Robot adaptive precision motion control method based on visual perception

Abstract

This invention relates to the field of industrial robot motion control technology, and discloses a vision-based adaptive precision motion control method for robots. This method performs rigid locking on non-test joints, constructs a single-axis reduced-order Jacobian matrix for the target test joint based on a standard Jacobian matrix, injects test signals into the target test joint, and simultaneously acquires multimodal data. Based on the single-axis reduced-order Jacobian matrix and multimodal data, it extracts static backlash and dynamic transmission stiffness, and performs mechanical wear and thermodynamic state assessments. This invention isolates motion crosstalk between multiple joints, decouples and extracts transmission errors of individual joints, and performs adaptive compensation in the dimensions of position, torque, and gain, effectively overcoming errors caused by performance degradation of mechanical transmission mechanisms and maintaining the long-term control accuracy of the robot.

Description

Technical Field

[0001] This invention relates to the field of industrial robot motion control technology, specifically a visual perception-based adaptive precision motion control method for robots. Background Technology

[0002] When industrial robots perform high-precision tasks, they primarily rely on internal mechanical transmission mechanisms, such as reducers, to achieve accurate trajectory following. During the robot's long-term service life, gear wear and frictional heat generation within the transmission mechanism are unavoidable physical phenomena. This leads to increased static backlash, decreased dynamic transmission stiffness, and thermally induced deformation. These nonlinear physical degradations directly translate into transmission errors, affecting the robot's end-effector positioning accuracy.

[0003] Existing robot servo control systems typically employ factory-set fixed control gains and static feedforward compensation strategies. This conventional control method is based on the assumption that mechanical transmission characteristics remain constant, lacking the ability to dynamically adapt to the evolution of the internal physical state of the machine. When the mechanical structure experiences performance degradation, the fixed control parameters cannot match the actual physical properties of the underlying hardware, thus causing motion lag and trajectory deviation.

[0004] To implement precise error compensation, accurate acquisition of the actual transmission error of specific joints is required. However, industrial robots are multi-axis, strongly coupled systems. When measuring the error characteristics of a single joint, displacement errors in other joints due to their own flexibility or minor external disturbances are superimposed on the spatial displacement of the end effector through the kinematic chain. This spatial motion crosstalk makes it difficult for existing methods to accurately decouple and extract the true backlash and stiffness data of a single joint. Furthermore, relying solely on sensor data in a single dimension is insufficient to comprehensively reflect the true wear and thermodynamic state of the transmission mechanism, and is prone to misjudgment of state due to measurement noise. Due to the lack of reliable single-axis error decoupling methods and comprehensive state assessment mechanisms, existing control methods cannot implement precise synchronous compensation in the position, torque, and gain dimensions for the degradation state of specific joints, ultimately leading to a significant decrease in the control accuracy and motion smoothness of the robot in the later stages of its service life. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a vision-based adaptive precision motion control method for robots. This method solves the problems of transmission errors caused by nonlinear hysteresis, dynamic stiffness decay, and thermal deformation in the internal mechanical transmission mechanism of existing industrial robots during long-term operation. Furthermore, the multi-axis motion coupling effect makes it difficult to extract the error characteristics of individual joints, and the underlying servo control strategy is unable to adapt to the degradation of the mechanical physical state, resulting in a decrease in the accuracy of end-effector trajectory control and motion smoothness.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a visual perception-based adaptive precision motion control method for robots, comprising the following steps:

[0007] Identify the target test joint and the non-test joint, perform rigid locking on the non-test joint, and construct a single-axis reduced-order Jacobian matrix for the target test joint based on the standard Jacobian matrix.

[0008] Test signals are injected into the target test joint and multimodal data is collected synchronously to generate a multimodal synchronous data set;

[0009] Based on the uniaxial reduced-order Jacobian matrix and the multimodal synchronization data set, the static backlash and dynamic transmission stiffness of the target test joint are extracted, and mechanical wear and thermodynamic state assessments are performed.

[0010] Based on the static backlash, the dynamic transmission stiffness, and the results of the state assessment, the commutation compensation angular displacement and feedforward compensation torque are calculated, and the feedback control gain is updated according to the dynamic transmission stiffness. These are then combined into control commands and issued for execution.

[0011] Furthermore, rigid locking is performed on the non-test joint, and a uniaxial reduced-order Jacobian matrix for the target test joint is constructed based on the standard Jacobian matrix, specifically including:

[0012] Send zero-speed hold commands to the servo drives corresponding to all non-test joints, and increase the gain of the position control loop and speed control loop to the maximum critical value allowed by the load inertia;

[0013] Extract the standard Jacobian matrix of the industrial robot, set the column vectors corresponding to the non-test joints in the standard Jacobian matrix to zero, retain the column vectors corresponding to the target test joints, and generate the single-axis reduced-order Jacobian matrix.

[0014] Furthermore, the multimodal data includes the actual spatial displacement vector of the end effector, theoretical angular displacement, quadrature-axis drive current, vibration acceleration signal, and temperature signal; multimodal data is acquired synchronously to generate a multimodal synchronous data set, specifically including:

[0015] Based on the precise time protocol, a synchronization message is broadcast and a hardwired trigger pulse is sent to synchronously trigger the corresponding sensing module.

[0016] The actual spatial displacement vector at the end point, the theoretical angular displacement, the cross-axis drive current, the vibration acceleration signal, and the temperature signal acquired in parallel are normalized and low-pass filtered to generate the multimodal synchronous data set with a unified time axis.

[0017] Furthermore, the static backlash and dynamic transmission stiffness of the target test joint are extracted, specifically including:

[0018] The actual spatial displacement vector at the end is inversely solved into the actual angular displacement change using the single-axis reduced-order Jacobian matrix, and the actual transmission error is calculated by combining the theoretical angular displacement.

[0019] Based on the quadrature axis drive current, the output drive torque converted to the reducer output end is obtained, and a hysteresis loop is constructed with the output drive torque as the abscissa and the actual transmission error as the ordinate.

[0020] The static backlash and dynamic transmission stiffness are extracted based on the hysteresis loop, and the nonlinear mechanical loss work is calculated by integrating the closed region of the hysteresis loop envelope.

[0021] Furthermore, the actual spatial displacement vector at the end point is inversely solved into the actual angular displacement change using the single-axis reduced-order Jacobian matrix, specifically including:

[0022] Calculate the product matrix of the single-axis reduced Jacobian matrix and its transpose, and obtain the minimum singular value of the product matrix;

[0023] When the minimum singular value is lower than a preset safety threshold, a damping factor is introduced. The damping factor is combined with the transpose of the single-axis reduced-order Jacobian matrix to construct a regularized pseudo-inverse matrix.

[0024] The actual spatial displacement vector at the end is subjected to a forward inverse kinematics solution using the regularized pseudo-inverse matrix to obtain the actual angular displacement change.

[0025] Furthermore, a mechanical wear condition assessment is performed, specifically including:

[0026] Bandpass filtering is performed on the vibration acceleration signal, and the root mean square value within a single cycle is calculated as the vibration energy characteristic.

[0027] The ratio of the static back clearance to the nominal static back clearance and the ratio of the vibration energy characteristic to the nominal reference value are assigned normalized weights and summed to calculate the comprehensive wear evaluation index.

[0028] When the comprehensive wear evaluation index exceeds the wear alarm threshold, the target test joint is determined to be in an excessive wear state.

[0029] Furthermore, a thermodynamic state assessment is performed, specifically including:

[0030] The average dissipation power is calculated based on the nonlinear mechanical loss power and the duration of a single cycle.

[0031] After subtracting the ambient temperature from the temperature signal, a linear trend fitting is performed to obtain the temperature rise gradient;

[0032] Calculate the cross ratio between the temperature rise gradient and the average power dissipation. When the cross ratio is greater than a preset reference threshold multiple, it is determined that there is an abnormal dry friction or lubrication deterioration state.

[0033] Furthermore, the commutation compensation angular displacement is calculated, specifically including:

[0034] Calculate the thermal back clearance increment based on the temperature rise gradient;

[0035] The effective back gap compensation amount is obtained by summing the static back gap and the thermal back gap increment.

[0036] Substituting the effective backlash compensation amount and the desired angular velocity of the target test joint into the smooth transition function generates the continuous reversing compensation angular displacement.

[0037] Further, the feedforward compensation torque is calculated, specifically including:

[0038] The desired angular displacement and desired angular velocity are input into the constructed radial basis function neural network, mapped through the hidden layer and linearly weighted by the output layer to output the feedforward compensation torque;

[0039] Obtain the actual feedback angular displacement and calculate the position following error;

[0040] The position following error is combined with the dynamic transmission stiffness and mapped to a torque deviation compensation amount. This compensation amount is then substituted into the online learning law to incrementally update the weight parameters of the output layer, and the updated weight parameters are clamped at upper and lower limits.

