Decoupling control methods, computer program products, and multi-degree-of-freedom test benches

By introducing a three-dimensional special Euclidean group and a geometric Jacobian matrix, combined with a physical information neural network and fractional rigid-flexible coupling dynamics, the singularity and non-physical jump problems of multi-degree-of-freedom test benches in large-angle motion and long-term experiments are solved, achieving stable and safe decoupled control.

CN122860987BActive Publication Date: 2026-11-17CHONGQING VEHICLE TEST & RES INST CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611339474.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-09-01
Publication Date
2026-11-17
Estimated Expiration
2046-09-01

AI Technical Summary

Technical Problem

Existing multi-degree-of-freedom test benches are prone to gimbal lock-up during complex large-angle motions, resulting in singular Jacobian matrices and non-physical drastic output jumps. Furthermore, prolonged testing can cause mechanical wear or heat generation, leading to the failure of fixed parameter matrices. Purely data-driven neural networks lack the constraints of physical laws and are prone to outputting dangerous commands.

Method used

A mapping model of the actuator joint space and table space velocity is established by using a three-dimensional special Euclidean group and geometric Jacobian matrix. Real-time feedback control is performed by combining physical information neural network. The decoupled driving parameters are optimized by deep feedforward neural network. Fractional rigid-flexible coupling dynamic equations are introduced to establish a total loss function to ensure physical constraints. The backpropagation algorithm is used to iteratively optimize the network weights.

Benefits of technology

Maintaining the stability of the drive matrix during complex large-angle movements and long-term tests, avoiding non-physical jumps, possessing the ability to resist mechanical wear and temperature drift, achieving high-precision online adaptive control, and ensuring equipment safety.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122860987B_ABST
    Figure CN122860987B_ABST
Patent Text Reader

Abstract

This invention discloses a decoupling control method, a computer program product, and a multi-degree-of-freedom test bench, comprising: establishing a mapping model between the actuator joint space velocity and the table space velocity using a three-dimensional special Euclidean group and a geometric Jacobian matrix; acquiring the physical feedback state and target spectrum of the multi-degree-of-freedom test bench in real time; and inferring theoretical decoupling driving parameters through a deep feedforward neural network in a physical information neural network; substituting the theoretical decoupling driving parameters into the mapping model and the physical excitation force conversion model to determine the theoretical coupling dynamic response; embedding the residual of the differential dynamic equation calculated based on the theoretical coupling dynamic response into the loss function of the deep feedforward neural network; and using a backpropagation algorithm to iteratively optimize the network weights of the neural network to minimize the loss function; and finally, solving the decoupling driving parameters that satisfy the physical safety boundary and actual actuation constraints through the optimized physical information neural network.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of data processing methods or equipment applicable to specific functions, specifically to a decoupling control method, a computer program product, and a multi-degree-of-freedom test bench. Background Technology

[0002] In automotive component noise and reliability testing (BSR) and testing, multi-degree-of-freedom test benches are widely used to simulate real-vehicle road patterns. Existing multi-axis motion decoupling mainly relies on forward / inverse kinematics calculations, that is, establishing a fixed Jacobian matrix through the system geometry to transform the multi-degree-of-freedom Cartesian coordinate target at the center point into displacement or force drive commands for multiple actuators. However, traditional decoupling algorithms typically use Euler angles to describe spatial rotation. When performing large-angle complex motions on the test bench, they are prone to gimbal lock-up, leading to singularities in the Jacobian matrix and causing drastic, non-physical jumps in the actuator output.

