A method and system for nonlinear decoupling of a triaxial force sensor

Abstract

The present application relates to the technical field of force sensor, disclose a kind of nonlinear decoupling method and system of triaxial force sensor, comprising: acquisition multi-dimensional original data, and carry out adaptive noise reduction, dynamically select wavelet base, construct filter matrix and obtain denoising signal;Dynamic PINN proxy model is constructed, material parameters and environmental variables are fused, and stiffness matrix is output, online parameter updating mechanism is established, and the optimized parameter calculated is constructed enhanced dynamic compensation matrix;The output stiffness matrix and enhanced dynamic compensation matrix are fused and the main singular value is reserved;Reconstruction obtains reconstructed strain field, establishes multi-physical field coupling model compensation quantum-thermal coupling error, and compensates the lead inductance coupling under high frequency;Multi-scale graph structure is constructed, loss function is reinforced by physical constraint, crystal plasticity hysteresis compensation is carried out, multi-scale features are fused, and the decoupling force value is calculated;Establish intelligent calibration decision, predict drift trend and dynamically adjust calibration cycle.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of force sensors, in particular to a nonlinear decoupling method and system of a three-axis force sensor. BACKGROUND

[0002] The three-axis force sensor is a precision instrument for measuring forces in two orthogonal directions simultaneously, and is widely used in industrial automation, robotics, medical devices and fitness equipment fields. When an external force acts on the elastic body, the strain gauge generates resistance changes with elastic deformation, and the small resistance changes are converted into voltage signals through the bridge circuit. Some high-end products use piezoelectric principle to realize dynamic force measurement through the positive piezoelectric effect of the piezoelectric quartz crystal wafer, and combine with the charge amplifier to improve the signal stability.

[0003] In the prior art, the nonlinear problems in a specific range are handled based on a linear model or a simple polynomial expansion. In high load or complex working conditions, high-order nonlinear effects will still cause significant errors. The traditional strain gauge layout or bridge circuit can only partially suppress cross interference through the hardware structure, and the coupling error is large under the action of dynamic force. The existing decoupling method relies on a large amount of static calibration data and needs offline training, resulting in long calibration period and high cost. The parameter drift of the sensor after long-term use causes the decoupling model to fail, and the existing technology lacks online self-calibration capability. The dynamic response bandwidth of the sensor required by the automobile power assembly test is usually higher than that of the traditional strain sensor, and there is phase distortion. The hardware decoupling scheme often sacrifices sensitivity or structural strength as a trade-off. The shaft pin type sensor is slotted on the surface to enhance the stress, while weakening the overload capacity. The MEMS sensor realizes miniaturization, but the decoupling structure is limited by the processing technology, and the lateral stiffness is insufficient.

[0004] Therefore, it is necessary to provide a nonlinear decoupling method and system of a three-axis force sensor. SUMMARY

[0005] The purpose of the present application is to provide a nonlinear decoupling method and system of a three-axis force sensor. To solve the above-mentioned problems in the prior art, the present application realizes the following technical solutions:

[0006] In a first aspect, the present application provides a nonlinear decoupling method of a three-axis force sensor, which specifically includes the following steps:

[0007] Step one: collect multi-dimensional original data, perform adaptive noise reduction through space-time correlation, dynamically select a wavelet basis, introduce a space-time correlation constraint, construct a filter matrix to calculate a noise reduction signal;

[0008] Step two: based on the obtained noise reduction signal, a dynamic PINN proxy model is constructed to fuse material parameters and environmental parameters, output a stiffness matrix, establish an online parameter updating mechanism, calculate the obtained optimization parameters, and construct an enhanced dynamic compensation matrix; the output stiffness matrix and the enhanced dynamic compensation matrix are fused to retain the main singular value to solve the matrix ill-conditioning problem;

[0009] Step three: based on the dynamic PINN proxy model, a super-resolution strain field is reconstructed to obtain a reconstructed strain field, a multi-physical field coupling model is established to compensate quantum-thermal coupling errors, and a wire inductance coupling under high frequency is compensated;

[0010] Step four: based on the reconstructed strain field, a multi-scale graph structure is constructed, a loss function is strengthened through physical constraints, a crystal plasticity hysteresis compensation is performed, multi-scale features are fused, and a decoupling force value is calculated and output;

[0011] Step five: based on the obtained decoupling force value, an intelligent calibration decision is established, a drift trend is predicted, and a calibration cycle is dynamically adjusted.

