Method for improving polishing brightness of jade carving by using multi-component abrasive proportioning

Abstract

The present application provides a kind of method for improving jade carving polishing brightness by multi-component abrasive proportioning, belong to jade carving technical field, the present application is by to the surface of jade carving workpiece high spectral imaging scanning, obtain hardness distribution matrix and texture area clustering chart, initialize the particle swarm of particle swarm-simulation annealing hybrid driving optimization algorithm, diamond micro powder, boron carbide and alumina are mixed according to initial proportioning, complete physical feasibility check by lattice boltzmann method and discrete element method macro-micro two-way coupling simulation mechanism, utilize fiber bragg grating array real-time acquisition wall surface shear force and by the operator network based on multi-modal physical information, rheological tensor is analyzed reversely, confirm low-dimensional control point by artificial intelligence agent model and execute parallel microcosmic simulation verification, finally complete high-brightness jade carving polishing processing, solve the technical problems that multi-component abrasive proportioning cannot be real-time self-adapting optimization according to jade anisotropic hardness field, leading to unstable polishing brightness.

Description

Technical Field

[0001] This invention belongs to the field of jade carving technology, and specifically relates to a method for improving the polishing brightness of jade carvings by using a multi-component abrasive ratio. Background Technology

[0002] Jade polishing is a crucial step in the surface treatment of jade products. Traditional processes rely on operator experience, using single-grain-size diamond powder or alumina abrasives, and employing fixed rotation speeds and suspension viscosities to complete the polishing operation. With the widespread adoption of CNC polishing equipment, some solutions have introduced preset abrasive ratio tables, matching different jade materials by looking up the table, which has become the mainstream method for automated jade polishing. However, jade materials such as jadeite and Hetian jade inherently possess significant mineral composition inhomogeneity, with soft areas and hard impurities coexisting on the same workpiece surface. This causes the required abrasive cutting force, suspension rheological state, and contact stress in different areas to dynamically change with location during the polishing process. In current fixed-ratio schemes, because the ratio parameters cannot perceive the spatial distribution of surface hardness or respond to the real-time drift of the rheological state during polishing, the matching of abrasive particle size and mass fraction remains suboptimal for a long time, ultimately leading to persistent problems such as uneven polishing brightness distribution and local over- or under-cutting. In other words, existing technologies suffer from the technical problem that multi-component abrasive ratios cannot be adaptively optimized in real time according to the anisotropic hardness field of jade, resulting in unstable polishing brightness. Summary of the Invention

[0003] In view of this, the present invention provides a method for improving the polishing brightness of jade carvings by using a multi-component abrasive ratio, which can solve the technical problem in the prior art that the multi-component abrasive ratio cannot be adaptively optimized in real time according to the anisotropic hardness field of jade, resulting in unstable polishing brightness.

[0004] This invention is implemented as follows: This invention provides a method for improving the polishing brightness of jade carvings using a multi-component abrasive formulation, comprising the following steps:

[0005] The surface of the jade carving workpiece is scanned by hyperspectral imaging to obtain the surface hardness distribution matrix and texture region clustering map. The particle swarm-simulated annealing hybrid driven optimization algorithm is then input to complete the particle swarm initialization of the abrasive initial ratio parameters.

[0006] Diamond micro powder, boron carbide and alumina were mixed in the initial ratio. The physical feasibility was verified by the micro-rheological constitutive equation of the macro-micro bidirectional coupling simulation mechanism of lattice Boltzmann method and discrete element method, and infeasible ratios were eliminated.

[0007] A fiber Bragg grating array is used to collect the shear force on the polishing disk wall in real time. Combined with an operator network based on multimodal physical information, the rheological tensor of the polishing zone is analyzed in reverse. The dynamic viscosity parameters of the suspension are dynamically corrected and fed back to the particle swarm-simulated annealing hybrid driving optimization algorithm.

[0008] The real-time rheological tensor, the microhardness matrix of the jade surface, and the evolution data of the micro-depth contour of polishing marks are input into the elastoplastic stress mapping model, and the normal dynamic pressure distribution field of the polishing wheel is output to drive the dynamic adjustment of the abrasive ratio gradient step size.

[0009] After the multi-source heterogeneous high-frequency sensor data is processed by the time stamp alignment network and the wavelet packet adaptive threshold denoising preprocessing architecture, the extracted weak feature vectors are fed back to the simulated annealing temperature decay mechanism of the particle swarm-simulated annealing hybrid driven optimization algorithm, triggering the Metropolis probability to jump out of local extrema.

[0010] After the particle swarm optimization algorithm converges, it outputs the globally optimal abrasive ratio. It then performs precise parallel microscopic simulation verification based on the low-dimensional control points confirmed by the Riemannian manifold embedded dimensionality reduction surrogate model. Finally, it is deployed to the polishing equipment to complete the high-brightness jade carving polishing process.

[0011] Specifically, the hyperspectral imaging scan uses an imaging spectrometer with a working wavelength covering 400–2500 nm to collect the reflection spectrum of the jade carving workpiece pixel by pixel. The mineral composition and microhardness at each location are inverted by the position and intensity of the spectral absorption peaks, forming a hardness distribution matrix with a spatial resolution of 5–20 μm.

[0012] Specifically, the generation of the texture region cluster map is achieved by using a topological manifold learning algorithm to cluster the hardness distribution matrix data. The number of cluster categories is determined by performing a series of clustering experiments on multiple representative jade samples and taking the number of categories corresponding to the maximum contour coefficient.

[0013] Specifically, the particle swarm-simulated annealing hybrid-driven optimization algorithm initializes the particle swarm using the particle size range, mass fraction range, suspension stirring speed range, and chemical additive concentration of the multi-component abrasive as particle position vectors. In each iteration, the particles update their speed and position based on their own historical best position and the group's historical best position. The updated ratio is then substituted into the constitutive equation of the hybrid abrasive suspension based on microrheology for fitness evaluation.

[0014] Specifically, the simulated annealing temperature decay mechanism is triggered when the fitness of the particle swarm does not significantly improve in consecutive iterations, with a Metropolis probability. Accepting inferior solutions, the temperature decreases logarithmically. The temperature is gradually reduced, and the algorithm terminates when the temperature falls below the termination temperature threshold.

[0015] Specifically, the macro-micro bidirectional coupling simulation mechanism of the lattice Boltzmann method and the discrete element method is implemented by solving the velocity and pressure fields of the polishing slurry liquid phase at the macro level using the lattice Boltzmann method, and tracking the collision, rolling and scratch trajectories of individual abrasive particles at the micro level using the discrete element method. The two layers are coupled through bidirectional data exchange of fluid resistance and particle reaction force.

[0016] Specifically, the fiber Bragg grating array consists of multiple fiber Bragg grating sensing elements arranged in an array on the edge wall of the polishing disk. The local strain of the wall is calculated by detecting the drift of the Bragg wavelength, and then the distribution of shear force on the wall is inverted.

[0017] Specifically, the operator network based on multimodal physical information uses wall pressure gradient data as input and the residuals of the Navier-Stokes equations embedded in the physical information neural network framework as physical constraint loss to reverse analyze the rheological tensor of the polished zone center that cannot be directly measured.

[0018] Specifically, the elastoplastic stress mapping model uses Fourier neural operators as the backbone structure of the network as a whole. Each Fourier operator layer embeds elastoplastic time-varying operator units. The elastoplastic time-varying operator units are composed of a von Mises yield criterion determination module and a kinematic hardening constitutive update module connected in series. The output of each Fourier operator layer is connected to a physical conservation constraint layer.

[0019] Specifically, the time-varying operator unit of elastoplastic mechanics reads the virtual stress tensor in the current feature map during each forward propagation, compares it with the analytical contact stress calculated based on Hertzian contact theory, and transmits the comparison error back to the current layer through Jacobian matrix iterative correction.

[0020] Specifically, the rheological state comprehensive evaluation function is based on the real-time dynamic viscosity deviation rate. Root mean square deviation of normal dynamic pressure and the normalized value of acoustic emission signal energy Calculate the rheological deviation index using three data points ,in accordance with The learning rate of the elastoplastic stress mapping model is adjusted within the specified interval.

[0021] Specifically, the timestamp alignment network is based on a hardware-level cascaded triggering mechanism to unify the clock source of multiple sensors. By writing global nanosecond-level timestamps into each sensor acquisition module, the heterogeneous data from multiple sources are interpolated and aligned on a unified time axis.