[0003] Moreover, mechanical wear or heat caused by long-term testing can cause the fixed parameter matrix to fail. If a purely data-driven neural network (such as a traditional BP or LSTM) is directly introduced for inverse mapping, due to the lack of physical law constraints, it is very easy to output dangerous commands that cause the equipment to "run away" or be damaged when encountering unseen test spectra. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a decoupling control method, a computer program product, and a multi-degree-of-freedom test bench. This method maintains the full rank and stability of the geometric Jacobian matrix during large-angle complex movements on the test bench, preventing drastic, non-physical jumps in the driver output. The specific technical solution is as follows: In a first aspect, a decoupling control method is provided, and in a first implementable mode of the first aspect, it includes: A mapping model between the joint space velocity of the actuator and the table space velocity is established by using a three-dimensional special Euclidean group and a geometric Jacobian matrix. The physical feedback state and target spectrum of the multi-degree-of-freedom test bench are acquired in real time. The physical feedback state is the physical feedback state output by the sensors configured on the multi-degree-of-freedom test bench, including the real-time displacement, velocity and acceleration of each actuator. The target spectrum is the excitation signal that the multi-degree-of-freedom test bench is expected to reproduce, including the acceleration time history or power spectral density spectrum collected through actual vehicle road tests. The physical feedback state and the target spectrum are input into the deep feedforward neural network in the physical information neural network to infer the theoretical decoupling driving parameters, as well as the modal displacement and modal acceleration of the multi-degree-of-freedom test bench. The theoretical decoupling driving parameters include the joint displacement vector and joint velocity vector of each actuator of the multi-degree-of-freedom test bench. The theoretical coupled dynamic response of the multi-degree-of-freedom test bench under the control of theoretical decoupling driving parameters is determined by the mapping model and the physical excitation force conversion model. The theoretical coupled dynamic response includes the physical excitation force vector and the kinematic screw of the center of the table. Substituting the theoretical decoupling driving parameters inferred from the deep feedforward neural network into the mapping model, the motion screw of the multi-degree-of-freedom test platform center is calculated; substituting the theoretical decoupling driving parameters inferred from the deep feedforward neural network into the physical excitation force conversion model, the physical excitation force vector to be applied to the platform is calculated. Based on the physical excitation force vector in the theoretically coupled dynamic response, combined with the modal displacement and modal acceleration, the equation residual of the differential dynamic equation is calculated, and the data fidelity loss of the deep feedforward neural network is calculated based on the motion spinor of the center of the platform in the theoretically coupled dynamic response. The equation residual is used as the physical law penalty loss, and the total loss function is established by combining the data fidelity loss. The backpropagation algorithm is used to iteratively optimize the network weights of the deep feedforward neural network so that the value of the total loss function is minimized. Based on the current physical feedback state and target spectrum of the multi-degree-of-freedom test bench, the decoupling driving parameters of the multi-degree-of-freedom test bench are predicted by an optimized physical information neural network.

[0005] In conjunction with the first feasible method of the first aspect, in the second feasible method of the first aspect, a mapping model between the joint space velocity of the actuator and the platform space velocity is established, including: The motion screw of each actuator is determined based on the corresponding joint displacement vector, and the motion screws of each actuator are stacked into the geometric Jacobian matrix.

[0006] In conjunction with the second possible implementation of the first aspect, in the third possible implementation of the first aspect, the specific expression of the mapping model is as follows: ; in, The kinematic spinor at the center of the table. Let be the pose transformation matrix at the center of the platform. Li Qun's position The corresponding adjoint operator, Let be the joint displacement vector of the actuator. Let be the joint velocity vector of the actuator. It is a geometric Jacobian matrix. This represents the number of actuators.

[0007] In conjunction with the first feasible method of the first aspect, in the fourth feasible method of the first aspect, the calculation of the equation residuals of the differential dynamic equations includes: Based on the flexible modal parameters of the multi-degree-of-freedom test rig, the fractional-order rigid-flexible coupling dynamic equation is established by introducing the Caputo fractional derivative, and the fractional-order rigid-flexible coupling dynamic equation is used as the differential dynamic equation.

[0008] In conjunction with the fourth possible implementation method of the first aspect, in the fifth possible implementation method of the first aspect, the specific expression of the fractional-order rigid-flexible coupling dynamic equation is as follows: ; in, Here is the modal mass matrix. The second derivative of the modal coordinates, Here is the modal damping matrix. For Caputo fractional differential operators, For flexible modal coordinate vectors, Here is the modal stiffness matrix. Let be the modal elastic restoring force vector. For the mode shape matrix, The physical excitation force vector, This is the order of fractional calculus.

[0009] In conjunction with the first possible implementation of the first aspect, in the sixth possible implementation of the first aspect, an adaptive penalty weight coefficient is introduced into the total loss function to balance the strength of data fitting and physical constraints.

[0010] In conjunction with the first possible implementation of the first aspect, the seventh possible implementation of the first aspect further includes: soft limiting the decoupling drive parameters based on the feedback state of the bottom-level sensor of the multi-degree-of-freedom test bench.

[0011] In a second aspect, a computer program product is provided, including a computer program / instructions that, when executed by a processor, implement the steps of the decoupling control method as described in any of the first to seventh implementable embodiments of the first aspect.

[0012] Thirdly, a multi-degree-of-freedom test bench is provided, including a central hardware control unit, which uses a decoupling control method as described in any of the first to seventh implementable methods of the first aspect to determine the decoupling drive parameters.

[0013] Beneficial Effects: By employing the decoupling control method, computer program product, and multi-degree-of-freedom test bench of this invention, and replacing Euler angles in traditional decoupling algorithms with a three-dimensional special Euclidean group, singularity-free rigid body kinematic benchmark modeling is achieved in full space. This eliminates Euler angle gimbal lock-up at the theoretical level, ensuring the multi-degree-of-freedom test bench maintains full rank and stability of the driving matrix even when simulating large-angle deflections under extreme road conditions, preventing sudden equipment protection or waveform distortion. Furthermore, the use of a physical information neural network endows the multi-degree-of-freedom test bench with the ability to resist mechanical wear and temperature drift during complex tests. Moreover, by embedding the equation residuals as prior physical laws into the loss function of the deep feedforward neural network of the physical information neural network, the output decoupling driving parameters are strictly locked within the physical safety boundaries formed by the differential dynamic equations. This achieves high-precision online adaptive control while possessing absolute physical operational safety, ensuring the operational safety of expensive test pieces and the equipment itself. Attached Figure Description

[0014] To more clearly illustrate the specific embodiments of the present invention, the accompanying drawings used in the specific embodiments will be briefly described below. In all the drawings, the elements or parts are not necessarily drawn to scale.