[0012] Further, the method for constructing the filter matrix is:

[0013] Synchronously collect multi-modal data, and synchronously obtain multi-dimensional original data based on a preset sampling rate;

[0014] Based on the obtained multi-dimensional original data, self-adaptive noise reduction is performed through space-time correlation;

[0015] Real-time detection of signal dominant frequency through FFT , dynamically select wavelet basis, obtain optimized wavelet basis , dynamically switch according to signal dominant frequency

[0016] Based on the obtained optimized wavelet basis, introduce space-time correlation constraints, construct a filter matrix, and calculate the noise reduction signal through the formula: , wherein, is the optimized wavelet basis, is the input original signal, is a spatial gradient operator, is a space-time correlation factor, is an infrared temperature field matrix, is a strain field matrix, is a matrix direct product;

[0017] Further, the method for outputting the stiffness matrix is:

[0018] Construct a dynamic PINN proxy model, design a 4-layer 128-node neural network, fuse material parameters and environmental parameters, and output a stiffness matrix :​​

[0019] ;

[0020] in, For neural network modules, For Young's modulus, Poisson's ratio, For geometric tolerance, The network training parameters are obtained through pre-training with entity labeling data. Humidity coupling factor, For temperature coupling factor, For ambient humidity, The ambient temperature;

[0021] Furthermore, the method for establishing the online parameter update mechanism is as follows:

[0022] An online parameter update mechanism is established through the dynamic PINN proxy model;

[0023] The model is pre-tuned every 1000 sets of measured data, optimizing parameters through gradient descent, using the formula: Calculate the optimization parameters ,in, For historical model parameters, For learning rate, The loss function with respect to the model parameters gradient vector, The L2 norm squared error between the model's predicted and measured values. To output the stiffness matrix, For the measured response signal, For standard force input;

[0024] Furthermore, the method for solving the matrix ill-conditioned problem is as follows:

[0025] Construct an enhanced dynamic compensation matrix;

[0026] Through enhanced dynamic compensation matrix:

[0027] ;

[0028] Add a coupling term between the humidity change rate and the temperature gradient, where, The vibration attenuation coefficient is... For temperature gradient, , and These represent the accelerations of the beam along the X-axis, Y-axis, and Z-axis, respectively. Let be the derivative of ambient humidity with respect to time, i.e., the rate of change of humidity. , , is a humidity coupling coefficient for humidity change rate, , , is a vibration coupling coefficient, , , is a temperature coupling coefficient;

[0029] fusion matrix of the output stiffness matrix and the enhanced dynamic compensation matrix truncated singular value decomposition (TSVD) is performed to retain the first 6 principal singular values wherein, is a left singular matrix, is a right singular matrix, is a principal singular value diagonal matrix;

[0030] Further, the method for performing super-resolution strain field reconstruction is:

[0031] ultrasonic surface waves are excited, and the sound wave flight time is collected through 4 acoustic emission sensors; combining the piezoelectric impedance and the sound wave flight time, the strain field is reconstructed by using the time reversal reconstruction operator to improve the spatial resolution, and the reconstructed strain field is obtained by the formula: wherein, is a time reversal reconstruction operator, is a piezoelectric impedance, is a sound wave flight time, is a Gaussian smoothing kernel;

[0032] Compensate for quantum-thermal coupling errors: Establish a multi-physical field coupling model of nanosilver wire resistance :

[0033] ;

[0034] wherein, is a resistance temperature coefficient, is a temperature gradient, is a quantum tunneling coefficient, is a silver wire spacing change, is a strain absolute value, is a thermal-hygro coupling coefficient, is a triple integral item representing the spatial coupling effect of strain, temperature and humidity;

[0035] Further, the method for compensating for wire inductance coupling at high frequencies is:

[0036] Compensate for wire inductance coupling at high frequencies. For the X-axis beam, the compensation voltage of the X-axis beam is calculated by the formula: wherein,​​ is the original voltage of the X-axis beam, is the kth adjacent channel distributed capacitance, is the time step, is the kth adjacent channel distributed inductance, is the kth adjacent channel distributed current, is the adjacent channel index, while compensating the wire inductance coupling under high frequency for the Y-axis beam and the Z-axis beam;

[0037] Further, the method for performing crystal plasticity hysteresis compensation is:

[0038] A multi-scale graph structure is constructed, and the multi-scale graph structure includes: an atomic layer, a micro layer, and a macro layer;

[0039] The loss function is reinforced by physical constraints, and the curl, divergence, and causal constraints are fused, through the formula:

[0040] ;

[0041] The reinforced loss function is obtained , wherein is the physical constraint, is the causal constraint, is the data loss;