[0022] Specifically, the wavelet packet adaptive threshold denoising method involves decomposing the aligned multi-source data into multi-resolution sub-bands using wavelet packet decomposition, adaptively determining the threshold based on noise energy estimation in each sub-band, and reconstructing the coefficients that exceed the threshold after soft thresholding.

[0023] Specifically, the Riemannian manifold embedded dimensionality reduction surrogate model uses a Riemannian geometric framework to perform nonlinear dimensionality reduction on the high-dimensional abrasive ratio space, performs adaptive Bayesian optimization on the low-dimensional manifold, and uses Gaussian process regression as the surrogate model to fit the response surface of the objective functions of brightness and roughness. The dimension of the low-dimensional manifold is determined by performing principal geodesic analysis on multiple ratio datasets in the preliminary experiment.

[0024] The initial particle count for the particle swarm optimization is 50–200; the coupling time step for the macro-micro bidirectional coupling simulation mechanism is 1–10 μs; the sampling rate for the high-frequency torque sensor is 10–100 kHz, and the sampling rate for the acoustic emission sensor is 100 kHz–1 MHz; the spacing between adjacent sensing elements in the fiber Bragg grating array is 2–10 mm, and the wavelength demodulation frequency is 1–10 kHz; the Fourier neural operator consists of 4–8 stacked Fourier operator layers, retaining 12–32 low-frequency modes; the Jacobian matrix iteration count is 3–8 times; the training, validation, and test sets for the elastoplastic stress mapping model are divided in an 8:1:1 ratio; the noise power spectral density threshold factor is 1.5–3; the low-dimensional manifold dimension is 2–10, retaining at least 90% of the explained variance.

[0025] This invention acquires the surface hardness distribution matrix and texture region clustering map of jade through hyperspectral imaging. A particle swarm optimization-simulated annealing hybrid-driven optimization algorithm is used to globally optimize the abrasive ratio within the physically feasible solution space. The physical feasibility of the ratio is verified using a macro-micro bidirectional coupling simulation mechanism of the lattice Boltzmann method and the discrete element method. Furthermore, the rheological tensor is analyzed in real time using a fiber Bragg grating array and an operator network based on multimodal physical information. The normal dynamic pressure distribution field is then output through an elastoplastic stress mapping model to drive the dynamic adjustment of the abrasive ratio gradient step size, thus solving the defect that fixed ratio schemes cannot respond to anisotropic hardness fields and real-time rheological state drift. This invention uses the hardness distribution matrix as input constraint and the physical constitutive equation as the core of fitness evaluation, ensuring that each ratio search is performed within a physically reasonable solution space. Real-time closed-loop feedback from sensor data ensures that the ratio parameters always track the actual mechanical state of the polishing interface. In summary, this invention solves the technical problem mentioned in the background art where the multi-component abrasive ratio cannot be adaptively optimized in real time according to the anisotropic hardness field of jade, leading to unstable polishing brightness. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention.

[0027] Figure 2 This is a schematic diagram of the hardness distribution on the surface of a jade carving workpiece using hyperspectral imaging. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0029] like Figure 1 The diagram shown is a flowchart of a method for improving the polishing brightness of jade carvings using a multi-component abrasive formulation, provided by the present invention. This method includes the following steps:

[0030] S01. Perform hyperspectral imaging scanning on the surface of the jade carving workpiece to obtain the surface hardness distribution matrix and texture region clustering map, and input the particle swarm-simulated annealing hybrid driven optimization algorithm to complete the particle swarm initialization of the abrasive initial ratio parameters.

[0031] S02. Diamond micro powder, boron carbide and alumina are mixed according to the initial ratio. The physical feasibility is verified by the micro-rheological constitutive equation of the macro-micro bidirectional coupling simulation mechanism of lattice Boltzmann method and discrete element method. Infeasible ratios are eliminated.

[0032] S03: The fiber Bragg grating array collects the shear force on the polishing disk wall in real time. Combined with the operator network based on multimodal physical information, the rheological tensor of the polishing zone is analyzed in reverse. The dynamic viscosity parameters of the suspension are dynamically corrected and fed back to the particle swarm-simulated annealing hybrid driving optimization algorithm.

[0033] S04. Input the real-time rheological tensor, the microhardness matrix of the jade surface and the evolution data of the micro-depth contour of the polishing marks into the elastoplastic stress mapping model, output the normal dynamic pressure distribution field of the polishing wheel, and drive the dynamic adjustment of the abrasive ratio gradient step size.

[0034] S05. After the multi-source heterogeneous high-frequency sensing data is processed by the time stamp alignment network and the wavelet packet adaptive threshold denoising preprocessing architecture, the extracted weak feature vectors are fed back to the simulated annealing temperature decay mechanism of the particle swarm-simulated annealing hybrid driven optimization algorithm to trigger the Metropolis probability to jump out of the local extremum.

[0035] After the S06 particle swarm-simulated annealing hybrid-driven optimization algorithm converges, it outputs the globally optimal abrasive ratio. It then performs precise parallel microscopic simulation verification based on the low-dimensional control points confirmed by the Riemannian manifold embedded dimensionality reduction surrogate model. Finally, it is deployed to the polishing equipment to complete the high-brightness jade carving polishing process.

[0036] Among them, hyperspectral imaging scanning refers to using an imaging spectrometer with a working wavelength covering 400-2500nm to collect the reflectance spectrum of the jade carving workpiece pixel by pixel. By inverting the position and intensity of the spectral absorption peaks, the mineral composition and microhardness at each position are obtained, forming a hardness distribution matrix with a spatial resolution of 5-20μm. The matrix data is then clustered using a topological manifold learning algorithm to delineate the texture region cluster map of weak areas and hard impurities in jadeite and Hetian jade. The number of cluster categories in the texture region cluster map is determined by a series of clustering experiments on multiple representative jade samples, and the number of categories corresponding to the largest contour coefficient is taken as the officially used value.

[0037] The specific implementation of the particle swarm optimization algorithm driven by simulated annealing is as follows: A particle swarm of 50-200 particles is initialized, using the particle size range of the multi-component abrasive (0.5–100 μm), mass fraction range (1%–40%), suspension stirring speed range (200–2000 rpm), and chemical additive concentration as the particle position vector. In each iteration, the particles update their velocity and position based on their own historical best position and the swarm's historical best position. The updated ratio is then substituted into the constitutive equation of the mixed abrasive suspension based on microrheology for fitness evaluation. The constitutive equation uses a power-law model to describe the non-Newtonian fluid behavior, and the viscosity coefficient as a function of shear rate is pre-calibrated using a macro-micro bidirectional coupling simulation mechanism of the lattice Boltzmann method and the discrete element method. When the fitness of the particle swarm does not significantly improve after 10–30 consecutive iterations, a simulated annealing mechanism is triggered, using Metropolis probability... Accepting inferior solutions, among which For fitness difference, initial control temperature The standard deviation of fitness fluctuation was obtained through calibration of 1000 randomized combinations in a preliminary experiment, with temperature decreasing logarithmically. Gradually reduce, This represents the current iteration round. The algorithm terminates when the temperature is below a preset termination temperature threshold, which is obtained by averaging the lower limits of the temperature range at convergence during multiple optimization experiments. The technical advantages of this algorithm are: using the micro-rheological constitutive equation as the core of fitness evaluation, ensuring that each search step is performed within the physically feasible solution space; the particle swarm mechanism provides efficient parallel global exploration capabilities; and the Metropolis probability acceptance criterion of the simulated annealing mechanism, based on the principles of thermodynamics and statistical mechanics, gives the algorithm the ability to probabilistically escape local extrema. The synergistic effect of the two makes the global optimization for high-dimensional abrasive ratios both physically constrained and robust to global search, effectively overcoming the multi-extremum problem caused by the anisotropic hardness field of jade.

[0038] The specific implementation of the macro-micro bidirectional coupling simulation mechanism of the lattice Boltzmann method and the discrete element method is as follows: at the macro level, the lattice Boltzmann method is used to solve the velocity field and pressure field of the polishing slurry liquid phase; at the micro level, the discrete element method is used to track the collision, rolling and scratch trajectories of individual abrasive particles in the flow field. The two layers are coupled through bidirectional data exchange of fluid resistance and particle reaction force, with a coupling time step of 1 to 10 μs. The sampling rate of the high-frequency torque sensor is set to 10 to 100 kHz, and the sampling rate of the acoustic emission sensor is set to 100 kHz to 1 MHz. The signals collected in real time by both are used to dynamically correct the edge shear damping coefficient of the liquid-solid two-phase flow. The correction step of the edge shear damping coefficient is obtained by residual analysis of multiple measurements of a reference liquid with known rheological properties in the calibration experiment.