[0015] Figure 1 A flowchart of a decoupling control method provided in an embodiment of the present invention; Figure 2 The simulation curves show the phase difference between the present invention and the traditional integer-order damping decoupling control method in the high-frequency band. Detailed Implementation

[0016] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.

[0017] It should be understood that the "multi-degree-of-freedom test bench" is also called the "multi-degree-of-freedom vibration table" and is widely used to simulate the road spectrum of real vehicles.

[0018] The "target spectrum" refers to the excitation signal that is expected to be reproduced on the test bench, which can usually be obtained from the acceleration time history or power spectral density spectrum collected from actual vehicle road tests.

[0019] Example 1

[0020] like Figure 1 The flowchart shown illustrates a decoupling control method, which includes: Step 1: Using a three-dimensional special Euclidean group and a geometric Jacobian matrix, establish a mapping model between the joint space velocity of the actuator and the table space velocity. Step 2: Real-time acquisition of the physical feedback state and target spectrum of the multi-degree-of-freedom test bench, and forward inference of the theoretical decoupling driving parameters through the deep feedforward neural network in the physical information neural network; Step 3: Determine the theoretical coupling dynamic response of the multi-degree-of-freedom test bench under the control of theoretical decoupling driving parameters through the mapping model and the physical excitation force conversion model; Step 4: Calculate the equation residuals of the differential dynamic equations and the data fidelity loss of the deep feedforward neural network based on the theoretical coupled dynamic response. Use the equation residuals as the physical law penalty loss and combine them with the data fidelity loss to establish the total loss function. Step 5: Iteratively optimize the network weights of the deep feedforward neural network using the backpropagation algorithm to minimize the value of the total loss function; Step 6: Based on the current physical feedback state and target spectrum of the multi-degree-of-freedom test bench, predict the decoupling driving parameters of the multi-degree-of-freedom test bench through the optimized physical information neural network.

[0021] Specifically, firstly, instead of using singular Euler angles, we can introduce a three-dimensional special Euclidean group and combine it with the geometric Jacobian matrix to establish a mapping model between the actuator joint space velocity and the table space velocity of the multi-degree-of-freedom test bench, so as to describe the absolute rigid body motion of the multi-degree-of-freedom test bench table.

[0022] Then, the physical feedback states output by sensors configured on the multi-degree-of-freedom test bench, such as displacement sensors and acceleration sensors, can be acquired in real time, including the real-time displacement, velocity, and acceleration of each actuator. Combined with a pre-set target spectrum, the theoretical decoupling driving parameters of the multi-degree-of-freedom test bench, as well as its modal displacement and modal acceleration, are derived through a deep feedforward neural network within the physical information neural network. The theoretical decoupling driving parameters include the joint displacement coordinate vectors of each actuator on the multi-degree-of-freedom test bench.

[0023] Subsequently, the theoretical decoupling driving parameters inferred from the deep feedforward neural network can be substituted into the previously constructed mapping model to obtain the motion screw of the multi-degree-of-freedom test platform's center. Simultaneously, the theoretical decoupling driving parameters are substituted into the physical excitation force conversion model to calculate the required physical excitation force vector for the platform.

[0024] Then, the physical excitation force vector is substituted into the differential dynamics equation as a known term. The modal displacement, modal velocity, and modal acceleration predicted synchronously by the deep feedforward neural network are substituted into the differential dynamics equation to calculate the equation residual. The equation residual is used as a penalty loss for the a priori physical laws. Combined with the data fidelity loss calculated based on the spinor, the loss function of the deep feedforward neural network is constructed, so that the output decoupling driving parameters are strictly locked within the physical safety boundary formed by the differential dynamics equation.

[0025] Subsequently, by applying physical laws to penalize predictions that do not conform to dynamic laws, the backpropagation algorithm is used to iteratively optimize the network weights of the deep feedforward neural network to minimize its loss function. This achieves on-the-fly learning and absolutely safe online decoupling, fundamentally ensuring the operational safety of the expensive test specimen and the equipment itself. Finally, the optimized physical information neural network can calculate the decoupling driving parameters that strictly satisfy the physical safety boundaries and actual motion constraints based on the current physical feedback state and target spectrum of the multi-degree-of-freedom test bench.