[0042] Performing crystal plasticity hysteresis compensation: introducing a fractional order hysteresis operator based on dislocation dynamics, through the formula: The hysteresis compensation strain force is calculated , wherein is the shear modulus, is the Burgers vector, is the dislocation density, is the fractional order, is the fractional order derivative of the strain force, is the strain force;

[0043] Further, the method for calculating the output decoupling force value is:

[0044] Multi-scale features are fused through an 8-head graph attention network GAT, and finally the decoupling force value is output, through the formula: The decoupling force value is calculated , wherein is the multi-scale feature, are respectively the node feature vectors of the atomic layer, the micro layer, and the macro layer, is the edge feature from the atomic layer to the micro layer, is the edge feature from the micro layer to the macro layer, is the hysteresis compensation strain force, is the strain force, is the inertial force.

[0045] In a second aspect, the embodiment of the present application provides a nonlinear decoupling system of a three-axis force sensor, which specifically comprises the following modules:

[0046] A data acquisition module: acquires multi-dimensional original data, performs adaptive noise reduction through space-time correlation, dynamically selects a wavelet basis, introduces a space-time correlation constraint, constructs a filter matrix to obtain a denoised signal;

[0047] A matrix enhancement module: based on the obtained denoised signal, a dynamic PINN proxy model is constructed, material parameters and environmental parameters are fused, a stiffness matrix is output, an online parameter updating mechanism is established, the obtained optimized parameters are calculated, and an enhanced dynamic compensation matrix is constructed; the output stiffness matrix and the enhanced dynamic compensation matrix are fused to retain the main singular values to solve the matrix ill-conditioning problem;

[0048] A reconstruction compensation module: based on the dynamic PINN proxy model, a super-resolution strain field reconstruction is performed to obtain a reconstructed strain field, a multi-physical field coupling model is established to compensate quantum-thermal coupling errors, and a wire inductance coupling under high frequency is compensated;

[0049] A fusion output module: based on the reconstructed strain field, a multi-scale graph structure is constructed, a loss function is reinforced through physical constraints, crystal plasticity hysteresis compensation is performed, multi-scale features are fused, and a decoupling force value is calculated and output;

[0050] An intelligent calibration module: based on the obtained decoupling force value, an intelligent calibration decision is established, a drift trend is predicted and a calibration period is dynamically adjusted.

[0051] The present application has the following advantages:

[0052] 1. Fractal geometry is combined with multi-modal sensing, strain transfer uniformity is optimized through fractal iteration, the time-space coupling characteristics of infrared temperature field and strain field are utilized to improve the signal-to-noise ratio, the limitations of traditional single strain sensing are broken through, multi-parameter collaborative sensing of force, temperature and humidity is realized, rich original information dimensions are provided for decoupling, the bottleneck of traditional static calibration model is broken through, a dynamic physical information neural network proxy model that fuses material parameters and environmental parameters is constructed, an online parameter updating mechanism and an enhanced dynamic compensation matrix are introduced, the model mismatch problem caused by aging and environmental changes of the sensor is solved, and dynamic self-adaptation is realized.

[0053] 2. Construct a three-level multi-scale graph structure, associate material lattice characteristics, micro-strain unit characteristics, and macro-sensing data, combine physical constraints and causal reasoning loss functions, distinguish real correlations from false correlations between physical fields, break through the limitations of traditional macro-modeling that ignore material micro-characteristics, combine atomic-level elastic parameters, crystal plasticity hysteresis compensation, and macro-force decoupling, realize deep fusion of cross-scale physical information, introduce a deep reinforcement learning driven adaptive excitation strategy, dynamically adjust the calibration period by combining LSTM to predict drift trends, break through the passivity of traditional fixed-period calibration, realize active calibration and maintenance cost optimization under all working conditions. BRIEF DESCRIPTION OF DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0055] Figure 1 is a step flow chart of a nonlinear decoupling method of a three-axis force sensor provided by embodiment 1 of the present application;

[0056] Figure 2 is a structural schematic diagram of a nonlinear decoupling system of a three-axis force sensor provided by embodiment 2 of the present application. DETAILED DESCRIPTION

[0057] In order to make the person skilled in the art better understand the present application scheme, the following will combine the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0058] Embodiment 1: as shown in the present application, the nonlinear decoupling method of a three-axis force sensor provided by the embodiment of the present application specifically includes the following steps: Figure 1

[0059] Step one: collect multi-dimensional original data, perform adaptive noise reduction through space-time correlation, dynamically select wavelet basis, introduce space-time correlation constraint, construct filter matrix to obtain denoised signal;