[0039] The fiber Bragg grating array refers to arranging multiple fiber Bragg grating sensing elements in an array on the edge wall of the polishing disk. By detecting the drift of the Bragg wavelength, the local strain of the wall is calculated, and then the distribution of wall shear force is inverted. The spacing between adjacent sensing elements is 2-10 mm, and the wavelength demodulation frequency is 1-10 kHz. The operator network based on multimodal physical information takes the wall pressure gradient data as input and uses the residual of the Navier-Stokes equation embedded in the physical information neural network framework as the physical constraint loss to reverse analyze the rheological tensor at the center of the polishing zone that cannot be directly measured. The weight coefficients of the physical constraint loss are determined by performing gradient descent optimization in multiple sets of reference liquid experiments with known rheological parameters.

[0040] The specific structure of the elastoplastic stress mapping model is as follows: the network as a whole adopts Fourier neural operators as the backbone structure. The input channels include the dynamic viscosity field of the abrasive suspension, the microhardness matrix of the jade surface, the evolution data of the micro-depth contour of the polishing marks, and the polishing wheel rotation speed. The output channel is the distribution field of the normal dynamic pressure of the polishing wheel. The Fourier neural operators are composed of 4 to 8 stacked Fourier operator layers. Each Fourier operator layer is composed of the sum of the parallel Fourier transform branch and the local linear transform branch. The number of low-frequency modes retained by the Fourier transform branch is 12 to 32. Each Fourier operator layer embeds an elastoplastic time-varying operator unit. The layer consists of a von Mises yield criterion determination module and a kinematic hardening constitutive update module connected in series. During each forward propagation, it reads the virtual stress tensor from the current feature map and compares it with the analytical contact stress calculated based on Hertzian contact theory. The comparison error is iteratively corrected using the Jacobian matrix and propagated back to the current layer. The number of iterations is 3–8, and the iteration convergence threshold is determined through residual statistical analysis of multiple sets of known elastoplastic deformation experimental data. The output of each Fourier operator layer is connected to a physical conservation constraint layer, which forcibly applies mass and momentum conservation penalties. The penalty weight coefficients are determined by grid search experiments during the pre-training phase. The hidden layer width is 64. ~256, using Gaussian error linear units as activation functions; the normal dynamic pressure distribution field of the polishing wheel output by the network is downsampled to the resolution corresponding to the abrasive ratio gradient step size adjustment grid through bilinear interpolation, driving the dynamic adjustment of the search step size in the Riemannian manifold embedding surrogate model; the memory allocation strategy is as follows: the Fourier transform module occupies an independent memory partition, the intermediate tensor of the time-varying operator unit of elastoplastic mechanics is stored in half-precision floating-point format to compress memory usage, and the gradient checkpointing mechanism is enabled for the gradient calculation of the physical conservation constraint layer to reduce backpropagation memory peaks; the memory allocation strategy is as follows: training data batches are loaded into memory asynchronously, with a batch size of 8~3 2. The prefetch queue depth is 2–4. The CUDA stream allocation strategy is as follows: Fourier transform calculations, iterative corrections of elastoplastic time-varying operator units, and data loading are allocated to 3 independent CUDA streams for concurrent execution. The CUDA thread block hierarchical allocation strategy is as follows: matrix operations of the Fourier operator layer are mapped to two-dimensional thread blocks with a block size of 16×16, and point-by-point iterative operations of the elastoplastic time-varying operator units are mapped to one-dimensional thread blocks with a block size of 128. The hierarchical allocation strategy is as follows: the low-frequency Fourier mode operation layer is allocated to the high-priority CUDA stream, and the gradient operation layer of the physical conservation constraint layer is allocated to the low-priority CUDA stream. The above allocation parameters are related to the input resolution. and batch size The relationship between them is ,in This is a scaling factor related to network width. The number of Fourier operator layers. The memory usage benchmark test was performed on the target hardware to determine different network widths; the steps for establishing the training dataset of the elasto-plastic stress mapping model specifically include: collecting polishing experimental data of various jade materials under different abrasive ratios, different polishing wheel speeds, and different normal loads, with measurement indicators covering dynamic viscosity, microhardness matrix, microscopic depth profile of polishing marks, and normal dynamic pressure, collecting no less than 5000 sets of experimental samples; spatially interpolating the collected data to a uniform resolution grid, and dividing it into training, validation, and test sets in an 8:1:1 ratio; the steps for training the elasto-plastic stress mapping model specifically include: using the Adam optimizer, with an initial learning rate of The learning rate decays by a factor of 0.5 every 50 training epochs; the total loss function is composed of a weighted sum of the data fitting loss, the mass conservation penalty term, and the momentum conservation penalty term, with each weight coefficient determined through grid search experiments during the pre-training phase; early stopping is performed when there is no improvement after 20 consecutive training epochs on the validation set; the initial learning rate and decay factor are determined by performing a learning rate range test on a small-scale pre-experiment dataset, and the number of consecutive epochs without improvement is determined through statistical analysis of the validation set loss convergence curve; the technical effect of the elastoplastic stress mapping model is that the frequency domain mapping mechanism of the Fourier neural operator enables the network to... The model efficiently captures the large-scale spatial distribution of the dynamic stress field of polishing. The embedded time-varying elastoplastic mechanical operator unit, through internal iterative correction of the von Mises yield criterion and Hertzian contact analytical solution, forces the stress tensor output by the network to be compatible with the constitutive equation of solid mechanics. The physical conservation constraint layer further ensures that mass and momentum do not violate the basic equation of continuum mechanics, thus giving the abrasive ratio driving signal predicted by the model real physical mechanism support. This eliminates the risk of physical distortion of pure data-driven models in small sample or out-of-distribution scenarios, and finally achieves high-fidelity online mapping of the dynamic stress field of the polishing interface.

[0041] Specifically, a rheological state comprehensive evaluation function is defined to adjust the learning rate of the elastoplastic stress mapping model. This function is calculated based on three data points: real-time dynamic viscosity deviation rate, root mean square deviation of normal dynamic pressure, and normalized value of acoustic emission signal energy, to obtain the rheological deviation index. , ;in , , The weighting coefficients were determined through regression analysis of the correlation between the three data points and the final brightness result in multiple rounds of polishing experiments. This represents the relative deviation between the measured dynamic viscosity and the constitutive equation prediction. The root mean square deviation of the normal dynamic pressure distribution field is given. Let be the normalized value of the acoustic emission signal energy; when When the learning rate remains unchanged, the learning rate remains at its current value; when At that time, the learning rate is multiplied by the decay factor. A gentle downward adjustment was made; when At that time, the learning rate is multiplied by the decay factor. Make a rapid downward adjustment; when When this happens, the network parameter rollback mechanism is triggered, restoring the model weights to the optimal checkpoint of the previous validation set and using the minimum learning rate. Restart training The threshold values ​​for each segment were determined through systematic experiments conducted under multiple representative polishing conditions, and statistical analysis was performed on different threshold values. The final surface brightness stability distribution corresponding to the interval, with a significant jump in brightness stability. The value is used as the threshold boundary and is determined by mean estimation from at least 30 sets of experiments.

[0042] The timestamp alignment network refers to a data preprocessing network based on a hardware-level cascaded triggering mechanism that unifies the clock sources of multiple sensors. By writing global nanosecond-level timestamps into each sensor acquisition module, it interpolates and aligns heterogeneous data from multiple sources with sampling rate differences ranging from Hz to MHz on a unified time axis. The wavelet packet adaptive threshold denoising refers to decomposing the aligned multi-source data into multi-resolution sub-bands using wavelet packet decomposition. The threshold is adaptively determined based on noise energy estimation in each sub-band. Coefficients exceeding the threshold are reconstructed after soft thresholding, thereby separating machine tool mechanical noise from the weak feature vector that truly reflects the abrasive cutting state. The threshold determination method is as follows: pure background noise data is collected in the shutdown state, and the noise power spectral density of each sub-band is statistically analyzed. The initial threshold is set at 1.5 to 3 times the standard deviation of the noise power spectral density. The threshold is iteratively corrected in the first 100 sampling windows of the formal machining process using online estimation. The multiplier coefficient is determined by the optimal analysis of signal-to-noise ratio in calibration experiments of multiple sets of known cutting states.

