An exoskeleton carrying assistance control method and system based on multi-modal fusion and double-machine cooperation

Abstract

The application provides a kind of exoskeleton carrying assistance control method and system based on multi-modal fusion and double-machine cooperation, relating to exoskeleton robot control technical field.The method includes real-time acquisition of multi-source perception data and preprocessing, extraction of motion assistance features, and estimation of the current carrying weight;Through the multi-modal hierarchical intention recognition model composed of spatial gate verification, biological electricity and foot pressure prior prediction and dynamics weighted verification, the action intention is predicted in advance;Dynamic double-machine coupling torque distribution mechanism is introduced, and the non-symmetrical compensation ratio of left and right legs is adjusted in real time according to the motion assistance characteristics;According to the real-time load and muscle fatigue degree, the assistance output is dynamically corrected;And through FSM, four state switches of standby monitoring, transition waiting, assistance execution and soft exit are realized;Applied to industrial carrying, logistics sorting and rehabilitation medical scene of lower limb exoskeleton robot, the human-machine cooperation flexibility and operation safety under complex working conditions are improved.

Description

Technical Field

[0001] This invention relates to the field of exoskeleton robot control technology, and in particular to an exoskeleton handling assistance control method and system based on multimodal fusion and dual-machine collaboration. Background Technology

[0002] Exoskeleton robots are widely used in industrial handling, logistics sorting, and rehabilitation medical scenarios. They are mainly used to assist the human body in carrying heavy objects and reduce the burden on the waist and lower limb muscles. However, human-machine interaction control in complex dynamic environments still faces many challenges.

[0003] Existing technologies suffer from the following shortcomings: 1. At the level of motion intention recognition, traditional control strategies mostly rely on single-dimensional surface electromyography (sEMG) signals or simple kinematic and dynamic thresholds for judgment. Relying solely on SEMG signals is highly susceptible to noise interference from sweating, electrode displacement, etc.; while relying solely on kinematic thresholds results in significant recognition lag and cannot accommodate different load and speed conditions. Existing algorithms fail to deeply integrate human bioelectrical signals with environmental mechanical parameters such as plantar pressure, making it difficult to achieve advanced and accurate perception of subtle force exertion trends. 2. At the level of bipedal coordinated control, existing algorithms often neglect the highly dynamic correlation of forces on both lower limbs during lifting movements. When facing extremely complex conditions such as uneven ground or asymmetrical eccentric load lifting, they cannot dynamically redistribute torque based on the actual force differences and joint torque differences between the left and right feet, easily leading to human imbalance or even human-machine conflict. 3. At the level of assisted execution, most exoskeletons currently use gravity compensation or sinusoidal assist models with preset constant stiffness. In actual operations, the weight of the objects being moved varies randomly, and the wearer's lower back muscles inevitably become fatigued during high-intensity work. Fixed-mode power assist output cannot adapt to real-time load changes, nor can it actively increase the assist depth when the body's muscles are fatigued, resulting in stiff power assist output and a lack of smooth, accompanying sensation. Summary of the Invention

[0004] The main objective of this invention is to provide an exoskeleton handling assistance control method based on multimodal fusion and dual-machine collaboration, which solves the problems of delayed intention recognition, poor collaboration, rigid output, and low control accuracy of existing exoskeleton systems in complex scenarios.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: an exoskeleton transport assistance control method based on multimodal fusion and dual-machine collaboration, comprising the following steps: Multi-source sensing data is acquired in real time and preprocessed to extract motion assistance features, including the difference in hip position, the difference in pressure between the left and right feet, the difference in interactive torque, and the root mean square value and average power frequency of electromyographic signals on the surface of the erector spinae muscles in the lumbar region; and the mass of the heavy object being moved is estimated online based on the dynamic observer. A multimodal hierarchical intent recognition model is constructed, and the motion assistance features are input. Spatial gating and basic collaborative verification, bioelectric and spatial prior prediction, and dynamic weighted verification are performed sequentially to make hierarchical judgments on the user's carrying starting intent. The user's carrying starting intent is determined to be true only when the bipedal posture is within the safe threshold, muscle exertion and forward shift of the center of gravity are detected, and the kinematic or dynamic features are greater than the set threshold. The finite state machine controls the exoskeleton to switch between four states: standby monitoring, transition waiting, assisted execution, and compliant exit, based on the intent recognition result. In assisted execution mode, the torque distribution coefficients of the left and right legs are calculated based on the difference in hip position, the difference in pressure between the left and right feet, and the difference in interactive torque. Combining the mass of the load and the muscle fatigue level extracted based on the average power frequency characteristics, a target assist torque is generated based on a load-fatigue dual adaptive bionic torque model. ;Will Multiply by the torque distribution coefficients for the left and right legs respectively to generate independent left and right leg assist commands and execute them; When the handling is detected to be finished, the control output torque gradually decreases to zero according to an exponential law.