[0060] Fractal micro-channel and multi-modal sensing fusion, through H-shaped tree-shaped bifurcation structure to improve strain transmission uniformity, introduce space-time correlation filtering mechanism, use the space-time coupling characteristics of infrared temperature field and strain field to improve the signal-to-noise ratio of dynamic signal; ​

[0061] In specific embodiments, the main body material of the triaxial force sensor is selected to be an aluminum alloy beam, which has high strength and good elastic recovery performance, ensuring good deformation linearity within the preset load range;

[0062] Adopt The ceramic isolation layer is constructed by ceramics, which effectively blocks the transmission of thermal strain and reduces the influence of temperature change on measurement;

[0063] Based on The ceramic isolation layer constructed by ceramics adopts a copper foil with a preset thickness including the ceramic substrate where the ceramic isolation layer is located, forming an electromagnetic shielding barrier to attenuate the intensity of external electromagnetic interference to within a preset range;

[0064] The geometric structure of the triaxial force sensor is designed, which includes an X-axis beam, a Y-axis beam and a Z-axis beam;

[0065] The fractal sensing array is designed, and the elastomer surface is laser etched with H-shaped tree-shaped bifurcated microchannels, with preset main channel width and bifurcated channel width, and fractal iteration 3 times;

[0066] The integration of the fractal sensing array includes 3-axis nano silver wire strain gauges distributed at the bifurcation nodes, 8-point thin film platinum resistors and micro infrared thermoelectric array distributed along the depth direction of the flow channel, realizing two-dimensional distribution monitoring of temperature field;

[0067] Synchronous acquisition of multi-modal data is performed, and light pulse triggering sampling is adopted to synchronously obtain multi-dimensional raw data based on a preset sampling rate , which includes infrared temperature field matrix , strain gauge output voltage , three-axis acceleration , environmental humidity and 8-point thin film platinum resistance temperature ;

[0068] Based on the obtained multi-dimensional raw data, adaptive noise reduction is performed through space-time correlation;

[0069] Specifically, the dominant frequency of the signal is detected in real time through FFT , and the wavelet basis is dynamically selected through the formula:

[0070] ;

[0071] The optimized wavelet basis is calculated , and is dynamically switched according to the dominant frequency of the signal for subsequent wavelet transform to reduce noise of the original signal, wherein is the Daubechies 8-order wavelet, which is more optimal for time-frequency decomposition of high-frequency signals, Symlets 6 order wavelet, balance the smoothness of low frequency signal and orthogonality;

[0072] Based on the obtained optimal wavelet basis, introduce the space-time correlation constraint, construct the filter matrix, and through the formula:

[0073]

[0074] The denoising signal is calculated , wherein, is the optimal wavelet basis, is the input original signal, is the spatial gradient operator, is the space-time correlation factor, is the infrared temperature field matrix, is the strain field matrix, is the matrix direct product, which further suppresses noise by using the spatial correlation of temperature field and strain field;

[0075] Step two: based on the obtained denoising signal, construct a dynamic PINN proxy model, fuse material parameters and environmental parameters, output the stiffness matrix, establish an online parameter updating mechanism, calculate the optimized parameters, construct an enhanced dynamic compensation matrix; fuse the output stiffness matrix and the enhanced dynamic compensation matrix to retain the main singular value to solve the matrix ill-conditioning problem;

[0076] The physical information neural network PINN proxy model is combined with online multi-dimensional original data backflow to realize dynamic parameter real-time updating, solve the problem of insufficient environmental adaptability of traditional static calibration model; introduce humidity-temperature-force three-dimensional coupling term, construct full working condition calibration matrix, and improve the decoupling precision under complex environment;

[0077] In specific embodiments, a dynamic PINN proxy model is constructed, a 4-layer 128-node neural network is designed, material parameters and environmental parameters are fused, and a stiffness matrix is outputted.

[0078]

[0079] , wherein, is the neural network module, is the Young's modulus, is the Poisson's ratio, is the geometric tolerance, is the network training parameter, which is obtained by pre-training with entity calibration data, is the humidity coupling factor, is the temperature coupling factor, is the environmental humidity, is the environmental temperature.