[0043] Among them, the Riemannian manifold embedding dimensionality reduction surrogate model refers to the nonlinear dimensionality reduction of the high-dimensional abrasive proportion space using the Riemannian geometric framework, embedding the high-dimensional proportion vector into a 2- to 10-dimensional low-dimensional manifold, performing adaptive Bayesian optimization on the low-dimensional manifold, using Gaussian process regression as the surrogate model to fit the response surface of the brightness and roughness objective functions, and dynamically adjusting the Bayesian acquisition function based on the Riemann curvature tensor at each control point on the manifold using an exploration-trade factor, triggering parallel microscopic simulations for accurate evaluation only for a few control points with the highest acquisition function values; the dimension of the low-dimensional manifold is determined by performing principal geodesic analysis on multiple proportion datasets in the pre-experiment, retaining the minimum dimension required to interpret at least 90% of the variance.

[0044] Among them, the time-varying operator unit of elastoplastic mechanics refers to the time-varying operator module embedded inside the Fourier operator layer, which is based on the von Mises yield criterion and the kinematic hardening constitutive equation. During each forward propagation, it performs elastoplastic state determination and constitutive update on the virtual stress tensor implicitly encoded in the current feature map, and compares the updated stress tensor with the analytical solution of Hertz contact theory. The residual is then corrected to the parameters of the current layer through Jacobi matrix iteration, thereby achieving strict embedding of the mechanical constraints of the continuous physical medium within the network layer.

[0045] Among them, Hertz contact theory refers to the classical solid mechanics theory that describes the stress and deformation distribution in the contact area of ​​two elastic bodies under normal load. In this scheme, it is used to calculate the analytical solution of the normal pressure distribution when abrasive particles are in contact with the jade surface, and serves as a reference benchmark for the internal iterative correction of the elastoplastic stress mapping model.

[0046] Metropolis probability refers to the acceptance probability criterion in the simulated annealing algorithm based on the thermodynamic Boltzmann distribution. When the fitness of a new solution is worse than that of the current solution, the algorithm uses an exponential probability to determine whether to accept the inferior solution, thereby giving the algorithm the ability to escape local extrema.

[0047] The specific implementation of step S01 is as follows: An imaging spectrometer with a working wavelength covering 400–2500 nm is used to scan the surface of the jade carving workpiece pixel by pixel, acquiring a complete reflectance spectrum curve for each pixel location. By identifying the position and intensity of the mineral characteristic absorption peaks in the spectral curve, the type and content of mineral components at that location are inverted. Then, based on the empirical mapping relationship between the mineral elastic modulus and Vickers hardness, the inversion result is converted into a microhardness value, ultimately forming a hardness distribution matrix with a spatial resolution of 5–20 μm. Subsequently, a topological manifold learning algorithm is used to perform nonlinear dimensionality reduction and clustering on the hardness distribution matrix. Using the silhouette coefficient as an evaluation index for the number of categories, a series of clustering experiments are performed on multiple representative jade samples. The number of categories corresponding to the largest silhouette coefficient is selected to delineate the texture region clustering map of weak areas and hard impurities in jadeite and Hetian jade. After obtaining the hardness distribution matrix and texture region clustering map, the two are used as constraint inputs for the particle swarm-simulated annealing hybrid driven optimization algorithm. The particle size range, mass fraction range, suspension stirring speed range and chemical additive concentration of the multi-component abrasive are used as particle position vectors to initialize 50 to 200 particles, thus completing the particle swarm initialization of the initial abrasive ratio parameters.

[0048] The specific implementation of step S02 is as follows: Diamond micron powder, boron carbide, and alumina are mixed according to the initial ratio output by the particle swarm initialization to prepare multiple sets of mixed abrasive suspensions. For each suspension, a macro-micro bidirectional coupling simulation mechanism of the lattice Boltzmann method and the discrete element method is used to verify its physical feasibility. At the macro level, the lattice Boltzmann method solves the velocity and pressure fields of the polishing slurry liquid phase in the polishing disk, with a coupling time step set to 1–10 μs; at the micro level, the discrete element method tracks the collision, rolling, and scratch trajectories of individual abrasive particles in the above flow field and feeds back the reaction force of the particles on the flow field to the lattice Boltzmann method, realizing macro-micro bidirectional data exchange. The micro-rheological constitutive equation results output by the simulation are evaluated. If a suspension corresponding to a certain ratio exhibits infeasible phenomena such as phase separation, particle agglomeration, or violation of power-law model constraints in the simulation, then that ratio is removed from the particle swarm, and only physically feasible ratio schemes are retained for subsequent iterations.

[0049] The specific implementation of step S03 is as follows: Multiple fiber Bragg grating sensing elements are arranged in an array with an adjacent spacing of 2-10 mm on the edge wall of the polishing disk. The wavelength demodulation frequency is set to 1-10 kHz to detect the drift of the Bragg wavelength in real time, calculate the local strain of the wall, and then invert the spatial distribution of the wall shear force. A high-frequency torque sensor with a sampling rate of 10-100 kHz and an acoustic emission sensor with a sampling rate of 100 kHz-1 MHz are used to synchronously acquire signals for inverse dynamic correction of the edge shear damping coefficient of the liquid-solid two-phase flow. The acquired wall pressure gradient data is input into an operator network based on multimodal physical information. This network uses the physical information neural network framework embedded with the Navier-Stokes equation residuals as physical constraint loss to inversely analyze the rheological tensor at the center of the polishing zone, which cannot be directly measured, dynamically correct the dynamic viscosity parameters of the suspension, and feed the correction results back to the particle swarm optimization-simulated annealing hybrid driven optimization algorithm in real time to update the constitutive equation parameters used for fitness evaluation.

[0050] The specific implementation of step S04 is as follows: The rheological tensor, the microhardness matrix of the jade surface, and the evolution data of the microscopic depth contour of polishing marks obtained through real-time analysis are used as inputs and fed into the elastoplastic stress mapping model along with the polishing wheel rotation speed. This model uses Fourier neural operators as its backbone and consists of 4–8 stacked Fourier operator layers, with each layer retaining 12–32 low-frequency modes. Each layer also embeds a time-varying elastoplastic mechanical unit. During each forward propagation, the time-varying elastoplastic mechanical unit reads the implicitly encoded virtual stress tensor in the current feature map, determines the elastoplastic state according to the von Mises yield criterion, and compares it with the analytical contact stress calculated based on Hertzian contact theory. The comparison error is iteratively corrected using the Jacobian matrix and propagated back to the current layer, with 3–8 iterations. Each layer's output is connected to a physical conservation constraint layer, which forcibly applies mass conservation penalties and momentum conservation penalties. The model ultimately outputs the dynamic pressure distribution field normal to the polishing wheel, which, after bilinear interpolation downsampling, drives the dynamic adjustment of the abrasive ratio gradient step size. Simultaneously, the rheological deviation index is calculated based on the rheological state comprehensive evaluation function. ,in accordance with The learning rate of the elastoplastic stress mapping model is adjusted accordingly within the specified interval. The network parameter rollback mechanism is triggered when the model weights are restored to the optimal checkpoint of the previous validation set and training is restarted with the minimum learning rate.

[0051] The specific implementation of step S05 is as follows: Multi-source heterogeneous high-frequency sensor data is first preprocessed by a timestamp alignment network. This network unifies the clock sources of each sensor based on a hardware-level cascaded triggering mechanism, writes a global nanosecond-level timestamp to each acquisition module, and interpolates and aligns multi-source data with sampling rate differences between Hz and MHz on a unified time axis. The aligned data is then fed into a wavelet packet adaptive threshold denoising preprocessing architecture. Wavelet packet decomposition decomposes the data into multi-resolution sub-bands. Pure background noise data is acquired in a stopped state, and the noise power spectral density of each sub-band is statistically analyzed. An initial threshold of 1.5 to 3 times the standard deviation of the noise power spectral density is used, and iterative correction is performed online within the first 100 sampling windows of the formal processing. Wavelet packet coefficients exceeding the threshold are reconstructed after soft thresholding to separate machine tool mechanical noise and extract weak feature vectors that truly reflect the abrasive cutting state. The extracted weak feature vectors are fed back to the simulated annealing temperature decay mechanism of the particle swarm optimization-simulated annealing hybrid driven optimization algorithm. When the fitness of the particle swarm does not significantly improve in continuous iterations, the Metropolis probability is triggered. To accept inferior solutions, the temperature decreases logarithmically. By gradually reducing the value, the algorithm can escape local optima.

