femoris muscle involved in its discrimination rule relative to their respective judgment threshold boundaries, and performing weighted summation. The transition confidence scores of each candidate mode are obtained by fusion. The transition confidence scores are checked in order of priority: sitting to standing, going up stairs, going down stairs, and walking on flat ground. The transition confidence scores are checked to see if they exceed the minimum confidence threshold in order to select the arbitration winning candidate mode identifier and its corresponding winning mode confidence score. S37, based on the pre-constructed transition de-jittering duration lookup matrix, the benchmark de-jittering duration is extracted using the currently locked motion mode identifier and the arbitration winning candidate mode identifier as a joint index. The benchmark de-jittering duration is inversely scaled by the ratio of the minimum confidence threshold to the winning mode confidence score to obtain the adaptive de-jittering duration. When the cumulative time for the arbitration winning candidate mode identifier to remain unchanged continuously reaches the adaptive de-jittering duration, the mode switching lock is executed to update the motion mode identifier.

[0049] First, step S35 is executed to perform directed transition reachability masking. The system pre-constructs a 4x4 directed transition reachability adjacency matrix. The row and column indices correspond to four modes: sitting-to-standing transition, climbing stairs, descending stairs, and walking on flat ground, respectively. A value of 1 in the matrix element indicates that the transition from the corresponding row mode to the column mode is reachable under the spatiotemporal continuity constraint of human lower limb movement, while a value of 0 indicates that it is not reachable. The values ​​of this matrix are determined based on the biomechanical empirical laws of human lower limb movement: after a sitting-to-standing transition, only a transition to walking on flat ground is allowed, because the human body needs to stand firmly before starting; walking on flat ground allows transitions to climbing stairs, descending stairs, and sitting-to-standing transitions, because walking on flat ground is the pivotal transition state between modes; climbing stairs only allows a transition to walking on flat ground, because after climbing stairs, one enters level walking; descending stairs only allows a transition to walking on flat ground, because after descending stairs, one enters level walking. In each control cycle, the system uses the currently locked movement mode identifier as the row index and extracts the corresponding row vector from the directed transition reachability adjacency matrix as the reachability mask vector. Simultaneously, each of the four discrimination rules independently calculates and outputs its own Boolean trigger state, which is combined to form a candidate mode trigger flag vector. Then, an element-wise logical AND operation is performed between the reachability mask vector and the candidate mode trigger flag vector. Any candidate mode whose corresponding element in the reachability adjacency matrix is ​​zero is forcibly masked as invalid, regardless of whether the condition combination of its discrimination rules is satisfied by the current sensor data. After masking, the reachability constraint candidate mode vector is obtained, retaining only the biomechanically reachable valid candidates under the current mode.

[0050] Then, step S36 is executed to perform weighted scoring of transition confidence and joint arbitration of priority. For each candidate mode that is still valid in the candidate mode vector of reachability constraints, the system calculates the normalized margin distance of the knee joint angle, knee joint angular velocity, thigh tilt angle, and instantaneous root mean square value of the rectus femoris muscle involved in its discrimination rule relative to their respective judgment threshold boundaries, and performs weighted fusion with pre-calibrated contribution weights to obtain the transition confidence score of each candidate mode. The calculation formula is as follows:

[0051] in, The transition confidence score is the value corresponding to the m-th valid candidate motion pattern. Let be the total number of sensing feature condition dimensions included in the discrimination rule for the m-th candidate pattern; d is the feature dimension index number, with a value range of . ; Let d be the pre-calibrated contribution weight coefficient of the d-th feature dimension in the pattern discrimination, satisfying that the sum of all weights is one; This represents the current actual measurement value of the d-th feature dimension; This represents the decision threshold boundary for the d-th feature dimension in the m-th candidate pattern discrimination rule. This represents the normalized range of the d-th feature dimension, used to eliminate scale bias caused by different physical dimensions in the weighted summation. The further the measured value is from the decision threshold boundary, the higher the score, indicating a stronger match between the current sensor observation data and the pattern.

[0052] After obtaining the transition confidence scores of each valid candidate mode, the system verifies them sequentially according to the priority order of sitting-to-standing transition, going up stairs, going down stairs, and walking on flat ground. If the transition confidence score of the current highest priority candidate mode exceeds the preset minimum confidence threshold, then that mode wins. If it does not exceed the threshold, the next lower priority candidate mode is selected for the same verification, until a winning candidate mode identifier and its corresponding winning mode confidence score are selected, or if all candidates fail to meet the admission conditions, the current mode remains unchanged. The typical range of the minimum confidence threshold is 0.15 to 0.30, and in this embodiment, it is set to 0.20.

[0053] Finally, step S37 is executed to perform adaptive debouncing window locking based on the transition type. The system pre-constructs a four-row, four-column transition debouncing duration lookup matrix, where the row index represents the currently locked mode, the column index represents the candidate target mode, and the matrix elements represent the baseline debouncing confirmation duration required for the corresponding transition type. The matrix is ​​assigned values ​​according to the following principles: gradual transitions correspond to longer baseline debouncing durations, such as the transition from walking up or down stairs on flat ground, with a baseline debouncing duration set to 300ms to 400ms; abrupt transitions correspond to shorter baseline debouncing durations, such as the sudden start of a transition from walking to sitting / standing on flat ground, with a baseline debouncing duration set to 100ms to 150ms.

[0054] The system uses the currently locked motion mode identifier as the row index and the arbitration winner candidate mode identifier as the column index to extract the corresponding baseline deshake duration from the transfer deshake duration lookup matrix. Furthermore, the system introduces the confidence score of the winning mode to inversely adjust the baseline deshake duration. The formula for calculating the adaptive deshake duration is as follows:

[0055] in, The adaptive debouncing confirmation duration actually used for the current transfer event; The baseline deshake confirmation time is extracted from the deshake duration lookup matrix based on the indexes of the currently locked mode and the winning candidate mode; The minimum confidence threshold preset for the arbitration process; This is the confidence score for the winning pattern output by the arbitration step. Since the confidence score for the winning pattern must be greater than or equal to the minimum confidence threshold, and The ratio is always less than or equal to 1, therefore the adaptive deshake duration never exceeds the baseline deshake duration. When the confidence level is far above the admission threshold, the ratio approaches zero, the deshake duration is significantly shortened, and the system quickly completes the mode switch; when the confidence level is only slightly higher than the admission threshold, the ratio approaches 1, the deshake duration is close to the baseline value, and the system waits cautiously to accumulate more timing confirmation evidence.

