Remote fault maintenance method and system for humanoid robot

Abstract

The present application relates to the technical field of manipulator control, and discloses a remote fault maintenance method and system for a human-shaped robot, comprising: in response to a joint abnormal signal, injecting a multi-frequency perturbation test torque, collecting a drive motor end angle position, a reducer output end angle position and a built-in torque sensor feedback torque, demodulating a transmission side gap angle and determining a performance degradation joint; driving a healthy redundant joint to apply a reverse approaching torque to establish a rigid contact of a tooth surface across a gap; under the condition of maintaining the rigid contact state, down-regulating a weighted Jacobian matrix motion weight and shifting a load; switching a compliant virtual impedance following mode and remodeling impedance parameters to smoothly shift the manipulator to a safe maintenance pose, so that the present application can demodulate mechanical wear parameters in situ, eliminate pose collapse and tooth surface collision impact in the load shifting process, and block limit cycle oscillation caused by transmission side gap.

Description

Technical Field

[0001] This invention belongs to the field of robotic arm control technology, and in particular relates to a remote fault maintenance method and system for humanoid robots. Background Technology

[0002] Multi-degree-of-freedom robotic arms are widely used for workpiece gripping and spatial trajectory operations. Each joint typically integrates a torque motor, precision reducer, dual encoders, and torque sensor. They maintain trajectory tracking and end-load posture through servo closed-loop control. Under long-term gravity loads and inertial impacts, the reducer tooth surface is continuously subjected to periodic alternating loads, which can easily lead to mechanical wear and gradually increase transmission backlash.

[0003] When existing remote maintenance systems detect that the joint following error exceeds the set range, they typically stop the machine and lock the brake, or issue a general servo reset command. However, if the drive force is directly cut off when the robot is in a non-zero position with a load, the end effector load may fall uncontrollably, causing collision damage to the mechanism. General servo reset can only restore the control state and cannot eliminate the transmission backlash already formed in the reducer. When the position closed loop is re-engaged, the mechanical backlash caused by wear will induce limit loop oscillations under high-stiffness feedback, affecting not only the end effector trajectory accuracy but also further aggravating fatigue damage to the reducer components. In response to the above situation, existing technologies also attempt to reduce the control weight of the performance-degraded joints by using the redundant degrees of freedom inverse Jacobian matrix and transfer the load to the healthy redundant joints. However, when the manipulator is under load, the tooth surface of the joint reducer bearing the gravity load is usually in contact with one side. If the stiffness of the performance-degraded joint is directly reduced and the load is transferred, the transmission backlash of the reducer will lose the rigid constraint of the motor end. Under the action of gravity at the end, the tooth surface may detach from the original force-bearing side and cross the mechanical clearance, causing a short-term sinking. This not only disrupts the end pose stability but also easily generates rigid collision impact when the tooth surface re-contacts.

[0004] Therefore, the technical problem to be solved by this invention is how to accurately identify the physical degradation parameters of the joint online while maintaining the stability of the end-effector's bearing posture, and how to smoothly transfer the load torque and reshape the joint impedance while eliminating the backlash of the reducer tooth surface transmission. Summary of the Invention

[0005] To address the problems in the background art, the technical solution of the present invention is as follows: A remote fault maintenance method for a humanoid robot, comprising the following steps: Step S1: In response to the abnormal signal of the robot joint, control the robot to maintain the current end-effector posture, inject multi-frequency micro-perturbation test torque into the feedforward channel of each joint, and simultaneously collect the encoder angular position at the drive motor end, the encoder angular position at the reducer output end, and the feedback torque from the built-in torque sensor. Step S2: Calculate the theoretical reference torque by combining the rigid body dynamics model of the manipulator, and demodulate the transmission backlash angle and torque sensing zero bias from the collected angular position and feedback torque; when the transmission backlash angle or torque sensing zero bias exceeds the first threshold, determine the performance deterioration joint. Step S3: Construct the zero-space torque vector of the side clearance preload and distribute it to each healthy redundant joint. Drive the healthy redundant joints to apply a reverse approximation torque to the performance-deteriorated joints along the kinematic chain while maintaining the fixed end pose, until the absolute value of the feedback torque variation rate crosses the contact threshold to establish a rigid contact across the clearance. Step S4: While maintaining rigid contact across the gap, reduce the motion weight distribution of the performance-degraded joint in the weighted Jacobian matrix, calculate the target driving torque using the null space projection algorithm, and transfer the main task load to the healthy redundant joint. Step S5: Switch the degraded joint to compliant virtual impedance following mode and reshape the impedance parameters to drive the healthy redundant joints to move in synergistic motion, thereby smoothly transferring the robotic arm to a safe maintenance position.

[0006] Preferably, step S1 includes the following sub-steps: Step S11, configuring the multi-frequency perturbation test torque as a sinusoidal sweep frequency torque signal, the frequency range of which is 1Hz to 20Hz, the amplitude is 5% to 10% of the rated continuous torque of the corresponding joint, and the duration is 3 complete cycles; Step S12, during the injection of the multi-frequency perturbation test torque, synchronously recording the encoder angular position at the drive motor end, the encoder angular position at the reducer output end, and the feedback torque from the built-in torque sensor at a sampling frequency of 1000Hz, and maintaining the 6-dimensional spatial pose of the robot end fixed.

[0007] Preferably, step S2 includes the following sub-steps: Step S21, calculate the transmission backlash angle of each joint based on the difference between the connecting rod angular position measured by the encoder at the reducer output end and the rotor angular position measured by the encoder at the drive motor end, divided by the reducer transmission ratio; Step S22, calculate the torque sensing zero bias based on the time integral average of the difference between the feedback torque from the built-in torque sensor and the theoretical reference torque within the frequency sweep cycle; Step S23, calculate the equivalent viscous friction coefficient of each joint based on the linear fitting slope between the angular velocity during the multi-frequency perturbation process and the dynamic torque required to overcome friction.

[0008] Preferably, step S3 includes the following sub-steps: Step S31, the backlash preload zero-space torque vector allocated to each healthy redundant joint generates internal force coupling in the kinematic closed chain inside the manipulator, so that the reducer tooth surface of the deteriorated joint is in contact with the reverse bearing side to eliminate the transmission backlash angle; Step S32, the built-in torque sensor of the deteriorated joint is monitored in real time to detect the feedback torque. When the absolute value of the feedback torque variation rate reaches 15 Nm / s and the measured feedback torque reaches 0.5 Nm, the transmission tooth surface closing gap is determined and a cross-gap rigid contact is established.

[0009] Preferably, step S5 includes the following sub-steps: Step S51, the virtual stiffness coefficient of the performance-degraded joint in the compliant virtual impedance following mode is reduced to 1% to 5% of the joint stiffness under normal operating conditions, and the virtual damping coefficient is configured to make the closed-loop response of the performance-degraded joint in an overdamped state with a damping ratio of 1.25; Step S52, the gravity feedforward compensation term is activated during the pose transfer process to keep the effective value of the motor phase current of the performance-degraded joint below 0.35A.

[0010] Preferably, the coordinated motion of the healthy redundant joints in step S5 includes: step S53, planning the healthy joints of the shoulder and elbow to move along a cubic polynomial interpolation trajectory, driving the robot arm to smoothly transfer to a safe retraction position within 3.5s, while the performance-degraded joints are in a low-stiffness passive following rotation state under the traction of the external kinematic closed chain, blocking the nonlinear coupling between the high-gain position closed loop and the transmission backlash to prevent the generation of limit loop oscillation.

[0011] Preferably, when the humanoid robot is in a bipedal standing balance state and the driving torque of the ankle joint on one side of the lower limb decreases, step S4 includes: step S41, adjusting the bipedal contact force distribution matrix according to the upper limit of the residual torque of the damaged ankle joint, shifting the projection of the humanoid robot's body center of mass in the horizontal plane by 85mm toward the center of the polygon of the healthy supporting foot, so that the vertical contact reaction force ratio of the healthy supporting foot is adjusted to 82% and the vertical contact reaction force ratio of the damaged supporting foot is adjusted to 18%; step S42, compensating for the torque deficiency of the damaged ankle joint by the torque margin of the hip and knee joints on the same side.

[0012] Preferably, after step S5, a parameter solidification step is also included: Step S54, after the robot arm reaches the safe maintenance posture, the step displacement test trajectory is sent to the motion controller, and the end position tracking error and steady-state oscillation amplitude are collected; when the measured end absolute positioning error is within 0.18mm and there is no high-frequency current oscillation in each joint, the updated weighted Jacobian matrix weight parameters, zero space optimization coefficients and transmission backlash feedforward compensation table are written into the non-volatile flash memory chip of the motion controller.