[0041] Furthermore, updating the feedback control gain based on the dynamic transmission stiffness specifically includes:

[0042] Obtain the initial position loop proportional gain and reference dynamic stiffness of the servo driver;

[0043] Calculate the attenuation ratio of the dynamic transmission stiffness relative to the reference dynamic stiffness, and combine it with the gain adjustment factor to calculate the reduced current position loop proportional gain.

[0044] When the current position loop proportional gain is lower than the set safety lower limit threshold, it is forced to clamp at the safety lower limit threshold.

[0045] A second aspect of the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described visual perception-based robot adaptive precision motion control method.

[0046] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described visual perception-based adaptive precision motion control method for robots.

[0047] This invention provides a vision-based adaptive precision motion control method for robots. It offers the following advantages:

[0048] 1. This invention employs a multi-axis kinematic decoupling and Jacobian matrix order reduction mechanism. By issuing a zero-velocity hold command to non-test joints to execute rigid locking, and by assigning zeros to the column vectors corresponding to the standard Jacobian matrix in the forward kinematic model to generate a single-axis reduced-order Jacobian matrix, this step effectively isolates the motion crosstalk caused by the superposition of minute displacements of non-test joints on the end-effector spatial displacement. It avoids the ill-conditioned inversion problem in traditional multi-axis simultaneous micro-motion, establishes a unique deterministic mapping between the angular displacement of a single test joint and the actual end-effector displacement, and improves the accuracy of single-joint transmission error extraction.

[0049] 2. This invention constructs a state separation evaluation system based on multimodal synchronized data and hysteresis loops. It utilizes a cross-modal time synchronization mechanism to acquire phase-difference-free multimodal data. By constructing a hysteresis loop between the output driving torque and the actual transmission error, it decouples and extracts static backlash, dynamic transmission stiffness, and nonlinear mechanical loss work. Simultaneously, it cross-maps and verifies the extracted mechanical transmission parameters with physical state features such as vibration root mean square values ​​and temperature rise gradients. This method avoids the measurement limitations of single-sensor data, achieves quantitative assessment of the internal wear and thermodynamic state of mechanical transmission mechanisms, and reduces the probability of state misjudgment.

[0050] 3. This invention achieves multi-dimensional adaptive compensation control that incorporates physical state degradation. Based on the extracted static and thermal backlash increments, continuous commutation compensation angular displacements are generated to avoid system impact. A radial basis function neural network is used to learn and output feedforward compensation torque online to overcome friction and damping. Finally, the position loop proportional gain is adaptively adjusted based on the degree of dynamic transmission stiffness decay. This control strategy dynamically corrects system control parameters in three dimensions: position, torque, and gain. This allows servo control commands to match the changes in the physical properties of the underlying mechanical transmission mechanism in real time, maintaining the robot's motion control accuracy and stability margin during the mechanical structure performance degradation cycle. Attached Figure Description

[0051] Figure 1 This is a system architecture topology diagram of an embodiment of the present invention;

[0052] Figure 2 This is a flowchart of the joint adaptive control method based on multimodal coupling of the present invention;

[0053] Figure 3This is a schematic diagram illustrating the principle of multi-axis kinematic decoupling and Jacobian matrix order reduction in an embodiment of the present invention.

[0054] Figure 4 This is a flowchart of the micro-vibration frequency domain feature extraction and excitation signal reconstruction according to an embodiment of the present invention;

[0055] Figure 5 This is a timing diagram of the multimodal clock synchronization acquisition mechanism according to an embodiment of the present invention;

[0056] Figure 6 This is a schematic diagram illustrating the hysteresis dynamics modeling principle of a single-axis reduced-order transmission according to an embodiment of the present invention.

[0057] Figure 7 This is a schematic diagram of the cross-coupling evaluation principle of cross-modal physical quantities in this invention;

[0058] Figure 8 This is a flowchart of the adaptive control compensation process combining hysteresis feature evolution according to an embodiment of the present invention.

[0059] Figure 9 This is a feature extraction diagram of a single-cycle hysteresis loop according to the present invention;

[0060] Figure 10 This is a time series comparison chart of the end-point trajectory tracking error of the present invention;

[0061] Figure 11 This is a comparison diagram of the vibration power spectral density of the joint in operation state according to the present invention.

[0062] Among them, 100 is the visual perception module; 200 is the encoder module; 300 is the electrical perception module; 400 is the vibration sensing module; 500 is the temperature sensing module; 600 is the central control module; 610 is the command domain; 620 is the analysis domain; and 630 is the servo domain. Detailed Implementation

[0063] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] See attached document Figure 1 , Figure 1 This is a system architecture topology diagram according to an embodiment of the present invention. The present invention provides a vision-based adaptive precision motion control system for robots, which may include: a vision perception module 100, an encoder module 200, an electrical perception module 300, a vibration sensing module 400, a temperature sensing module 500, and a central control module 600.

[0065] The vision perception module 100 is mounted above a fixed reference base outside the industrial robot. The vision perception module 100 covers the movement range of the industrial robot's end effector within a preset test station or preset static posture area, collects optical calibration features on the end effector, and extracts the actual spatial pose information of the end effector in the reference coordinate system. In a preferred embodiment, the vision perception module 100 uses a high-resolution industrial camera and combines it with a sub-pixel-level edge extraction algorithm to obtain minute displacement features. Simultaneously, the vision perception module 100 has pre-performed high-precision hand-eye calibration with the robot's reference coordinate system to ensure high-resolution measurement capability of the end effector's micrometer-level physical spatial displacement.

[0066] The encoder module 200 is at least located on the target test joint motor side of the industrial robot to measure the theoretical angular displacement of the rotor side of the target test joint motor in real time. In a preferred embodiment, the encoder module 200 may also be located on the reducer output side to obtain output angular displacement information for auxiliary verification or redundancy comparison of the visual inverse kinematics results. In embodiments without an output-side encoder, the system uses the theoretical angular displacement output by the motor-side encoder as the basic measurement.

[0067] The electrical sensing module 300 is connected to the servo driver current loop sampling link of the industrial robot. The electrical sensing module 300 is used to collect the three-phase drive current of the motor corresponding to the target test joint and convert the three-phase drive current into quadrature axis current.

[0068] The vibration sensing module 400 is rigidly connected to the side of the external reducer housing of the target test joint. The vibration sensing module 400 is used to acquire vibration acceleration signals of the target test joint during operation.

[0069] The temperature sensing module 500 is used to collect the temperature signal of the target test joint. In a preferred embodiment, the temperature sensing module 500 is embedded inside the reducer of the target test joint; in another embodiment, the temperature sensing module 500 can also be placed on the inner side of the reducer housing near the core friction pair, or placed on the outer surface of the reducer housing and combined with a preset heat conduction model to calculate the internal temperature rise.

[0070] The central control module 600 establishes communication connections with the visual sensing module 100, the encoder module 200, the electrical sensing module 300, the vibration sensing module 400, and the temperature sensing module 500, respectively.

[0071] The central control module 600 is logically divided into an instruction domain 610, an analysis domain 620, and a servo domain 630.

[0072] The instruction field 610 is used to generate standard path planning instructions and high-frequency micro-amplitude commutation test signals.

[0073] The analysis domain 620 synchronously acquires data collected by each sensing module based on a hardware distributed clock mechanism, and performs Jacobian matrix reduction and hysteresis loop fitting.

[0074] The servo domain 630 is used to receive the dynamic parameters, wear state parameters and thermodynamic state parameters output by the analysis domain 620, and generate compensation control quantity and servo gain adjustment quantity according to the parameters to adjust the position loop command, feedforward torque command and control gain parameter of the underlying servo controller.

[0075] In one implementation, the instruction domain 610 is responsible for generating basic motion commands and test excitation signals, the analysis domain 620 is responsible for parameter identification and state evaluation, and the servo domain 630 is responsible for sending the compensated control sequence to the servo driver of the target test joint.

[0076] See attached document Figure 2 , Figure 2 This is a flowchart of a joint adaptive control method based on multimodal coupling according to an embodiment of the present invention. The present invention provides a vision-based adaptive precision motion control method for robots, comprising the following steps:

[0077] S10, when the industrial robot is in a stationary posture during non-working intervals, the central control module 600 outputs a zero-speed high-gain hold command to the non-test joints to keep the non-test joints locked and constructs a single-axis reduced-order Jacobian matrix for the target test joint.

[0078] S20, the central control module 600 acquires the environmental micro-vibration signal of the target test joint in a static state through the vibration sensing module 400, extracts the inherent resonant frequency drift component of the environmental micro-vibration signal through the autoregressive moving average model, and adjusts the carrier frequency of the excitation signal according to the inherent resonant frequency drift component to generate a micro-amplitude commutation test signal.