[0056] When the cumulative time for which the arbitration winner candidate mode identifier remains unchanged reaches the adaptive debouncing duration, the system performs mode switching locking, updates the currently locked motion mode identifier to the arbitration winner candidate mode identifier, and outputs the new motion mode identifier. If the arbitration winner candidate mode identifier changes within the adaptive debouncing duration, the cumulative timer is reset, and the currently locked mode remains unchanged.

[0057] In one specific embodiment, the user wears a knee exoskeleton to perform a standing motion from a seated position. During a certain control cycle, the system obtains the instantaneous root mean square value of the rectus femoris muscle from step S2 as 95 μV, the maximum root mean square reference value as 192 μV, and the resting root mean square baseline as 5.2 μV. The system calculates the normalized electromyographic activation characteristic as (95-5.2) / (192-5.2) = 0.48. Simultaneously, it reads the knee joint angle (92 degrees), knee joint angular velocity (-12 degrees per second, indicating the angle is decreasing, i.e., the joint is moving in the extension direction), and thigh tilt angle (35 degrees) from the output of the inertial measurement unit. Within the last 200 ms, the instantaneous root mean square value of the rectus femoris muscle increased from 18 μV to 95 μV, with a growth rate of (95-18) / 192 = 40.1%, exceeding the burst threshold of 40%. The electromyographic feedforward trigger signal is active. The system enters the decision tree discrimination process, first verifying the highest priority sit-to-stand transition rule: feedforward trigger valid, knee angle greater than 92 degrees (sitting flexion angle 90 degrees) and in extension, thigh tilt angle exceeding the 30-degree forward tilt threshold by 35 degrees, and growth rate exceeding the 40.1% burst threshold by 40%. If all four conditions are met, the decision tree outputs the sit-to-stand transition mode as a candidate identifier. In the first embodiment, this candidate identifier must be continuously maintained for 200ms before being locked for output. In the second embodiment, the system first queries the reachability adjacency matrix to confirm that the transition from the current stationary mode to sit-to-stand mode is valid. Then, it calculates the transition confidence score of this mode as 0.45, exceeding the minimum confidence threshold of 0.20, and the arbitration wins. The system extracts the baseline de-jitter duration from the transition de-jitter duration lookup matrix as 120ms, and calculates the adaptive de-jitter duration as 120 multiplied by 0.20 divided by 0.45, which equals 53ms. Due to the high confidence level, the debouncing time is shortened to 53ms. After the winning candidate in the arbitration is held for 53ms, the system completes the mode switching lock and outputs the sitting / standing transfer mode identifier to step S4.

[0058] S4 selects the corresponding torque branch based on the movement mode identifier and performs a coordinated torque strategy to allocate the instantaneous root mean square value of the rectus femoris muscle, the maximum root mean square reference value, the knee joint angle, and the knee joint angular velocity to obtain the baseline auxiliary torque. It should be understood that the requirements for exoskeleton torque output in different movement modes differ fundamentally in terms of direction, amplitude, time-varying characteristics, and safety constraints. When walking on flat ground, the user's knee joint muscle strength is usually sufficient to autonomously complete the gait cycle, and the main task of the exoskeleton is to eliminate its own mechanical transmission friction to avoid hindering normal walking. When climbing stairs, the knee joint needs to overcome the gravitational torque generated by body weight to complete the knee extension and lifting action, requiring the exoskeleton to provide active assistance proportional to the muscle activation level. When descending stairs, the knee joint is in an eccentric braking state, requiring the exoskeleton to provide damping buffering force adapted to the degree of muscle weakness to prevent uncontrolled knee flexion. During the sit-to-stand transition, the knee joint has different load characteristics in different angle ranges, requiring the exoskeleton to gradually reduce the assistance amplitude in stages to balance the initial assistance requirements with the recovery of end-effectoral control capabilities. Therefore, the function of step S4 is to guide the control flow to the corresponding torque calculation branch based on the motion mode identifier, execute the mode-specific torque mapping algorithm on the instantaneous root mean square value of the rectus femoris muscle, the maximum root mean square reference value, the knee joint angle and the knee joint angular velocity, generate the benchmark auxiliary torque that matches the current biomechanical requirements of the movement, and perform anti-hyperextension safety reconstruction before torque output to ensure that the output torque will not cause hyperextension injury when the joint is close to the limit position.

[0059] In one of the possible implementation methods of this application, such as Figure 5 As shown, step S4 includes: S41, when the movement mode is identified as walking on flat ground, feedforward friction compensation is performed on the knee joint angular velocity based on the pre-identified Coulomb friction torque and viscous friction coefficient to obtain a reference auxiliary torque that makes the follow-up resistance approach zero; S42, when the movement mode is identified as climbing stairs, the instantaneous root mean square value of the rectus femoris muscle is subtracted from the resting root mean square baseline, then normalized by the difference between the maximum root mean square reference value and the resting root mean square baseline, and then multiplied by the proportional gain coefficient and the system's maximum torque limit value to obtain a reference torque that is proportional to the electromyographic intensity. Quasi-assisted torque; S43, when the movement mode is identified as going down stairs, an inverse proportional damping operator is constructed by superimposing the basic damping coefficient with the reciprocal of the normalized electromyography intensity, and the damping is calculated on the knee joint angular velocity to obtain the benchmark auxiliary torque in which the weaker the muscle, the stronger the damping; S44, when the movement mode is identified as sitting to standing transfer, the standing process is divided into three stages according to the knee joint angle: full assistance interval, linear decrease interval, and force withdrawal interval. The maximum torque limit of the system is reduced in segments according to a preset ratio to obtain the benchmark auxiliary torque that varies with the angle.

[0060] First, the system parses the motion mode identifier output in step S3 and guides the torque calculation of the current control cycle to the corresponding algorithm branch based on its value. The specific algorithm structure of the four torque branches is as follows.