[0026] By introducing a three-dimensional special Euclidean group, the Euler angle gimbal lock phenomenon is eliminated at the theoretical level, effectively solving the singularity problem in attitude description. While avoiding the multi-degree-of-freedom test rig from entering singular kinematic configurations, the multi-degree-of-freedom test rig can maintain the full rank and stability of the drive matrix when simulating large-angle deflections under extreme road conditions, greatly reducing the probability of sudden equipment protection or waveform distortion. Experimental verification shows that under extreme combined working conditions where the actuator stroke approaches its physical extreme value, the matrix condition number can be stably maintained. The benign range.

[0027] In this embodiment, optionally, a mapping model is established between the actuator joint space velocity and the table space velocity, including: The motion screw of each actuator is determined based on the corresponding joint displacement vector, and the motion screws of each actuator are stacked into the geometric Jacobian matrix.

[0028] Specifically, in this embodiment, the unit local motion screw of each actuator along its axis in the current posture can be determined based on the joint displacement vectors of each actuator in the multi-degree-of-freedom test bench. The unit local motion screw can be calculated by extracting the Plück coordinates of each actuator's axis in the base coordinate system. The unit local motion screw is then used as the motion screw of the actuator, and the motion screws of all actuators are stacked column-wise into a geometric Jacobian matrix, as shown in the following expression: ; in, The local rotational property is expressed as a unit.

[0029] It should be understood that this embodiment is only used as an example to illustrate the mapping of displacement vectors to the geometric Jacobian matrix. Other parameters, such as acceleration or driving force, can also be used to map the geometric Jacobian matrix. However, since the electro-hydraulic servo / electric cylinders of multi-degree-of-freedom test benches typically use a position closed loop as the basic control loop, and the spatial geometry of the system itself is determined by joint displacements, this invention preferentially selects joint displacement vectors to establish the geometric Jacobian matrix.

[0030] In this embodiment, optionally, the specific expression of the mapping model is as follows: ; in, The kinematic spinor at the center of the table. Let be the pose transformation matrix at the center of the platform. Li Qun's position The corresponding adjoint operator, Let be the joint displacement vector of the actuator. Let be the joint velocity vector of the actuator. It is a geometric Jacobian matrix. This represents the number of actuators.

[0031] Furthermore, to establish the mapping relationship between the joint displacement vector and the physical excitation force vector, this embodiment supplements the construction of a physical excitation force conversion model based on the equivalent structural characteristics of the actuator. Its specific expression is as follows:

[0032] in, This is converted into the physical excitation force vector required to be applied to the platform by each actuator. The equivalent stiffness matrix of an actuator system (such as an electro-hydraulic servo actuator). This is the equivalent damping matrix of the actuator system. Through this physical excitation force conversion model, the decoupled driving parameters (joint displacement and joint velocity) inferred from the neural network can be directly mapped to the physical excitation force input required for subsequent dynamic equations.

[0033] Specifically, the theoretically coupled dynamic response includes the physical excitation force vector and the motion screw of the platform center. The aforementioned mapping model establishes a singular reference mapping for the low-frequency, large-displacement motion of the multi-degree-of-freedom test bench. The output of this mapping model provides the foundation for subsequent forward inference using a physical information neural network. By substituting the theoretical decoupling driving parameters of the multi-degree-of-freedom test bench, i.e., the inferred joint displacement vectors of the actuators, derived through forward inference by a deep feedforward neural network, into the aforementioned mapping model, the motion screw of the multi-degree-of-freedom test bench platform center can be calculated.

[0034] The theoretical decoupling driving parameters of the multi-degree-of-freedom test bench, derived from the forward inference of the deep feedforward neural network, namely the joint displacement vector and joint velocity vector of the actuators, are substituted into the physical excitation force conversion model mentioned above. Combined with the equivalent stiffness and damping characteristics of each actuator's own structure, the physical excitation force vector required to be applied to the test bench for each actuator is calculated.

[0035] It should be understood that this embodiment is only used as an example to illustrate the mapping model established by the geometric Jacobian matrix. In this invention, the mapping model can also be constructed using traditional Euler angles or quaternions, but this invention prefers the non-singular Lie group adjoint operator to avoid computational divergence under large poses.

[0036] In this embodiment, optionally, determining the theoretically coupled dynamic response through the mapping model and the differential dynamic equation includes: Based on the flexible modal parameters of the multi-degree-of-freedom test rig, the fractional-order rigid-flexible coupling dynamic equation is established by introducing the Caputo fractional derivative, and the fractional-order rigid-flexible coupling dynamic equation is used as the differential dynamic equation.

[0037] Specifically, the theoretically coupled dynamic response can be determined using the established mapping model and traditional integer-order ordinary differential dynamic equations. However, the geometric Jacobian matrix is ​​based on the assumption that the test platform and fixture are absolutely rigid bodies. In the high-frequency range of abnormal noise testing (e.g., >50Hz), the multi-degree-of-freedom test platform and fixture will inevitably undergo elastic deformation. The fixed Jacobian matrix cannot describe this flexible deformation, leading to severe inter-axis parasitic vibrations in the high-frequency range.