[0079] S30, the central control module 600 injects the micro-amplitude reversal test signal into the servo driver of the target test joint. During this micro-amplitude excitation cycle, it synchronously collects the actual spatial displacement output by the visual perception module 100, the theoretical angular displacement output by the encoder module 200, the quadrature axis current output by the electrical perception module 300, the vibration acceleration output by the vibration sensing module 400, and the temperature signal output by the temperature sensing module 500.

[0080] S40, the central control module 600 uses the pseudo-inverse matrix of the single-axis reduced-order Jacobian matrix to inversely solve the actual displacement of the end space into the actual transmission error of the target test joint. The actual transmission error is fitted with the output torque obtained by converting the cross-axis current to form a single-cycle hysteresis loop, and dynamic parameters are extracted from the single-cycle hysteresis loop. The dynamic parameters include static backlash, dynamic transmission stiffness and nonlinear mechanical loss work.

[0081] S50, the central control module 600 performs weighted calculations on the static backlash and the extracted high-frequency vibration energy value, and outputs the mechanical wear state judgment result; and calculates the average dissipation power based on the nonlinear mechanical loss work and the period of the micro-amplitude commutation test signal, and then constructs a thermodynamic state evaluation index in combination with the temperature rise gradient, and outputs the joint lubrication thermodynamic state judgment result.

[0082] S60 After the industrial robot switches to a continuous working cycle, the central control module 600 performs feedforward compensation on the motion command of the target test joint based on the static backlash extracted in step S40 and the thermal backlash increment derived from the temperature change or temperature rise gradient. It also uses dynamic transmission stiffness to correct the position loop proportional gain parameter of the servo controller to improve the trajectory following accuracy and motion smoothness under continuous working cycle.

[0083] See attached document Figure 3 , Figure 3 This is a schematic diagram illustrating the principle of multi-axis kinematic decoupling and Jacobian matrix order reduction according to an embodiment of the present invention. In the joint adaptive control method based on multimodal coupling provided by the present invention, step S10 regarding multi-axis kinematic decoupling and non-test axis rigidification locking specifically includes the following sub-steps:

[0084] S101, in this embodiment, the central control module 600 monitors the global motion state parameters of the industrial robot in real time. When the industrial robot completes the current path planning trajectory and the motion speed of each joint returns to zero and enters the preset program waiting period, the central control module 600 determines that the industrial robot is in a stationary posture during non-working intervals. It should be noted that, considering the continuous operation requirements of the industrial robot, the above-mentioned program waiting period is usually set to a time window of milliseconds to seconds, and the specific duration is determined based on the stabilization and tuning time of the end effector. In this state, the central control module 600, according to the preset maintenance schedule, designates the joint that currently needs to be calibrated for transmission error as the target test joint, defined as the first joint. Axis, where The range of values ​​is , This represents the total degrees of freedom of the system.

[0085] Furthermore, before executing the test, the central control module 600 also verifies whether the current end effector is located within a preset safety test window. The safety test window must meet at least one of the following conditions: the minimum safety clearance between the end effector and surrounding tooling, workpiece, fence, or ground is greater than a preset threshold; the end effector is not in contact with the workpiece; and the target test joint does not cause interference to cables, air pipes, or auxiliary actuators when performing micro-amplitude reversing movements. If the current stationary posture does not meet the requirements of the safety test window, the test is skipped.

[0086] In an industrial robot, all other motion joints besides the target test joint are defined as non-test joints, denoted as the nth joint. Axis, where .

[0087] S102, to eliminate displacement errors caused by the flexibility or minor disturbances of other joints, the central control module 600 sends isolation control commands to the underlying servo system, causing all non-tested joints to enter a rigid locking state. As a preferred method, the central control module 600 sends isolation control commands to all... The servo drive corresponding to the axis sends a zero-speed high-gain hold command. In practice, after receiving this command, the servo drive adjusts the proportional gain and integral gain parameters of its internal position control loop and speed control loop to the maximum critical value allowed by the current load inertia. This maximum critical value is generally pre-calibrated by the inertia identification program at the factory to ensure that the system does not experience high-frequency oscillations. By increasing the gain, the motor can output sufficiently large electromagnetic stiffness to resist small external disturbance torques. With a hardware configuration that includes a mechanical brake, the central control module 600 can also directly output discrete electrical signals to close all the... The shaft uses an electromagnetic brake. Through the dual mechanisms of electrical gain adjustment and mechanical braking, the non-tested joint is equivalent to a high-rigidity stationary link in terms of physical connection and control response. This treatment helps to isolate spatial motion crosstalk caused by the flexibility of other joints in the transmission chain, and largely avoids the superposition of microscopic backlash from the non-tested joint into the overall spatial displacement of the end effector.

[0088] S103, after completing the non-test axis locking, the central control module 600 extracts the standard Jacobian matrix based on the current static posture of the industrial robot. For the forward kinematics solution of the standard Jacobian matrix under the current static posture of the robotic arm, those skilled in the art can directly calculate and obtain it using the standard DH parameter method or screw theory combined with the current angle of the joint encoder. The calculation derivation process is well-known in the field and will not be elaborated here. To ensure the numerical stability of the subsequent inverse matrix solution, the system also checks whether the current pose is near a kinematic singularity. If it is close to a singularity, the current test is skipped and the system waits for the next preset static posture, or the robotic arm is adjusted to a pre-calibrated safe test posture before executing the subsequent identification process. The total number of system degrees of freedom of the industrial robot is set to... The extracted standard Jacobian matrix is ​​denoted as .

[0089] S104, from a physical perspective, the standard Jacobian matrix describes the combined effect of all minute joint rotations on the end-effector displacement. Under the rigid locking mechanism of step S102, the angular displacement change of all non-tested joints... The constraint is forcibly set to zero. Based on this boundary constraint, the central control module 600 constructs a single-axis reduced-order Jacobian matrix for the target test joint. The central control module 600 then applies this standard Jacobian matrix... In the middle, retain the joint number of the target test. The first axis corresponding to Column vectors, and corresponding to all non-test joints. All column vector values ​​for the axis are assigned to zero. By extracting the retained non-zero column vectors, the central control module 600 generates a 6×1 dimension single-axis reduced-order Jacobian matrix, denoted as... .

[0090] The actual three-dimensional displacement vector of the end effector Actual angular displacement of the target test joint The positive kinematic mapping model is reconstructed into the following formula based on the order reduction mechanism:

[0091] ;

[0092] In the formula, This represents the end-space displacement vector containing three linear displacement components and three angular displacement components. It is a single-axis reduced-order Jacobian matrix. The target test joint is measured by the single actual angular displacement change. This order reduction process establishes a unique and deterministic mapping relationship between the end-effector spatial displacement change and the single joint angle change, mathematically avoiding the ill-conditioned inverse kinematics problem inherent in traditional multi-axis simultaneous micro-motion system inverse kinematics solutions.

[0093] See attached document Figure 4 , Figure 4 This is a flowchart of micro-vibration frequency domain feature extraction and excitation signal reconstruction according to an embodiment of the present invention. In a joint adaptive control method based on multimodal coupling provided by the present invention, step S20 regarding micro-vibration frequency domain feature extraction and excitation signal reconstruction may specifically include the following sub-steps:

[0094] S201, when the target test joint is in the aforementioned isolated and stationary state, the central control module 600 acquires the environmental micro-vibration signal of the target test joint through the vibration sensing module 400. Since the non-test joint has entered a rigid locked state at this time, the environmental micro-vibration signal mainly originates from the minor excitation of the target test joint by the background noise of the industrial environment. In this embodiment, the central control module 600 opens a preset time window. The continuous vibration acceleration data within this time window are extracted using a sampling rate that satisfies the Nyquist sampling theorem, forming a set of environmental micro-vibration signals, denoted as . To effectively capture the high-frequency micro-vibration characteristics caused by microscopic stiffness changes within the reducer, the sampling rate is typically set to the kilohertz level or higher. The length of this time window is generally set to include at least several complete low-frequency vibration cycles to ensure the frequency resolution of subsequent frequency domain analysis.

[0095] S202, after acquiring a data sequence of sufficient length, the system needs to perform frequency domain transformation to extract characteristic frequencies. Based on the acquired environmental micro-vibration signals, the central control module 600 performs frequency domain estimation using an autoregressive moving average model. Since environmental micro-vibration signals in industrial settings are often accompanied by broadband white noise, traditional discrete Fourier transforms are prone to spectral leakage within short time windows. Introducing an autoregressive moving average model helps transform noisy discrete-time series into a parameterized system model with poles and zeros, thereby obtaining a relatively smooth and higher-resolution power spectral density estimate. For determining the autoregressive order and moving average order in the autoregressive moving average model, those skilled in the art can use the Akaike information criterion to determine the model order and use conventional algorithms such as the Yule-Walker equation to solve the model coefficients. The specific parameter estimation and equation solving processes are well-known techniques in the field and will not be elaborated here.