[0061] When the movement mode is identified as flat ground walking, the system executes a zero-impedance servo control strategy. The goal of this strategy is to make the exoskeleton appear as a transparent, resistance-free servo device to the user, achieved through feedforward friction and inertia compensation. Based on pre-identified Coulomb friction torque, viscous friction coefficient, and equivalent moment of inertia of the exoskeleton joint transmission mechanism, the system performs feedforward compensation calculations on the knee joint angular velocity and angular acceleration. The torque formula is as follows:

[0062] in, The Coulomb friction torque of the exoskeleton joint transmission mechanism is determined through offline identification experiments. Its typical range is 0.1 Nm to 0.5 Nm, and in this embodiment, it is 0.2 Nm. sgn(·) is the standard sign function, which takes the direction sign of the knee joint angular velocity. The identification value of the viscous friction coefficient of the exoskeleton joint, determined by offline identification experiments, typically ranges from 0.005 Nm·s / deg to 0.02 Nm·s / deg, and is 0.01 Nm·s / deg in this embodiment. Real-time knee joint angular velocity; This is the equivalent moment of inertia of the exoskeleton joint actuator, converted to the output end via a reducer, with a typical value range of 0.01 kg·m. 2 Up to 0.05 kg·m 2 In this embodiment, the value is taken as 0.02 kg·m 2 ; The real-time knee joint angular acceleration is obtained by differentiating the knee joint angular velocity with respect to time. The first two terms in this formula compensate for the resistance generated by Coulomb friction and viscous friction, respectively, while the third term compensates for the inertial resistance generated by the exoskeleton's own rotational inertia during acceleration and deceleration. After these three compensations, the user experiences a follow-up resistance of no more than 0.3 Nm, and the exoskeleton behaves as a follow-up device transparent to human movement.

[0063] When the movement mode is identified as climbing stairs, the system executes a proportional electromyography (EMG)-assisted control strategy. This strategy establishes a continuous proportional mapping between normalized EMG intensity and output torque, allowing the assistance amplitude to dynamically change in real time according to the actual muscle activation level. The system first normalizes the instantaneous root mean square value of the rectus femoris muscle using the following normalization formula:

[0064] in, This represents the instantaneous root mean square value of the rectus femoris muscle. The rectus femoris channel component in the resting root mean square baseline calibrated in step S1; This is the maximum root mean square reference value calibrated in step S1. The normalized result is constrained to the range of 0 to 1 by amplitude limiting; that is, a value less than 0 is taken as 0, and a value greater than 1 is taken as 1. A normalized electromyography (EMG) intensity value of 0 indicates that the muscle is at rest, and a value of 1 indicates that the maximum voluntary contraction level has been reached. Based on this, the assist torque is calculated:

[0065] in, The adjustable proportional gain coefficient is set to 5 levels in this embodiment, with values ​​of 0.2, 0.4, 0.6, 0.8 and 1.0 respectively. The therapist can select the appropriate level according to the user's rehabilitation stage and muscle strength level. The maximum assist torque limit of the system is set at 16 Nm in this embodiment. This formula ensures that the output torque is continuously proportional to the user's current electromyographic activation level: when the user increases the force, the electromyographic amplitude increases, the normalized electromyographic intensity increases, and the system output torque increases synchronously; when the user decreases the force, the normalized electromyographic intensity decreases, and the system assist decreases synchronously, avoiding excessive compensation that inhibits active force exertion.

[0066] When the movement mode is identified as descending stairs, the system executes an angular velocity-EMG inverse proportional damping control strategy. The design principle of this strategy is that the knee joint is in an eccentric braking state when descending stairs; the weaker the muscles, the higher the risk of uncontrolled joint flexion, and the stronger the exoskeleton needs to provide damping cushioning. The system constructs an inverse proportional damping operator by superimposing the reciprocal of the normalized EMG intensity with the baseline damping coefficient, and calculates the damping torque for the knee joint angular velocity using the following formula:

[0067] in, ε is the basic damping coefficient under the buffer state of going down stairs, and its typical value range is from 0.05 Nm·s / deg to 0.15 Nm·s / deg. In this embodiment, the value is 0.08 Nm·s / deg; ε is a minimum factor constant to prevent the denominator from being zero. In this embodiment, it is set to 0.1. The aforementioned normalized electromyographic intensity; For the real-time knee joint angular velocity. The characteristic of this formula is: when... When the value is small (low muscle activation level, indicating weak muscle strength), the fractional term has a larger value, the total damping coefficient increases, and a stronger damping torque is output to provide more eccentric braking protection; when When the value is large (indicating high muscle activation level, meaning the muscle has sufficient eccentric braking capacity), the fractional term is small, the total damping coefficient decreases, and the exoskeleton reduces intervention, allowing the user to autonomously control the descent speed. This mechanism ensures an adaptive protective characteristic where the lower the electromyographic level, the stronger the damping.

[0068] When the movement mode is identified as sit-to-stand transition, the system executes a multi-stage progressively decreasing assist control strategy. This strategy divides the standing process into three stages based on the knee joint angle, and within each stage, the system's maximum torque limit is progressively reduced in segments according to a preset ratio. The full assist range corresponds to a knee joint angle from 90 degrees to 60 degrees. Within this range, the user is in the initial stage of standing up, where the knee joint torque demand is at its maximum, and the system output is 80% of the system's maximum torque limit, i.e., 0.80 multiplied by [presumably a factor]. The linear decreasing range corresponds to the knee joint angle from 60 degrees to 30 degrees. Within this range, as the knee joint gradually extends, the gravitational torque arm shortens, and the load gradually decreases, with the system output torque linearly decreasing from 80% to 30%. The force withdrawal range corresponds to the knee joint angle from 30 degrees to 0 degrees. Within this range, the user is close to a standing position, and the system output torque linearly decreases from 30% to zero, prompting the user to rely on their own muscle strength to complete the final straightening movement in the final stage. Taking the force withdrawal range as an example, the torque calculation formula in this stage is:

[0069] in This represents the current real-time knee joint angle. The formula for calculating the torque within the linearly decreasing interval is:

[0070] This segmented decreasing strategy takes into account both the need for assistance in the early stages of standing up and the training goal of restoring active control ability in the later stages of standing up, avoiding the suppression of active muscle participation by applying too much assistance within the angle range where the user already has the ability to stand independently.

[0071] In the sub-stage of anti-hyperextension safety reconstruction and final reference torque output, the system performs anti-hyperextension protection processing on the torque values ​​calculated by the aforementioned branches. The system sets two key parameters: the initial force-removal angle and the full force-removal boundary angle. The typical range of the initial force-removal angle is 5 to 15 degrees, and in this embodiment, it is set to 10 degrees; the typical value of the full force-removal boundary angle is 0 degrees, i.e., the knee joint is fully extended. When the real-time knee joint angle is greater than the initial force-removal angle, the system does not make any correction to the torque, and the output values ​​of the aforementioned branches are directly used as the reference auxiliary torque.

[0072] When the real-time knee joint angle is between the initial angle of force withdrawal and the boundary angle of complete force withdrawal, the system performs linear force withdrawal correction on the torque. The correction formula is as follows:

[0073] in, The original torque values ​​calculated for each mode branch; This represents the current real-time knee joint angle. The initial angle for force withdrawal is 10 degrees in this embodiment; The complete force-removal boundary angle is 0 degrees in this embodiment. This formula ensures that the torque gradually decreases to zero in a linear proportion as the joint approaches full extension, avoiding the application of extension torque near the extension limit, which could lead to knee hyperextension.