[0038] Furthermore, the vibration isolation rubber and hydraulic / electric servo system mechanical hinges widely used in multi-degree-of-freedom test benches exhibit highly nonlinear and "historical memory effect" high-frequency damping characteristics. Using integer-order differential equations (i.e., linear functions of velocity) as differential dynamic equations cannot accurately characterize this hysteresis damping, leading to inherent deviations in phase decoupling at high frequencies.

[0039] To address this, the Caputo fractional derivative can be introduced to establish a fractional rigid-flexible coupling dynamic equation as a differential dynamic equation. The Caputo fractional derivative is used to accurately describe high-frequency viscoelastic damping, and the flexible modal parameters of the multi-degree-of-freedom test bench are incorporated into the kinematic equations, thus eliminating high-frequency structural deformation at the algorithmic level. The fractional rigid-flexible coupling dynamic equation can accurately predict and eliminate the flexible resonance response of the multi-degree-of-freedom test bench surface and fixtures, significantly reducing inter-axis parasitic crosstalk in the mid-to-high frequency range (50Hz-200Hz).

[0040] In this embodiment, optionally, the specific expression of the fractional-order rigid-flexible coupling dynamic equation is as follows: ; in, Here is the modal mass matrix. The second derivative of the flexible modal coordinate vector, such as modal acceleration. Here is the modal damping matrix. For Caputo fractional differential operators, Here is the modal stiffness matrix. Let be the modal elastic restoring force vector. For the mode shape matrix, To obtain the physical excitation force vector at the current moment through the above physical excitation force conversion model, The order of the fractional calculus is determined by parameter identification based on the dynamic stiffness sweep frequency test results of the test bench. The typical value range is [range missing]. . and Each represents the current time. Modal displacement and modal acceleration.

[0041] Specifically, For multi-degree-of-freedom test bench The first-order flexible modal coordinate vector can be a modal displacement. Modal mass matrix. It can be obtained by extracting the finite element modal structure of the multi-degree-of-freedom test bench and its fixtures, specifically represented as a diagonal matrix: ; For the first Modal mass. Modal stiffness matrix. Specifically, it is represented by a diagonal matrix: , , For the first First natural circular frequency. Modal damping matrix. It has the characteristics of being represented as a diagonal matrix. , For the first Viscous dissipation coefficients of the first mode. Mode shape matrix. , represents the number of degrees of freedom at the physical nodes. In the equation... The method accurately characterizes the viscoelastic hysteresis damping properties of the tabletop connector and clamping system under high-frequency vibration.

[0042] In this embodiment, optionally, the equation residual is used as the physical law penalty loss to establish the total loss function, including: superimposing the physical law penalty loss and the data fidelity loss of the deep feedforward neural network to establish the total loss function.

[0043] Specifically, the topology of the deep feedforward neural network includes an input layer, multiple fully connected hidden layers, and an output layer. The input layer receives physical feedback states from various sensors (such as real-time displacement, velocity, and acceleration of each actuator) and target spectra. The hidden layers perform deep feature extraction on the input data through nonlinear activation functions to obtain modal displacement and modal acceleration. The output layer outputs the corresponding theoretical decoupling driving parameters (i.e., joint displacement vectors and joint velocity vectors of each actuator), thereby realizing the nonlinear mapping function from multi-source physical feedback states to decoupling driving parameters.

[0044] During the model training phase, specific steps and parameter settings include: using an adaptive optimizer such as Adam for gradient descent optimization, setting the initial learning rate to 0.001 and dynamically decaying it according to the training batches, and setting the maximum iteration period to ensure that the total loss function converges to a stable threshold.

[0045] In this embodiment, optionally, an adaptive penalty weight coefficient is introduced into the total loss function to balance the strength of data fitting and physical constraints.

[0046] Specifically, this embodiment uses the equation residuals as a physical law penalty loss, plus the data fidelity loss calculated based on the motion spinor at the center of the platform to characterize the data fitting loss of the deep feedforward neural network, to construct the total loss function of the deep feedforward neural network. To balance the strength of data fitting and physical constraints, and to prevent the network from overfitting the data and violating physical safety, or rigidly adhering to physical constraints without tracking the input target signal, an adaptive penalty weight coefficient is introduced into the total loss function. This adaptive penalty weight coefficient effectively avoids the network overfitting the target trajectory noise during training, thus violating the underlying mechanical laws, or excessively satisfying the physical equations, leading to increased trajectory tracking errors. The specific expression of the constructed total loss function is as follows:

[0047] ; ; in, To compensate for data fidelity loss, To utilize the aforementioned mapping model, the motion spinor of the table center is obtained by converting the theoretical decoupled driving parameters (joint displacement and joint velocity) predicted by the deep feedforward neural network. The actual motion screw corresponding to the physical trajectory of the platform measured by the sensor or the spatial target spectrum set externally can be constrained to ensure the waveform reproduction of the test platform. The loss is penalized by physical laws (which is essentially the unbalanced quantity of the equations generated by substituting the parameters predicted by the neural network and the physical excitation force vector into the dynamic equations, and using this as a physical prior constraint without real data labels). The adaptive penalty weight coefficient is typically set between 0.1 and 100, with a preferred value of 10.