[0013] Preferably, in step S1, the triggering conditions for the abnormal signal of the robot joint include: step S13, monitoring the steady-state angle tracking error of the robot during the target rotation tracking process, and when the steady-state angle tracking error reaches 0.024 rad and exceeds the alarm threshold of 0.015 rad, generating an abnormal signal of the robot joint and triggering a remote in-situ fault diagnosis command.

[0014] A remote fault maintenance system for a humanoid robot, comprising a method for remote fault maintenance of a humanoid robot, including: The module includes a perturbation test acquisition module, a parameter demodulation and identification module, a side gap pre-tightening closure module, a load distribution reconstruction module, and an impedance reshaping migration module. The perturbation test acquisition module is used to inject multi-frequency perturbation test torque into the feedforward channels of each joint, and simultaneously acquire the encoder angular position at the drive motor end, the encoder angular position at the reducer output end, and the feedback torque from the built-in torque sensor. The parameter demodulation and identification module is used to combine the dynamic theory reference torque with the acquired data to demodulate the transmission backlash angle and torque sensing zero bias, and to identify the performance-degraded joints. The side clearance preload closing module is used to construct a side clearance preload zero-space torque vector to distribute to each healthy redundant joint, drive the healthy redundant joint to apply a reverse approximation torque until the absolute value of the feedback torque variation rate crosses the contact threshold to establish rigid contact across the clearance. The load distribution and reconstruction module is used to reduce the motion weight of the performance-degraded joint in the weighted Jacobian matrix while maintaining rigid contact across the gap, and to solve the target driving torque and transfer the main task load through the null projection algorithm. The impedance reshaping and migration module is used to switch the performance-degraded joints to a compliant virtual impedance following mode and reshape the impedance parameters, driving the healthy redundant joints to move in synergistic motion, and enabling the robot arm to smoothly transfer to a safe maintenance position.

[0015] Compared with existing technologies, the remote fault maintenance method and system for humanoid robots of the present invention have the following advantages: 1. In the remote fault maintenance of a humanoid robot, the reducer transmission backlash angle, torque sensing zero bias and equivalent friction parameters can be identified online while the manipulator is in the end-effector working position. The actual wear state of the joint transmission components can be grasped without disassembling the joint or using external disassembly tools. Thus, fault diagnosis can be completed without interrupting the current load state, and a reliable state basis can be provided for subsequent adjustment of the weighted Jacobian matrix and redistribution of joint load.

[0016] 2. Before reducing the motion weight of the degraded joint and transferring the load, apply a reverse approximation torque using a healthy redundant joint to make the reducer tooth surface cross the transmission backlash and re-fit. After confirming the formation of rigid contact across the backlash, transfer the gravity and dynamic inertial load. This avoids end sinking caused by the tooth surface suddenly crossing the backlash when the degraded joint is unloaded, and reduces secondary collisions caused by tooth surface re-contact, making the load transfer of the robot arm under load more stable.

[0017] 3. By switching the degraded joint to a compliant virtual impedance following mode, allowing it to move with healthy redundant joints at a lower stiffness and maintaining an overdamped response, the repetitive interaction between the high-stiffness position closed loop and the transmission backlash can be reduced, suppressing the resulting limit loop oscillation. When the robot moves to a safe maintenance position, the high-frequency jitter of the drive motor phase current and mechanical howling can be reduced, and the joints that have suffered mechanical damage can be prevented from being repeatedly impacted by high-gain position control. At the same time, the positional stability and operational continuity of the robot during its movement can be maintained. Attached Figure Description

[0018] Figure 1 This is a flowchart of the online demodulation of backlash in the robotic arm transmission and the compliant posture transfer of the present invention; Figure 2 This is a block diagram showing the module composition and collaboration of the remote fault maintenance system of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0020] A remote fault maintenance method for a humanoid robot includes the following steps: Step S1: In response to the abnormal signal of the robot joint, control the robot to maintain the current end-effector posture, inject multi-frequency micro-perturbation test torque into the feedforward channel of each joint, and simultaneously collect the encoder angular position at the drive motor end, the encoder angular position at the reducer output end, and the feedback torque from the built-in torque sensor. Step S2: Calculate the theoretical reference torque by combining the rigid body dynamics model of the manipulator, and demodulate the transmission backlash angle and torque sensing zero bias from the collected angular position and feedback torque; when the transmission backlash angle or torque sensing zero bias exceeds the first threshold, determine the performance deterioration joint. Step S3: Construct the zero-space torque vector of the side clearance preload and distribute it to each healthy redundant joint. Drive the healthy redundant joints to apply a reverse approximation torque to the performance-deteriorated joints along the kinematic chain while maintaining the fixed end pose, until the absolute value of the feedback torque variation rate crosses the contact threshold to establish a rigid contact across the clearance. Step S4: While maintaining rigid contact across the gap, reduce the motion weight distribution of the performance-degraded joint in the weighted Jacobian matrix, calculate the target driving torque using the null space projection algorithm, and transfer the main task load to the healthy redundant joint. Step S5: Switch the degraded joint to compliant virtual impedance following mode and reshape the impedance parameters to drive the healthy redundant joints to move in synergistic motion, thereby smoothly transferring the robotic arm to a safe maintenance position.

[0021] Preferably, step S1 includes the following sub-steps: Step S11, configuring the multi-frequency perturbation test torque as a sinusoidal sweep frequency torque signal, the frequency range of which is 1Hz to 20Hz, the amplitude is 5% to 10% of the rated continuous torque of the corresponding joint, and the duration is 3 complete cycles; Step S12, during the injection of the multi-frequency perturbation test torque, synchronously recording the encoder angular position at the drive motor end, the encoder angular position at the reducer output end, and the feedback torque from the built-in torque sensor at a sampling frequency of 1000Hz, and maintaining the 6-dimensional spatial pose of the robot end fixed.

[0022] Preferably, step S2 includes the following sub-steps: Step S21, calculate the transmission backlash angle of each joint based on the difference between the connecting rod angular position measured by the encoder at the reducer output end and the rotor angular position measured by the encoder at the drive motor end, divided by the reducer transmission ratio; Step S22, calculate the torque sensing zero bias based on the time integral average of the difference between the feedback torque from the built-in torque sensor and the theoretical reference torque within the frequency sweep cycle; Step S23, calculate the equivalent viscous friction coefficient of each joint based on the linear fitting slope between the angular velocity during the multi-frequency perturbation process and the dynamic torque required to overcome friction.

[0023] Preferably, step S3 includes the following sub-steps: Step S31, the backlash preload zero-space torque vector allocated to each healthy redundant joint generates internal force coupling in the kinematic closed chain inside the manipulator, so that the reducer tooth surface of the deteriorated joint is in contact with the reverse bearing side to eliminate the transmission backlash angle; Step S32, the built-in torque sensor of the deteriorated joint is monitored in real time to detect the feedback torque. When the absolute value of the feedback torque variation rate reaches 15 Nm / s and the measured feedback torque reaches 0.5 Nm, the transmission tooth surface closing gap is determined and a cross-gap rigid contact is established.

[0024] Preferably, step S5 includes the following sub-steps: Step S51, the virtual stiffness coefficient of the performance-degraded joint in the compliant virtual impedance following mode is reduced to 1% to 5% of the joint stiffness under normal operating conditions, and the virtual damping coefficient is configured to make the closed-loop response of the performance-degraded joint in an overdamped state with a damping ratio of 1.25; Step S52, the gravity feedforward compensation term is activated during the pose transfer process to keep the effective value of the motor phase current of the performance-degraded joint below 0.35A.

[0025] Preferably, the coordinated motion of the healthy redundant joints in step S5 includes: step S53, planning the healthy joints of the shoulder and elbow to move along a cubic polynomial interpolation trajectory, driving the robot arm to smoothly transfer to a safe retraction position within 3.5s, while the performance-degraded joints are in a low-stiffness passive following rotation state under the traction of the external kinematic closed chain, blocking the nonlinear coupling between the high-gain position closed loop and the transmission backlash to prevent the generation of limit loop oscillation.

[0026] Preferably, when the humanoid robot is in a bipedal standing balance state and the driving torque of the ankle joint on one side of the lower limb decreases, step S4 includes: step S41, adjusting the bipedal contact force distribution matrix according to the upper limit of the residual torque of the damaged ankle joint, shifting the projection of the humanoid robot's body center of mass in the horizontal plane by 85mm toward the center of the polygon of the healthy supporting foot, so that the vertical contact reaction force ratio of the healthy supporting foot is adjusted to 82% and the vertical contact reaction force ratio of the damaged supporting foot is adjusted to 18%; step S42, compensating for the torque deficiency of the damaged ankle joint by the torque margin of the hip and knee joints on the same side.