[0052] The specific implementation of step S06 is as follows: After the particle swarm optimization algorithm driven by simulated annealing cools to below the termination temperature threshold, it converges and outputs the globally optimal abrasive mix ratio. Subsequently, a Riemannian manifold embedding dimensionality reduction surrogate model is used to perform nonlinear dimensionality reduction on the high-dimensional abrasive mix ratio space using a Riemannian geometric framework. The high-dimensional mix ratio vector is embedded into a 2-10 dimensional low-dimensional manifold. Adaptive Bayesian optimization is performed on the low-dimensional manifold. Gaussian process regression is used as the response surface for fitting the brightness and roughness objective functions of the surrogate model. Based on the Riemannian curvature tensor at each control point on the manifold, the Bayesian acquisition function is dynamically adjusted using a tradeoff coefficient. Parallel microscopic simulation is triggered only for a few control points with the highest acquisition function values ​​for precise evaluation. After verification, the optimal mix ratio is deployed to the polishing equipment to complete the high-brightness jade carving polishing process. The dimension of the low-dimensional manifold is determined by performing principal geodesic analysis on multiple mix ratio datasets in the pre-experiment, retaining the minimum dimension required to explain at least 90% of the variance.

[0053] It should be noted that the key technologies of this invention include: a surface hardness spatial perception technology based on hyperspectral imaging and topological manifold learning algorithms, which transforms the spectral information of mineral components on the jade surface into spatially resolved hardness constraints, providing a physical basis for ratio optimization and overcoming the fundamental defect of traditional methods that rely on empirical estimation and ignore anisotropic hardness fields; a physical feasibility verification technology based on the macro-micro bidirectional coupling simulation mechanism of the lattice Boltzmann method and the discrete element method, which verifies the physical rationality of the ratio scheme at two scales: macroscopic flow field and microscopic particle dynamics, ensuring that each step of optimization is carried out within the physically feasible solution space; and an elastoplastic stress mapping model with Fourier neural operators as the backbone and embedded with the analytical solutions of the von Mises yield criterion and Hertzian contact theory, which forces the network output to be compatible with the basic equations of solid mechanics through physical constitutive constraints, eliminating the risk of physical distortion in purely data-driven models. The three key technologies mentioned above work together to form a complete closed loop from surface hardness perception and physical verification of the ratio to high-fidelity mapping of the stress field. This ensures that the abrasive ratio remains consistent with the actual mechanical state of the jade surface throughout the polishing process, thus eliminating the systematic mismatch of a fixed ratio scheme under anisotropic hardness field in principle.

[0054] It should be noted that under conditions of long-term continuous operation of jade carving polishing equipment and continuous drift of the rheological state of the polishing slurry due to abrasive consumption and temperature rise, the dynamic viscosity and normal dynamic pressure distribution at the polishing interface will significantly deviate from their initial calibration values ​​over time. This results in a continuous deterioration of polishing brightness in the later stages of processing with a fixed mixing scheme. This problem is particularly prominent when processing jade workpieces with large areas of weak zones and alternating distributions of hard impurities. The above technical problem arises because the abrasive particles in the polishing slurry break down and become smaller during processing. At the same time, the heat generated by the friction between the polishing disc and the workpiece raises the temperature of the suspension. The combined effect of these two factors causes the non-Newtonian fluid viscosity coefficient of the suspension to deviate from the initial power-law model calibration value due to the shear rate. Furthermore, the rheological tensor at the center of the polishing zone cannot be directly measured by conventional sensors. Therefore, the fixed mixing scheme cannot detect this dynamic drift, and the normal dynamic pressure distribution field continuously deviates from its optimal value, ultimately leading to a monotonous deterioration of polishing brightness as processing time increases. The usual solutions to the above technical problems are to periodically stop the machine to recalibrate the viscosity of the suspension, or to add a viscosity stabilizer to the polishing fluid to suppress the drift of the rheological state. However, the former will interrupt the continuity of processing and cause brightness fluctuations after each restart, while the latter may change the distribution of abrasive cutting force and introduce new brightness non-uniformity because of the adsorption competition effect between chemical additives and abrasive particles. Neither of these methods can achieve real-time closed-loop correction of the rheological state without interrupting processing. This invention effectively solves this technical problem. A fiber Bragg grating array continuously acquires the shear force on the polishing disc wall at a wavelength demodulation frequency of 1–10 kHz. Combined with multimodal signals from a high-frequency torque sensor and an acoustic emission sensor, an operator network based on multimodal physical information, using the Navier-Stokes equation residuals as physical constraint loss, performs real-time inverse analysis of the rheological tensor at the center of the polishing zone, which cannot be directly measured. This dynamically corrects the dynamic viscosity parameters of the suspension, and the correction results are immediately fed back to the particle swarm optimization-simulated annealing hybrid-driven optimization algorithm. After updating the constitutive equation parameters, the ratio fitness is reassessed. The elastoplastic stress mapping model, based on the updated dynamic viscosity and microhardness matrix, re-outputs the normal dynamic pressure distribution field of the polishing wheel, driving the dynamic adjustment of the abrasive ratio gradient step size. The entire correction process continues without interrupting processing, fundamentally eliminating the influence of rheological state drift on polishing brightness during long-term continuous processing.

[0055] This invention solves the aforementioned technical problems, and its inherent logic is reflected in the following aspects. First, the anisotropic hardness field on the jade surface is the fundamental cause of the failure of fixed-ratio schemes. Hyperspectral imaging scanning can invert mineral composition and microhardness pixel by pixel, forming a spatially resolved hardness distribution matrix. Then, a topological manifold learning algorithm is used to cluster the matrix data, dividing the workpiece surface into several texture regions with different mechanical properties. This step transforms the anisotropic information of the workpiece surface into calculable spatial constraints, providing a physical basis for subsequent ratio optimization, rather than relying on empirical estimation.

[0056] Secondly, the optimization space for multi-component abrasive formulations is high-dimensional and contains numerous extrema, making it easy for conventional optimization algorithms to get trapped in local optima. The particle swarm optimization-simulated annealing hybrid-driven optimization algorithm organically combines the parallel global search capability of the particle swarm optimization algorithm with the Metropolis probabilistic escape mechanism of the simulated annealing algorithm. Fitness evaluation directly invokes the constitutive equation of the mixed abrasive suspension based on micro-rheology, ensuring that the search results at each step are physically feasible. The macro-micro bidirectional coupling simulation mechanism of the lattice Boltzmann method and the discrete element method further verifies the physical feasibility of the formulation at both the macro-flow field and micro-particle trajectory levels, eliminating schemes that do not meet rheological constraints, thereby strictly limiting the search space to a physically reasonable subset.

[0057] Furthermore, the rheological state of the suspension continuously changes dynamically during polishing, and static proportions cannot adapt to this drift. A fiber Bragg grating array acquires the wall shear force in real time. An operator network based on multimodal physical information uses the Navier-Stokes equation residuals as physical constraint losses to inversely analyze the rheological tensor of the polishing zone, which cannot be directly measured. This dynamically corrects the dynamic viscosity parameters of the suspension and feeds them back to the optimization algorithm. An elastoplastic stress mapping model, with Fourier neural operators as its backbone, embeds internal iterative corrections of the analytical solutions of the von Mises yield criterion and Hertzian contact theory. It outputs the normal dynamic pressure distribution field of the polishing wheel, driving the dynamic adjustment of the abrasive proportion gradient step size. A physical conservation constraint layer forcibly applies mass and momentum conservation penalties to ensure that the model output is compatible with the fundamental equations of continuum mechanics, avoiding physical distortions in small sample or out-of-distribution scenarios caused by purely data-driven models.