[0048] The adaptive penalty weight coefficient can also be dynamically updated and adjusted during network iteration based on the changes in the residual gradient of the equation. The specific adjustment process is as follows: evaluate the gradient of the physical law penalty loss during each backpropagation. If the gradient descent is slow or tends to stagnate, increase the adaptive weight coefficient by a set step size to strengthen the physical constraint. If the gradient descent is too fast and the data fidelity loss is difficult to converge, decrease the value of the weight coefficient accordingly.

[0049] , These are the modal displacements and modal accelerations obtained by the deep feedforward neural network simultaneously predicting (or combining with an automatic differentiation mechanism) the joint displacement vectors and joint velocity vectors, based on the multi-task branch structure within the network. The theoretically decoupled driving parameters can include joint displacement vectors and joint velocity vectors. Let L2 be the norm, representing the residual energy of the unbalanced equation.

[0050] By utilizing the aforementioned total loss function, even if unknown disturbances affect the data fidelity of the deep feedforward neural network, the network weights can be iteratively optimized by penalizing predictions that do not conform to dynamic laws through physical loss. This ensures that the theoretical decoupling driving parameters inferred by the deep feedforward neural network are strictly locked within a safe physical boundary constituted by dynamic partial differential equations, fundamentally guaranteeing the operational safety of expensive test pieces and equipment.

[0051] Step 1: The multi-degree-of-freedom test bench state data collected in real time by the sensor is sent into the deep feedforward neural network constructed above. After forward inference calculation, the neural network outputs the theoretical decoupling driving parameters. Step 2: Using the physical excitation force conversion model of the electro-hydraulic servo system constructed in Example 1, the joint displacement vector and joint velocity vector in the theoretical decoupled drive parameters are converted into the physical excitation force vector output by the actuator, and the motion screw of the center of the table is calculated by the mapping model based on the joint displacement vector and joint velocity vector. Step 3: Substitute the modal displacement and modal acceleration into the left side of the fractional rigid-flexible coupling dynamic equation to calculate the predicted excitation force. Calculate the residual of the above fractional rigid-flexible coupling dynamic equation under the current theoretical decoupling driving parameters based on the predicted excitation force and the physical excitation force vector.

[0052] Step 4: Multiply the residual of the equation by the dynamic adaptive penalty weight, and use this as a physical penalty term based on the prior physical laws. This is then hard-embedded into the total loss function of the deep feedforward neural network given in Example 1. During the training phase, the algorithm performs backpropagation based on this total loss function to update the neural network parameters, ultimately enabling the network to output optimal decoupling driving parameters that satisfy the constraints of the physical laws.

[0053] Example 2

[0054] Example 2 is largely the same as Example 1, except that the control method further includes: soft limiting of the decoupling drive parameters based on the feedback state of the bottom sensor of the multi-degree-of-freedom test bench.

[0055] Specifically, to prevent the black-box model from outputting erroneous commands that cause acceleration or displacement to exceed limits when encountering extreme real-vehicle road spectra (without visible spectra), thereby avoiding damage to the expensive BSR fixture or test bench itself, after obtaining the current decoupling drive parameters of the multi-degree-of-freedom test bench through a physical information neural network, the decoupling drive parameters can also be soft-limited by using the feedback status of the displacement and acceleration sensors at the bottom layer of the multi-degree-of-freedom test bench.

[0056] Specifically, the amplitude of the current decoupled drive parameters is compared in real time with the set hardware safety operating threshold. This hardware safety operating threshold can usually be obtained by consulting the manufacturer's technical specifications of the actuators at the bottom layer of the multi-degree-of-freedom test bench (e.g., the actuator's maximum physical mechanical stroke, maximum permissible thrust, or maximum safe acceleration limit). If the command is within the hardware safety operating threshold range, the output is retained as is. If the command exceeds the hardware safety operating threshold range, a smoothing saturation function (e.g., the sigmoid function or hyperbolic tangent function) is used to nonlinearly and smoothly decay the excess portion, gradually limiting it within the safety boundary. The limited decoupled drive parameters then drive each actuator to perform the corresponding action.

[0057] A computer program product includes a computer program / instructions that, when executed by a processor, implement the steps of the decoupling control method as described in Embodiment 1 or Embodiment 2.

[0058] A multi-degree-of-freedom test bench includes a central hardware control unit, which uses a decoupling control method as described in Embodiment 1 or Embodiment 2 to determine decoupling drive parameters.