[0006] In the preferred embodiment, preprocessing includes: The smoothed angular velocity and angular acceleration of the joints are calculated using a Kalman state observer. By removing the gravity, inertia, and friction terms from the system dynamics model, a pure human-machine interaction torque is obtained through decoupling. ; Bandpass filtering was performed on the sEMG signal, and the root mean square value representing muscle exertion intensity and the average power frequency representing muscle fatigue were extracted. Based on plantar pressure data, the CoP trajectory of the left and right plantar pressure center points and the plantar pressure difference are calculated. .

[0007] In the preferred scheme, an extended Kalman filter is used to dynamically predict the mass M of the heavy object online, specifically including: The discrete state transition equations are constructed as follows: ; in, , for The estimated mass of the heavy object at all times. This is process noise; The dynamic nonlinear observation equation is: ; in, They are respectively The position of the hip joint, smooth angular velocity, and angular acceleration at any given moment; This is the dynamic regression matrix corresponding to the heavy load, which mainly maps the gravity and inertia parameters of the load; To measure the noise, it follows a mean of 0 and a covariance matrix of... The Gaussian distribution, i.e. ; Within each control cycle, the system iteratively performs state prediction, covariance prediction, calculates the Jacobian matrix of the observation function, calculates the Kalman gain, updates the optimal state estimate, and updates the error covariance, outputting the optimal smoothed load quality estimate M.

[0008] In the preferred embodiment, the multimodal hierarchical intent recognition model specifically includes: First layer: Only when the hip joint is within the preset effective carrying angle range and the absolute value of the positional difference between the left and right hip joints. When the value is below the set security threshold, the subsequent identification process is activated; The second layer monitors the root mean square value of the electromyography signal on the surface of the erector spinae muscles in the waist and the center of pressure on the foot in real time. When a step change in the root mean square value is detected and the center of pressure on the foot shows a characteristic forward shift, it is determined that the user has a tendency to start carrying, and the bottom layer driver is ready in advance. The third layer: When either the smooth joint speed or the pure human-machine interaction torque exceeds the corresponding set threshold, the intention to start the transport is finally determined to be genuine.

[0009] In the preferred scheme, the finite state machine flow control is specifically as follows: Standby monitoring state: The motor output torque is zero, and the underlying layer continues to perform multi-source perception and multi-modal hierarchical intent recognition; Transition waiting state: When the intent is determined to be true, a state transition is triggered to enter this state, and the joint movement direction and speed are monitored in real time. If the user is detected to be moving in the opposite direction or remaining still for a long time, the state returns to the standby monitoring state. If the optimal assist angle is reached, the state switches to the assist execution state. The optimal assist angle is when the hip joint flexion angle is greater than a preset threshold and is within the effective range of the transport stroke. Assisted execution status: The dynamic dual-machine coupling collaborative torque allocation and dual adaptive torque execution module is invoked to perform the assist; Compliant Exit State: This state is activated when the system is detected to be upright or the speed has returned to zero. The output torque is then unloaded exponentially before returning to the standby monitoring state.

[0010] In the preferred scheme, during the compliant exit state, the output torque is unloaded according to the following exponential law: ; in, The real-time torque at the instant of entering the compliant exit state. is the damping time constant.

[0011] In the preferred embodiment, the dynamic dual-machine coupling cooperative torque distribution in step S4 specifically includes: The difference in hip position is uniformly defined by subtracting the data from the data on the left side. Plantar pressure difference and the difference in left and right interaction torque Construct a coupling strategy to dynamically calculate the initial weights on the left side: ; in, For each feature, the normalized feedback gain coefficient is used. A saturation limiting protection mechanism is introduced to calculate the final actual allocation coefficients for the left and right sides: ; ; in, and The preset security boundary threshold for the system.

[0012] In the preferred scheme, the load-fatigue dual adaptive sinusoidal biomimetic torque model is as follows: ; in, The basic auxiliary stiffness coefficient set for the system; For load adaptive compensation, For the EKF online estimated weight of the object, For load gain, For fatigue adaptive compensation term, For fatigue gain, This is a muscle fatigue function extracted based on the average power frequency. This is a hip joint angle mapping strategy.

[0013] In the preferred embodiment, the muscle fatigue function The calculation method is as follows: ; in, The average power frequency of the electromyographic signal on the surface of the erector spinae muscle at the initial moment. The average power frequency of the electromyographic signal on the surface of the erector spinae muscle detected in real time.