[0058] Finally, after preprocessing the multi-source heterogeneous high-frequency sensor data through a timestamp alignment network and wavelet packet adaptive threshold denoising, the weak feature vectors are fed back to the simulated annealing temperature decay mechanism, triggering the Metropolis probability to escape local extrema. The Riemannian manifold embedding dimensionality reduction surrogate model nonlinearly embeds the high-dimensional ratio space into a low-dimensional manifold, performing adaptive Bayesian optimization on the low-dimensional manifold. Parallel microscopic simulations are triggered only for a few control points with the highest acquired function values, balancing global search efficiency with reasonable allocation of computational resources. These steps form a complete closed-loop feedback system, ensuring that the abrasive ratio continuously tracks the actual mechanical state of the jade surface throughout the polishing process, fundamentally eliminating the systematic mismatch of a fixed ratio scheme under anisotropic hardness fields.

[0059] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0060] The specific implementation of step S01 is to utilize the operating band covering 400–2500 The imaging spectrometer collects the reflectance spectrum of the jade carving workpiece pixel by pixel. By inverting the position and intensity of the spectral absorption peaks, the mineral composition and microhardness at each location are obtained, forming a spatial resolution of 5–20. Hardness distribution matrix , The formula is expressed as follows:

[0061] ;

[0062] In the formula, For the first Line 1 Microhardness value at each column pixel, in units of , , These represent the number of rows and columns of pixels in the scanned area. A topological manifold learning algorithm is used to cluster the matrix data, defining a texture region cluster map. The number of cluster categories is determined through a series of clustering experiments on multiple representative jade samples, using contour coefficients. The maximum number of categories As a formally used value, the profile coefficient For a dimensionless quantity, the formula is as follows:

[0063] ;

[0064] In the formula, The average distance between a sample point and other sample points in the same cluster, in units of 1. , The average distance between a sample point and all sample points in the nearest heterogeneous cluster, in units of 1. The two have the same dimensions. The range is Then the hardness distribution matrix was... The particle swarm optimization algorithm, which is a hybrid drive of particle swarm and simulated annealing, is used to initialize the initial abrasive mix parameters.

[0065] The specific implementation of step S02 involves mixing diamond micron powder, boron carbide, and alumina according to an initial ratio, and then verifying the physical feasibility of the micro-rheological constitutive equation established by a macro-micro bidirectional coupling simulation mechanism using the lattice Boltzmann method and the discrete element method. The constitutive equation uses a power-law model to describe the non-Newtonian fluid behavior, and the formula is as follows:

[0066] ;

[0067] In the formula, Shear stress, unit: , This is the consistency coefficient, in units of... , Shear rate, in units of , The rheological index is a dimensionless quantity. The initial yield stress is given by . The aforementioned parameters are pre-calibrated using a macro-micro bidirectional coupling simulation mechanism combining the lattice Boltzmann method and the discrete element method. In the coupled simulation, the lattice Boltzmann method solves for the macroscopic velocity vector field. With pressure field Discrete Element Method (DEM) tracks the position vector of each abrasive particle. and velocity vector ,in The granules are numbered, and the bidirectional data exchange coupling time step between the two layers is 1 to 10. Infeasible combinations were eliminated.

[0068] The specific implementation of step S03 is that a fiber Bragg grating array is arranged on the edge wall of the polishing disk, and the spacing between adjacent sensing elements is 2 to 10. The wavelength demodulation frequency is 1 to 10. By detecting the Bragg wavelength shift The formula for inverting wall shear force is as follows:

[0069] ;

[0070] In the formula, This is the Bragg wavelength shift, in units of... , The initial Bragg center wavelength, in units of , denoted as the effective elastic-optical coefficient of the optical fiber, a dimensionless quantity with an empirical value of 0.22. Let be the local strain of the wall, and be a dimensionless quantity. The dimensions of both sides of the formula are . The operator network based on multimodal physical information takes wall pressure gradient data as input and embeds the Navier-Stokes equation residuals. As a physical constraint loss, the inverse analytical rheological tensor Total loss function The formula is expressed as follows:

[0071] ;

[0072] In the formula, This is the total loss function value. For data fitting loss, The physical constraint loss weighting coefficient is a dimensionless quantity, determined by performing gradient descent optimization on multiple sets of reference liquid experiments with known rheological parameters. The residuals of the Navier-Stokes equations are represented by three terms unified to dimensionless loss values. Dynamic viscosity parameters. The dynamic corrections are fed back to the particle swarm optimization algorithm driven by simulated annealing.

[0073] The specific implementation of step S04 is to convert the real-time rheological tensor Microhardness matrix The microscopic depth contour evolution data of polishing marks are input into the elastoplastic stress mapping model. The network uses 4 to 8 stacked Fourier operator layers, each consisting of a parallel sum of a Fourier transform branch and a local linear transform branch, retaining 12 to 32 low-frequency modes. The embedded time-varying elastoplastic mechanical operator units are used to simulate the virtual stress tensor. The von Mises yield test is performed, and the formula is expressed as follows:

[0074] ;

[0075] In the formula, This is the yield function value, in units of , These are the components of the deviatoric stress tensor, in units of . , The yield stress of the material is given by units of 1000 ppm. The dimensions on both sides of the formula are ,when When the material is in an elastic state, It enters the plastic state at this time. The virtual stress tensor... Analysis of contact stress using Hertzian contact theory The residuals are compared using the Jacobian matrix. Iterative correction, with 3 to 8 iterations, is expressed by the following iterative formula:

[0076] ;

[0077] In the formula, This represents the iteration round number. For the first The virtual stress tensor of the next iteration, in units of , The contact stress tensor is analyzed using Hertzian contact theory, with units of . , Let be the Jacobian matrix, and let be the first-order partial derivative matrix of the stress residual with respect to the stress tensor. Both sides of the formula are dimensionless. The physical conservation constraint layer imposes a mass conservation penalty term. With momentum conservation penalty term Total loss function The formula is expressed as follows:

[0078] ;

[0079] In the formula, This represents the total loss function value of the elastoplastic stress mapping model. For data fitting loss, , The penalty weight coefficient is a dimensionless quantity, determined through grid search experiments during the pre-training phase. This is a penalty term for the conservation of mass. As a momentum conservation penalty term, the four dimensions are unified into dimensionless loss values. Output: Normal dynamic pressure distribution field of the polishing wheel. The unit is This drives the dynamic adjustment of the gradient step size of the abrasive ratio.

[0080] The specific implementation of step S05 is as follows: multi-source heterogeneous high-frequency sensor data are synchronized with a unified clock source via a timestamp alignment network, and each sensor acquisition module writes a global nanosecond-level timestamp, thus merging the sampling rate differences. to Multi-source data of varying magnitudes are interpolated and aligned on a unified time axis. Subsequently, a wavelet packet adaptive thresholding denoising preprocessing architecture decomposes the aligned data into multi-resolution sub-bands, and the threshold is adaptively determined based on the noise power spectral density in each sub-band. The formula is expressed as follows:

[0081] ;

[0082] In the formula, Determine the noise threshold for each sub-band. This represents the average background noise power spectral density of each subband. The standard deviation of the background noise power spectral density for each subband is given. All three dimensions are consistent with the power spectral density, and the unit is 1. , The coefficients are dimensionless multiples, determined through signal-to-noise ratio optimization analysis in multiple calibration experiments under known cutting conditions, with empirical values ​​ranging from 1.5 to 3. For coefficients exceeding a threshold, soft-thresholding shrinkage is performed followed by reconstruction, extracting weak feature vectors. Feedback to the simulated annealing temperature decay mechanism of the particle swarm optimization-simulated annealing hybrid driven optimization algorithm triggers the Metropolis probability to escape local extrema.