[0059] The effectiveness and technical advantages of the fractional-order adaptive decoupling method described in this invention are demonstrated by applying it to a multi-degree-of-freedom vehicle-mounted abnormal noise test bench. This bench includes a multi-axis high-performance excitation actuation assembly and high-frequency joint hinges and elastic vibration-damping rubber connectors vector-connected to it. The experiment was conducted in a central hardware control unit equipped with a GPU acceleration chip, and quantitative evaluation was performed under three typical operating conditions: 1. Geometric boundary limit composite pose condition (used to verify the full rank and singularity of Lie groups) The control system provides a set of multi-axis composite motion trajectory commands that approach the physical limits of the test bench hardware, driving the worktable to synchronously perform multi-axis large-range composite deflections. At this time, the dynamic displacement outputs of the multiple actuators are continuously and synchronously approaching the limit positions of the test bench's own designed mechanical travel boundaries.

[0060] Compared with existing technology 1 (traditional Euler angles + geometric Jacobian decoupling): under extreme composite attitudes where multiple axial displacements simultaneously reach the travel boundary, the Euler angle description matrix degenerates and infinitely approaches its mathematical singularity region, causing the condition number of the geometric Jacobian matrix of the system to increase sharply to above 100%. This causes severe numerical divergence in the decoupling algorithm, resulting in the control system outputting non-physical sudden change commands, directly triggering hardware overcurrent and over-limit protection shutdown, and causing waveform distortion and breakage during testing.

[0061] Applying the technical solution of this invention: Because it employs Lie group manifolds and adjoint operators at the underlying level, it does not rely on Euler angles. Even in extreme composite poses where the strokes of each actuator approach their own physical extrema, the kinematic description maintains non-singularity characteristics throughout space. The adjoint operator and the system Jacobian matrix condition number remain stably maintained throughout the entire stroke range. In the extremely low good-state range, the exciter drive command is continuous and smooth, completely eliminating non-physical jumps and achieving smooth control under extreme mechanical boundaries.

[0062] 2. Wideband (50Hz-200Hz) mid-to-high frequency abnormal noise (BSR) test condition (used to verify the fractional-order dynamic decoupling accuracy) Load a typical automotive dashboard tooling fixture onto the table, given... The multi-axis high-frequency random excitation spectrum was used to assess its inter-axis parasitic crosstalk index. The first three flexible modal parameters of the test bench and fixture were extracted (the first modal resonance frequency is approximately 100Hz). In this invention, the fractional calculus order is set as... (with appendix) Figure 2 (Simulation parameters are consistent). Table 1 records the ratio of the parasitic vibration acceleration response amplitude in the undesired axis to the amplitude in the main excitation direction at different central excitation frequencies (inter-axis crosstalk, unit: dB): Table 1. Quantitative comparison (dB) of interaxial parasitic crosstalk of different decoupling methods in the mid-to-high frequency band.

[0063] From the quantitative data in Table 1, combined with the appendix Figure 2 It is known that near the 100Hz modal resonance region, existing technology 2 (traditional geometric rigid body Jacobian matrix) suffers from a sharp amplification of crosstalk to +2 dB due to the failure of the rigid body assumption; existing technology 3 (traditional integer-order constant damping model) provides some relief, but due to neglecting the "historical memory effect" of high-frequency viscoelastic damping, it produces a significant phase compensation deviation, with crosstalk only reaching -8 dB. The present invention's solution precisely cancels out the high-frequency phase difference and hysteresis damping response through fractional derivatives. The interaxial parasitic crosstalk is stably controlled below -21dB across the entire frequency band, and is reduced by a full 20dB at the resonance peak compared to integer order technology, which greatly improves the decoupling accuracy of the high-frequency flexible dominant region.

[0064] 3. Long-term continuous testing and blind testing without visible spectrum (used to verify adaptive evolution and physical absolute safety) The test bench was run continuously for 12 hours to simulate hysteretic nonlinear drift caused by mechanical wear and heat. Subsequently, without recalibration, a set of high-peak, non-stationary extreme real vehicle road spectra (spectrals not seen) not included in the system's historical training set was suddenly input.

[0065] Compared with existing technology 4 (pure data-driven traditional LSTM neural network): when faced with mechanical parameter heating drift and unseen burst spectrum, due to the lack of physical law constraints in the black box network, a set of "flying car" excitation force commands that do not conform to energy conservation are output during forward inference. The actuator's instantaneous acceleration reaches 150% of the maximum limit, directly causing the test piece to overload and be damaged.

[0066] By applying the technical solution of this invention, thanks to the hard-coded physical law penalty loss in the adaptive loss function, even when the network has no spectrum seen during forward inference, its output command is still strictly locked within the safe physical boundary formed by the dynamic partial differential equations, even if there are unknown perturbations in the data fidelity loss. After 12 hours of thermal drift, the system still controls the mean square error of target spectrum trajectory tracking within 3.2% through "learning by testing," and no risk of exceeding the limit occurs. This quantitatively proves that the invention achieves high-precision online adaptation while possessing absolute physical operational safety.