[0014] Secondly, a multimodal fusion and dual-machine collaboration-based exoskeleton handling assistance control system is provided, applicable to the aforementioned multimodal fusion and dual-machine collaboration-based exoskeleton handling assistance control method, including: The data acquisition and processing module is used to acquire and preprocess multi-source sensing data in real time, extract motion assistance features, including the difference in hip position, the difference in pressure between the left and right feet, the difference in interactive torque, and the root mean square value and average power frequency of the electromyographic signal on the surface of the erector spinae muscles in the lumbar region; and to predict the mass of the heavy object being carried online based on the dynamic observer. The intent recognition model module is used to construct a multimodal hierarchical intent recognition model and input the motion assistance features. It sequentially performs spatial gating and basic co-verification, bioelectrical and spatial prior prediction, and dynamic weighted verification to perform hierarchical judgment on the user's carrying starting intent. The user's carrying starting intent is determined to be true only when the bipedal posture is within the safe threshold, muscle exertion and forward shift of the center of gravity are detected, and the kinematic or dynamic characteristics are greater than the set threshold. The state machine switching module is used to control the exoskeleton to switch between four states—standby monitoring, transition waiting, assisted execution, and compliant exit—based on the intent recognition result using a finite state machine. The torque calculation module is used to calculate the torque distribution coefficient between the left and right legs based on the difference in hip position, the difference in pressure between the left and right feet, and the difference in interactive torque during assisted execution. It also combines the mass of the load with muscle fatigue extracted based on the average power frequency characteristics to generate the target assist torque using a biomimetic torque model. ;Will Multiply by the torque distribution coefficients for the left and right legs respectively to generate independent left and right leg assist commands and execute them; The cyclic standby module is used to control the output torque to gradually decrease to zero exponentially when the end of the handling is detected.

[0015] This invention provides a multimodal fusion and dual-machine collaboration-based exoskeleton handling assistance control method and system. Addressing the problems of delayed intention recognition, poor leg coordination, and stiff assistance output in complex dynamic environments, existing handling exoskeletons construct a multi-source sensing network to synchronously acquire hip joint kinematics, human-machine interaction torque, surface electromyographic signals of the erector spinae muscles in the lumbar region, and plantar pressure distribution. This is combined with an extended Kalman filter to predict the weight of the object online. A multimodal hierarchical intention recognition model, composed of spatial gating verification, prior prediction of bioelectricity and plantar pressure, and dynamic weighted verification, is used to achieve advanced prediction of movement intentions. A dynamic dual-machine coupling collaborative torque distribution mechanism is introduced to adjust the asymmetric compensation ratio of the left and right legs in real time based on the difference in hip position, plantar pressure difference, and interaction torque difference. A load-fatigue dual adaptive bionic torque model is established to dynamically correct the assistance output based on real-time load and muscle fatigue. Finally, a finite state machine is used to achieve full-cycle flow control of four states: standby monitoring, transition waiting, assistance execution, and compliant exit. This invention can be widely applied to industrial handling, logistics sorting, and rehabilitation medical scenarios of lower limb exoskeleton robots, significantly improving human-machine collaboration compliance and operational safety under complex working conditions. Attached Figure Description

[0016] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is an overall flowchart of the method of the present invention; Figure 2 This is a flowchart of the control method of the present invention; Figure 3 This is a comparison chart of the intended recognition response speed and hysteresis elimination effect of the present invention; Figure 4 This is a comparison diagram of the torque distribution of dual-machine coupling cooperation under complex eccentric working conditions according to the present invention; Figure 5 This is a graph showing the changing trends of working time, muscle fatigue, and adaptive torque compensation in this invention. Detailed Implementation

[0017] Example 1 like Figure 2 As shown, an exoskeleton handling assistance control method based on multimodal fusion and dual-machine collaboration includes the following steps: S1: Real-time acquisition and preprocessing of multi-source sensing data, extraction of motion assistance features, including the difference in hip position, the difference in pressure between the left and right feet, the difference in interactive torque, and the root mean square value and average power frequency of electromyographic signals on the surface of the erector spinae muscles in the lumbar region; and online prediction of the mass of the heavy object being moved based on the dynamic observer. S2: Construct a multimodal hierarchical intention recognition model and input the motion assistance features. Then, sequentially perform spatial gating and basic collaborative verification, bioelectric and spatial prior prediction, and dynamic weighted verification to make hierarchical judgments on the user's carrying starting intention. Only when the posture of both feet is within the safe threshold, muscle exertion and the tendency of the center of gravity to move forward are detected, and the kinematic or dynamic characteristics are greater than the set threshold, the user's carrying starting intention is determined to be true. S3: The finite state machine controls the exoskeleton to switch between four states: standby monitoring, transition waiting, assisted execution, and compliant exit, based on the intent recognition result. S4: In assisted execution mode, calculate the torque distribution coefficient of the left and right legs based on the difference in hip position, the difference in pressure between the left and right feet and the difference in interaction torque; S5: Combining the mass of the object with the muscle fatigue level extracted based on the average power frequency characteristics, a target assist torque is generated based on a biomimetic torque model. ;Will Multiply by the torque distribution coefficients for the left and right legs respectively to generate independent left and right leg assist commands and execute them; S6: When the end of the handling is detected, the control output torque gradually decreases to zero according to an exponential law.