[0096] S203, the central control module 600 outputs a power spectral density curve using a constructed autoregressive moving average model and extracts the inherent resonant frequency drift component from it. As the industrial robot continues to operate throughout its lifecycle, wear on the reducer teeth or increased bearing clearance within the target test joint can lead to a decrease in local mechanical stiffness, causing a low-frequency shift in the joint's actual inherent resonant frequency relative to the factory calibration reference value. The central control module 600 uses a peak-finding algorithm to extract the frequency coordinates corresponding to the global maximum energy peak in the aforementioned power spectral density curve, defining it as the inherent resonant frequency drift component in the current state. Considering that multiple local peaks with similar amplitudes may occur in actual industrial noise environments, when the energy difference between multiple peaks is detected to be within a preset tolerance range, the central control module 600 selects the peak with the lower frequency as the [energy difference]. This is to accommodate the low-frequency offset physical characteristics caused by mechanical degradation. Furthermore, if no single peak energy is detected within the set analysis frequency band, the system determines that the stiffness of the current joint has not significantly degraded. The value is directly taken as the theoretical resonant frequency pre-stored inside the servo controller.

[0097] S204, after confirming the resonance characteristics of the current joint, the system enters the adaptive generation phase of the test signal. The central control module 600 generates the test signal based on the extracted inherent resonant frequency drift component. The carrier frequency of the excitation signal is adjusted in reverse to generate a reconstructed micro-amplitude commutation test signal. In subsequent steps, a specific frequency excitation signal needs to be injected into the joint to measure transmission errors. To reduce the risk of the injected excitation signal inducing nonlinear mechanical resonance within the joint, the central control module 600 establishes a safe bandwidth. This safe bandwidth The value is calculated based on the half-power bandwidth of the resonance peak in the power spectral density curve, that is, the bandwidth where the amplitude drops to 0.707 times the peak value.

[0098] Based on this, the central control module 600 tests the system's default frequency. Perform verification. If the default test frequency meets the following requirements:

[0099] ;

[0100] If the default test frequency falls into the resonance risk range, frequency shifting is required.

[0101] As a preferred implementation, the system first proceeds according to a preset step size. Generate candidate frequencies in the low-frequency direction:

[0102] ;

[0103] If a candidate frequency in a certain low-frequency direction simultaneously satisfies the servo bandwidth constraint, displacement response constraint, and test safety constraint, then take:

[0104] ;

[0105] in, The smallest positive integer that satisfies the condition.

[0106] If no candidate frequency satisfying the constraints exists in the low-frequency direction, then candidate frequencies in the high-frequency direction are generated instead.

[0107] ;

[0108] If a candidate frequency in a certain high-frequency direction satisfies the servo bandwidth constraint, displacement response constraint, and test safety constraint, then take:

[0109] ;

[0110] If no candidate frequencies satisfying the constraints exist in either the high or low direction, the system uses a preset safety fallback frequency. ,Right now:

[0111] ;

[0112] Alternatively, abandon this test and wait for the next preset static posture.

[0113] As a preferred approach, the servo bandwidth constraint can be expressed as:

[0114] ;

[0115] in, This represents the upper limit of the executable bandwidth of the servo driver's position ring.

[0116] Based on the reconstructed excitation carrier frequency The central control module 600 reconstructs the micro-amplitude commutation test signal. Its mathematical expression is constructed as follows:

[0117] ;

[0118] In the formula, The excitation signal carrier frequency is the final frequency determined after adjustment using a safe frequency band avoidance algorithm. Give instructions for angular displacement that changes continuously over time. Time is the independent variable. The preset micro-amplitude angular displacement constant, The initial phase is used to ensure the smoothness of the speed during the instruction initiation phase. Typically, the value is zero. Among them, the micro-amplitude angular displacement constant... The value range is limited to 1.5 to 3 times the nominal backlash of the target test joint. This range ensures that the joint motor can drive the input shaft to rotate forward and backward to fully overcome the backlash dead zone, while reducing the probability of large-scale spatial motion at the output end causing cross-boundary collisions. This reconstruction process avoids mechanical resonance points from the control source, providing a foundation for the stability of physical quantities in the subsequent multimodal data acquisition process.

[0119] See attached document Figure 5 , Figure 5 This is a timing diagram of a multimodal clock synchronization acquisition mechanism according to an embodiment of the present invention. In a joint adaptive control method based on multimodal coupling provided by the present invention, step S30 regarding the multimodal synchronization acquisition mechanism may specifically include the following sub-steps:

[0120] S301, in this embodiment, the instruction field 610 within the central control module 600 injects the reconstructed micro-amplitude commutation test signal as a position command into the servo driver of the target test joint. During the cycle of driving the joint to perform continuous micro-commutation movements, to ensure the alignment of cross-modal data in the time domain, the central control module 600 initiates a hardware-level cross-modal time synchronization mechanism. For sensor nodes with different sampling frequencies and underlying communication interfaces, the system relies on the IEEE 1588 precise time protocol to construct an underlying data synchronization network. The central control module 600, acting as the master clock node, periodically broadcasts synchronization messages to each sensing module. Each sensing module uses an internal hardware timer to record the timestamps of message transmission and reception, thereby calculating the transmission delay of the communication link and dynamically fine-tuning its local clock. Considering that some visual sensing modules 100 or vibration sensing modules 400 in actual industrial settings may not have communication interfaces compatible with the Ethernet clock protocol, as a preferred redundant synchronization method, the central control module 600 synchronously uses high-speed digital I / O ports to send nanosecond-level hardwired trigger pulses to these sensor nodes. The synchronization mechanism combining the aforementioned network protocol with hard-wired connections ensures that the data sampling actions of each heterogeneous sensing module are consistent in the time domain, reducing the data phase misalignment error caused by communication jitter under the traditional asynchronous polling mechanism.

[0121] S302, after establishing a unified time reference, the analysis domain 620 within the central control module 600 executes parallel acquisition commands for multiple physical parameters within a micro-amplitude excitation cycle. Specifically, the visual perception module 100 controls the camera shutter to expose upon receiving a synchronization trigger pulse, capturing the image of the optical calibration board of the end effector. After matrix mapping from the pixel coordinate system to the reference world coordinate system, it outputs the actual three-dimensional displacement vector of the end effector at this moment. Simultaneously, the encoder module 200 reads the position pulse count value from the internal register of the joint motor and converts it into the theoretical angular displacement of the motor rotor side based on its resolution. The electrical perception module 300 continuously acquires the three-phase stator current data from the underlying servo driver. For the decoupling conversion of the three-phase AC signal to the quadrature-axis current, those skilled in the art can use the Clark coordinate transformation and Park coordinate transformation algorithms to obtain the quadrature-axis current. This quadrature-axis current is directly proportional to the output electromagnetic torque of the motor. The specific derivation process of the relevant coordinate transformation is well-known in the art and will not be elaborated here. The vibration sensing module 400 and the temperature sensing module 500 are also triggered by the same clock command to convert analog to digital, and output the real-time vibration acceleration value and temperature value respectively.

[0122] S303, considering the differences in data type and communication units between the output data of different sensing modules, the analysis domain 620 performs format normalization processing on the discrete data streams obtained from the parallel acquisition and constructs a data set based on a unified time axis. During format normalization, the system uniformly converts linear displacement in the spatial displacement vector to meters (SI), angular displacement to radians, encoder theoretical displacement to radians, quadrature-axis current to amperes, and acceleration to gravitational acceleration equivalents or meters per second squared. The central control module 600 performs synchronous sampling at any given time. The generated multimodal synchronization data set can be represented in matrix form as follows:

[0123] ;

[0124] In the formula, Represents a multimodal synchronization data set. The 6×1D end-effector actual spatial displacement vector captured by the visual perception module 100. The theoretical angular displacement measured by encoder module 200. The quadrature-axis drive current is extracted by the electrical sensing module 300 through coordinate transformation. The vibration acceleration signal acquired by the vibration sensing module 400 The temperature signal collected by the temperature sensing module 500.

[0125] As a preferred approach, in order to avoid sudden changes in measurement data caused by high-frequency electromagnetic interference in the industrial field, the central control module 600 performs zero-phase low-pass filtering on displacement, encoder, current and temperature signals in the buffer area; for vibration acceleration signals, the original synchronous sampling data is retained as a high-frequency feature analysis branch, and pre-processed low-frequency vibration data can be generated in parallel as a state stability judgment branch.

[0126] The cutoff frequency of the low-pass filter is preferably set to between 5 and 10 times the carrier frequency of the micro-amplitude commutation test signal, so as to reduce the high-frequency white noise at the bottom layer of the sensor circuit while preserving the low-frequency rigid body response characteristics of the target test joint caused by the input command; and the wear assessment of the vibration channel is based on bandpass analysis performed on the original vibration data that has not been truncated by the above low-pass filter.