[0027] Preferably, after step S5, a parameter solidification step is also included: Step S54, after the robot arm reaches the safe maintenance posture, the step displacement test trajectory is sent to the motion controller, and the end position tracking error and steady-state oscillation amplitude are collected; when the measured end absolute positioning error is within 0.18mm and there is no high-frequency current oscillation in each joint, the updated weighted Jacobian matrix weight parameters, zero space optimization coefficients and transmission backlash feedforward compensation table are written into the non-volatile flash memory chip of the motion controller.

[0028] Preferably, in step S1, the triggering conditions for the abnormal signal of the robot joint include: step S13, monitoring the steady-state angle tracking error of the robot during the target rotation tracking process, and when the steady-state angle tracking error reaches 0.024 rad and exceeds the alarm threshold of 0.015 rad, generating an abnormal signal of the robot joint and triggering a remote in-situ fault diagnosis command.

[0029] A remote fault maintenance system for a humanoid robot, comprising a method for remote fault maintenance of a humanoid robot, including: The module includes a perturbation test acquisition module, a parameter demodulation and identification module, a side gap pre-tightening closure module, a load distribution reconstruction module, and an impedance reshaping migration module. The perturbation test acquisition module is used to inject multi-frequency perturbation test torque into the feedforward channels of each joint, and simultaneously acquire the encoder angular position at the drive motor end, the encoder angular position at the reducer output end, and the feedback torque from the built-in torque sensor. The parameter demodulation and identification module is used to combine the dynamic theory reference torque with the acquired data to demodulate the transmission backlash angle and torque sensing zero bias, and to identify the performance-degraded joints. The side clearance preload closing module is used to construct a side clearance preload zero-space torque vector to distribute to each healthy redundant joint, drive the healthy redundant joint to apply a reverse approximation torque until the absolute value of the feedback torque variation rate crosses the contact threshold to establish rigid contact across the clearance. The load distribution and reconstruction module is used to reduce the motion weight of the performance-degraded joint in the weighted Jacobian matrix while maintaining rigid contact across the gap, and to solve the target driving torque and transfer the main task load through the null projection algorithm. The impedance reshaping and migration module is used to switch the performance-degraded joints to a compliant virtual impedance following mode and reshape the impedance parameters, driving the healthy redundant joints to move in synergistic motion, and enabling the robot arm to smoothly transfer to a safe maintenance position.

[0030] Example 1: Taking the right robotic arm of a 7-DOF humanoid robot in a continuous positioning operation at the assembly station, where it is holding a 2kg workpiece at the end, as an example, the robotic arm is composed of the first to third rotary joints of the shoulder, the fourth rotary joint of the elbow, and the fifth to seventh rotary joints of the wrist connected in series. The fifth joint of the wrist has a built-in harmonic reducer with a transmission ratio of 100, an encoder at the drive motor end, an absolute encoder at the reducer output end, and a built-in torque sensor. Under the condition of continuously bearing the cantilever gravitational torque and dynamic alternating load, the motion controller detects that the steady-state angle tracking error of the fifth joint during the target rotation tracking process reaches 0.024rad, which exceeds the preset alarm threshold of 0.015rad, triggering the remote in-situ fault diagnosis process.

[0031] While maintaining the 6D spatial working posture and end-load clamping state of the robotic arm's end effector constant, the motion controller keeps the closed-loop reference command for each joint position unchanged. It sequentially superimposes multi-frequency perturbation test torque onto the servo loop feedforward channel of each joint, and simultaneously acquires the encoder angular position at the drive motor end, the encoder angular position at the reducer output end, and the feedback torque from the built-in torque sensor for the corresponding joint. The multi-frequency perturbation test torque uses a sinusoidal sweep frequency torque signal. ,satisfy ,in, The amplitude of the perturbation torque. This is the current frequency sweep test frequency. The sampling time is specified; the sweep frequency test frequency is set within the range of 1Hz to 20Hz, and each test frequency lasts for 3 complete cycles. The amplitude of the perturbation torque is controlled within 5% to 10% of the rated continuous torque of the corresponding joint. Taking the 5Hz test frequency of the 5th joint as an example... The torque was set to 1.2 N·m, accounting for 8% of the rated continuous torque of the corresponding joint, and the duration of 3 complete cycles was 0.6 s; during the test, the angular position of the encoder at the drive motor end was synchronously recorded at a sampling frequency of 1000 Hz. Angular position of absolute encoder at reducer output end and built-in torque sensor feedback torque The theoretical reference torque calculated by combining the rigid body dynamics model According to the formula Calculate the transmission backlash angle ,in, The gear ratio of the reducer is 100; the calculated average transmission backlash angle is 0.0084 rad, exceeding the standard factory tolerance of 0.003 rad. The first threshold is determined according to the factory tolerance or measurement tolerance corresponding to the parameter being judged. The first threshold corresponding to the transmission backlash angle is 0.003 rad, and the first threshold corresponding to zero bias of the torque sensor adopts the normal measurement tolerance of the built-in torque sensor. Simultaneously, according to the formula... Solve for the sweep frequency period Internal torque sensing zero bias The obtained value was 0.0025 N·m, which was within the normal measurement tolerance. Therefore, it was determined that the mechanical fault characteristic of the 5th joint was excessive transmission backlash caused by wear of the tooth surface of the harmonic reducer. In the process of extracting the equivalent viscous friction coefficient of each joint, the motion controller performed second-order numerical differentiation on the angular position of the encoder at the drive motor end of the high-frequency synchronous acquisition to obtain the angular acceleration signal. Combined with the rigid body dynamics model, the inertial torque generated by the rotational inertia of the rotor and the reducer and the static torque of the connecting rod were calculated in real time. The theoretical inertial torque, gravity torque and torque sensing zero bias were removed from the torque feedback from the built-in torque sensor to obtain the dynamic residual torque signal. Then, a low-pass digital filter with a cutoff frequency of 35 Hz was applied to the residual signal. In the sampling interval where the angular velocity of the perturbation motion is non-zero, the filtered dynamic friction torque and the corresponding angular velocity sampling points were extracted. The least squares method was used to perform linear regression fitting. The slope of the fitted line was used as the equivalent viscous friction coefficient of the joint, and the calculated value was 0.12 N·m·s / rad.

[0032] The robotic arm end effector maintains a 2kg workpiece clamping force rotation. When the 5th joint reducer is in a non-zero load state, the meshing tooth surface is in a unilateral pressure-fitting state. Directly reducing the motion weight of this performance-degraded joint and switching to compliant control will cause the external gravity load to cause the transmission tooth surface to detach from the original force-bearing side, resulting in mechanism sinking and secondary collision impact when crossing the 0.0084rad transmission backlash. To eliminate the pose drop caused by backlash release, before adjusting the motion weight distribution in the weighted Jacobian matrix, the motion controller uses the fixed end-effector 6-dimensional spatial working pose as a constraint. The preload torque used to push the 5th joint tooth surface across the backlash is decomposed into the healthy redundant joints of the shoulder and elbow according to the orthogonal zero-space projection relationship of the end-effector task Jacobian matrix. The torque components corresponding to each healthy redundant joint form the backlash preload zero-space torque vector. Under this constraint, the physical connections of the serial robotic arms remain unchanged, and the fixed end-effector pose constraint and the self-motion relationship of the healthy redundant joints form a closed kinematic chain within the robotic arm during the control process; the motion controller will... The healthy redundant joints allocated to the shoulder and elbow are driven to apply a reverse approximation torque to the 5th joint along the internal kinematic closed chain of the robot. This causes the meshing tooth surface of the reducer flexure to actively cross the transmission backlash of 0.0084 rad and closely engage with the reverse load side. The built-in torque sensor collects the differential signal of the feedback torque in real time. When the absolute value of the feedback torque variation rate is... When the speed reaches 15 N·m / s and the measured feedback torque amplitude reaches 0.5 N·m, it is determined that the internal tooth surface of the reducer has completed the clearance crossing and established a rigid contact force chain across the clearance. Since the 7-DOF manipulator has one redundant degree of freedom relative to the 6-dimensional end-effector pose task, its joint velocity and torque space has a dynamic zero space orthogonal to the Jacobian matrix of the end-effector task. When the end-effector clamps a 2kg suspended workpiece, the side clearance preload zero-space torque vector, after being orthogonally projected by the Jacobian transpose matrix of the end-effector task, generates an equivalent generalized external force spinor in the end-effector space that is zero. The displacement and attitude of the end-effector workpiece remain unchanged. This torque vector forms a self-motion internal torque balance between the joints, driving the rotor of the 5th joint motor to overcome its own rotational inertia and rotate across the 0.0084 rad mechanical tooth surface clearance, so that the tooth surface of the harmonic reducer flexure is in contact with the reverse bearing side, thereby completing the side clearance preload and establishing the rigid contact force chain across the clearance under the kinematic closed chain constraint of the manipulator.