[0018] In this embodiment, multi-source sensing data is acquired in real time and preprocessed to extract motion assistance features. A multimodal hierarchical intent recognition model is constructed to determine the user's intention to start moving objects in layers. A finite state machine controls the exoskeleton to switch states based on the intent recognition results, obtains the target assist torque, and multiplies it by the torque distribution coefficients of the left and right legs to generate independent left and right leg assist commands and start the motors. When the moving is finished, the output torque is controlled to exit smoothly according to an exponential law. This realizes a complete intelligent control link for exoskeleton moving assistance from intent capture to execution to smooth unloading, which greatly improves the integrity and coherence of human-machine collaborative control in complex dynamic environments and improves the accuracy of assist control.

[0019] This embodiment is based on a lower limb exoskeleton system. The system mainly includes a microcontroller such as a DSP or STM32, a motor and drive controller mounted on the hip joint, an encoder, a torque sensor, an sEMG sensor attached to the erector spinae muscles of the lower back, and a smart insole with built-in multi-point thin-film pressure sensors. The control cycle is set to a fixed time step (e.g., 10ms). Multi-source data is transmitted to the main controller in real time via CAN bus or EtherCAT. The control system mainly consists of a multi-source sensing network module, a multimodal hierarchical intent recognition module, a finite state machine (FSM) main control module, a dynamic dual-machine coupled collaborative torque distribution module, and a dual adaptive torque execution module.

[0020] S1: Construction of Multi-Source Sensing Network and Data Preprocessing S11: This embodiment constructs a multi-dimensional synchronous sensing network to achieve comprehensive capture of human physiological electrical signals, kinematics, dynamics, and environmental characteristics, including: 1) By reading the motor encoders built into the left and right hip joint actuators of the exoskeleton, the raw position data is obtained, and the smooth angular velocity and angular acceleration of the joints are calculated using a Kalman state observer; 2) Combining the torque sensor built into the hip joint actuator, the motor torque is decoupled and calculated based on the system's built-in dynamic model, accurately separating the pre-calibrated gravity term from the total output torque. Inertial term and friction terms To obtain pure human-computer interaction torque .

[0021] 3) Sensors are used to synchronously acquire sEMG signals of the erector spinae muscles in the human lumbar region. After bandpass filtering and time-frequency domain feature extraction, the root mean square (RMS) value representing muscle contraction intensity and the average power frequency (MPF) representing muscle fatigue are output in real time. The smart insole acquires the trajectory of the center of pressure (CoP) of the forefoot and heel of both feet and the total pressure value, and calculates the pressure difference between the left and right feet. .

[0022] This embodiment suppresses the amplification effect of quantization noise and measurement noise, thereby improving data accuracy.

[0023] Based on the dynamic observer, this system constructs an extended Kalman filter (EKF) to achieve online dynamic prediction of the mass M of the currently being transported heavy object.

[0024] Since human-machine coupling systems are highly nonlinear, standard quality estimation is easily affected by sensor noise. Therefore, EKF is used for optimal estimation, and the specific model is constructed as follows.

[0025] S12: An extended Kalman filter is used to perform online dynamic estimation of the mass M of the heavy object, specifically including: S121: Assuming that the mass of the heavy object remains constant or changes only very slowly during a single handling cycle, the state variable of the system is defined as follows: The discrete state transition equation of the system is expressed as: ; in, for The estimated mass of the heavy object at all times. The noise is process noise; it follows a mean of 0 and a covariance matrix of... The Gaussian distribution, i.e. .

[0026] S122: Constructing the dynamic nonlinear observation equations Human-computer interaction torque after exoskeleton decoupling As the system's observed value Based on the system's dynamic equilibrium relationship, nonlinear dynamic regression terms related to the load are extracted, and observation equations are established: ; in, They are respectively The position of the hip joint, smooth angular velocity, and angular acceleration at any given moment; This is the dynamic regression matrix corresponding to the heavy load, which mainly maps the gravity and inertia parameters of the load; To measure the noise, it follows a mean of 0 and a covariance matrix of... The Gaussian distribution, i.e. ; S123: EKF Iterative Update Mechanism: In each control cycle, the following steps are performed in a loop: state prediction, covariance prediction, calculation of the Jacobian matrix of the observation function, calculation of Kalman gain, state optimal estimation update, and error covariance update, outputting the optimal smoothed load quality prediction value M.

[0027] State prediction: ; Covariance prediction: ; Calculate the Jacobian matrix of the observation function: since the observation equation is linear with respect to the state variables, this Jacobian matrix is ​​the regression vector. ; Calculate the Kalman gain: ; State-optimal estimation update: ; Error covariance update: ; Through the aforementioned EKF closed-loop iteration, the system can continuously utilize noisy interactive torques and kinematic data during motion to converge in real time and output the optimal smooth load mass prediction. .