[0127] See attached document Figure 6 , Figure 6 This is a schematic diagram illustrating the principle of single-axis reduced-order transmission hysteresis dynamics modeling according to an embodiment of the present invention. In a joint adaptive control method based on multimodal coupling provided by the present invention, step S40 regarding single-axis reduced-order transmission hysteresis dynamics modeling may specifically include the following sub-steps:

[0128] S401, in this embodiment, the analysis domain 620 within the central control module 600 uses the pseudo-inverse matrix of the single-axis reduced-order Jacobian matrix to inversely solve the actual spatial displacement of the end effector captured by the visual perception module 100 into the actual angular displacement of the target test joint. Since the single-axis reduced-order Jacobian matrix is ​​a non-square matrix, direct inversion lacks a mathematical definition; therefore, the system employs the Moore-Penrose generalized inverse matrix algorithm. To reduce the risk of numerical divergence caused by excessive matrix condition numbers due to micro-movements near singular points, the system introduces a damping factor to construct a regularized pseudo-inverse matrix, the mathematical expression of which is:

[0129] ;

[0130] In the formula, Let be the regularized pseudo-inverse of a single-axis reduced-order Jacobian matrix. This refers to the single-axis reduced-order Jacobian matrix constructed in step S10. This is the transpose of the matrix. It is the identity matrix. This is the damping factor. As a preferred embodiment, this damping factor... It is not a fixed constant; the system calculates the square matrix in real time. The minimum singular value is determined, and when this singular value is lower than a preset safety threshold, it adaptively increases within the range of 0.01 to 0.1. The value of is thus used to achieve a balance between computational accuracy and numerical stability.

[0131] From the physical essence of kinematic mapping, the Jacobian matrix relates to minute displacement variables. Therefore, the central control module 600 extracts the start time of the minute excitation cycle. spatial location Theoretical angular displacement measured by encoder module 200 Using the zero-point as a reference, the relative displacement variable at each sampling time is calculated, and then the actual angular displacement change of the target test joint is calculated based on the above-mentioned regularized pseudo-inverse matrix. :

[0132] ;

[0133] In the formula, This represents the vector difference between the actual spatial displacement at the end point and the initial displacement at the same time cross-section. It is combined with the theoretical angular displacement difference output by encoder module 200. The system further calculates the actual transmission error converted to the output of the reducer. :

[0134] ;

[0135] In the formula, The nominal reduction ratio constant of the target test joint reducer is determined.

[0136] S402, based on the acquired kinematic errors, the system synchronously analyzes the corresponding dynamic parameters. The central control module 600, based on the quadrature-axis drive current extracted by the electrical sensing module 300, converts it into an ideal electromagnetic drive torque referred to the reducer output. The difference between this ideal electromagnetic drive torque and the actual torque required to overcome external displacement essentially reflects internal friction damping and nonlinear losses. The specific conversion model is constructed as follows:

[0137] ;

[0138] In the formula, For ideal electromagnetic drive torque, The electromagnetic torque constant of the motor is... This is the quadrature axis drive current. For mechanical transmission efficiency. Considering that the frequency of the micro-amplitude commutation test signal is in the low-frequency range and the amplitude is extremely small, the inertial torque component caused by the resulting angular acceleration is small and can be approximately ignored here; or, as a more accurate implementation method, when obtaining the ideal electromagnetic drive torque, the nominal inertial torque term calculated based on the known inertia of the motor and reducer rotors is simultaneously subtracted. Wherein, the electromagnetic torque constant... With mechanical transmission efficiency The specific values ​​can be obtained by reading the configuration table on the factory nameplate of the motor and reducer, or by the offline parameter identification program of the servo drive under no-load conditions.

[0139] 403. To intuitively reflect the nonlinear transmission characteristics of the joint, the system needs to construct a geometric mapping between torque and displacement difference. The central control module 600 uses a time series as the correlation link to construct a two-dimensional phase space with the output drive torque as the abscissa and the actual transmission error as the ordinate. With the periodic excitation of the micro-amplitude commutation test signal, the discrete data points in this phase space form a hysteresis trajectory. Considering that baseline drift may occur during the test cycle due to slight temperature rise or electrical noise, causing the beginning and end of a single cycle's data to not close, the system performs a ordinate sequence adjustment before fitting. Perform linear detrending processing to subtract the constant drift component that accumulates over time.

[0140] After drift subtraction, the system uses a least-squares polynomial fitting algorithm to perform continuous envelope fitting on the discrete point set, generating a continuous single-period hysteresis loop function. For the specific optimization and iterative process of the least-squares polynomial fitting algorithm, those skilled in the art can use the conventional Gauss-Newton method or gradient descent method; the parameter optimization and iterative approximation process are well-known techniques in the field and will not be elaborated upon here.

[0141] S404, relying on the single-cycle hysteresis loop after continuous processing, the central control module 600 extracts key dynamic parameters. These dynamic parameters specifically include static backlash, dynamic transmission stiffness, and nonlinear mechanical loss work.

[0142] In the static backlash extraction logic, the system locates two intersection points of the hysteresis loop and the vertical axis, which represent the forward and reverse transmission error values ​​when the output drive torque is zero. The central control module 600 calculates the absolute distance between these two points and defines it as the static backlash. This numerical value physically represents the size of the pure geometric clearance when the tooth surface is under no stress.

[0143] For calculating the dynamic transmission stiffness, the system preferably selects a preset number of discrete points for local linear fitting within a local interval near the maximum positive and negative values ​​of the output driving torque, in order to reduce the sensitivity of the single-point derivative to noise and the fitting order. The central control module 600 calculates the equivalent slope of the corresponding interval based on the local linear fitting results, and defines the average of the reciprocals of the slopes as the current dynamic transmission stiffness. The reason for taking the reciprocal of the slope is that the vertical axis of the phase space represents displacement and the horizontal axis represents force, and the reciprocal of the slope conforms to the physical stiffness definition of the force-to-displacement ratio.

[0144] To solve for the nonlinear mechanical loss work, the central control module 600 uses a numerical integration algorithm to calculate the area of ​​the closed region enclosed by the entire hysteresis loop. As a preferred method, the system uses the trapezoidal integral rule or Simpson's integral rule to accumulate the area of ​​the discretized closed boundary, defining it as the nonlinear mechanical loss work. The physical significance of this enclosed area lies in reflecting the total energy consumed inside the reducer due to frictional damping and internal material dissipation within a single micro-amplitude commutation cycle.

[0145] See attached document Figure 7 , Figure 7 This is a schematic diagram illustrating the principle of cross-modal physical quantity cross-coupling evaluation according to an embodiment of the present invention. In a joint adaptive control method based on multimodal coupling provided by the present invention, step S50 regarding cross-modal physical quantity cross-coupling evaluation may specifically include the following sub-steps:

[0146] S501, in this embodiment, the central control module 600 extracts the vibration energy characteristics of the target test joint under micro-amplitude excitation based on a multi-modal synchronous data set. Considering that a single vibration amplitude is usually affected by broadband background noise in industrial environments, the system needs to process the vibration acceleration signal... Bandpass filtering is performed. Under the test condition of alternating forward and reverse rotation with slight amplitude, the broadband excitation generated by the micro-friction of the tooth surface is more likely to excite the inherent resonance of the system. Therefore, as a preferred approach, the passband frequency range of the bandpass filter is based on the inherent resonance frequency drift component extracted in step S20. The frequency band signal containing local wear fault information is determined by its high-frequency harmonic distribution range. After filtering and extracting the characteristic frequency band signal, the central control module 600 calculates the root mean square energy value of the signal in a single excitation cycle, and records it as... The root mean square value helps characterize the level of excitation energy generated inside the transmission chain due to microscopic spalling or increased roughness of the tooth surface.

[0147] S502, to avoid misjudgments due to reliance on a single physical quantity, the central control module 600 constructs a wear assessment function that cross-maps static backlash and vibration energy. In actual working conditions, the increase in pure geometric clearance and the increase in high-frequency excitation energy are correlated in reflecting the physical process of mechanical wear. The system constructs the following dimensionless comprehensive wear mapping function through normalization:

[0148] ;

[0149] In the formula, To provide a comprehensive wear and tear evaluation index, The static back gap extracted in step S40 The root mean square value of the extracted vibrational energy. and These are the nominal static backlash and nominal root mean square value of vibration, respectively, used as the benchmarks for the target test joint during factory calibration. and Let be the normalized weight constant, and satisfy... The specific values ​​of these two weighting constants can be obtained by extracting historical sample data of the same type of reducer at different wear stages in accelerated life bench tests, and calculating them using principal component analysis or multiple linear regression algorithms.