[0074] When the real-time knee joint angle is less than 0 degrees, indicating that the knee joint has exceeded the fully extended position and is showing a tendency to hyperextension, the system activates hardware-level limit locking protection, forcibly resetting the reference auxiliary torque to zero and triggering the mechanical limit locking mechanism. The torque value after anti-hyperextension safety reconstruction is output as the final reference auxiliary torque to step S5 for subsequent fatigue compensation and co-contraction protection adjustment.

[0075] In one specific embodiment, a user wearing a knee exoskeleton is performing a stair-climbing motion. During a certain control cycle, the motion mode output in step S3 is identified as the stair-climbing mode, and the system enters the proportional electromyography (EMG)-assisted calculation branch. The instantaneous root mean square value of the rectus femoris muscle is obtained from step S2 as 85 microvolts, and the resting root mean square baseline rectus femoris component is obtained from step S1 as 5.2 microvolts, with a maximum root mean square reference value of 192 microvolts. The system calculates the normalized EMG intensity as (85-5.2) / (192-5.2) equals 0.427, which is confirmed to be within the range of 0 to 1 after amplitude limiting calculation. The therapist selects the 3rd proportional gain level for this user. The value is set to 0.6, and the maximum system torque is 16 Nm. The calculated assist torque is 0.6 multiplied by 16 multiplied by 0.427, which equals 4.10 Nm. The current knee joint angle is 45 degrees, which is 10 degrees greater than the initial angle of force withdrawal, so the anti-hyperextension correction is not triggered. The system outputs 4.10 Nm as the baseline assist torque to step S5.

[0076] At another moment, while the same user is walking on flat ground, the angular velocity of the knee joint during the swing phase is -150 degrees per second (the joint moves in the direction of extension), and the angular acceleration of the knee joint is -200 degrees per second squared. The system enters the zero-impedance follower branch. The friction and inertia compensation torque is calculated as 0.2 multiplied by 1 plus 0.01 multiplied by 150 plus 0.02 multiplied by 200, which equals 5.70 Nm. This torque is in the same direction as the direction of movement and is used to counteract the friction and inertial resistance of the exoskeleton. The follower resistance actually felt by the user is less than 0.3 Nm. This value is used as the reference auxiliary torque output to step S5.

[0077] At another moment, the user performs a seat-to-stand transition with the current knee angle at 75 degrees, within the full-assist range (90 to 60 degrees), and the system output torque is 16 multiplied by 0.80, equaling 12.8 Nm. When the knee extends to 45 degrees, it enters the linearly decreasing range, and the system output torque is 16 multiplied by (0.30 + (0.80 - 0.30) multiplied by (45 - 30) / (60 - 30)) equaling 16 multiplied by 0.55, equaling 8.8 Nm. When the knee continues to extend to 15 degrees, it enters the force-reducing range, and the system output torque is 16 multiplied by 0.30 multiplied by (15 - 0) / (30 - 0) equaling 2.4 Nm. The torque values ​​calculated at each of the above moments are all verified by anti-hyperextension safety reconstruction and then used as the baseline auxiliary torque to be transmitted to step S5.

[0078] S5 uses a dual-criteria approach—the sliding window descent rate of the rectus femoris instantaneous root mean square value and the leftward shift rate of the median frequency value of the rectus femoris—to detect fatigue. It also monitors abnormal co-contraction based on the instantaneous root mean square values ​​of the rectus femoris and biceps femoris, and performs adaptive safety adjustments to the baseline auxiliary torque to obtain adaptive driving torque commands. It is understandable that during actual rehabilitation training, the user's muscle physiological state dynamically changes over time. When the rectus femoris gradually enters a fatigued state, its ability to generate active torque decreases. If the exoskeleton continues to output the nominal torque at the non-fatigue state, the assistance may be insufficient to compensate for the declining muscle strength, preventing the user from safely completing the movement. Alternatively, continuing to apply the pre-fatigue resistance level during resistance training may accelerate muscle damage. Furthermore, when abnormal co-contraction patterns occur between the agonist (rectus femoris) and antagonist (biceps femoris), the knee joint is simultaneously subjected to bidirectional tension in both flexion and extension directions, causing a sharp increase in intra-articular pressure. If the system fails to identify and intervene in time, it may lead to articular cartilage damage or ligament sprains. Existing knee exoskeletons primarily rely on passive protection measures such as angle limiting and torque limiting, lacking the ability to actively detect and intervene in the aforementioned muscle physiological degeneration. Therefore, step S5 introduces an active safety monitoring layer based on electromyographic indicators at the final stage of torque output. This layer detects muscle fatigue in real time through a time-frequency dual-domain joint criterion, monitors abnormal synergistic patterns in real time through the antagonist muscle co-contraction ratio, and performs adaptive safety adjustments to the baseline auxiliary torque based on the detection results. Ultimately, it generates an adaptive drive torque command that can directly drive the servo motor.

[0079] The specific execution process of step S5 includes three parallel sub-stages that eventually converge: joint fatigue detection of sliding window descent rate and median frequency left shift rate, abnormal monitoring of antagonistic muscle co-contraction ratio, and adaptive torque safety adjustment based on the detection results.

[0080] In one possible implementation of this application, step S5 includes: S51, continuously calculating the proportional change of the instantaneous root mean square value of the rectus femoris muscle relative to the initial reference period mean using a sliding window of a preset length as the sliding window descent rate, and simultaneously calculating the proportional change of the median frequency value of the rectus femoris muscle relative to the initial reference period mean as the median frequency left shift rate; S52, when the sliding window descent rate is lower than the descent threshold and the median frequency left shift rate is lower than the left shift threshold, and both indicators are simultaneously satisfied and continue to exceed the fatigue confirmation time, a muscle fatigue state is determined, and incremental compensation for the baseline auxiliary torque is performed; S53, the ratio of the instantaneous root mean square value of the biceps femoris muscle to the sum of the instantaneous root mean square values ​​of the rectus femoris muscle and the instantaneous root mean square value of the biceps femoris muscle is calculated in real time as the co-contraction ratio. When the co-contraction ratio exceeds the abnormal threshold and continues to exceed the co-contraction confirmation time, an abnormal co-contraction mode is determined, and a protective reduction of the baseline auxiliary torque is performed according to a preset ratio. After the above fatigue compensation and co-contraction protection, an adaptive driving torque command is obtained.