[0028] This embodiment makes full use of noisy interactive torque and kinematic observation data, thereby improving data accuracy.

[0029] S2: The multimodal hierarchical intent recognition model specifically includes: First layer: Spatial gating and basic collaborative verification. This embodiment retains and optimizes the physical configuration's safety lock mechanism. The intent detection algorithm is only activated when the hip joint is within the preset effective handling angle area. Simultaneously, the system calculates the positional difference between the left and right hip joints in real time. Only when the absolute value Only when the current posture falls below the set safety threshold is it considered to conform to the carrying characteristics of bipedal support. The current level, as a rigid physical boundary, can filter out non-carrying noises such as large-amplitude single-leg swings, walking, or sitting postures.

[0030] The second layer: Bioelectric and spatial prior prediction. After passing the first layer of verification, the system monitors the RMS value of the erector spinae muscles (sEMG) in the lumbar region and the center of plantar pressure (CoP) in real time. When a step change in the RMS signal is detected and the center of plantar pressure (CoP) shows a characteristic forward shift, the system determines that the user has a "strong tendency to start moving." At this time, the bottom-level actuator is ready in advance to eliminate the mechanical response dead zone.

[0031] The third layer: dynamic weighted verification. After passing the second layer verification, when the smooth joint velocity is satisfied... Or pure human-computer interaction torque For any one of these conditions, the multimodal weighted decision ultimately determines the true intent.

[0032] The effective handling angle range is set to 20° to 90°, and the empirical range for the position difference safety threshold is 10° to 15°. and The empirical values ​​range from 15° / s to 30° / s and from 2 N·m to 5 N·m, respectively.

[0033] like Figure 3 As shown, the target assist torque output curve of this embodiment is compared with that of the traditional single kinematic threshold model; the dashed line marks the starting moment of human force exertion, reflecting that the intention recognition response of this embodiment is faster and the recognition lag is smaller, which can achieve advanced prediction.

[0034] This embodiment reduces recognition lag, false touch rate, and response delay, while improving the recognition's anticipation, accuracy, and anti-interference ability, and enhancing the timeliness and reliability of the assist response.

[0035] S3: Flow control based on FSM In this embodiment, FSM divides the entire exoskeleton's work cycle into four core states, including: State I (Standby monitoring state), State II (Transitional waiting state), State III (Assisted execution state), and State IV (Smooth exit state).

[0036] The execution flow is as follows: State I is the starting point of the task cycle. When the algorithm determines that the intention is true, FSM immediately triggers a state transition, flowing to State II. For situations where the intention has been triggered but the joint has not yet reached the optimal assistance angle, an action termination monitoring mechanism is introduced in FSM. In this state, the direction and speed of joint movement are monitored in real time. If the user is detected to be moving in the opposite direction or remaining still for a long time, it is determined that the user has a tendency to abandon the lifting action, and FSM returns to State I. If the optimal assistance angle is successfully reached (i.e., the hip flexion angle reaches a preset threshold and is within the effective lifting range), FSM officially switches to State III. When FSM enters State III, it calls the dynamic dual-machine coupled collaborative torque distribution module and the dual adaptive torque execution module to perform assistance. When termination conditions such as upright posture or zero speed are detected, FSM will switch to State IV.

[0037] In the preferred scheme, after the FSM enters state IV, to avoid discomfort such as a jolt caused by directly cutting off the power assist, a virtual working environment with high damping is simulated, causing the output torque to unload according to an exponential law: ; in, The real-time torque at the instant of entering the compliant exit state. is the damping time constant. Its empirical value range is 0.1~0.5s.

[0038] In this embodiment, the assist torque is ensured to reach zero quickly and smoothly, eliminating the impact on the human body caused by sudden torque changes. Then, the single cycle ends and the system safely returns to state I.

[0039] This embodiment improves system operational security and enhances decision-making robustness under complex operating conditions.

[0040] S4: Dynamic Dual-Machine Coupling and Cooperative Torque Distribution To address the issue of human imbalance easily caused by independent single-leg control or simple symmetrical control in complex dynamic environments such as uneven ground or eccentric heavy object handling, this embodiment introduces a dynamic dual-machine coupled collaborative torque distribution mechanism when providing assistance, establishing a dynamic torque distribution matrix. Let the torque distribution coefficients for the left and right legs be... and (satisfy Real-time calculation of the difference in position between the two hips Plantar pressure difference and the difference in left and right interaction torque .

[0041] The difference in hip position is uniformly defined by subtracting the data from the data on the left side. Plantar pressure difference and the difference in left and right interaction torque Construct a coupling strategy to dynamically calculate the initial weights on the left side: ; in, The normalized feedback gain coefficient for each feature is empirically taken to range from 0.05 to 0.2. When the wearer carries a load eccentrically on the right side or steps on an uneven surface with their right foot, the pressure on the right foot surges, leading to... The system will automatically lower This means that the weights on the right side are increased accordingly to provide asymmetric collaborative compensation.