[0150] After calculating the comprehensive wear index, the system compares it with the set wear alarm threshold. When If the wear level exceeds the set threshold continuously, the system determines that the target test joint is in an excessive wear state and generates a maintenance mark. This threshold is usually set between 1.5 and 2.0, and the specific value is calibrated according to the robot's tolerance limit for end-effector displacement accuracy in different application scenarios.

[0151] S503, while performing mechanical wear assessment, the system simultaneously analyzes the thermodynamic state inside the target test joint. The central control module 600 utilizes the single-cycle nonlinear mechanical loss work calculated in step S40. Calculate its average power dissipation. Average power dissipation The calculation formula is:

[0152] ;

[0153] In the formula, This represents the duration of a single cycle of the micro-amplitude commutation test signal. This power dissipation physically characterizes the rate at which the reducer overcomes internal nonlinear damping per unit time.

[0154] At the same time, the central control module 600 collects the temperature signal from the temperature sensing module 500. The temperature rise gradient is obtained by performing time-domain differentiation. Considering that the large thermal inertia of the metal casing will cause a lag in the external temperature measurement response, the temperature sensing module 500 preferably uses a fast-response thin-film thermistor attached to the core friction pair inside the reducer (such as the outer ring of the main bearing), and is equipped with a weak signal amplification circuit to capture the transient temperature rise characteristics during the micro-amplitude test. To reduce the probability of differential abrupt changes caused by single-point measurement noise, the system uses the least squares method to perform linear trend fitting on the temperature data sequence within the set time window, and the slope of the fitted line is the temperature rise gradient. To account for potential interference from ambient temperature fluctuations, the system uses data from ambient temperature sensors deployed outside the base for differential subtraction before calculating the temperature rise gradient, thereby obtaining the relative temperature rise rate caused solely by internal mechanical work.

[0155] S504, based on the extracted thermodynamic parameters, the central control module 600 constructs a cross-ratio of power consumption and temperature rise gradient to perform a thermodynamic assessment of internal lubrication failure or abnormal dry friction conditions. Under normal lubrication conditions, there is a relatively stable thermal conduction balance between mechanical power dissipation and generated heat. The system establishes the following cross-ratio function:

[0156] ;

[0157] In the formula, The thermodynamic cross ratio is used to characterize the rate of temperature rise caused by unit mechanical energy dissipation.

[0158] In actual operation, when the internal lubricating oil film ruptures or the grease dries up, the coefficient of friction tends to increase, and a significant portion of the consumed mechanical work is converted into concentrated heat energy, leading to an increase in the thermodynamic cross ratio. The central control module 600 calculates the resulting ratio. Compared with the pre-established reference threshold under nominal lubrication conditions Comparison. When Greater than the benchmark threshold When the friction reaches a certain multiple (such as 1.3 to 1.5 times), the system determines that there is abnormal dry friction or lubrication deterioration inside, and triggers the underlying logic to lock or prompts for lubricant replenishment.

[0159] Furthermore, when the system determines that the target test joint has abnormal dry friction or lubrication deterioration, the central control module 600 can prohibit or reduce the adaptive update rate of the online compensation model, and can simultaneously reduce the test excitation amplitude, pause the online identification process of the current joint, or send maintenance warning information to the host computer to avoid continuing to update control parameters based on distorted data under abnormal hot conditions.

[0160] See attached document Figure 8 , Figure 8 This is a flowchart of adaptive control compensation incorporating hysteresis feature evolution according to an embodiment of the present invention. In a joint adaptive control method based on multimodal coupling provided by the present invention, step S60 regarding adaptive control compensation incorporating hysteresis feature evolution may specifically include the following sub-steps:

[0161] S601, in this embodiment, the instruction domain 610 within the central control module 600 constructs position compensation logic based on velocity-feedback according to the static backlash extracted in step S40. When the desired velocity of the target test joint crosses zero and reverses, due to the presence of pure geometric backlash inside the reducer, the motion command of the motor rotor often cannot immediately cause the actual physical displacement at the reducer output. To address the aforementioned transmission backlash phenomenon, the system needs to superimpose a compensation term related to the static backlash size into the desired position command. Considering that directly applying step position compensation can easily cause overshoot and mechanical shock in the driver position loop, the system uses a smooth transition function to generate continuous backlash compensation angular displacement. :

[0162] Furthermore, to reflect the gap variation law of the reducer under different thermal states, the central control module 600 calculates the thermal back clearance increment based on the temperature change or temperature rise gradient of the target test joint during operation. As a preferred embodiment, the thermal back gap increment It can be obtained based on offline thermal calibration results using a linear or piecewise linear relationship. In the linear implementation, its expression can be written as:

[0163] ;

[0164] in, The thermal back clearance coefficient, The temperature at the current moment. For calibration reference temperature.

[0165] In another implementation, to take into account the dynamic heating process, an incremental thermal back gap model can be constructed based on the temperature rise gradient:

[0166] ;

[0167] in, This is the temperature rise gradient correction factor. This represents the temperature gradient at the current moment.

[0168] Based on this, the system defines the effective backlash compensation as:

[0169] ;

[0170] The effective backlash compensation amount is then substituted into the smoothing transition function to generate the commutation compensation angular displacement. Preferably, the commutation compensation angular displacement can be expressed as:

[0171] .

[0172] in, Test joints for target The expected angular velocity at time t. This is the speed smoothing coefficient. Through the above method, the compensation control can simultaneously adapt to changes in mechanical geometric clearance and thermal state.

[0173] S602 addresses the nonlinear hysteresis deformation and dynamic friction errors that are difficult to cover by static backlash compensation mechanisms. The system utilizes a radial basis function neural network (RBFNN) to construct a hysteresis dynamic feedforward compensation model. This RBFNN topology includes an input layer, hidden layers, and an output layer. The input layer has a dimension of 2 and receives the theoretically expected angular displacement and angular velocity from the motor side after Z-score normalization preprocessing, thereby reducing the impact of the dimensional differences between the two physical quantities on numerical calculations. Specifically, the input vector is defined as... ,in Test joints for target The theoretical expected angular displacement at time t. Test joints for target The theoretical expected angular velocity at time t.

[0174] Hidden layer contains The system has several neuron nodes, using a Gaussian kernel function as the activation function to achieve a nonlinear mapping from the two-dimensional kinematic input space to the high-dimensional feature space. The output layer is a single node, which outputs the hysteresis feedforward compensation torque through a linear weighted summation of the hidden layer features. This output physical quantity characterizes the additional electromagnetic torque required to overcome the changes in dynamic stiffness and nonlinear losses within the reducer under the current motion state.

[0175] During model computation, the hidden layer... The output of each neuron The calculation formula is:

[0176] ;

[0177] In the formula, The input vector of the neural network is a two-dimensional vector composed of the theoretical expected angular displacement and the theoretical expected angular velocity of the target test joint in this embodiment. and The first The center vector and width parameter of a Gaussian basis.

[0178] Hysteresis feedforward compensation torque of the output layer Calculated by the following formula:

[0179] ;

[0180] In the formula, These are the weight parameters corresponding to the hidden layer to the output layer.

[0181] As a preferred approach, the model training process employs a strategy combining offline training and online fine-tuning. The training dataset is derived from the multimodal synchronous dataset constructed in step S30 and the historical operation database of the same robot model. The input feature samples for model training are the theoretical angular displacement and angular velocity sequences of the motor; the corresponding label data is defined as the difference sequence between the output driving torque calculated in step S40 and the theoretical nominal torque derived based on the ideal rigid body dynamics equations. In the offline initialization phase, the system uses the K-means clustering algorithm to perform unsupervised clustering of the input feature sample space, thereby determining the number of hidden layer nodes. and the center vectors of each Gaussian basis With initial width To mitigate overfitting when fitting complex hysteresis loops, the model training employs a joint loss function combining mean squared error and L2 regularization. :

[0182] ;

[0183] In the formula, This represents the total number of samples in a single training batch. This is the index number of the current sample. , for the first Hysteresis feedforward compensation torque of each sample output For the first The torque difference label corresponding to each sample is the regularization coefficient.

[0184] After entering the online operation phase, as the mechanical state slowly drifts over time, the central control module 600 collects the actual position feedback data of the target test joint in real time and calculates the current position following error:

[0185] ;

[0186] in, Test the joint at time 1000 rpm The actual feedback angular displacement. Test the joint at time 1000 rpm The desired target angular displacement.