[0081] First, in step S51, the system establishes a reference baseline during the initial phase of exoskeleton operation. The initial reference period is the first 30 seconds after training begins. The temporal average of the instantaneous root mean square (RMS) values ​​of the rectus femoris muscle during this period is calculated and used as the RMS reference baseline. Simultaneously, the temporal average of the median frequency values ​​of the rectus femoris muscle during this period is calculated and used as the median frequency reference baseline. The initial reference period is chosen to be 30 seconds because this duration is sufficient to cover multiple complete gait cycles or functional movement cycles, allowing for the acquisition of statistically representative baseline values. Furthermore, within the first 30 seconds of training, the muscles are still in a non-fatigued state, ensuring that the baseline values ​​accurately reflect the user's initial electrophysiological level.

[0082] After establishing the reference baseline, the system opens a sliding analysis window with a length of 5 seconds and a step size of 1 second, continuously updating the two proportional changes throughout the training process. The formula for calculating the sliding window descent rate is:

[0083] in, This is the window mean of the instantaneous root mean square value of the rectus femoris muscle within the current 5-second sliding window; The root mean square reference standard is established for the initial reference period. The formula for calculating the left shift rate of the median frequency is:

[0084] in, This is the window mean of the median frequency values ​​of the rectus femoris muscle within the current 5-second sliding window. The median frequency reference base established for the initial reference period.

[0085] The fatigue determination criteria are as follows: the sliding window descent rate is lower than the descent threshold and the median frequency left shift rate is lower than the left shift threshold. When both indicators are met simultaneously and persist for more than the fatigue confirmation duration, the system determines that the rectus femoris muscle has entered a state of physiological fatigue. The typical value for the descent threshold is -0.15, meaning the root mean square value decreases by more than 15% relative to the initial baseline; the typical value for the left shift threshold is -0.10, meaning the median frequency shifts to a lower frequency by more than 10% relative to the initial baseline; the typical value for the fatigue confirmation duration is 30 seconds, meaning that the state of both indicators exceeding the limit must last for 30 seconds to confirm fatigue, avoiding misjudgments caused by brief changes in exercise rhythm. The reason for using a dual-indicator criterion is that: a single decrease in the root mean square value may be caused by the user actively reducing exercise intensity or changing exercise rhythm, and is not true fatigue; a single left shift in the median frequency may be affected by changes in electrode-skin interface impedance caused by sweat penetration; only when the time-domain energy index and the frequency-domain spectral index show characteristic changes simultaneously do they have highly specific physiological indicators of muscle fatigue. When the dual-indicator joint criterion is met and the duration reaches 30 seconds, the system generates a fatigue trigger sub-flag, which is set to 1, indicating that the fatigue state has been confirmed.

[0086] Then, in step S52, the system calculates in real time the ratio of the instantaneous root mean square value of the biceps femoris to the sum of the instantaneous root mean square values ​​of the rectus femoris and the biceps femoris, i.e., the co-contraction ratio. The calculation formula is as follows:

[0087] in, This represents the instantaneous root mean square value of the biceps femoris muscle. This is the instantaneous root mean square value of the rectus femoris muscle. The physical significance of this ratio lies in quantifying the proportion of the antagonist muscle (biceps femoris) in the total muscle electrical activity of the knee joint. During normal knee extension, the agonist muscle (rectus femoris) dominates, while the antagonist muscle only provides necessary joint stability and synergistic activation; the typical co-contraction ratio is 0.15 to 0.25. When the user overactivates the biceps femoris due to pain-protective reflexes, neurological control disorders, or incorrect compensatory movement strategies, the co-contraction ratio will abnormally increase. In this case, the knee joint simultaneously bears opposing torques in the flexion and extension directions, significantly increasing intra-articular pressure.

[0088] The criteria for determining abnormal co-contraction are as follows: when the co-contraction ratio exceeds the abnormal threshold and persists beyond the co-contraction confirmation duration, the system determines it as an abnormal co-contraction pattern. A typical value for the abnormal threshold is 0.5, meaning the antagonist muscle activation level reaches more than 50% of the total activation level, far exceeding the antagonistic synergy level in normal knee extension. A typical value for the co-contraction confirmation duration is 500 ms, meaning the excessive co-contraction ratio must persist for 500 ms to confirm the abnormality, excluding transient high ratio events caused by the muscle relaxation-contraction transition. The abnormal threshold can be adjusted by the therapist within the range of 0.3 to 0.7 based on the user's specific condition and rehabilitation stage. When the duration of the excessive co-contraction ratio reaches the confirmation duration, the system generates a synergy abnormality sub-flag, set to 1, indicating that the abnormal co-contraction pattern has been confirmed.

[0089] Finally, in step S53, the system performs differentiated adaptive adjustment processing on the reference auxiliary torque output in step S4 based on the state of the fatigue trigger sub-flag and the cooperative anomaly sub-flag.

[0090] When the fatigue trigger flag is 1, the system determines that the user's rectus femoris muscle has entered a state of fatigue, and its active force production capacity has decreased. At this time, the system performs incremental assistance compensation on the baseline assist torque, that is, it adds a preset proportion of compensation increment to the current baseline assist torque, so that the exoskeleton automatically increases the assist output to maintain the safety of completing the movement when the user's muscle strength declines. The typical value range of the compensation ratio is 10% to 30%, and in this embodiment, it is set to 20%, that is, the baseline assist torque increases by 20% in the state of fatigue.

[0091] When the coordination anomaly sub-flag is 1, the system determines that the user is experiencing an abnormal co-contraction pattern, with the joints subjected to inappropriate bidirectional antagonistic loads. At this time, the system protectively reduces the baseline auxiliary torque by a preset ratio, immediately decreasing the torque applied by the exoskeleton to alleviate the overall load on the joints. A typical reduction ratio is 25%, meaning that in the abnormal co-contraction state, the system immediately reduces the current torque output by 25% and simultaneously issues a warning signal to notify the therapist or user.

[0092] When both the fatigue trigger sub-flag and the synergistic anomaly sub-flag are 0, the system does not make any correction to the baseline auxiliary torque, and the baseline auxiliary torque is directly output as the adaptive drive torque command. When both sub-flags are 1 at the same time (i.e., fatigue and abnormal co-contraction coexist), the system performs a protective downshift of co-contraction based on the principle of safety priority, because the risk of immediate joint damage caused by abnormal co-contraction is higher than the risk of insufficient assistance caused by fatigue.