[0080] ​​It should be noted that the material parameters include: Young's modulus, Poisson's ratio and geometric tolerance; environmental parameters include: environmental humidity and environmental temperature;

[0081] Based on the dynamic PINN proxy model, an online parameter updating mechanism is established; The preset model fine-tuning is triggered once every 1000 groups of measured data, the parameters are optimized through gradient descent, and the formula is:

[0083] ;

[0084] The optimized parameters are calculated , wherein is the historical model parameter, is the learning rate, is the gradient vector of the loss function with respect to the model parameter , is the L2 norm square error of the model predicted value and the measured value, is the output stiffness matrix, is the measured response signal, is the standard force input, and the model is dynamically adapted to the sensor aging;

[0085] Based on the calculated optimized parameters, an enhanced dynamic compensation matrix is constructed;

[0086] Through the enhanced dynamic compensation matrix:

[0087] ;

[0088] Coupling terms of humidity change rate and temperature gradient are supplemented, wherein is the vibration attenuation coefficient, and the preset value is 0.02, is the temperature gradient, , and are the accelerations of the X-axis beam, the Y-axis beam and the Z-axis beam, is the derivative of the environmental humidity with respect to time, i.e. the humidity change rate, , , is the humidity change rate humidity coupling coefficient, , , is the vibration coupling coefficient, , , is the temperature coupling coefficient;

[0089] The fusion matrix of the output stiffness matrix and the enhanced dynamic compensation matrix is subjected to truncated singular value decomposition TSVD, and the first 6 principal singular values are retained​ :

[0090] ;

[0091] wherein, is a left singular matrix, is a right singular matrix, is a main singular value diagonal matrix, solving the matrix ill-posed problem;

[0092] Step three: based on the dynamic PINN proxy model, the super-resolution strain field is reconstructed to obtain the reconstructed strain field, a multi-physical field coupling model is established to compensate the quantum-thermal coupling error, and the lead inductance coupling under high frequency is compensated;

[0093] In specific implementation, the strain field super-resolution reconstruction of the acoustic emission-impedance tomography fusion combines the time reversal algorithm to improve the spatial resolution; the quantum tunneling-thermal coupling compensation model reduces the nonlinear error of the nano silver wire strain gauge under micro strain;

[0094] Super-resolution strain field reconstruction: excite ultrasonic surface waves, and collect the acoustic wave flight time through four acoustic emission sensors; combine the piezoelectric impedance and the acoustic wave flight time, use the time reversal reconstruction operator to reconstruct the strain field to improve the spatial resolution, through the formula:

[0095] ;

[0096] obtain the reconstructed strain field wherein, is a time reversal reconstruction operator, is a piezoelectric impedance, is an acoustic wave flight time, is a Gaussian smoothing kernel, and the standard deviation The preset value is 0.05;

[0097] Compensate the quantum-thermal coupling error:

[0098] Establish a multi-physical field coupling model of the nano silver wire resistance :

[0099] ;

[0100] wherein, is a resistance temperature coefficient, is a temperature gradient, is a quantum tunneling coefficient, is a silver wire spacing change, is a strain absolute value, is a thermal-hygro coupling coefficient, is a triple integral item representing the spatial coupling effect of strain, temperature and humidity;

[0101] Compensate the wire inductance coupling at high frequency, eliminate the wire cross interference:

[0102] Based on the transmission line theory, compensate the wire inductance coupling at high frequency, for the X-axis beam, through the formula:

[0103]

[0104] The compensation voltage of the X-axis beam is calculated , wherein is the original voltage of the X-axis beam, is the distributed capacitance of the kth adjacent channel, is the time step, is the distributed inductance of the kth adjacent channel, is the distributed current of the kth adjacent channel, is the adjacent channel index, and the Y-axis beam and the Z-axis beam are compensated for the wire inductance coupling at high frequency;

[0105] Step four: based on the reconstructed strain field, construct a multi-scale graph structure, strengthen the loss function through physical constraints, perform crystal plasticity hysteresis compensation, fuse multi-scale features, and calculate the output decoupling force value;

[0106] The multi-scale graph attention network MS-GAT fuses atomic-micro-macro three-level structure information, solves the problem of ignoring material micro characteristics in traditional graph networks, introduces a causal reasoning loss function to distinguish the causal correlation and false correlation between physical fields, and improves the robustness of the decoupling model;

[0107] A multi-scale graph structure is constructed, which includes an atomic layer, a micro layer and a macro layer;

[0108] Atomic layer: taking the lattice nodes of the elastomer material as units, the feature vector , is the atomic density, is the atomic elastic modulus, is the Poisson's ratio;

[0109] Micro layer: taking the micro strain unit as the node, the feature vector , wherein is the micro strain force, is the strain gradient, is the micro temperature;