[0187] The system follows the error at the current position. As a basis for online adjustment, the output layer weights of the network model (or controller) are adjusted. Perform incremental updates. Preferably, the output layer weight update formula can be written as:

[0188] ;

[0189] in, Test joints for target Output layer weight parameters at time 10:00 In order to be in Output layer weight parameters at time 10:00 The online learning rate is preferably set between 0.0001 and 0.01. For the first The activation output value of each Gaussian basis at the current moment. This refers to the dynamic transmission stiffness extracted in step S40. It should be noted that position following error is taken into account. (Dimensions are in radians) and output layer weights The dimensional differences between (measured in the dimension of torque) are addressed by introducing dynamic transmission stiffness. This directly maps the position error to an equivalent lost torque compensation in a physical sense, making the online weight update formula completely unified in terms of physical dimensions. Simultaneously, the online learning rate... It is purely used as a dimensionless update step size, thereby ensuring the rigor of the dynamic closed-loop feedback logic and the stability of network convergence.

[0190] To prevent weight divergence during the online update process, the system sets upper and lower limit clamping thresholds for the updated output layer weights, namely:

[0191] ;

[0192] in, and They represent the first The upper and lower limits of the weights of each output layer can be preset based on the weight distribution characteristics obtained from offline training or to control safety limits.

[0193] Furthermore, when the absolute value of the current position following error satisfies:

[0194] ;

[0195] And continuously preset duration Time (of which) (Using a preset error tolerance threshold), the system freezes output layer weight updates to avoid introducing unnecessary parameter jitter during the stabilization phase. As a preferred implementation, online fine-tuning is only applied to the output layer weights. Execution, and the center vector of each Gaussian basis. With width parameter The weights remain unchanged after offline training to reduce online computational complexity and improve control stability. The system generates control compensation quantities in real time based on the updated output layer weights and applies them to the target test joint, thereby eliminating the effects of mechanical drift in a closed loop.

[0196] In step S603, based on feedforward compensation, the system further adjusts the underlying feedback loop because the degradation of mechanical stiffness synchronously alters the stability margin of the system's control closed loop. The central control module 600 then adjusts the dynamic transmission stiffness extracted in step S40. The feedback control gain at the servo driver's underlying layer is adaptively and dynamically tuned. A dynamic transmission stiffness lower than the factory-set reference stiffness typically reflects a weakening of the joint's resistance to load disturbances. To reduce the risk of high-frequency oscillations induced by the original high-stiffness control parameters in a low-stiffness mechanical system, the system synchronously lowers the servo position loop proportional gain based on the stiffness attenuation ratio. :

[0197] ;

[0198] In the formula, and These are the position loop proportional gain initially set for the servo driver and the factory-calibrated reference dynamic stiffness, respectively. This is the gain adjustment factor. By... Setting the gain between 0.5 and 1.0 allows the system to appropriately increase its compliance while maintaining basic tracking performance. Simultaneously, to prevent excessive gain downsizing that could lead to system malfunction due to severe stiffness degradation, the central control module 600 has an internal safety lower limit threshold. After each calculation, the system... and A numerical comparison is performed; if the calculated value is lower than the threshold, then... Forced clamp located It also sends a maintenance warning to the host computer indicating a serious lack of mechanical rigidity.

[0199] S604, combining the adjustment results of the above open-loop feedforward and closed-loop feedback, the instruction domain 610 within the central control module 600 will calculate the backlash compensation angular displacement. Add it to the basic position command and include the hysteresis feedforward compensation torque. As the feedforward torque setpoint, along with the updated feedback control gain A comprehensive adaptive control command sequence is packaged and synthesized. Through the underlying fieldbus, the command domain 610 sends this control command sequence to the servo driver of the target test joint in real time for execution. This multi-dimensional coupling compensation mechanism helps the robot control system automatically adapt to the physical evolution of the underlying mechanical transmission characteristics at the software algorithm level, thereby maintaining good accuracy and smoothness of the end-effector trajectory throughout the entire lifecycle of the device.

[0200] To more clearly demonstrate the practical application effect of the joint adaptive control method based on multimodal coupling of this invention, refer to... Figure 11The following is a detailed explanation using the scenario of "high-precision micro-dispensing operation for 3C electronic products".

[0201] In this embodiment, a 6-DOF industrial robot with a rated load of 10kg and an arm span of 1.2m is used to perform the micro-dispensing task, with an end-effector trajectory accuracy required to be within ±0.05mm. A 12-megapixel industrial camera (visual perception module 100) is fixedly mounted on the top of the industrial robot base. Due to frequent starts and stops over a long period, the reducer of the robot's second joint (the upper arm pitch joint, defined as the target test joint J2) is subjected to heavy loads, posing a risk of wear and increased backlash.

[0202] During the 2-second non-working interval of production material changeover, the system triggers the online adaptive calibration process:

[0203] (1) The central control module 600 rigidly locks the joints J1, J3 to J6 through the zero-speed high-gain command (S10).

[0204] (2) Analysis by vibration sensing module 400 shows that the natural resonant frequency of the environmental micro-vibration of joint 12 has drifted from the factory-calibrated low frequency of 45Hz to 41.5Hz. In order to avoid this resonance zone, the system adaptively reconstructs the carrier frequency of the micro-amplitude commutation test signal to 50Hz and sets the micro-amplitude angular displacement to 0.05 degrees (S20).

[0205] (3) Under 50Hz excitation, the system synchronously acquires visual displacement, motor encoder data, servo quadrature current and local temperature based on the IEEE1588 protocol (S30).

[0206] (4) The static backlash of the J2 joint is calculated by using the reduced-order Jacobian inverse solution and hysteresis dynamics extraction. The dynamic transmission stiffness is 0.025 degrees (an increase of 0.01 degrees compared to the new machine). The temperature drops by 12% (S40). At the same time, based on the temperature rise gradient measured by the internal temperature sensor, the system determines that the lubrication is in a normal state, but lower backlash compensation is required (S50).

[0207] (5) When entering the next dispensing cycle, the system starts feedforward and feedback compensation. Based on the static backlash of 0.025 degrees and the thermal increment, the commutation compensation angular displacement is generated to eliminate pure hysteresis; at the same time, the radial basis function neural network outputs the hysteresis feedforward compensation torque according to the current desired position and speed, and synchronously reduces the servo position loop gain by 10% to adapt to the current stiffness (S60).

[0208] To verify the effectiveness of the adaptive precision motion control system of the present invention, an end-effector circular trajectory following test was conducted on the platform of the above embodiment (test speed was 100 mm / s, and the radius of the circular arc was 50 mm). The actual spatial tracking error of the end effector and the joint vibration acceleration were extracted and compared and analyzed.

[0209] Two control experimental groups were set up:

[0210] Control group (traditional control method): adopts factory-fixed servo gain configuration and only includes rigid body feedforward control based on the motor encoder side, without visual inverse solution and hysteresis state adaptive update.

[0211] Experimental group (method of this invention): The multimodal hysteresis parameters extracted by this system are used to activate the nonlinear feedforward compensation of the RBF neural network and the dynamic tuning of the closed-loop servo gain.

[0212] Comparison of end-point trajectory tracking accuracy:

[0213] Because the J2 joint undergoes a zero-crossing reversal (reversal) when the dispensing trajectory switches between the four quadrants, traditional control methods, due to the lack of compensation for increased static backlash and nonlinear friction, result in significant quadrant bulge errors at the reversal point. The maximum peak value of the end-effector absolute tracking error reaches 0.085 mm, exceeding the ±0.05 mm process tolerance limit. However, by adopting the method of this invention, thanks to the speed reversal feedforward position compensation (based on accurately extracted...) With the dynamic compensation torque provided by the RBF neural network, the lag during the commutation zero-crossing stage is greatly reduced, the peak absolute tracking error at the end of the trajectory is reduced to 0.021mm, and the root mean square error (RMSE) decreases from 0.038mm to 0.012mm, improving the trajectory tracking accuracy by approximately 68%. The comparison effect of the end-of-trace trajectory tracking accuracy is shown below. Figure 9 As shown.

[0214] Comparison of high-frequency micro-vibration suppression effects:

[0215] Traditional control methods, which maintain a high gain ratio even with stiffness degradation, are prone to servo-mechanical coupling resonance. Power spectral density (PSD) analysis of the vibration acceleration signal from the housing side of the 12-joint reducer reveals that the control group exhibits a significant resonance peak in the 40Hz–50Hz frequency band, accompanied by abnormal noise during operation. After applying the method of this invention, the system... The position loop gain was adaptively reduced during attenuation. Furthermore, it smooths out internal impacts by relying on dynamic feedforward torque. Spectrum comparisons show that the method of this invention attenuates the resonant peak energy by nearly 75%, and the root mean square acceleration (RMSE)... Significantly reduced [the risk of vibration], greatly improved the smoothness of joint movement, extended the remaining service life of the mechanical transmission chain, and demonstrated a comparative effect in suppressing high-frequency micro-vibrations. Figure 10 As shown in the figure, the horizontal axis represents the running time. (Unit: s), the vertical axis represents the absolute tracking error at the end (unit: mm), and the two horizontal dashed lines in the figure represent the process tolerance boundary of ±0.05 mm.