[0067] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A decoupling control method, characterized in that, include: A mapping model between the joint space velocity of the actuator and the table space velocity is established by using a three-dimensional special Euclidean group and a geometric Jacobian matrix. The physical feedback state and target spectrum of the multi-degree-of-freedom test bench are acquired in real time. The physical feedback state is the physical feedback state output by the sensors configured on the multi-degree-of-freedom test bench, including the real-time displacement, velocity and acceleration of each actuator. The target spectrum is the excitation signal that the multi-degree-of-freedom test bench is expected to reproduce, including the acceleration time history or power spectral density spectrum collected through actual vehicle road tests. The physical feedback state and the target spectrum are input into the deep feedforward neural network in the physical information neural network to infer the theoretical decoupling driving parameters, as well as the modal displacement and modal acceleration of the multi-degree-of-freedom test bench. The theoretical decoupling driving parameters include the joint displacement vector and joint velocity vector of each actuator of the multi-degree-of-freedom test bench. The theoretical coupled dynamic response of the multi-degree-of-freedom test bench under the control of theoretical decoupling driving parameters is determined by the mapping model and the physical excitation force conversion model. The theoretical coupled dynamic response includes the physical excitation force vector and the motion screw of the table center. The theoretical decoupling driving parameters inferred by the deep feedforward neural network are substituted into the mapping model to calculate the motion screw of the multi-degree-of-freedom test bench table center. The theoretical decoupling driving parameters inferred by the deep feedforward neural network are substituted into the physical excitation force conversion model to calculate the physical excitation force vector to be applied to the table. Based on the physical excitation force vector in the theoretically coupled dynamic response, combined with the modal displacement and modal acceleration, the equation residual of the differential dynamic equation is calculated, and the data fidelity loss of the deep feedforward neural network is calculated based on the motion spinor of the center of the platform in the theoretically coupled dynamic response. The equation residual is used as the physical law penalty loss, and the total loss function is established by combining the data fidelity loss. The backpropagation algorithm is used to iteratively optimize the network weights of the deep feedforward neural network so that the value of the total loss function is minimized. Based on the current physical feedback state and target spectrum of the multi-degree-of-freedom test bench, the decoupling driving parameters of the multi-degree-of-freedom test bench are predicted by an optimized physical information neural network.

2. The decoupling control method according to claim 1, characterized in that, Establish a mapping model between the joint space velocity of the actuator and the platform space velocity, including: The motion screw of each actuator is determined based on the corresponding joint displacement vector, and the motion screws of each actuator are stacked into the geometric Jacobian matrix.

3. The decoupling control method according to claim 2, characterized in that, The specific expression of the mapping model is as follows: ; in, The kinematic spinor at the center of the table. Let be the pose transformation matrix at the center of the platform. Li Qun's position The corresponding adjoint operator, Let be the joint displacement vector of the actuator. This is the joint velocity vector of the actuator. It is a geometric Jacobian matrix. This represents the number of actuators.

4. The decoupling control method according to claim 1, characterized in that, Calculate the residuals of the differential dynamic equations, including: Based on the flexible modal parameters of the multi-degree-of-freedom test rig, the fractional-order rigid-flexible coupling dynamic equation is established by introducing the Caputo fractional derivative, and the fractional-order rigid-flexible coupling dynamic equation is used as the differential dynamic equation.

5. The decoupling control method according to claim 4, characterized in that, The specific expression for the fractional-order rigid-flexible coupling dynamic equation is as follows: ; in, Here is the modal mass matrix. The second derivative of the modal coordinates. Here is the modal damping matrix. For Caputo fractional differential operators, For flexible modal coordinate vectors, Here is the modal stiffness matrix. The modal elastic restoring force vector, For the mode shape matrix, The physical excitation force vector. This is the order of fractional calculus.

6. The decoupling control method according to claim 1, characterized in that: The total loss function incorporates an adaptive penalty weight coefficient to balance the strength of the data fit with the physical constraints.

7. The decoupling control method according to claim 1, characterized in that, Also includes: The decoupled driving parameters are softly limited based on the feedback status of the bottom-level sensors of the multi-degree-of-freedom test bench.

8. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the decoupling control method as described in any one of claims 1-7.

9. A multi-degree-of-freedom test bench, comprising a central hardware control unit, characterized in that: The central hardware control unit uses the decoupling control method as described in any one of claims 1-7 to determine the decoupling drive parameters.

Citation Information

Patent Citations

  • Multi-rotor unmanned aerial vehicle adaptive attitude stability control algorithm and device

    CN120523039A

  • Industrial robot joint vibration solving method based on current information and physical information neural network

    CN121670639A