[0110] Macro layer: taking the physical node of the sensing array as a unit, the feature vector , wherein is the macro strain force, is the infrared temperature field matrix, is the environmental humidity, and the interlayer edge feature is defined as a scale conversion coefficient ​, the preset value of atom-microscopic is , the preset value of micro-macroscopic is ;

[0111] The loss function is reinforced by physical constraints, and the curl, divergence and causal constraints are fused, through the formula:

[0112] ;

[0113] The reinforced loss function is obtained , wherein is the physical constraint, is the curl of the strain force, is the strain force, is the theoretical curl, is the divergence of the Cauchy stress tensor, is the Cauchy stress tensor, is the external force vector; is the causal constraint, is the KL divergence function, is the probability distribution function, is the strain force, is the ambient temperature, is the ambient humidity, the conditional independence of force and environmental parameters is measured by KL divergence, is the data loss, is the predicted strain force, is the measured strain force, is the influence coefficient, and the preset value is 10;

[0114] Crystal plasticity hysteresis compensation is performed:

[0115] A fractional order hysteresis operator based on dislocation dynamics is introduced, and the formula is:

[0116] ;

[0117] The hysteresis compensation strain force is calculated , wherein is the shear modulus, is the Burgers vector, is the dislocation density, which is estimated in real time by the PINN model, is the fractional order, is the fractional order derivative of the strain force, is the strain force;

[0118] Decoupling calculation process:

[0119] Multi-scale features are fused through an 8-head graph attention network GAT, and the final output is the decoupling force value, through the formula:

[0120] ;

[0121] The decoupling force value is calculated , wherein, is a multi-scale feature, is a node feature vector of the atomic layer, the micro layer, and the macro layer, respectively, is an edge feature from the atomic layer to the micro layer, is an edge feature from the micro layer to the macro layer, is a hysteresis compensation strain force, is a strain force, is an inertial force;

[0122] Step five: based on the obtained decoupling force value, an intelligent calibration decision is established, the drift trend is predicted, and the calibration period is dynamically adjusted;

[0123] The adaptive excitation strategy driven by deep reinforcement learning DQN combines laser thermal excitation and piezoelectric mechanical excitation to realize full-condition no-dead-angle calibration; the drift trend prediction based on LSTM dynamically adjusts the calibration period, thereby reducing maintenance costs;

[0124] An intelligent calibration decision driven by deep reinforcement learning DQN is established;

[0125] A state space is constructed: ;

[0126] , wherein, is a drift amplitude, is a derivative of the ambient temperature with respect to time, i.e., a temperature change rate, is a signal signal-to-noise ratio, is an ambient humidity;

[0127] An action space is constructed, and the action space includes: step mechanical excitation, sweep mechanical excitation, random mechanical excitation, and laser thermal excitation;

[0128] A reward function is constructed: , wherein, is an excitation energy consumption, and the optimal action is selected through an ε-greedy strategy;

[0129] Based on the obtained action space, the laser thermal excitation is calibrated:

[0130] Local heating is performed by using a fiber laser to generate a controllable thermal strain field :

[0131] ;

[0132] , wherein, is a thermal expansion coefficient, is a laser power, is a spot radius; TC4 thermal conductivity, heating center coordinates, spatial coordinates of thermal strain;

[0133] predicting the drift trend and dynamically adjusting the calibration period;

[0134] Specifically, the LSTM network is trained to predict the drift amount in the next N hours, and the calibration period is dynamically adjusted;

[0135] For example, if the absolute value of the drift amount in the next N hours is less than 0.3% , the calibration period is extended to 1.2 times the current period, otherwise the calibration period remains unchanged.

[0136] Embodiment 2: as Figure 2 shown, the nonlinear decoupling system of the triaxial force sensor provided by the embodiment of the application specifically comprises the following modules:

[0137] Data acquisition module: acquire multi-dimensional original data, perform adaptive noise reduction through space-time correlation, dynamically select wavelet basis, introduce space-time correlation constraint, construct a filter matrix to obtain a denoised signal;

[0138] Matrix enhancement module: based on the obtained denoised signal, construct a dynamic PINN proxy model, fuse material parameters and environmental parameters, output a stiffness matrix, establish an online parameter updating mechanism, calculate the optimized parameters, construct an enhanced dynamic compensation matrix; fuse the output stiffness matrix and the enhanced dynamic compensation matrix to retain the main singular values to solve the matrix ill-conditioning problem;

[0139] Reconstruction compensation module: based on the dynamic PINN proxy model, perform super-resolution strain field reconstruction to obtain a reconstructed strain field, establish a multi-physical field coupling model to compensate quantum-thermal coupling errors, and compensate the lead inductance coupling under high frequency;

[0140] Fusion output module: based on the reconstructed strain field, construct a multi-scale graph structure, strengthen the loss function through physical constraints, perform crystal plasticity hysteresis compensation, fuse multi-scale features, and calculate the output decoupling force value;

[0141] Intelligent calibration module: based on the obtained decoupling force value, establish an intelligent calibration decision, predict the drift trend and dynamically adjust the calibration period.