[0033] While maintaining the stability of the rigid contact force chain across the gap, the motion controller uses a diagonally weighted matrix of joint health states. The weight factor corresponding to the 5th joint is reduced from 1 to 0.05, forming a weighted matrix. And solve for the weighted Jacobian generalized inverse matrix. ,in, The Jacobian matrix of order 6x7 at the end effector of the robotic arm is determined through the dynamic null space projection relationship. Solve for the target driving torque vector of each joint. ;in, Indicates by and The weighted Jacobian transpose formed by multiplication This is the null-space projection moment vector used to minimize the output load of the 5th joint. To obtain the null space optimization coefficients for positive scalars, Joint angle vector, To output torque for the 5th joint, during the calculation, first based on... and Obtain the weighted Jacobian generalized inverse matrix, and then clamp the external force spinor at the end. Mapped to the joint torque space, and superimposed with the null-space projection torque aimed at minimizing the output torque of the 5th joint, the target driving torque vector of each joint is obtained. This solution reduces the static gravity load borne by the 5th joint from 4.85 N·m before reconstruction to 0.32 N·m, and the main task load is transferred to the healthy redundant joints of the shoulder and elbow. In the process of calculating the target driving torque, the null-space optimization target vector adopts the gradient direction that minimizes the sum of squares of the output torque of the 5th joint, which is used to unload the main task gravity and inertial load borne by the joint. At the same time, the controller constrains the lower limit of the torque after null-space projection to the reference preload torque of 0.32 N·m required to maintain tooth surface contact. When the main task gravity load is transferred from the original 4.85 N·m to the healthy redundant joints of the shoulder and elbow, the 5th joint still retains the contact offset that matches the contact threshold, maintaining the contact state of the tooth surface on the opposite bearing side.

[0034] After completing the dynamic load transfer, the control system changes the control mode of the 5th joint from the original position loop stiffness. and differential damping The rigid position closed loop switches to a compliant virtual impedance follower mode, and its output torque is based on the formula. Calculate, where, The virtual stiffness coefficient after reshaping is set to 25 N·m / rad, which is within the range of 1% to 5% of the joint stiffness under normal working conditions. The virtual damping coefficient after reshaping is tuned according to the overdamped state where the closed-loop response of the performance-degraded joint reaches a damping ratio of 1.25. In this embodiment, it is set to 3.5 N·m·s / rad. For the target to follow position, This represents the actual angular position of the connecting rod end. This represents the actual angular velocity at the end of the connecting rod. To compensate for the gravity feedforward torque in the current pose of the joint, the motion controller plans the healthy redundant joints of the shoulder and elbow to move collaboratively along a cubic polynomial interpolation trajectory, smoothly transferring the robotic arm to a preset safe retraction pose within 3.5 seconds. This safe retraction pose serves as the safety maintenance pose in this embodiment. During the pose transfer, the active trajectory of the healthy redundant joint and the low-stiffness passive following relationship of the 5th joint form an external kinematic closed chain through end-effector motion constraints. Under the traction of this external kinematic closed chain, the 5th joint maintains low-stiffness passive following rotation while activating gravity feedforward compensation. The compensation term reduces the effective value of the phase current of the drive motor from 4.2A to 0.35A, thereby decoupling the nonlinearity of the transmission backlash and the high-gain position closed loop. In the dynamic pole configuration of the virtual impedance parameters, the equivalent rotational inertia of the rotor and connecting rod of the 5th joint is 0.04 kg·m². When the virtual stiffness coefficient is set to 25 N·m / rad, the undamped natural angular frequency of this second-order virtual impedance system is 25 rad / s. Based on this, the overdamped state where the closed-loop response of the performance-degraded joint reaches a damping ratio of 1.25 is used as the tuning condition for the virtual damping coefficient. The value is set to 3.5 N·m·s / rad, so that the closed-loop characteristic pole of the system is located on the negative real axis. The joint exhibits a passively compliant response with monotonically decaying during the external kinematic closed-chain traction process, thus suppressing angular displacement overshoot and limit cycle oscillation.

[0035] After the robotic arm smoothly reaches the safe maintenance position, the motion controller sends the step displacement test trajectory to the underlying servo driver. The absolute positioning error of the end effector, measured by an external laser tracker, converges to within 0.18mm, and there are no high-frequency harmonic oscillations in the current of each joint phase. Subsequently, the updated weighted Jacobian matrix weight parameters are... The zero-space optimization coefficient and the transmission backlash feedforward compensation table are written into the non-volatile flash memory chip of the motion controller to complete the in-situ fault maintenance closed-loop control of the robot under load conditions.

[0036] Example 2: The experimental verification platform for the right-side robotic arm of a current 7-DOF humanoid robot includes a base frame, a series joint drive assembly, a 6-dimensional force sensor at the end effector, a spatial laser tracker, and a main control computer. Each joint of the robotic arm has a built-in harmonic reducer with a transmission ratio of 100, a 23-bit encoder at the drive motor end, a 19-bit absolute encoder at the reducer output end, and a strain gauge torque sensor with a range of 50 N·m and a measurement resolution of 0.005 N·m. A 6-dimensional external force sensor with a measurement range of 100 N and an accuracy of 0.02 N is installed at the end effector. An externally arranged laser tracker with a spatial measurement accuracy of 0.01 mm is used. The system communication bus adopts a synchronous real-time Ethernet with a period of 1000 Hz. During the test, Gaussian white noise with a signal-to-noise ratio of 20 dB and power frequency harmonic interference with a frequency of 50 Hz are superimposed in the torque acquisition signal and current feedback channel to simulate the electromagnetic and mechanical vibration environment during operation.

[0037] The multi-frequency perturbation test torque adopts a sinusoidal frequency sweep method. During the frequency sweep, the test frequency is changed, and the encoder angular position at the drive motor end is read simultaneously. 1. Encoder angular position at the reducer output end and built-in torque sensor feedback torque When the test frequency is below 1Hz, the perturbation response period exceeds the servo refresh window, and the dynamic residual demodulation is delayed. When the test frequency is above 20Hz, the mechanical low-pass effect at the connecting rod end reduces the amplitude of the angular displacement response. When the amplitude of the perturbation torque is less than 5% of the rated continuous torque, the joint friction dead zone will obscure the transmission backlash displacement characteristics. When the amplitude is more than 10% of the rated continuous torque, the end-effector spatial pose fluctuation exceeds the clamping assembly tolerance. Under the condition of a rated continuous torque of 15N·m, the standard test process takes the frequency sweep test segment corresponding to 5Hz, and the amplitude of the perturbation torque is taken as 1.2N·m, which is 8% of the rated continuous torque.

[0038] Each test began with the robotic arm's end effector holding a 2kg workpiece in an initial loaded position, with the force rotation being... The controller demodulates the transmission backlash angle based on the synchronously acquired data. Zero bias with torque sensing Then apply the zero-space torque vector for side clearance preload. A reverse approximation torque is applied by a healthy redundant joint to establish rigid contact across the gap; subsequently, the weighting matrix is ​​adjusted. The target driving torque is calculated using the dynamic null space projection relationship. After completing the main task load transfer, then according to the virtual stiffness coefficient and virtual damping coefficient Switch to compliant virtual impedance following mode. During the test, record the transmission backlash angle demodulation error, the cross-gap sinking displacement during load transfer, the peak value of tooth surface collision torque, the effective value of phase current, and the end absolute positioning error under safe maintenance posture. Test groups A to E are used to examine the operation results of the complete maintenance process under standard working conditions, different wear levels, and parameter range endpoints. Control groups 1 and 2 remove the pre-tightening link or impedance reshaping link respectively. Out-of-range control groups 1 to 4 make the perturbation frequency, perturbation amplitude, or virtual stiffness coefficient exceed the set range respectively.