[0042] Furthermore, to prevent the calculated distribution coefficient from becoming incorrect under extreme eccentric loads or severe sensor fluctuations... or In case of physical out-of-bounds conditions, the system introduces a saturation limiting protection mechanism to calculate the final actual allocation coefficient. and : ; ; in, and The system's preset security boundary threshold, The value range is 0.15~0.25 (0.2 is optimal). The value range is 0.75~0.85 (0.8 is optimal).

[0043] The dynamic calculation and amplitude limit allocation mechanism in this embodiment provides asymmetric collaborative compensation to the greatest extent while improving the safety boundary of the underlying motor torque output, thus improving the balance of the human body's center of gravity and operational safety.

[0044] like Figure 4 As shown, this embodiment dynamically distributes the torque of the left and right legs compared to the traditional fixed torque distribution scheme. Under eccentric load conditions, the output torque of the left and right legs can be dynamically adjusted, achieving asymmetric compensation.

[0045] This embodiment ensures that the distribution coefficient does not exceed the limit under extreme working conditions, fundamentally guaranteeing the safety boundary of the underlying motor torque output and the balance of the human body's center of gravity.

[0046] S5: Load-fatigue dual adaptive torque execution This embodiment also incorporates a load-fatigue dual adaptive sinusoidal biomimetic torque model during assist execution, establishing dynamic calculation equations that deeply adapt to human biomechanics and physical states. Combined with real-time load and muscle fatigue estimation, the final generated target total assist torque command... The calculation formula is as follows: ; in, The basic auxiliary stiffness coefficient set for the system. For load adaptive compensation, For the EKF online estimated weight of the object, For load gain, the heavier the object, the greater the peak compensation torque provided by this item; For fatigue adaptive compensation term, For fatigue gain, This is a muscle fatigue function extracted based on the average power frequency. The hip joint angle mapping strategy closely matches the biomechanical characteristics of the human waist muscles, maximizing the output of assistance during deep bending and then smoothly and non-linearly decreasing as the body returns to an upright position.

[0047] In this embodiment, the average power frequency (MPF) of the erector spinae muscle sEMG is continuously monitored. As the working time increases, the high-frequency component decreases when the muscle is fatigued, and the MPF drifts to the low frequency.

[0048] In the preferred scheme, the muscle fatigue function The calculation method is as follows: ; in, The average power frequency of the electromyographic signal on the surface of the erector spinae muscle at the initial moment. The average power frequency of the electromyographic signal on the surface of the erector spinae muscle detected in real time.

[0049] This embodiment detects an increase in fatigue and actively increases... Weighting provides deeper intervention and compensation.

[0050] The final calculation Combining the aforementioned torque distribution coefficients, commands for the left leg will be generated. With right leg command After passing through safety limiting protection, the data is sent to the motor drive controller for closed-loop control. (The aforementioned...) Set to 10 N·m / rad ~ 30 N·m / rad. and The initial values ​​were all obtained through offline calibration by collecting experimental data from multiple testers under standard loads.

[0051] like Figure 5 As shown in the figure above, the average power frequency of the erector spinae muscle sEMG decreases with increasing working time, indicating gradual muscle fatigue. The figure below shows that the target assist torque in this embodiment automatically increases with the degree of muscle fatigue, verifying that the load-fatigue dual adaptive torque model can automatically increase the assist output as fatigue increases.

[0052] This embodiment eliminates the influence of individual differences and differences in electrode placement, ensuring the consistency and reliability of fatigue assessment.

[0053] S6: Smooth unloading and exiting the closed loop.

[0054] This embodiment monitors the progress of the handling action in real time. When the system detects that the human body has returned to an upright position or the movement speed has returned to zero, the FSM switches to state IV. A virtual damping control mechanism is introduced to make the current output torque gradually decay to zero according to an exponential law, eliminating the impact of sudden torque changes. Then the system ends the single cycle and safely returns to state I.

[0055] This embodiment achieves full-cycle autonomous intelligent assistance through a complete chain of perception → identification → decision-making → allocation → execution → exit, comprehensively improving compliance and safety under complex working conditions.