[0083] The specific implementation of step S06 is as follows: The complete process of the particle swarm optimization algorithm driven by simulated annealing is as follows. Particle position vector Defined as:

[0084] ;

[0085] In the formula, For the first Abrasive particle size, in units of . , Mass fraction, dimensionless percentage. The stirring speed of the suspension is expressed in units of 1000 rpm. , This refers to the concentration of chemical additives, in units of... The formulas for particle velocity update and position update are expressed as follows:

[0086] ;

[0087] ;

[0088] In the formula, For the first The particle in the first The velocity vector at the next iteration, with dimensions and same, This is the contraction factor, a dimensionless quantity, with a default value of 0.729. , For interval The random number within is a dimensionless quantity. , This is the acceleration factor, a dimensionless quantity, with a default value of 2.05. For the first The historical optimal position vector of each particle The vector representing the group's historical optimal position. This is the current iteration round number; the number of particles is initialized to 50-200. When the fitness does not significantly improve after 10-30 consecutive iterations, simulated annealing is triggered, and the Metropolitan Probability (MP) is adjusted. The formula is expressed as follows:

[0089] ;

[0090] In the formula, This represents the fitness difference between the new solution and the current solution. For the first The temperature is controlled in each cycle, and both have the same dimensions, making... It is a dimensionless quantity. The value is a dimensionless probability value, and the control temperature is calculated using the logarithmic decay formula:

[0091] ;

[0092] In the formula, The initial control temperature was obtained through calibration using the fitness fluctuation standard deviation of 1000 randomized combinations in a preliminary experiment. This is the current iteration round number. and The units are the same. The algorithm terminates under the condition that... Below the preset termination temperature threshold , The optimal temperature range was obtained by averaging the lower bounds of the convergence intervals through multiple optimization experiments. After convergence, the globally optimal allocation vector was output. The high-dimensional matching vectors are embedded into low-dimensional manifolds of 2–10 dimensions using a Riemannian manifold embedding surrogate model. The dimension of the low-dimensional manifold is determined through principal geodesic analysis, retaining the minimum dimension required to explain at least 90% of the variance. Gaussian process regression is used as a surrogate model on the low-dimensional manifold, based on the Riemann curvature tensor. Dynamically adjust the Bayesian acquisition function exploration - using a trade-off coefficient, only trigger parallel microscopic simulation for precise evaluation of a few control points with the highest acquisition function values, and finally deploy it to the polishing equipment.

[0093] In the explanation section, the rheological state comprehensive evaluation function is used to adjust the learning rate of the elastoplastic stress mapping model, and the rheological bias index... The formula is expressed as follows:

[0094] ;

[0095] In the formula, , , , where is the weighting coefficient, and is a dimensionless quantity. It was determined through regression analysis of the correlation between the three data points and the final brightness result during multiple rounds of polishing experiments. The relative deviation between the measured dynamic viscosity and the predicted value from the constitutive equation is calculated as follows: ,in This is the measured value of dynamic viscosity, in units of... , These are constitutive equation predictions, in units of , It is a dimensionless quantity. Normal dynamic pressure distribution field The root mean square deviation is a dimensionless normalized quantity. Let be the normalized value of the acoustic emission signal energy, which is a dimensionless quantity. It is a dimensionless comprehensive evaluation index. When... The learning rate remains constant when Learning rate multiplied by decay factor ,when Learning rate multiplied by decay factor ,when The network parameter rollback mechanism is triggered when the network weights are restored to the optimal checkpoint of the previous validation set and the minimum learning rate is used. Restart training, among which The minimum learning rate allowed during training is a dimensionless positive number. The threshold values ​​for each segment were determined through systematic experiments conducted under multiple representative polishing conditions, statistically analyzing the differences. The final surface brightness stability distribution corresponding to the interval, with a significant jump in brightness stability. The value is used as the threshold boundary and is determined by mean estimation from at least 30 sets of experiments. Memory allocation satisfies the following relationship:

[0096] ;

[0097] In the formula, This represents video memory usage, in bytes. This is a scaling factor related to network width, measured in bytes, determined by performing benchmark tests on different network widths on the target hardware to account for video memory usage. Input resolution, in pixels. For batch size, The number of Fourier operator layers. , , All are dimensionless positive integers, and the units on both sides of the formula are bytes.

[0098] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0099] This embodiment uses a jade carving ornament with dimensions of 85×60×15mm and material of Grade A jadeite as the processing object. The surface of the workpiece has a typical anisotropic hardness field with multiple hard pyroxene impurity points and soft clay mineral enrichment areas alternating. The goal is to complete the high-brightness surface treatment without interrupting polishing, and to verify the complete workflow of the method of the present invention.

[0100] First, step S01 is executed. An imaging spectrometer with a working wavelength coverage of 400–2500 nm is used to perform a pixel-by-pixel hyperspectral scan of the workpiece surface. The acquisition resolution is set to 10 μm, resulting in the acquisition of complete reflectance spectrum curves for approximately 510,000 pixels. By analyzing the positions and intensities of the mineral characteristic absorption peaks near 1000 nm, 1400 nm, and 2200 nm for each pixel, the content ratios of jadeite, tremolite, and clay minerals at each location are inverted. Based on the empirical mapping relationship between mineral elastic modulus and Vickers hardness, a hardness distribution matrix covering the entire workpiece surface is generated, with a hardness range of 5.8–8.3. Subsequently, the topological manifold learning algorithm was used to perform nonlinear dimensionality reduction and clustering on the hardness distribution matrix. A series of clustering experiments were performed on five representative jade samples. The silhouette coefficient reached its maximum value of 0.74 when the number of clusters was 5, and the number of clusters used in the formal application was determined to be 5. The hardness distribution matrix and texture region clustering map were input into the particle swarm optimization algorithm driven by simulated annealing. Using the diamond micron particle size, boron carbide mass fraction, alumina mass fraction, stirring speed, and dispersant concentration as particle position vectors, 100 particles were initialized to complete the particle swarm initialization of the abrasive initial ratio parameters. The initial ratio is shown in Table 1.

[0101] Table 1 Initial Abrasive Mixing Parameters

[0102]

[0103] In step S02, 100 initial mix proportions are substituted into a macro-micro bidirectional coupling simulation mechanism using the lattice Boltzmann method and the discrete element method, with a coupling time step of 5 μs. The macro-level lattice Boltzmann method solves for the liquid velocity and pressure fields within the polishing disk, while the micro-level discrete element method tracks the collision and scratch trajectories of individual abrasive particles. Bidirectional data exchange is achieved between the two layers through fluid resistance and particle reaction forces. After evaluation by the micro-rheological constitutive equations, 8 infeasible mix proportions that exhibited particle agglomeration or violated power-law model constraints in the simulation are eliminated, leaving 92 physically feasible mix proportions for subsequent iterations.

[0104] In step S03, 12 fiber Bragg grating sensing elements are arranged at a 5mm spacing on the edge wall of the polishing disk. The wavelength demodulation frequency is set to 5kHz, the sampling rate of the high-frequency torque sensor is set to 50kHz, and the sampling rate of the acoustic emission sensor is set to 500kHz. After processing begins, the fiber Bragg grating array continuously outputs the wall shear force distribution. Combined with the multimodal signals from the high-frequency torque sensor and the acoustic emission sensor, the signal is input to an operator network based on multimodal physical information. This network uses the Navier-Stokes equation residuals as physical constraint losses to inversely analyze and obtain the rheological tensor at the center of the polishing zone. At the 45-minute mark of processing, the inverse analysis results show that the dynamic viscosity of the suspension is 12% higher than the initial calibration value. The algorithm automatically feeds back the corrected dynamic viscosity parameters to the particle swarm optimization-simulated annealing hybrid driven optimization algorithm to update the constitutive equation parameters.

[0105] In step S04, the real-time rheological tensor, microhardness matrix, and microscopic depth profile evolution data of polishing marks are input into the elastoplastic stress mapping model. This model consists of six stacked Fourier operator layers, with each layer retaining 20 low-frequency modes. Each layer embeds time-varying elastoplastic mechanical operator units. The von Mises yield criterion determination module is connected in series with the kinematic hardening constitutive update module, and the residuals are compared with the analytical solution of Hertz contact theory. The Jacobian matrix iteration count is set to 5 times, and each layer's output is connected to a physical conservation constraint layer. The model outputs the normal dynamic pressure distribution field of the polishing wheel, driving the dynamic adjustment of the abrasive ratio gradient step size. Rheological deviation index The calculation results are shown in Table 2, with each weight coefficient... , , The correlation regression analysis of multiple polishing experiments was used to determine this.

[0106] Table 2 Calculation Results of Rheological Deviation Index at Each Processing Stage

[0107]

[0108] In step S05, multi-source heterogeneous high-frequency sensor data is preprocessed by a timestamp alignment network. Each sensor acquisition module writes a global nanosecond-level timestamp, and multi-source data with sampling rate differences between 5kHz and 500kHz are interpolated and aligned on a unified time axis. The aligned data is then fed into a wavelet packet adaptive threshold denoising preprocessing architecture. Pure background noise data is collected in a stopped state, and the noise power spectral density of each sub-band is statistically analyzed. The initial threshold is set to twice the standard deviation of the noise power spectral density of each sub-band, and iteratively corrected using online estimation within the first 100 sampling windows of the formal processing. This extracts weak feature vectors that truly reflect the abrasive cutting state. These weak feature vectors are fed back to the simulated annealing temperature decay mechanism, which initially controls the temperature. The fitness fluctuation standard deviation was calibrated to 128.6 based on 1000 randomized combinations, with temperature decreasing logarithmically. As the fitness of the particle swarm gradually decreases, it triggers the Metropolis probability acceptance mechanism after 15 consecutive iterations without significant improvement, successfully escaping local optima 3 times.