[0039] Under standard operating conditions, test group A used a perturbation torque amplitude of 8% of the rated continuous torque when the frequency sweep test reached 5Hz. After establishing pre-tightened contact, the virtual stiffness coefficient was set to 25 N·m / rad, accounting for 1.4% of the joint stiffness under normal operating conditions. After completing parameter demodulation, pre-tightening, load transfer, and compliant migration according to the above process, the demodulation error of the transmission backlash angle was 0.0002 rad, the cross clearance subsidence displacement was 0.03 mm, the peak value of the tooth surface collision torque was 0.62 N·m, the effective value of the phase current was 0.35 A, and the absolute positioning error of the end effector under the safe maintenance posture was 0.18 mm.

[0040] Test group B corresponds to the mild wear condition. Under the test conditions of 5Hz and 8%, pre-tight contact was established. The virtual stiffness coefficient was also taken as 25 N·m / rad, accounting for 1.4% of the joint stiffness under normal working conditions. After the same maintenance process, the transmission backlash angle demodulation error was 0.0001 rad, the cross clearance subsidence displacement was 0.02 mm, the peak value of the tooth surface collision torque was 0.55 N·m, the effective value of the phase current was 0.31 A, and the absolute positioning error of the end was 0.14 mm.

[0041] Test group C corresponds to the severe wear condition. The test frequency is 5Hz, the amplitude of the micro-perturbation torque accounts for 8%, and a virtual stiffness coefficient of 25 N·m / rad is used after establishing pre-tight contact, which accounts for 1.4%. Under the condition of increased wear, the demodulation error of the transmission side clearance angle is 0.0003rad, the clearance sinking displacement is 0.04mm, the peak value of the tooth surface collision torque is 0.71 N·m, the effective value of the phase current is 0.38A, and the absolute positioning error of the end is 0.21mm.

[0042] To examine the operating status at both ends of the set parameter range, test group D set the sweep frequency to 1Hz, the perturbation torque amplitude to 5% of the rated continuous torque, maintained pre-tightened contact, and the virtual stiffness coefficient to 18 N·m / rad, accounting for 1% of the joint stiffness under normal operating conditions. Under this lower frequency limit, the transmission backlash angle demodulation error was 0.0004 rad, the cross-clearance subsidence displacement was 0.04 mm, the peak value of the tooth surface collision torque was 0.66 N·m, the effective value of the phase current was 0.36 A, and the end-effector... For a positioning error of 0.19 mm, test group E set the sweep frequency to 20 Hz, the amplitude of the micro-perturbation torque to 10%, and established pre-tightened contact. The virtual stiffness coefficient was set to 90 N·m / rad, accounting for 5% of the joint stiffness under normal working conditions. The corresponding transmission backlash angle demodulation error was 0.0003 rad, the cross-clearance subsidence displacement was 0.03 mm, the peak value of the tooth surface collision torque was 0.68 N·m, the effective value of the phase current was 0.37 A, and the end-point absolute positioning error was 0.19 mm.

[0043] In control group 1, the test frequency remained at 5 Hz, the amplitude of the perturbation torque remained at 8%, and the virtual stiffness coefficient was 25 N·m / rad, accounting for 1.4% of the joint stiffness under normal working conditions. However, the load was unloaded directly without establishing pre-tightening contact. The demodulation error of the transmission backlash angle in this group remained at 0.0002 rad. Since the 5th joint directly crossed the transmission backlash of 0.0084 rad under the action of external gravity load, a cross-gap sinking displacement of 0.86 mm occurred during the load transfer process. The peak value of the tooth surface collision torque increased to 12.45 N·m, the effective value of the phase current was 0.42 A, and the absolute positioning error of the end in the safe maintenance posture was 0.78 mm. Compared with the test procedure of establishing rigid contact across the gap before transferring the load, the parameter demodulation results of this group remained basically unchanged, while the mechanical sinking and tooth surface collision during the load transfer stage increased significantly.

[0044] Control group 2 also used the 5Hz, 8% perturbation test conditions, but did not establish pre-tightening contact. Instead, it performed servo reset and maintained the position closed loop, without performing impedance reshaping in compliant virtual impedance following mode. The demodulation error of the transmission backlash angle in this group was 0.0002rad, the gap sinking displacement was 1.12mm, the peak value of the tooth surface collision torque was 18.6N·m, the effective value of the phase current reached 4.35A, and the absolute positioning error at the end was 1.65mm. Under the condition of maintaining a rigid position closed loop at the 5th joint, the transmission backlash and high-gain position control continued to couple. The corresponding phase current oscillation and end positioning error were higher than those of the test procedure that completed impedance reshaping.

[0045] In the out-of-range control group 1, the sweep frequency test frequency was reduced to 0.5Hz, and the perturbation torque amplitude was reduced to 3% of the rated continuous torque. The pre-tight contact was maintained for the rest of the process, and the virtual stiffness coefficient was taken as 25 N·m / rad, which accounts for 1.4% of the joint stiffness under normal working conditions. Under these conditions, the low-frequency response and small perturbation amplitude caused the backlash displacement characteristics to be affected by the friction dead zone, and the demodulation error of the transmission backlash angle increased to 0.0028 rad. During the subsequent load transfer and compliant migration process, the cross-clearance sinking displacement was 0.15 mm, the peak value of the tooth surface collision torque was 1.85 N·m, the effective value of the phase current was 0.36 A, and the absolute positioning error at the end was 0.39 mm.

[0046] The out-of-range control group 2 increased the sweep frequency test frequency to 25Hz and the perturbation torque amplitude to 15% of the rated continuous torque. After establishing pre-tight contact, a virtual stiffness coefficient of 25N·m / rad was still used, accounting for 1.4%. Under these high-frequency and high-amplitude conditions, the demodulation error of the transmission backlash angle was 0.0019rad, the cross-clearance subsidence displacement was 0.05mm, the peak value of the tooth surface collision torque was 0.75N·m, the effective value of the phase current increased to 0.82A, and the absolute positioning error at the end was 0.48mm.

[0047] The out-of-range control group 3 maintained the perturbation test conditions of 5Hz and 8% and established pre-tight contact, reducing the virtual stiffness coefficient to 9N·m / rad, which is only 0.5% of the joint stiffness under normal working conditions. The demodulation error of the transmission backlash angle was 0.0002rad, the peak value of the tooth surface collision torque was 0.65N·m, and the effective value of the phase current was 0.34A. Due to the virtual stiffness being lower than the set range, the cross-gap sinking displacement during the pose migration reached 0.42mm, and the absolute positioning error of the end effector under the safe maintenance pose was 0.56mm.

[0048] The out-of-range control group 4 also used the 5Hz, 8% perturbation test conditions and established pre-tight contact, but increased the virtual stiffness coefficient to 180 N·m / rad, accounting for 10% of the joint stiffness under normal working conditions. The demodulation error of the transmission backlash angle in this group was 0.0002 rad, the cross clearance subsidence displacement was 0.03 mm, and the peak value of the tooth surface collision torque was 0.69 N·m. Under the higher virtual stiffness, the oscillation coupling between the 5th joint and the transmission backlash was enhanced, the effective value of the phase current increased to 2.15 A, and the end absolute positioning error was 0.32 mm.

[0049] The above groups completed the test under the same sensor background noise and power frequency interference conditions. The test groups in the frequency range of 1Hz to 20Hz, the amplitude range of 5% to 10% of the perturbation torque, and the range of 1% to 5% of the virtual stiffness ratio all completed the zero-bias demodulation of the transmission backlash and torque sensing, and completed the main task load transfer and compliant virtual impedance following migration after establishing rigid contact across the gap. After removing the backlash preload, maintaining the rigid position closed loop, or making the test parameters go beyond the above range, the test results showed that the gap sank, tooth surface collision, phase current oscillation, or end positioning error increased, respectively.

[0050] Example 3: The current humanoid robot remote fault maintenance system is deployed in a 7-DOF serial manipulator control architecture. This system includes a main controller, a real-time industrial Ethernet communication bus, integrated control boards for each joint, dual encoders, and built-in torque sensors. The main controller integrates a perturbation testing and acquisition module, a parameter demodulation and identification module, a side clearance pre-tightening closure module, a load distribution reconstruction module, and an impedance reshaping and migration module. When the manipulator is in the working state of holding a 2kg workpiece at its end effector, the rotational force is... Each joint's internal harmonic reducer, drive motor encoder, reducer output absolute encoder, and built-in torque sensor transmit operating data to the main controller at a servo cycle of 1000Hz. When the main controller detects that the steady-state tracking error of the target angle of the wrist pitch joint (i.e., the 5th joint) reaches 0.024rad and exceeds the preset alarm threshold of 0.015rad, it triggers a remote in-situ fault diagnosis command.