[0056] Example 2 To further illustrate with reference to Embodiment 1, an exoskeleton handling assistance control system based on multimodal fusion and dual-machine collaboration, applicable to the method of Embodiment 1, includes: The data acquisition and processing module is used to acquire and preprocess multi-source sensing data in real time, extract motion assistance features, including the difference in hip position, the difference in pressure between the left and right feet, the difference in interactive torque, and the root mean square value and average power frequency of the electromyographic signal on the surface of the erector spinae muscles in the lumbar region; and to predict the mass of the heavy object being carried online based on the dynamic observer. The intent recognition model module is used to construct a multimodal hierarchical intent recognition model and input the motion assistance features. It sequentially performs spatial gating and basic co-verification, bioelectrical and spatial prior prediction, and dynamic weighted verification to perform hierarchical judgment on the user's carrying starting intent. The user's carrying starting intent is determined to be true only when the bipedal posture is within the safe threshold, muscle exertion and forward shift of the center of gravity are detected, and the kinematic or dynamic characteristics are greater than the set threshold. The state machine switching module is used to control the exoskeleton to switch between four states—standby monitoring, transition waiting, assisted execution, and compliant exit—based on the intent recognition result using a finite state machine. The torque calculation module is used to calculate the torque distribution coefficient between the left and right legs based on the difference in hip position, the difference in pressure between the left and right feet, and the difference in interactive torque during assisted execution. It also combines the mass of the load with muscle fatigue extracted based on the average power frequency characteristics to generate the target assist torque using a biomimetic torque model. ;Will Multiply by the torque distribution coefficients for the left and right legs respectively to generate independent left and right leg assist commands and execute them; The cyclic standby module is used to control the output torque to gradually decrease to zero exponentially when the end of the handling is detected.

[0057] This embodiment provides the working process, working details and technical effects of an exoskeleton transport assistance control method based on multimodal fusion and dual-machine collaboration, which can be referred to in Embodiment 1 and will not be repeated here.

[0058] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for controlling the transport assistance of an exoskeleton based on multimodal fusion and dual-machine collaboration, characterized in that, Includes the following steps: Multi-source sensing data is acquired in real time and preprocessed to extract motion assistance features, including the difference in hip position, the difference in pressure between the left and right feet, the difference in interactive torque, and the root mean square value and average power frequency of electromyographic signals on the surface of the erector spinae muscles in the lumbar region; and the mass of the heavy object being moved is estimated online based on the dynamic observer. A multimodal hierarchical intent recognition model is constructed, and the motion assistance features are input. Spatial gating and basic collaborative verification, bioelectric and spatial prior prediction, and dynamic weighted verification are performed sequentially to make hierarchical judgments on the user's carrying starting intent. The user's carrying starting intent is determined to be true only when the bipedal posture is within the safe threshold, muscle exertion and forward shift of the center of gravity are detected, and the kinematic or dynamic features are greater than the set threshold. The finite state machine controls the exoskeleton to switch between four states: standby monitoring, transition waiting, assisted execution, and compliant exit, based on the intent recognition result. In assisted execution mode, the torque distribution coefficients of the left and right legs are calculated based on the difference in hip position, the difference in pressure between the left and right feet, and the difference in interactive torque. Combining the mass of the object with muscle fatigue extracted based on the average power frequency characteristics, a target assist torque is generated based on a load-fatigue dual adaptive biomimetic torque model. ;Will Multiply by the torque distribution coefficients for the left and right legs respectively to generate independent left and right leg assist commands and execute them; When the handling is detected to be finished, the control output torque gradually decreases to zero according to an exponential law.

2. The exoskeleton transport assistance control method based on multimodal fusion and dual-machine collaboration according to claim 1, characterized in that, Preprocessing includes: The smoothed angular velocity and angular acceleration of the joints are calculated using a Kalman state observer. By removing the gravity, inertia, and friction terms from the system dynamics model, a pure human-machine interaction torque is obtained through decoupling. ; Bandpass filtering was performed on the sEMG signal, and the root mean square value representing muscle exertion intensity and the average power frequency representing muscle fatigue were extracted. Based on plantar pressure data, the CoP trajectory of the left and right plantar pressure center points and the plantar pressure difference are calculated. .

3. The exoskeleton transport assistance control method based on multimodal fusion and dual-machine collaboration according to claim 1, characterized in that, An extended Kalman filter is used to perform online dynamic estimation of the mass M of the heavy object, specifically including: The discrete state transition equations are constructed as follows: ； in, , for The estimated mass of the heavy object at all times. This is process noise; The dynamic nonlinear observation equation is: ； in, They are respectively The position of the hip joint, smooth angular velocity, and angular acceleration at any given moment; This is the dynamic regression matrix corresponding to the heavy load, which mainly maps the gravity and inertia parameters of the load; To measure the noise, it follows a mean of 0 and a covariance matrix of... The Gaussian distribution, i.e. ; Within each control cycle, the system iteratively performs state prediction, covariance prediction, calculates the Jacobian matrix of the observation function, calculates the Kalman gain, updates the optimal state estimate, and updates the error covariance, outputting the optimal smoothed load quality estimate M.

4. The exoskeleton transport assistance control method based on multimodal fusion and dual-machine collaboration according to claim 1, characterized in that, The multimodal hierarchical intent recognition model specifically includes: First layer: Only when the hip joint is within the preset effective carrying angle range and the absolute value of the positional difference between the left and right hip joints. When the value is below the set security threshold, the subsequent identification process is activated; The second layer monitors the root mean square value of the electromyography signal on the surface of the erector spinae muscles in the waist and the center of pressure on the foot in real time. When a step change in the root mean square value is detected and the center of pressure on the foot shows a characteristic forward shift, it is determined that the user has a tendency to start carrying, and the bottom layer driver is ready in advance. The third layer: When either the smooth joint speed or the pure human-machine interaction torque exceeds the corresponding set threshold, the intention to start the transport is finally determined to be genuine.