[0216] Depend on Figure 10 The curve data shows that the control group (gray dashed line) using traditional fixed gain control exhibits significant error abrupt changes at approximately 1.0s and 3.0s (corresponding to the zero-crossing reversal point of the J2 joint velocity). Its peak forward tracking error reaches approximately 0.07mm, and its peak reverse tracking error reaches approximately -0.07mm, both exceeding the ±0.05mm process tolerance boundary. This error mainly stems from the increased static backlash caused by reducer wear and the lack of systematic compensation for nonlinear friction.

[0217] In contrast, the experimental group (solid black line) employing the adaptive hysteresis compensation control of this invention showed significantly reduced error fluctuations at the same speed reversal point. Throughout the entire operating cycle, the peak value of the absolute tracking error at the end of the experimental group was stably controlled within approximately ±0.021 mm, remaining within the safe range of the process tolerance boundary throughout the entire time period. These data directly verify that this invention, by combining static backlash calculation with reversal compensation angular displacement and utilizing a radial basis function neural network to output feedforward compensation torque, effectively overcomes the idle lag and nonlinear friction interference of the mechanical transmission chain, ensuring the control accuracy of the end trajectory.

[0218] See attached document Figure 11 In the figure, the horizontal axis represents frequency (unit: Hz), and the vertical axis represents the power spectral density of vibration acceleration (unit: dB / Hz).

[0219] Depend on Figure 11 As shown in the curve distribution, the control group using traditional control (the gray dashed line marked with squares) exhibits a sharp resonance peak in the 40Hz to 50Hz frequency band, with a peak power spectral density of approximately -18dB / Hz. The physical mechanism behind this phenomenon lies in the fact that the dynamic transmission stiffness inside the reducer has decreased by 12%, while the underlying servo controller still maintains the factory-set high gain. The excessive electromagnetic stiffness is mismatched with the degraded mechanical stiffness, thus inducing high-frequency servo-mechanical coupling resonance.

[0220] The experimental group (solid black line) using the control method of this invention, under the same operating conditions, completely suppressed the resonance peak in the 40Hz to 50Hz frequency band, and the power spectral density in this band was generally limited to approximately -30dB / Hz, with a flat energy distribution. This comparative result demonstrates that the control mechanism of this invention, which synchronously adjusts the proportional gain of the servo position loop based on the dynamic transmission stiffness attenuation ratio, enables the underlying electrical parameters to dynamically adapt to the current mechanical and physical properties, effectively preventing the system from inducing high-frequency destructive oscillations under low stiffness conditions and improving the dynamic smoothness of the robot system.

Claims

1. A visual perception-based adaptive precision motion control method for robots, characterized in that, Includes the following steps: Identify the target test joint and the non-test joint, perform rigid locking on the non-test joint, and construct a single-axis reduced-order Jacobian matrix for the target test joint based on the standard Jacobian matrix. Test signals are injected into the target test joint and multimodal data is collected synchronously to generate a multimodal synchronous data set; Based on the uniaxial reduced Jacobian matrix and the multimodal synchronization data set, the static backlash and dynamic transmission stiffness of the target test joint are extracted, and mechanical wear and thermodynamic state assessments are performed. Based on the static backlash, the dynamic transmission stiffness, and the results of the state assessment, the commutation compensation angular displacement and feedforward compensation torque are calculated, and the feedback control gain is updated according to the dynamic transmission stiffness. These are then combined into control commands and issued for execution.

2. The visual perception-based adaptive precision motion control method for robots according to claim 1, characterized in that, The process of performing rigid locking on the non-test joint, and constructing a uniaxial reduced-order Jacobian matrix for the target test joint based on the standard Jacobian matrix, specifically includes: Send zero-speed hold commands to the servo drives corresponding to all the non-test joints, and increase the gain of the position control loop and speed control loop to the maximum critical value allowed by the load inertia; Extract the standard Jacobian matrix of the industrial robot, set the column vectors corresponding to the non-test joints in the standard Jacobian matrix to zero, retain the column vectors corresponding to the target test joints, and generate the single-axis reduced-order Jacobian matrix.

3. The visual perception-based adaptive precision motion control method for robots according to claim 1, characterized in that, The multimodal data includes the actual spatial displacement vector of the end effector, theoretical angular displacement, quadrature-axis drive current, vibration acceleration signal, and temperature signal; the synchronous acquisition of multimodal data to generate a multimodal synchronous data set specifically includes: Based on the precise time protocol, a synchronization message is broadcast and a hardwired trigger pulse is sent to synchronously trigger the corresponding sensing module. The actual spatial displacement vector at the end point, the theoretical angular displacement, the cross-axis drive current, the vibration acceleration signal, and the temperature signal acquired in parallel are normalized and low-pass filtered to generate the multimodal synchronous data set with a unified time axis.

4. The visual perception-based adaptive precision motion control method for robots according to claim 3, characterized in that, The extraction of the static backlash and dynamic transmission stiffness of the target test joint specifically includes: The actual spatial displacement vector at the end is inversely solved into the actual angular displacement change using the single-axis reduced-order Jacobian matrix, and the actual transmission error is calculated by combining the theoretical angular displacement. Based on the quadrature axis drive current, the output drive torque converted to the reducer output end is obtained, and a hysteresis loop is constructed with the output drive torque as the abscissa and the actual transmission error as the ordinate. The static backlash and dynamic transmission stiffness are extracted based on the hysteresis loop, and the nonlinear mechanical loss work is calculated by integrating the closed region of the hysteresis loop envelope.

5. The visual perception-based adaptive precision motion control method for robots according to claim 4, characterized in that, The step of using the single-axis reduced-order Jacobian matrix to inversely solve the actual spatial displacement vector at the end point into the actual angular displacement change specifically includes: Calculate the product matrix of the single-axis reduced Jacobian matrix and its transpose, and obtain the minimum singular value of the product matrix; When the minimum singular value is lower than a preset safety threshold, a damping factor is introduced. The damping factor is combined with the transpose of the single-axis reduced-order Jacobian matrix to construct a regularized pseudo-inverse matrix. The actual spatial displacement vector at the end is subjected to a forward inverse kinematics solution using the regularized pseudo-inverse matrix to obtain the actual angular displacement change.

6. The visual perception-based adaptive precision motion control method for robots according to claim 4, characterized in that, The mechanical wear condition assessment includes, in particular, the following: Bandpass filtering is performed on the vibration acceleration signal, and the root mean square value within a single cycle is calculated as the vibration energy characteristic. The ratio of the static back clearance to the nominal static back clearance and the ratio of the vibration energy characteristic to the nominal reference value are assigned normalized weights and summed to calculate the comprehensive wear evaluation index. When the comprehensive wear evaluation index exceeds the wear alarm threshold, the target test joint is determined to be in an excessive wear state.

7. The visual perception-based adaptive precision motion control method for robots according to claim 4, characterized in that, The thermodynamic state assessment includes, in particular: The average dissipation power is calculated based on the nonlinear mechanical loss power and the duration of a single cycle. After subtracting the ambient temperature from the temperature signal, a linear trend fitting is performed to obtain the temperature rise gradient; Calculate the cross ratio between the temperature rise gradient and the average power dissipation. When the cross ratio is greater than a preset reference threshold multiple, it is determined that there is an abnormal dry friction or lubrication deterioration state.

8. The visual perception-based adaptive precision motion control method for robots according to claim 7, characterized in that, The calculation of the reversal compensation angular displacement specifically includes: Calculate the thermal back clearance increment based on the temperature rise gradient; The effective back gap compensation amount is obtained by summing the static back gap and the thermal back gap increment. Substituting the effective backlash compensation amount and the desired angular velocity of the target test joint into the smooth transition function generates the continuous reversing compensation angular displacement.

9. The visual perception-based adaptive precision motion control method for robots according to claim 4, characterized in that, The calculation of the feedforward compensation torque specifically includes: The desired angular displacement and desired angular velocity are input into the constructed radial basis function neural network, mapped through the hidden layer and linearly weighted by the output layer to output the feedforward compensation torque; Obtain the actual feedback angular displacement and calculate the position following error; The position following error is combined with the dynamic transmission stiffness and mapped to a lost torque compensation amount. This is then substituted into the online learning law to incrementally update the weight parameters of the output layer, and the updated weight parameters are clamped at upper and lower limits.

10. The visual perception-based adaptive precision motion control method for robots according to claim 4, characterized in that, The step of updating the feedback control gain based on the dynamic transmission stiffness specifically includes: Obtain the initial position loop proportional gain and reference dynamic stiffness of the servo driver; Calculate the attenuation ratio of the dynamic transmission stiffness relative to the reference dynamic stiffness, and combine it with the gain adjustment factor to calculate the reduced current position loop proportional gain. When the current position loop proportional gain is lower than the set safety lower limit threshold, it is forced to clamp at the safety lower limit threshold.

Images