[0051] The perturbation test acquisition module maintains the robot arm's end-effector's 6D spatial pose and position closed-loop reference command unchanged, and writes sinusoidal sweep frequency perturbation torque signals to the feedforward control channels of each joint actuator. The frequency sweep test was set within the range of 1Hz to 20Hz, and the perturbation torque amplitude was controlled at 5% to 10% of the rated continuous torque of the corresponding joint. Each test frequency lasted for 3 complete cycles. Among them, the perturbation torque amplitude of the 5th joint at a test frequency of 5Hz was... The torque is set to 1.2 N·m, which is 8% of the rated continuous torque. The duration of three complete cycles is 0.6 s. The parameter demodulation and identification module synchronously reads the encoder angular position at the drive motor end via the communication bus. Angular position of absolute encoder at reducer output end and built-in torque sensor feedback torque Call the rigid body dynamics model to calculate the theoretical reference torque According to the formula Calculate the gear ratio of the reducer The transmission backlash angle at each sampling point is calculated, and the average value is obtained. According to the formula Calculate the sweep frequency period The torque sensor inside has zero bias, and thus obtains Simultaneously, angular velocity signals and dynamic residual torques are extracted, and the equivalent viscous friction coefficient is demodulated through least-squares linear fitting. If the transmission backlash angle exceeds the factory tolerance threshold of 0.003 rad, the system determines that the 5th joint is a performance deterioration joint.

[0052] After receiving the performance degradation joint index number and the demodulated transmission backlash angle of 0.0084 rad, the backlash preload closing module constructs the backlash preload zero-space torque vector before adjusting the control weight of the 5th joint. Under the constraint of maintaining the unchanged end-effector pose, the module issues internal torque offset commands to the healthy redundant joints, such as the first to third joints of the shoulder and the fourth joint of the elbow, at an increasing slope of 0.1 N·m / ms. This causes the self-motion of the healthy redundant joints and the pre-tensioning motion of the fifth joint to form a closed kinematic chain constraint within the manipulator. A reverse approximation torque is then applied to the fifth joint along this constraint, causing the meshing tooth surface of its reducer flexure to cross a mechanical clearance of 0.0084 rad and engage with the reverse bearing side. The module performs first-order difference calculations on the feedback signal from the built-in torque sensor of the fifth joint at a sampling rate of 1000 Hz. The absolute value of the feedback torque variation rate is... When the torque reaches 15 N·m / s and the measured feedback torque amplitude reaches 0.5 N·m, it is determined that the reducer tooth surface has completed the clearance crossing and established rigid contact across the clearance. In the contact criterion, the torque amplitude threshold of 0.5 N·m is 3.3% of the rated continuous torque of 15 N·m of the 5th joint, which is 10 times higher than the peak noise floor of 0.05 N·m of the sensor under vibration conditions, and lower than the 4.85 N·m cantilever gravity torque caused by the 2 kg load at the end. The torque variation rate threshold is 15 N·m / s. When the preload torque is linearly loaded with a slope of 0.1 N·m / ms, the torque variation rate in the tooth surface clearance stage is less than 2 N·m / s. When the tooth surface makes physical contact, the differential torque slope of the sensor rises to more than 15 N·m / s.

[0053] The load distribution reconstruction module, while maintaining rigid contact across the gap, uses a diagonally weighted matrix representing the joint health status. The weighting factor for the fifth joint was adjusted from 1 to 0.05, resulting in... And solve for the weighted Jacobian generalized inverse matrix. ,in, The end effector is a 6x7 Jacobian matrix; then, based on the dynamic null space projection relationship... Recalculate the target driving torque vector for each joint. ,in It is the identity matrix. To minimize the null-space projection moment vector of the output load at the 5th joint, With a zero-space optimization coefficient, after load reconstruction, the static gravity load borne by the 5th joint decreased from 4.85 N·m to 0.32 N·m.

[0054] The impedance reshaping migration module switches the 5th joint from a position closed-loop mode to a compliant virtual impedance following mode, with the virtual stiffness coefficient... The virtual damping coefficient is set at 25 N·m / rad. The value is set to 3.5 N·m·s / rad, and the closed-loop response of the performance-degraded joint is configured as an overdamped state with a damping ratio of 1.25; the module follows the formula... The output driving torque drives the healthy redundant joints to move collaboratively along a cubic polynomial trajectory, smoothly transferring the robotic arm to the preset safe maintenance position within 3.5s. The effective value of the motor phase current drops from 4.2A to 0.35A, eliminating high-frequency limit cycle oscillation. During the joint torque distribution process, the diagonal weighted matrix of the joint health status is used to characterize the distribution ratio of the ability of each joint to bear the end external load. After the weighting factor of the 5th joint is set to 0.05, its load-bearing capacity corresponds to 5% of the normal state. When the end space force spinor is mapped to the joint torque space based on the weighted Jacobian generalized inverse matrix, the torque mapping components of the first 4 healthy joints of the shoulder and elbow increase. The gravitational torque generated by the 2kg workpiece at the end is borne by the healthy joints, and the 5th joint retains the torque component required to maintain low stiffness balance.

[0055] After the robotic arm smoothly reaches the safe maintenance position, the main controller sends the step displacement test trajectory to the motion controller. The absolute positioning error at the end effector, measured by an external laser tracker, converges to within 0.18mm, and there is no resonant ripple in the current of each driver phase. Subsequently, the updated weighted Jacobian matrix weight parameters are... The zero-space optimization coefficient and transmission backlash feedforward compensation table are written into the non-volatile flash memory chip storage area of ​​each joint drive and control integrated board to complete the in-situ fault remote maintenance and parameter solidification of the multi-degree-of-freedom manipulator under load conditions.

[0056] Example 4: This example combines Figures 1 to 2 This paper describes a remote fault maintenance method and system for a humanoid robot, such as... Figure 1 As shown, step S1 responds to abnormal signals from the manipulator joints, controlling the manipulator to maintain its current end-effector posture, injecting multi-frequency micro-perturbation test torque into the feedforward channels of each joint, and simultaneously acquiring the encoder angular position at the drive motor end, the encoder angular position at the reducer output end, and the feedback torque from the built-in torque sensor. Step S2 calculates the theoretical reference torque using the manipulator's rigid body dynamics model, and demodulates the transmission backlash angle and torque sensing zero bias from the acquired angular positions and feedback torque. When the transmission backlash angle or torque sensing zero bias exceeds a first threshold, a performance-degraded joint is identified. Step S3 constructs a backlash preload zero-space torque vector and distributes it to each healthy redundant joint, driving... While maintaining the fixed end-effector pose, the healthy redundant joint applies a reverse approximation torque along the kinematic chain to the degraded joint until the absolute value of the feedback torque change rate crosses the contact threshold to establish a rigid contact across the gap. In step S4, while maintaining the rigid contact across the gap, the motion weight distribution of the degraded joint in the weighted Jacobian matrix is ​​reduced, the target driving torque is calculated through the null space projection algorithm, and the main task load is transferred to the healthy redundant joint. In step S5, the degraded joint is switched to a compliant virtual impedance following mode and the impedance parameters are reshaped to drive the healthy redundant joint to move in synergy, thereby smoothly transferring the robot arm to a safe maintenance pose.

[0057] like Figure 2 As shown, the remote fault maintenance system includes a perturbation test acquisition module, a parameter demodulation and identification module, a backlash preload closure module, a load distribution reconstruction module, and an impedance reshaping migration module. The perturbation test acquisition module injects multi-frequency perturbation test torque and transmits the acquired signal to the parameter demodulation and identification module. The parameter demodulation and identification module demodulates the transmission backlash angle and torque sensing zero bias and transmits the output signal to the backlash preload closure module. The backlash preload closure module applies a reverse approximation torque to the drive healthy redundant joint and transmits the output signal to the load distribution reconstruction module. The load distribution reconstruction module lowers the motion weights in the weighted Jacobian matrix and transmits the output signal to the impedance reshaping migration module. The impedance reshaping migration module switches to a compliant virtual impedance following mode and transmits the feedback signal to the perturbation test acquisition module to form an interactive closed loop.