[0142] The above describes one embodiment of the present application in detail, but the content is only the preferred embodiment of the present application and cannot be considered to limit the scope of the present application; the above formulas are all dimensionless values, and the formulas are obtained by collecting a large amount of data to simulate a formula of the most recent real situation, and the preset parameters in the formula are set by the person skilled in the art according to the actual situation and historical experience, and can be adjusted according to the actual situation; the above is only the preferred embodiment of the present application and cannot be used to limit the present application, and all equivalent changes and improvements made according to the scope of the present application should still belong to the patent coverage range of the present application.

Claims

1. A method of nonlinear decoupling of a triaxial force sensor, characterized by, The method comprises the following steps: Collecting multi-dimensional original data, performing adaptive noise reduction through space-time correlation, dynamically selecting a wavelet base, introducing a space-time correlation constraint, constructing a filter matrix, and calculating a denoised signal; Based on the obtained denoised signal, a dynamic PINN proxy model is constructed, material parameters and environmental parameters are fused, a stiffness matrix is output, an online parameter updating mechanism is established, and the obtained optimized parameters are calculated to construct an enhanced dynamic compensation matrix; The output stiffness matrix and the enhanced dynamic compensation matrix are fused to retain the main singular values to solve the matrix ill-conditioning problem; Based on the dynamic PINN proxy model, a super-resolution strain field is reconstructed to obtain a reconstructed strain field, a multi-physical field coupling model is established to compensate quantum-thermal coupling errors, and wire inductance coupling at high frequencies is compensated; Based on the reconstructed strain field, a multi-scale graph structure is constructed, a loss function is reinforced through physical constraints, crystal plasticity hysteresis compensation is performed, multi-scale features are fused, and decoupling force values are calculated and output; Based on the obtained decoupling force values, an intelligent calibration decision is established, a drift trend is predicted, and a calibration cycle is dynamically adjusted.

2. The method of claim 1, wherein, The method for constructing the filter matrix is: Synchronously collecting multi-modal data, and synchronously obtaining multi-dimensional original data based on a preset sampling rate; Based on the obtained multi-dimensional original data, adaptive noise reduction is performed through space-time correlation; Real-time detection of signal dominant frequency by FFT Dynamic selection of wavelet base to obtain optimized wavelet base Dynamic switching according to signal dominant frequency Dynamic switching; Based on the obtained optimized wavelet basis, the spatiotemporal correlation constraint is introduced, a filtering matrix is constructed, and the formula is: The denoised signal is calculated , wherein is the optimized wavelet basis, is the input original signal, is the spatial gradient operator, is the spatiotemporal correlation factor, is the infrared temperature field matrix, is the strain field matrix, is the matrix direct product.

3. The method of claim 1, wherein, The method for outputting the stiffness matrix is: A dynamic PINN surrogate model is constructed, a 4-layer 128-node neural network is designed, and material parameters and environmental parameters are fused to output the stiffness matrix : ； wherein, is a neural network module, is a Young's modulus, is a Poisson's ratio, is a geometric tolerance, is a network training parameter, pre-trained by entity calibration data, is a humidity coupling factor, is a temperature coupling factor, is an ambient humidity, is an ambient temperature.

4. The method of claim 1, wherein, The method for establishing the online parameter updating mechanism is: An online parameter updating mechanism is established through the dynamic PINN proxy model; The model is pre-tuned every 1000 sets of measured data, optimizing parameters through gradient descent, using the formula: Calculate the optimization parameters ,in, For historical model parameters, For learning rate, The loss function with respect to the model parameters gradient vector, The L2 norm squared error between the model's predicted and measured values. To output the stiffness matrix, For the measured response signal, This is the standard force input.

5. The method of claim 1, wherein, The method for solving the matrix ill-conditioning problem is: An enhanced dynamic compensation matrix is constructed; Through the enhanced dynamic compensation matrix: ； The humidity change rate is coupled with a temperature gradient term, wherein, is a vibration attenuation coefficient, is a temperature gradient, , and are accelerations of the X-axis beam, the Y-axis beam, and the Z-axis beam, respectively, is a derivative of the ambient humidity with respect to time, i.e., a humidity change rate, , , is a humidity change rate humidity coupling coefficient, , , is a vibration coupling coefficient, , , is a temperature coupling coefficient; Fusion matrix of output stiffness matrix and enhanced dynamic compensation matrix Truncated singular value decomposition (TSVD) is performed to retain the first six dominant singular values wherein, is a left singular matrix, is a right singular matrix, is a diagonal matrix of dominant singular values.