[0093] Based on the above logic, the final calculation of the adaptive driving torque command can be expressed as: when the fatigue trigger sub-flag is 1 and the cooperative anomaly sub-flag is 0:

[0094] in, The reference auxiliary torque output in step S4; The fatigue compensation increment ratio coefficient is set to 0.20 in this embodiment.

[0095] When the collaborative anomaly sub-flag is 1 (regardless of the fatigue trigger sub-flag status):

[0096] in, The coefficient for reducing the shrinkage protection ratio is set to 0.25 in this embodiment.

[0097] When both sub-flags are 0:

[0098] The adaptive drive torque command obtained after the above adaptive safety adjustment is the final output of step S5. This value is directly sent to the current loop of the exoskeleton servo driver to drive the brushless DC servo motor to output the corresponding joint torque.

[0099] In one specific embodiment, a user wears a knee exoskeleton for resistance training up stairs, and the training has lasted for 15 minutes. The system uses the time average of the instantaneous root mean square value of the rectus femoris muscle (RMS value) of 82 microvolts within the first 30 seconds after the start of training as the RMS reference benchmark, and the time average of the median frequency value of the rectus femoris muscle within the first 30 seconds as the median frequency reference benchmark. Within the current 5-second sliding window, the window mean of the instantaneous RMS value of the rectus femoris muscle decreases to 67 microvolts, and the system calculates the sliding window decrease rate as (67-82) / 82, which equals -0.183, lower than the decrease threshold of -0.15; simultaneously, the window mean of the median frequency value of the rectus femoris muscle within the current window decreases to 71 Hz, and the system calculates the median frequency left shift rate as (71-80) / 80, which equals -0.113, lower than the left shift threshold of -0.10. The state of both indicators exceeding the limit has lasted for 35 seconds, exceeding the fatigue confirmation time of 30 seconds, and the system confirms the fatigue state, setting the fatigue trigger sub-flag to 1. Meanwhile, the current instantaneous root mean square value of the rectus femoris is 65 microvolts, and the instantaneous root mean square value of the biceps femoris is 18 microvolts. The co-contraction ratio is 18 / (65+18) equals 0.217, which is lower than the abnormal threshold of 0.5. The synergistic abnormality sub-flag remains at 0. The baseline auxiliary torque output in step S4 is 4.10 Nm. The system performs fatigue compensation: 4.10 multiplied by (1+0.20) equals 4.92 Nm. The 4.92 Nm is output as an adaptive drive torque command to the servo driver. In another training scenario, the user experienced protective muscle spasms due to pain. The instantaneous root mean square value of the rectus femoris was 70 microvolts, and the instantaneous root mean square value of the biceps femoris was 78 microvolts. The co-contraction ratio was 78 / (70+78) equal to 0.527, which exceeded the abnormal threshold of 0.5 and lasted for 600ms, exceeding the confirmation time of 500ms. The synergistic abnormal sub-flag was set to 1, and the system immediately performed a protective reduction on the current baseline auxiliary torque of 5.2 Nm: 5.2 multiplied by (1-0.25) equals 3.9 Nm. 3.9 Nm was output as the adaptive drive torque command, and an alarm signal was sent to the therapist's terminal at the same time.

[0100] In summary, the knee exoskeleton collaborative control method based on the fusion of surface electromyography (SEMG) feedforward based on the embodiments of this application is elucidated. First, utilizing the time lead characteristic of SEMG signals relative to joint mechanical movement, a normalized reference frame between the resting state and the maximum contraction state is established through individualized baseline calibration. On this basis, a statistical dynamic threshold is used to realize the feedforward trigger determination of movement intention, thereby advancing the exoskeleton's response timing from the traditional posterior judgment of movement to the muscle activation initiation stage. Second, the SEMG feedforward trigger signal is used as the admission enable threshold for pattern recognition. Knee joint angle, angular velocity, and thigh tilt angle output by the inertial measurement unit are fused. A priority logic decision tree is used to distinguish four movement modes: walking, climbing stairs, descending stairs, and sitting-standing transfer. Then, based on the biomechanical requirements of different modes, differentiated torque strategies such as friction compensation follow-up, proportional SEMG assistance, inverse proportional damping buffer, and segmented decreasing assistance are executed to establish a continuous proportional mapping relationship between SEMG amplitude and output torque. Finally, a dual-index joint fatigue detection mechanism based on the root mean square descent rate in the time domain and the left shift rate of the median frequency in the frequency domain, as well as an abnormal collaborative monitoring mechanism based on the co-contraction ratio of antagonistic muscles, are introduced in the torque output stage to implement adaptive safety adjustment of the reference auxiliary torque, thereby embedding active physiological state perception and protection capabilities into the torque control loop.

[0101] Figure 6 The diagram below shows a block diagram of a knee exoskeleton collaborative control system incorporating surface electromyography feedforward according to an embodiment of this application. Figure 6As shown, the knee exoskeleton collaborative control system 600 integrating surface electromyography (EMG) feedforward includes: a baseline calibration module 610, used to perform individualized baseline calibration on dual-channel EMG signals acquired by the user in resting and maximally contracted states to obtain the resting root mean square baseline, resting standard deviation, and maximum root mean square reference value; an intent trigger determination module 620, used to perform EMG signal time-frequency state analysis and feedforward intent trigger determination on real-time dual-channel EMG signals based on the resting root mean square baseline and resting standard deviation to obtain the instantaneous root mean square value of the rectus femoris, the instantaneous root mean square value of the biceps femoris, the median frequency value of the rectus femoris, and the EMG feedforward trigger signal; and a pattern discrimination module 630, used as the EMG feedforward trigger signal as an access enable, integrating knee joint angle, knee joint angular velocity, and thigh... The tilt angle and the growth rate of the instantaneous root mean square value of the rectus femoris relative to the maximum root mean square reference value are used to determine the motion mode through a priority logic decision tree to obtain the motion mode identifier; the torque allocation module 640 is used to select the corresponding torque branch based on the motion mode identifier, and to perform a coordinated torque strategy allocation on the instantaneous root mean square value of the rectus femoris, the maximum root mean square reference value, the knee joint angle, and the knee joint angular velocity to obtain the baseline auxiliary torque; the safety adjustment module 650 is used to perform fatigue detection based on the dual-indicator joint criterion of the sliding window descent rate of the instantaneous root mean square value of the rectus femoris and the leftward shift rate of the median frequency value of the rectus femoris, and to perform co-contraction abnormality monitoring based on the instantaneous root mean square values ​​of the rectus femoris and the instantaneous root mean square values ​​of the biceps femoris, and to perform adaptive safety adjustment on the baseline auxiliary torque to obtain the adaptive drive torque command.