[0058] Example 5: When a multi-degree-of-freedom humanoid robot arm is put into operation on the production line and undergoes periodic maintenance, the main controller performs offline multi-axis collaborative initialization calibration based on the zero-point temperature drift and transmission backlash calibration data of the built-in torque sensors of each joint. Under the reference static condition of the robot arm being unloaded and freely suspended, with the motors of each joint de-energized and the brakes released, 1000 static measurement values ​​of the built-in torque sensors of each joint are continuously collected. The average difference is calculated by combining the theoretical gravitational torque calculation values ​​of each link of the robot arm, and the obtained bias torque is written into the underlying drive and control integrated board. The bias register completes the initial calibration of zero bias for torque sensing. After the servo loops of each joint are enabled, the robotic arm drives a single target joint in sequence, reciprocating in a small angle range in both the forward and reverse directions at a constant angular velocity of 0.05 rad / s. Based on the difference between the encoder angular position at the reducer output end and the encoder angular position at the drive motor end divided by the reducer transmission ratio, the initial reference grid table of the transmission backlash of the joint in the full stroke range is measured and fitted point by point, serving as the health baseline reference data for subsequent online dynamic demodulation of the deteriorated state characteristic tensor.

[0059] During the bipedal balance parameter calibration, the lateral distance between the centers of the two feet when the robot stands with its feet together is 200mm, and the total weight is 490N. When standing symmetrically, the left and right feet each bear a vertical contact reaction force of 245N. After the driving torque of the right ankle joint decays, its residual torque upper limit is 8.75N·m. Based on the support boundaries of the right foot in the front-back and left-right directions, the maximum vertical contact reaction force that the right foot can safely maintain is set to 88.2N, accounting for 18% of the robot's total weight. Correspondingly, the target value of the vertical contact reaction force of the healthy supporting foot is 401.8N, accounting for 82% of the robot's total weight. The main controller adjusts the bipedal contact force distribution matrix accordingly and translates the projection of the robot's body center of mass in the horizontal plane by 85mm toward the center of the polygon of the healthy supporting foot, so that the vertical contact reaction force of the two feet is redistributed according to the target ratios of 82% and 18%.

[0060] When the driving torque of one ankle joint of the lower limb of the bipedal humanoid robot decreases and it is standing on an unstructured ground, the maximum continuous output torque of the right ankle joint decreases from the rated peak of 35 N·m to 8.75 N·m due to a partial short circuit in the motor winding. The lower limb posture balance control module reads the position of the foot pressure center collected by the contact reaction force sensor and adjusts the support state of both sides according to the bipedal contact force distribution matrix. The vertical contact reaction force of the healthy side support foot corresponding to the robot's total weight of 490 N is adjusted to 401.8 N, and the vertical contact reaction force of the damaged side support foot is adjusted to 88.2 N. At the same time, the projection of the body's center of mass in the horizontal plane is translated 85 mm towards the center of the polygon of the healthy side support foot.

[0061] The robot's left and right lower limbs each have a 6-DOF kinematic chain, consisting of 3 DDOF at the hip, 1 DDOF at the knee, and 2 DDOF at the ankle. The lower limb posture balance control module establishes the rigid contact force balance relationship between the feet and the ground. Within the foot support polygon, the robot's total weight of 490N is used as the total vertical balance load. The target values ​​of the vertical contact reaction force between the healthy and damaged supporting feet are determined using the foot contact force distribution matrix. The force generated by the output torque of the right ankle joint decreasing from the rated peak of 35 N·m to 8.75 N·m is also considered. To compensate for the torque deficit, the control system, based on the Jacobian torque mapping matrix of the ipsilateral lower limb, distributes the pitching torque required to maintain the foot posture along the kinematic chain to the ipsilateral hip and knee joints. Specifically, the right knee joint increases the extension support torque by 4.2 N·m, and the right hip joint increases the pitching torque by 3.1 N·m. By compensating for the torque deficit of the damaged ankle joint through the torque margin of the ipsilateral hip and knee joints, the target maintaining torque of the pitching joint of the damaged right ankle is reduced to 5.4 N·m, which is lower than its residual torque upper limit of 8.75 N·m.

[0062] After completing the torque redistribution, the damaged right ankle pitch joint was switched from rigid position servo mode to low impedance mode with an output torque saturation upper limit of 7.5 N·m, and the virtual damping coefficient was configured to make its closed-loop response in an overdamped state with a damping ratio of 1.25. During the control process, the tilt angular velocity of the robot body to the right converged from 0.12 rad / s at the beginning of the loss of control to within 0.002 rad / s, and the vertical fluctuation of the body's center of mass height was maintained within 1.5 mm.

[0063] Example 6: Before the 7-DOF humanoid robot arm performs long-cycle automated assembly operations, the system performs pre-calibration and online adaptive parameter matrix generation based on historical operating data and sample working condition boundaries. The main controller continuously reads the original sampled values ​​of the phase current of each joint drive motor and the feedback torque of the built-in torque sensor through the serial industrial Ethernet interface at a sampling frequency of 1000Hz, and sets the load mass parameter of the current batch of work to 2kg. Then, it calls the built-in least squares fitting algorithm to perform offline calculation on the difference between the encoder angular position at the drive motor end and the encoder angular position at the reducer output end after conversion according to the reducer transmission ratio for 1000 consecutive sampling cycles, calibrates the transmission backlash feedforward compensation table corresponding to the current wear state of the robot arm, and writes the compensation table into the non-volatile flash memory chip of the drive and control integrated board.

[0064] During the online dynamic operation phase, the main controller maps the spinor of the end effector space to the target driving torque of each joint using the kinematic Jacobian matrix. When the system detects that the steady-state angle tracking error of the wrist pitch joint (the 5th joint) reaches 0.024 rad and triggers a remote in-situ fault diagnosis command, the perturbation test acquisition module, while maintaining the end effector's working posture, sequentially injects multi-frequency perturbation test torque into the feedforward channel of each joint, and simultaneously acquires the encoder angular position at the drive motor end, the encoder angular position at the reducer output end, and the feedback torque from the built-in torque sensor. Among them, the 5th joint uses the original sinusoidal sweep frequency perturbation torque signal in the 5Hz test segment. The parameter demodulation and identification module is based on the measured encoder angular position at the drive motor end. Angular position of encoder at reducer output The transmission backlash angle is calculated in real time based on the difference relationship after the gear ratio is converted. Meanwhile, to record the influence of the elastic torsion of the flexure of the harmonic reducer on the compensation parameters under cantilever load conditions, the module loads the rated torsional stiffness parameter of the harmonic reducer of this model, 12000 N·m / rad, reads the actual joint bearing torque of 4.85 N·m measured by the built-in torque sensor, and obtains the elastic torsional angle deformation of the flexure under this load as 0.0004 rad. This deformation is saved as the correction amount of the transmission backlash feedforward compensation table under load conditions. The online fault judgment still uses the difference between the encoder angular position at the drive motor end and the encoder angular position at the reducer output end after conversion according to the reducer transmission ratio as the transmission backlash angle.

[0065] After receiving the demodulated transmission backlash angle, the backlash preload closure module determines the backlash preload direction of the 5th joint. Under the constraint of the fixed end-effector pose, it distributes the preload torque to the healthy redundant joints of the shoulder and elbow according to the null-space projection relationship of the end-effector task Jacobian matrix, forming a backlash preload null-space torque vector. Each healthy redundant joint applies a reverse approximation torque to the 5th joint along the internal kinematic closed chain of the manipulator; when the built-in torque sensor feeds back the absolute value of the torque variation rate... When the speed reaches 15 N·m / s and the measured feedback torque reaches 0.5 N·m, it is determined that the reducer tooth surface crosses the transmission side clearance and fits against the reverse bearing side, establishing a rigid contact across the clearance.

[0066] During the transition from the establishment of rigid contact across the gap to the completion of the main task load transfer, the reverse approximation torque applied by the healthy redundant joint remains constant before the motion weight of the 5th joint is reduced, so that the tooth surface continues to fit against the reverse bearing side; within the 50ms transition window after the start of the weighted Jacobian load transfer, the main task gravity torque borne by the 5th joint is unloaded at a linear slope, while the reverse preload torque provided by the healthy redundant joint gradually decreases and smoothly switches to maintaining a positive clamping force of 0.32 N·m, so that the tooth surface remains under unilateral pressure during the load reversal and transfer.

[0067] The load allocation reconstruction module reduces the motion weight factor of the 5th joint in the weighted Jacobian matrix from 1 to 0.05, and updates the weighted Jacobian generalized inverse matrix accordingly. The target driving torque vector of each joint is recalculated using the null projection algorithm. The static gravity load borne by the 5th joint was reduced from 4.85 N·m to 0.32 N·m, and the main task load was transferred to the healthy redundant joint.