[0109] In step S06, the algorithm converges after the temperature drops below the termination temperature threshold of 0.87, outputting the globally optimal abrasive mix: diamond powder particle size 8μm, boron carbide mass fraction 11%, alumina mass fraction 18%, stirring speed 1100rpm, and dispersant concentration 3.2. The Riemannian manifold embedding dimensionality reduction surrogate model, through principal geodesic analysis, determined that the dimension of the low-dimensional manifold was 4 (explaining 92.3% of the variance). Adaptive Bayesian optimization was performed on the low-dimensional manifold, and parallel microscopic simulations were triggered for the six control points with the highest acquired function values. After successful verification, the optimal ratio was deployed to the polishing equipment to complete the high-brightness polishing process of the jade workpiece.

[0110] Compared to traditional fixed-ratio methods, the advancements of this invention are reflected in the following aspects: Traditional methods rely on operator experience to set the abrasive ratio, failing to perceive the spatial distribution differences of various mineral components on the workpiece surface. This invention, through hyperspectral imaging, transforms the anisotropic information of the surface hardness field into calculable spatial constraints, fundamentally solving the problem of the disconnect between the ratio and the actual mechanical properties of the workpiece surface. Traditional methods cannot respond to the dynamic drift of the rheological state of the suspension during long-term continuous processing. This invention, through a fiber Bragg grating array and an operator network based on multimodal physical information, forms a real-time closed loop, enabling the ratio parameters to continuously track the actual rheological state of the polishing interface, eliminating the brightness degradation mechanism caused by rheological drift. The elastoplastic stress mapping model, by embedding the analytical solutions of the von Mises yield criterion and Hertzian contact theory into the internal iteration of the Fourier neural operator, forces the model output to be compatible with the constitutive equations of solid mechanics, eliminating the risk of physical distortion in out-of-distribution scenarios for purely data-driven models, and giving the ratio-driven signal support from real physical mechanisms.

[0111] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4.

[0112] Table 3. Variable Explanation Table (Part 1)

[0113]

[0114] Table 4. Variable Explanation Table (Part Two)

[0115]

[0116] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for improving the polishing brightness of jade carvings using a multi-component abrasive formulation, characterized in that, Includes the following steps: The surface of the jade carving workpiece is scanned by hyperspectral imaging to obtain the surface hardness distribution matrix and texture region clustering map. The particle swarm-simulated annealing hybrid driven optimization algorithm is then input to complete the particle swarm initialization of the abrasive initial ratio parameters. Diamond micro powder, boron carbide and alumina were mixed in the initial ratio. The physical feasibility was verified by the micro-rheological constitutive equation of the macro-micro bidirectional coupling simulation mechanism of lattice Boltzmann method and discrete element method, and infeasible ratios were eliminated. A fiber Bragg grating array is used to collect the shear force on the polishing disk wall in real time. Combined with an operator network based on multimodal physical information, the rheological tensor of the polishing zone is analyzed in reverse. The dynamic viscosity parameters of the suspension are dynamically corrected and fed back to the particle swarm-simulated annealing hybrid driving optimization algorithm. The real-time rheological tensor, the microhardness matrix of the jade surface, and the evolution data of the micro-depth contour of polishing marks are input into the elastoplastic stress mapping model, and the normal dynamic pressure distribution field of the polishing wheel is output to drive the dynamic adjustment of the abrasive ratio gradient step size. After the multi-source heterogeneous high-frequency sensor data is processed by the time stamp alignment network and the wavelet packet adaptive threshold denoising preprocessing architecture, the extracted weak feature vectors are fed back to the simulated annealing temperature decay mechanism of the particle swarm-simulated annealing hybrid driven optimization algorithm, triggering the Metropolis probability to jump out of local extrema. After the particle swarm optimization algorithm converges, it outputs the globally optimal abrasive ratio. It then performs precise parallel microscopic simulation verification based on the low-dimensional control points confirmed by the Riemannian manifold embedded dimensionality reduction surrogate model. Finally, it is deployed to the polishing equipment to complete the high-brightness jade carving polishing process.

2. The method for improving the polishing brightness of jade carvings using a multi-component abrasive formulation according to claim 1, characterized in that, The hyperspectral imaging scan specifically involves using an imaging spectrometer with a working wavelength range of 400–2500 nm to collect the reflection spectrum of the jade carving workpiece pixel by pixel. By inverting the position and intensity of the spectral absorption peaks, the mineral composition and microhardness at each position are obtained, forming a hardness distribution matrix with a spatial resolution of 5–20 μm.

3. The method for improving the polishing brightness of jade carvings using a multi-component abrasive formulation according to claim 2, characterized in that, The generation of the texture region cluster map is specifically achieved by using a topological manifold learning algorithm to cluster the hardness distribution matrix data. The number of cluster categories is determined by performing a series of clustering experiments on multiple representative jade samples and taking the number of categories corresponding to the maximum contour coefficient.

4. The method for improving the polishing brightness of jade carvings using a multi-component abrasive formulation according to claim 3, characterized in that, The particle swarm-simulated annealing hybrid-driven optimization algorithm specifically initializes the particle swarm with particle position vectors based on the particle size range, mass fraction range, suspension stirring speed range, and chemical additive concentration of the multi-component abrasive. In each iteration, the particles update their speed and position based on their own historical best position and the group's historical best position. The updated ratio is then substituted into the constitutive equation of the hybrid abrasive suspension based on microrheology for fitness evaluation.

5. The method for improving the polishing brightness of jade carvings using a multi-component abrasive ratio according to claim 4, characterized in that, The timestamp alignment network is specifically based on a hardware-level cascaded triggering mechanism to unify the clock source of multiple sensors. By writing global nanosecond-level timestamps into each sensor acquisition module, the heterogeneous data from multiple sources are interpolated and aligned on a unified time axis.

6. The method for improving the polishing brightness of jade carvings using a multi-component abrasive formulation according to claim 5, characterized in that, The proposed macro-micro bidirectional coupling simulation mechanism of the lattice Boltzmann method and the discrete element method specifically uses the lattice Boltzmann method to solve the velocity and pressure fields of the polishing slurry at the macro level, and the discrete element method to track the collision, rolling and scratch trajectories of individual abrasive particles at the micro level. The two layers are coupled through bidirectional data exchange of fluid resistance and particle reaction force.

7. The method for improving the polishing brightness of jade carvings using a multi-component abrasive formulation according to claim 6, characterized in that, The fiber Bragg grating array specifically involves arranging multiple fiber Bragg grating sensing elements in an array on the edge wall of the polishing disk. By detecting the drift of the Bragg wavelength, the local strain of the wall is calculated, and then the distribution of shear force on the wall is inverted.

8. The method for improving the polishing brightness of jade carvings using a multi-component abrasive formulation according to claim 7, characterized in that, The operator network based on multimodal physical information specifically uses wall pressure gradient data as input, and uses the residuals of the Navier-Stokes equations embedded in the physical information neural network framework as physical constraint loss to reverse analyze the rheological tensor of the polished zone center that cannot be directly measured.

9. The method for improving the polishing brightness of jade carvings using a multi-component abrasive ratio according to claim 8, characterized in that, The elastoplastic stress mapping model specifically uses Fourier neural operators as the backbone structure of the network. Each Fourier operator layer embeds elastoplastic time-varying operator units. These elastoplastic time-varying operator units are composed of a von Mises yield criterion determination module and a kinematic hardening constitutive update module connected in series. The output of each Fourier operator layer is connected to a physical conservation constraint layer.

10. The method for improving the polishing brightness of jade carvings using a multi-component abrasive ratio according to claim 9, characterized in that, Specifically, the time-varying operator unit of the elastoplastic mechanics reads the virtual stress tensor in the current feature map during each forward propagation, compares it with the analytical contact stress calculated based on Hertzian contact theory, and transmits the comparison error back to the current layer through Jacobian matrix iterative correction.

Images