6. The method of claim 1, wherein, The method for reconstructing the super-resolution strain field is: The excited ultrasonic surface wave is collected by four acoustic emission sensors in terms of time of flight. The strain field is reconstructed by using a time reversal reconstruction operator to improve the spatial resolution in combination with the piezoelectric impedance and the time of flight, and the reconstructed strain field is obtained by a formula: wherein, is the time of flight, is the time reversal reconstruction operator, is the piezoelectric impedance, is the time of flight, is a Gaussian smoothing kernel. Compensating quantum-thermal coupling error: establish the multi-physical field coupling model of nanometer silver wire resistance : ； wherein, is the temperature coefficient of resistance, is the temperature gradient, is the quantum tunneling coefficient, is the silver wire spacing variation, is the strain absolute value, is the thermal-moisture coupling coefficient, is a triple integral term representing the spatial coupling effect of strain, temperature, and humidity.

7. The method of claim 1, wherein, The method for compensating the wire inductance coupling at high frequencies is: Compensate the wire inductance coupling at high frequency, for X-axis beam, through formula: The compensation voltage of X-axis beam is calculated , wherein is the original voltage of X-axis beam, is the kth adjacent channel distributed capacitance, is the time step, is the kth adjacent channel distributed inductance, is the kth adjacent channel distributed current, is the adjacent channel index, and the wire inductance coupling at high frequency is compensated for Y-axis beam and Z-axis beam.

8. The method of claim 1, wherein, The method for performing crystal plasticity hysteresis compensation is: A multi-scale graph structure is constructed, and the multi-scale graph structure comprises: an atomic layer, a microscopic layer, and a macroscopic layer; A loss function is reinforced through physical constraints, curl, divergence, and causal constraints are fused, and the formula is: ； reinforced loss function wherein, is a physical constraint, is a causal constraint, is a data loss; Crystal plasticity hysteresis compensation: A fractional order hysteresis operator based on dislocation dynamics is introduced by the formula: The hysteresis compensated strain force is calculated as where is the shear modulus, is the Burgers vector, is the dislocation density, is the fractional order, is the fractional order derivative of the strain force, is the strain force.

9. The method of claim 1, wherein, The method for calculating and outputting the decoupling force values is: Through the 8-head graph attention network GAT to fuse multi-scale features, the final output decoupling force value is obtained through the formula: The decoupling force value is calculated , wherein is a multi-scale feature, is a node feature vector of an atomic layer, a microscopic layer, and a macroscopic layer, respectively, is an edge feature from the atomic layer to the microscopic layer, is an edge feature from the microscopic layer to the macroscopic layer, is a hysteresis compensation strain force, is a strain force, is an inertial force.

10. A nonlinear decoupling system of a triaxial force sensor, the system being configured to perform the decoupling method of any one of claims 1 to 9, characterized in that, It comprises: A data acquisition module: collecting multi-dimensional original data, performing adaptive noise reduction through space-time correlation, dynamically selecting a wavelet base, introducing a space-time correlation constraint, constructing a filter matrix, and calculating a denoised signal; A matrix enhancement module: based on the obtained denoised signal, a dynamic PINN proxy model is constructed, material parameters and environmental parameters are fused, a stiffness matrix is output, an online parameter updating mechanism is established, and the obtained optimized parameters are calculated to construct an enhanced dynamic compensation matrix; The output stiffness matrix and the enhanced dynamic compensation matrix are fused to retain the main singular values to solve the matrix ill-conditioning problem; A reconstruction compensation module: based on the dynamic PINN proxy model, a super-resolution strain field is reconstructed to obtain a reconstructed strain field, a multi-physical field coupling model is established to compensate quantum-thermal coupling errors, and wire inductance coupling at high frequencies is compensated; A fusion output module: based on the reconstructed strain field, a multi-scale graph structure is constructed, a loss function is reinforced through physical constraints, crystal plasticity hysteresis compensation is performed, multi-scale features are fused, and decoupling force values are calculated and output; An intelligent calibration module: based on the obtained decoupling force values, an intelligent calibration decision is established, a drift trend is predicted, and a calibration cycle is dynamically adjusted.

Images