[0068] The impedance reshaping migration module then switches the fifth joint from a rigid position closed loop to a compliant virtual impedance following mode, with the virtual stiffness coefficient... The virtual damping coefficient is set at 25 N·m / rad. Set to 3.5 N·m·s / rad and activate the gravity feedforward compensation term. During the pose transfer process, the healthy redundant joint actively moves according to the cubic polynomial interpolation trajectory. Its movement and the pose constraint of the robot end effector together form an external kinematic closed chain. Under the traction of this external kinematic closed chain, the fifth joint maintains a low-stiffness passive following rotation state. Within 3.5s, the healthy redundant joint drives the robot to smoothly transfer to the safe retraction pose. The effective value of the motor phase current of the fifth joint decreases from 4.2A to 0.35A, and the high-frequency limit cycle oscillation is suppressed accordingly.

[0069] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.

Claims

1. A remote fault maintenance method for a humanoid robot, characterized in that, Includes the following steps: Step S1: In response to the abnormal signal of the robot joint, control the robot to maintain the current end-effector posture, inject multi-frequency micro-perturbation test torque into the feedforward channel of each joint, and simultaneously collect the encoder angular position at the drive motor end, the encoder angular position at the reducer output end, and the feedback torque from the built-in torque sensor. Step S2: Calculate the theoretical reference torque by combining the rigid body dynamics model of the manipulator, and demodulate the transmission backlash angle and torque sensing zero bias from the collected angular position and feedback torque; when the transmission backlash angle or torque sensing zero bias exceeds the first threshold, determine the performance deterioration joint. Step S3: Construct the zero-space torque vector of the side clearance preload and distribute it to each healthy redundant joint. Drive the healthy redundant joints to apply a reverse approximation torque to the performance-deteriorated joints along the kinematic chain while maintaining the fixed end pose, until the absolute value of the feedback torque variation rate crosses the contact threshold to establish a rigid contact across the clearance. Step S4: While maintaining rigid contact across the gap, reduce the motion weight distribution of the performance-degraded joint in the weighted Jacobian matrix, calculate the target driving torque using the null space projection algorithm, and transfer the main task load to the healthy redundant joint. Step S5: Switch the degraded joint to compliant virtual impedance following mode and reshape the impedance parameters to drive the healthy redundant joints to move in synergistic motion, so as to smoothly transfer the robot arm to a safe maintenance position.

2. The remote fault maintenance method for a humanoid robot according to claim 1, characterized in that, Step S1 includes the following sub-steps: Step S11, configure the multi-frequency perturbation test torque as a sinusoidal sweep frequency torque signal, with a frequency range of 1Hz to 20Hz, an amplitude of 5% to 10% of the rated continuous torque of the corresponding joint, and a duration of 3 complete cycles; Step S12, during the injection of the multi-frequency perturbation test torque, synchronously record the encoder angular position at the drive motor end, the encoder angular position at the reducer output end, and the feedback torque from the built-in torque sensor at a sampling frequency of 1000Hz, and maintain the 6-dimensional spatial pose of the robot end fixed.

3. The remote fault maintenance method for a humanoid robot according to claim 1, characterized in that, Step S2 includes the following sub-steps: Step S21, calculate the transmission backlash angle of each joint based on the difference between the connecting rod angular position measured by the encoder at the reducer output end and the rotor angular position measured by the encoder at the drive motor end, divided by the reducer transmission ratio. Step S22: Calculate the torque sensing zero bias based on the time integral average of the difference between the torque fed back by the built-in torque sensor and the theoretical reference torque within the frequency sweep cycle. Step S23: Calculate the equivalent viscous friction coefficient of each joint based on the linear fitting slope between the angular velocity and the dynamic torque required to overcome friction during the multi-frequency perturbation process.

4. The remote fault maintenance method for a humanoid robot according to claim 1, characterized in that, Step S3 includes the following sub-steps: Step S31, the backlash preload zero-space torque vector allocated to each healthy redundant joint generates internal force coupling in the kinematic closed chain inside the manipulator, so that the reducer tooth surface of the deteriorated joint is in contact with the reverse bearing side to eliminate the transmission backlash angle; Step S32, the built-in torque sensor of the deteriorated joint is monitored in real time to detect the feedback torque. When the absolute value of the feedback torque variation rate reaches 15 Nm / s and the measured feedback torque reaches 0.5 Nm, the transmission tooth surface closing gap is determined and a rigid contact across the gap is established.

5. The remote fault maintenance method for a humanoid robot according to claim 1, characterized in that, Step S5 includes the following sub-steps: Step S51, the virtual stiffness coefficient of the performance-degraded joint in the compliant virtual impedance following mode is reduced to 1% to 5% of the joint stiffness under normal working conditions, and the virtual damping coefficient is configured to make the closed-loop response of the performance-degraded joint in an overdamped state with a damping ratio of 1.

25. Step S52: During the pose transfer process, the gravity feedforward compensation term is activated to keep the effective value of the motor phase current of the degraded joint below 0.35A.

6. The remote fault maintenance method for a humanoid robot according to claim 1, characterized in that, Step S5 involves driving the coordinated motion of healthy redundant joints, including: Step S53, planning the healthy joints of the shoulder and elbow to move along a cubic polynomial interpolation trajectory, smoothly transferring the robotic arm to a safe retraction position within 3.5s. The performance-degraded joints are in a low-stiffness passive following rotation state under the traction of the external kinematic closed chain, blocking the nonlinear coupling between the high-gain position closed loop and the transmission backlash to prevent the generation of limit loop oscillation.

7. The remote fault maintenance method for a humanoid robot according to claim 1, characterized in that, When the humanoid robot is in a bipedal standing balance state and the driving torque of the ankle joint on one side of the lower limb decreases, step S4 includes: step S41, adjusting the bipedal contact force distribution matrix according to the upper limit of the residual torque of the damaged ankle joint, shifting the projection of the humanoid robot's body center of mass in the horizontal plane by 85mm towards the center of the polygon of the healthy supporting foot, so that the vertical contact reaction force ratio of the healthy supporting foot is adjusted to 82% and the vertical contact reaction force ratio of the damaged supporting foot is adjusted to 18%; step S42, compensating for the torque deficiency of the damaged ankle joint by the torque margin of the hip and knee joints on the same side.

8. The remote fault maintenance method for a humanoid robot according to claim 1, characterized in that, After step S5, a parameter solidification step is also included: Step S54, after the robot arm reaches the safe maintenance posture, the step displacement test trajectory is sent to the motion controller, and the end position tracking error and steady-state oscillation amplitude are collected; when the measured end absolute positioning error is within 0.18mm and there is no high-frequency current oscillation in each joint, the updated weighted Jacobian matrix weight parameters, zero space optimization coefficients and transmission backlash feedforward compensation table are written into the non-volatile flash memory chip of the motion controller.

9. A remote fault maintenance method for a humanoid robot according to claim 1, characterized in that, In step S1, the triggering conditions for the abnormal signal of the robot joint include: step S13, monitoring the steady-state angle tracking error of the robot during the target rotation tracking process, when the steady-state angle tracking error reaches 0.024 rad and exceeds the alarm threshold of 0.015 rad, generating the abnormal signal of the robot joint and triggering the remote in-situ fault diagnosis command.

10. A remote fault maintenance system for a humanoid robot, used to implement the remote fault maintenance method for the humanoid robot as described in claim 1, characterized in that, include: The module includes a perturbation test acquisition module, a parameter demodulation and identification module, a side gap pre-tightening closure module, a load distribution reconstruction module, and an impedance reshaping migration module. The perturbation test acquisition module is used to inject multi-frequency perturbation test torque into the feedforward channels of each joint, and simultaneously acquire the encoder angular position at the drive motor end, the encoder angular position at the reducer output end, and the feedback torque from the built-in torque sensor. The parameter demodulation and identification module is used to combine the dynamic theory reference torque with the acquired data to demodulate the transmission backlash angle and torque sensing zero bias, and to identify the performance-degraded joints. The side clearance preload closing module is used to construct a side clearance preload zero-space torque vector to distribute to each healthy redundant joint, drive the healthy redundant joint to apply a reverse approximation torque until the absolute value of the feedback torque variation rate crosses the contact threshold to establish rigid contact across the clearance. The load distribution and reconstruction module is used to reduce the motion weight of the performance-degraded joint in the weighted Jacobian matrix while maintaining rigid contact across the gap, and to solve the target driving torque and transfer the main task load through the null projection algorithm. The impedance reshaping and migration module is used to switch the performance-degraded joints to a compliant virtual impedance following mode and reshape the impedance parameters, driving the healthy redundant joints to move in synergistic motion, and enabling the robot arm to smoothly transfer to a safe maintenance position.

Images