[0102] Here, those skilled in the art will understand that the specific operations of each step in the above-described knee exoskeleton collaborative control system incorporating surface electromyography feedforward have been referenced above. Figures 1 to 5 The description of the fusion surface electromyography feedforward method for coordinated control of the knee exoskeleton is detailed here, and therefore, its repeated description will be omitted.

[0103] The terms "first," "second," etc., are used only to distinguish different technical features and do not imply their importance or number. The designation of a feature as "first" or "second" does not preclude the possibility of additional similar features. The scope of protection of this application shall be determined by the claims; any modifications, substitutions, or combinations that do not depart from the core idea of ​​this application shall fall within the scope of protection of this application.

[0104] The above description is merely a preferred embodiment of this application and is not intended to limit this application in any way. Although this application has been disclosed above with reference to preferred embodiments, it is not intended to limit this application. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the technical solution of this application. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of this application without departing from the scope of the technical solution of this application shall still fall within the scope of the technical solution of this application.

Claims

1. A method for coordinated control of a knee exoskeleton incorporating surface electromyography feedforward, characterized in that, include: S1, Individualized baseline calibration of dual-channel surface electromyography signals collected from the user in the resting and maximal contraction states is performed to obtain the resting root mean square baseline, resting standard deviation, and maximum root mean square reference value; S2, based on the resting root mean square baseline and resting standard deviation, performs time-frequency situational analysis and feedforward intent triggering determination on the real-time dual-channel surface electromyography (EMG) signals to obtain the instantaneous root mean square value of the rectus femoris, the instantaneous root mean square value of the biceps femoris, the median frequency value of the rectus femoris, and the EMG feedforward trigger signal. S3 uses electromyography feedforward trigger signal as the access enable, integrates knee joint angle, knee joint angular velocity, thigh tilt angle and the growth rate of instantaneous root mean square value of rectus femoris relative to the maximum root mean square reference value, and performs motion pattern discrimination through priority logic decision tree to obtain motion pattern identifier. S4, selects the stress torque branch based on the motion mode identifier, and performs a coordinated torque strategy to adjust the instantaneous root mean square value of the rectus femoris muscle, the maximum root mean square reference value, the knee joint angle and the knee joint angular velocity to obtain the reference auxiliary torque; S5 performs fatigue detection based on a dual-indicator criterion of the sliding window descent rate of the instantaneous root mean square value of the rectus femoris and the leftward shift rate of the median frequency value of the rectus femoris. It also monitors co-contraction abnormalities based on the instantaneous root mean square values ​​of the rectus femoris and the biceps femoris. It performs adaptive safety adjustment on the baseline auxiliary torque to obtain an adaptive driving torque command.

2. The knee exoskeleton collaborative control method integrating surface electromyography feedforward according to claim 1, characterized in that, The dual-channel surface electromyography (EMG) signal includes EMG signals from the rectus femoris muscle and the biceps femoris muscle, wherein step S1 includes: S11, perform dual digital filtering on the dual-channel surface electromyography signal to remove motion artifacts and grid interference, and obtain the resting state rectus femoris filtered signal and the resting state biceps femoris filtered signal. S12, perform non-overlapping sliding window statistical calculations on the resting rectus femoris filtered signal and the resting biceps femoris filtered signal to obtain the resting root mean square baseline and resting standard deviation; S13, during the user's multiple maximal isometric knee extension contractions, the root mean square extreme value is retrieved from the contraction segment of the maximum contraction state rectus femoris filtered signal using a peak sliding window, and the absolute maximum value among the multiple contractions is taken as the maximum root mean square reference value.

3. The knee exoskeleton collaborative control method integrating surface electromyography feedforward according to claim 1, characterized in that, Step S2 includes: S21. After bandpass filtering and power frequency notch preprocessing of the real-time dual-channel surface electromyography signal, the instantaneous root mean square values ​​of the rectus femoris channel and the biceps femoris channel are calculated using a short-time sliding window to obtain the instantaneous root mean square values ​​of the rectus femoris and the biceps femoris. S22, the resting root mean square baseline is added to the resting standard deviation by a preset multiple to construct a dynamic trigger threshold. When the instantaneous root mean square value of the rectus femoris muscle exceeds the dynamic trigger threshold and the continuous holding time reaches the confirmation window duration, it is determined to be an active movement intention trigger and an electromyographic feedforward trigger signal is generated. S23, perform a fast Fourier transform on the filtered signal of the rectus femoris channel using a long-time sliding window and extract the median frequency of the power spectral density to obtain the median frequency value of the rectus femoris.

4. The knee exoskeleton collaborative control method integrating surface electromyography feedforward according to claim 1, characterized in that, The decision rule for the priority logic decision tree in step S3 is as follows: S31. When the electromyographic feedforward trigger signal is effective and the knee joint angle is greater than the preset sitting flexion angle and begins to extend, while the thigh tilt angle exceeds the forward tilt threshold and the instantaneous root mean square value of the rectus femoris muscle increases relative to the maximum root mean square reference value within the preset time period, it is determined to be a sitting-to-standing transfer mode. S32, when the electromyographic feedforward trigger signal is valid and the knee joint angle is greater than the preset upstairs flexion angle and moves in the extension direction, while the knee joint angular velocity is negative and its absolute value exceeds the extension velocity threshold, and the thigh tilt angle exceeds the upstairs tilt threshold, it is determined to be the upstairs mode. S33, when the electromyography feedforward trigger signal is valid and the knee joint angular velocity is positive and its absolute value exceeds the flexion velocity threshold, and the knee joint angle shows a continuous increasing trend within the preset short-term observation window, it is determined to be the stair-going mode. S34. When the electromyographic feedforward trigger signal is valid and the knee joint angle is within the normal gait range, and the knee joint angular velocity changes alternately between positive and negative and the absolute value of the thigh tilt angle is lower than the forward tilt discrimination threshold, it is determined to be a flat ground walking mode.