5. The exoskeleton transport assistance control method based on multimodal fusion and dual-machine collaboration according to claim 1, characterized in that, Finite state machine flow control, specifically: Standby monitoring state: The motor output torque is zero, and the underlying layer continues to perform multi-source perception and multi-modal hierarchical intent recognition; Transition waiting state: When the intent is determined to be true, a state transition is triggered to enter this state, and the joint movement direction and speed are monitored in real time. If the user is detected to be moving in the opposite direction or remaining still for a long time, the state returns to the standby monitoring state. If the optimal assist angle is reached, the state switches to the assist execution state. The optimal assist angle is when the hip joint flexion angle is greater than a preset threshold and is within the effective range of the transport stroke. Assisted execution status: The dynamic dual-machine coupling collaborative torque allocation and dual adaptive torque execution module is invoked to perform the assist. Compliant Exit State: This state is activated when the system is detected to be upright or the speed has returned to zero. The output torque is then unloaded exponentially before returning to the standby monitoring state.

6. The exoskeleton transport assistance control method based on multimodal fusion and dual-machine collaboration according to claim 5, characterized in that, In the compliant exit state, the output torque is unloaded according to the following exponential law: ； in, The real-time torque at the instant of entering the compliant exit state. is the damping time constant.

7. The exoskeleton transport assistance control method based on multimodal fusion and dual-machine collaboration according to claim 1, characterized in that, The dynamic dual-machine coupling cooperative torque distribution in step S4 specifically includes: The difference in hip position is uniformly defined by subtracting the data from the data on the left side. Plantar pressure difference and the difference in left and right interaction torque Construct a coupling strategy to dynamically calculate the initial weights on the left side: ； in, For each feature, the normalized feedback gain coefficient is denoted as . A saturation limiting protection mechanism is introduced to calculate the final actual allocation coefficients for the left and right sides: ； ； in, and The preset security boundary threshold for the system.

8. The exoskeleton transport assistance control method based on multimodal fusion and dual-machine collaboration according to claim 1, characterized in that, The load-fatigue dual adaptive sinusoidal biomimetic torque model is as follows: ； in, The basic auxiliary stiffness coefficient set for the system; For load adaptive compensation, For the EKF online estimated weight of the object, For load gain, This is a fatigue adaptive compensation term. For fatigue gain, This is a muscle fatigue function extracted based on the average power frequency. This is a hip joint angle mapping strategy.

9. The exoskeleton handling assistance control method based on multimodal fusion and dual-machine collaboration according to claim 8, characterized in that, The muscle fatigue function The calculation method is as follows: ； in, The average power frequency of the electromyographic signal on the surface of the erector spinae muscle at the initial moment. The average power frequency of the electromyographic signal on the surface of the erector spinae muscle detected in real time.

10. A multimodal fusion and dual-machine collaboration-based exoskeleton handling assistance control system, characterized in that, The method applicable to any one of claims 1-9, comprising: The data acquisition and processing module is used to acquire and preprocess multi-source sensing data in real time, extract motion assistance features, including the difference in hip position, the difference in pressure between the left and right feet, the difference in interactive torque, and the root mean square value and average power frequency of the electromyographic signal on the surface of the erector spinae muscles in the lumbar region; and to predict the mass of the heavy object being carried online based on the dynamic observer. The intent recognition model module is used to construct a multimodal hierarchical intent recognition model and input the motion assistance features. It sequentially performs spatial gating and basic co-verification, bioelectrical and spatial prior prediction, and dynamic weighted verification to perform hierarchical judgment on the user's carrying starting intent. The user's carrying starting intent is determined to be true only when the bipedal posture is within the safe threshold, muscle exertion and forward shift of the center of gravity are detected, and the kinematic or dynamic characteristics are greater than the set threshold. The state machine switching module is used to control the exoskeleton to switch between four states—standby monitoring, transition waiting, assisted execution, and compliant exit—based on the intent recognition result using a finite state machine. The torque calculation module is used to calculate the torque distribution coefficient between the left and right legs based on the difference in hip position, the difference in pressure between the left and right feet, and the difference in interactive torque during assisted execution. It also combines the mass of the load with muscle fatigue extracted based on the average power frequency characteristics to generate the target assist torque using a biomimetic torque model. ;Will Multiply by the torque distribution coefficients for the left and right legs respectively to generate independent left and right leg assist commands and execute them; The cyclic standby module is used to control the output torque to gradually decrease to zero exponentially when the end of the handling is detected.

Images