5. The knee exoskeleton collaborative control method integrating surface electromyography feedforward according to claim 1, characterized in that, Step S4 includes: S41, when the movement mode is identified as walking on flat ground, feedforward friction compensation is performed on the knee joint angular velocity based on the pre-identified Coulomb friction torque and viscous friction coefficient to obtain a reference auxiliary torque that makes the follow-up resistance approach zero. S42, when the movement mode is identified as climbing stairs, the instantaneous root mean square value of the rectus femoris muscle is subtracted from the resting root mean square baseline, and then normalized by the difference between the maximum root mean square reference value and the resting root mean square baseline. Then, it is multiplied by the proportional gain coefficient and the maximum torque limit value of the system to obtain the reference auxiliary torque that is proportional to the electromyographic intensity. S43, when the movement mode is identified as going down stairs, an inverse proportional damping operator is constructed by superimposing the basic damping coefficient with the reciprocal of the normalized electromyography intensity, and the damping is calculated on the knee joint angular velocity to obtain the reference auxiliary torque that the weaker the muscle, the stronger the damping. S44, when the movement mode is identified as sit-to-stand transition, the standing process is divided into three stages based on the knee joint angle: full assistance interval, linear decrease interval, and force withdrawal interval. The maximum torque limit of the system is reduced in segments according to a preset ratio to obtain the reference auxiliary torque that changes with the angle.

6. The knee exoskeleton collaborative control method incorporating surface electromyography feedforward according to claim 1, characterized in that, Step S5 includes: S51, continuously calculate the change in the ratio of the instantaneous root mean square value of the rectus femoris muscle to the mean value of the initial reference period using a sliding window of a preset length as the sliding window descent rate, and simultaneously calculate the change in the ratio of the median frequency value of the rectus femoris muscle to the mean value of the initial reference period as the median frequency left shift rate. S52, when the sliding window descent rate is lower than the descent threshold and the median frequency left shift rate is lower than the left shift threshold, and both indicators are met and continue to exceed the fatigue confirmation time, it is determined to be a muscle fatigue state, and assist incremental compensation is performed on the baseline assist torque. S53, the ratio of the instantaneous root mean square value of the biceps femoris muscle to the sum of the instantaneous root mean square values ​​of the rectus femoris muscle and the instantaneous root mean square values ​​of the biceps femoris muscle is calculated in real time as the co-contraction ratio. When the co-contraction ratio exceeds the abnormal threshold and continues to exceed the co-contraction confirmation time, it is determined to be an abnormal co-contraction mode. The reference auxiliary torque is protectively reduced according to a preset ratio. After the above fatigue compensation and co-contraction protection, the adaptive driving torque command is obtained.

7. The knee exoskeleton collaborative control method integrating surface electromyography feedforward according to claim 5, characterized in that, In step S44, the full-assist interval corresponds to the knee joint angle from 90 degrees to 60 degrees and the output is 80% of the system's maximum torque limit value; the linear decrease interval corresponds to the knee joint angle from 60 degrees to 30 degrees and the output decreases linearly from 80% to 30%; and the force withdrawal interval corresponds to the knee joint angle from 30 degrees to 0 degrees and the output decreases linearly from 30% to zero.

8. The knee exoskeleton collaborative control method integrating surface electromyography feedforward according to claim 4, characterized in that, Each discrimination rule is arbitrated in the priority order of sitting to standing, going up stairs, going down stairs, and walking on flat ground, and a de-jittering window is set to lock the motion mode identifier in time.

9. The knee exoskeleton collaborative control method integrating surface electromyography feedforward according to claim 4, characterized in that, The timing lock for arbitration of each discrimination rule and motion pattern identification includes: S35, based on the pre-constructed directed transition reachability adjacency matrix, extracts the reachability mask vector with the current locked motion mode identifier as the row index, performs element-wise logical AND masking on the candidate mode trigger flag vectors output independently by each discrimination rule, forces the candidate modes that are biomechanically inaccessible to be invalidated, and obtains the reachability constraint candidate mode vector. S36. For each candidate mode that is still valid in the accessibility constraint candidate mode vector, calculate the normalized margin distance of the knee joint angle, knee joint angular velocity, thigh tilt angle and instantaneous root mean square value of rectus femoris muscle involved in its discrimination rule relative to its respective judgment threshold boundary, and perform weighted fusion to obtain the transfer confidence score of each candidate mode. Check whether the transfer confidence score exceeds the minimum confidence admission threshold in the order of priority of sitting to standing transfer, going up stairs, going down stairs and walking on flat ground in order to select the arbitration winning candidate mode identifier and its corresponding winning mode confidence score. S37. Based on the pre-constructed transfer deshake duration lookup matrix, the baseline deshake duration is extracted using the currently locked motion mode identifier and the arbitration winner candidate mode identifier as a joint index. The baseline deshake duration is inversely scaled using the ratio of the lowest confidence admission threshold to the confidence score of the winner mode to obtain the adaptive deshake duration. When the cumulative time for the arbitration winner candidate mode identifier to remain unchanged continuously reaches the adaptive deshake duration, mode switching lock is performed to update the motion mode identifier.

10. A knee exoskeleton collaborative control system incorporating surface electromyography (SEM) feedforward, used to execute the knee exoskeleton collaborative control method incorporating SEM feedforward as described in any one of claims 1-9, characterized in that, include: The baseline calibration module is used to perform individualized baseline calibration on the dual-channel surface electromyography signals collected by the user in the resting and maximal contraction states to obtain the resting root mean square baseline, resting standard deviation, and maximum root mean square reference value. The intent trigger determination module is used to perform time-frequency situational analysis and feedforward intent trigger determination on real-time dual-channel surface electromyography signals based on the resting root mean square baseline and resting standard deviation to obtain the instantaneous root mean square value of rectus femoris, the instantaneous root mean square value of biceps femoris, the median frequency value of rectus femoris, and the electromyography feedforward trigger signal. The pattern discrimination module uses electromyography feedforward trigger signal as the access enable, integrates the knee joint angle, knee joint angular velocity, thigh tilt angle and the growth rate of the instantaneous root mean square value of the rectus femoris relative to the maximum root mean square reference value, and performs motion pattern discrimination through priority logic decision tree to obtain motion pattern identifier. The torque allocation module is used to select the corresponding torque branch based on the motion mode identifier, and to coordinate the instantaneous root mean square value of the rectus femoris muscle, the maximum root mean square reference value, the knee joint angle, and the knee joint angular velocity to obtain the reference auxiliary torque. The safety adjustment module is used to perform fatigue detection based on a dual-indicator joint criterion of the sliding window descent rate of the instantaneous root mean square value of the rectus femoris muscle and the leftward shift rate of the median frequency value of the rectus femoris muscle. It also performs co-contraction abnormality monitoring based on the instantaneous root mean square values ​​of the rectus femoris muscle and the instantaneous root mean square values ​​of the biceps femoris muscle, and performs adaptive safety adjustment on the reference auxiliary torque to obtain an adaptive driving torque command.

Images