Alumina ceramic grinding multidimensional crack prevention adaptive control method

Abstract

The application belongs to the technical field of precision machining control, and specifically discloses an alumina ceramic grinding multidimensional crack prevention adaptive control method. The method establishes a fracture mechanics constitutive model, quantifies the mapping relationship between the grinding force, temperature, grain size and crack threshold value; synchronously collects acoustic emission and temperature field data, generates a fusion tensor through linear frequency modulation wavelet transformation, tensor decomposition and dynamic time warping, extracts crack precursor features through graph convolution and Transformer, and outputs crack risk probability with a confidence interval through a Bayesian network; a non-singular terminal sliding mode controller is constructed based on the risk probability and threshold constraint, quantum particle swarm optimization is used to optimize the approaching law parameters, and the sliding mode margin is allocated through opportunity constraint programming, so that the three grinding parameters are constructed into a non-cooperative game to solve the Nash equilibrium, and the gradient aggregation and ant colony mechanism are used to suppress the control chattering output linkage control quantity and the grinding wheel dressing parameters. The application reduces grinding cracks.

Description

Technical Field

[0001] This invention belongs to the field of precision machining control technology, specifically relating to a multi-dimensional anti-cracking adaptive control method for grinding alumina ceramics. Background Technology

[0002] In the grinding of alumina ceramics, the material is brittle and prone to cracking. Existing conventional crack prevention control methods typically employ a single physical quantity threshold triggering mechanism. Specifically, a force gauge is installed on the grinding spindle or worktable, or a thermal imager is placed in the machining area to acquire the grinding force or surface temperature in real time. The acquired physical quantity is compared with a preset safety threshold. When the physical quantity exceeds the threshold, a control command is triggered. The controller uses a proportional-integral-derivative algorithm to adjust the grinding spindle speed or feed rate independently based on the deviation value. Meanwhile, wheel dressing relies on manual experience and is performed offline according to a fixed machining duration.

[0003] The core problem with the above-mentioned existing technical solutions is that: single physical quantity threshold comparison can only reflect the mechanical or thermal abrupt changes that have occurred, and cannot extract the precursor features of multi-physics field coupling before crack initiation, resulting in time lag in crack prevention control; the proportional-integral-derivative algorithm can only perform single-dimensional decoupling adjustment of rotational speed or feed rate, and cannot achieve multi-dimensional linkage adaptive adjustment of spindle speed, feed rate and grinding depth under the mapping constraints of grinding force, temperature and material grain size. Single parameter adjustment is prone to breaking the grinding thermo-mechanical coupling balance, causing control overshoot and leading to workpiece cracking. Summary of the Invention

[0004] The purpose of this invention is to provide a solution to the problems described in the background section.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A multi-dimensional crack prevention adaptive control method for grinding alumina ceramics includes: A fracture mechanics constitutive model for the grinding process of alumina ceramics was established to quantify the mapping relationship between grinding force, grinding temperature, material grain size and crack initiation and propagation threshold. Acoustic emission signals and surface temperature field data during the grinding process are collected synchronously, and after preprocessing, they are input into the Transformer temporal feature prediction network to extract the crack initiation precursor feature sequence and output the crack risk probability for multiple future control cycles. Using the crack risk probability and the threshold constraint of the fracture mechanics constitutive model as the sliding surface design parameters, a non-singular terminal sliding mode adaptive controller is constructed to solve and output the linkage control quantities of grinding spindle speed, feed rate and grinding depth, and simultaneously correct the grinding wheel dressing parameters. By updating the sensing data and prediction results according to a fixed control cycle, and iteratively correcting the control parameters, adaptive closed-loop control for multi-dimensional crack prevention is achieved.

[0006] Preferably, the process of establishing a fracture mechanics constitutive model includes: introducing thermo-elastic-plastic mechanics finite element simulation to obtain the dynamic coupling distribution of stress field and temperature field in the grinding zone; Combining the anisotropic characteristics of alumina ceramic microcrystals, a random polygonal grain topology is generated using a Vinio diagram. The dynamic coupling distribution is input as a boundary condition into the random polygonal grain topology. The extended finite element method is used to simulate the crack propagation path along grain boundaries or transgranular paths. By traversing different grain sizes and grinding parameter combinations, the mapping relationship is fitted and generated.

[0007] Preferably, the preprocessing of acoustic emission signals and surface temperature field data includes: using linear frequency modulated wavelet transform in radar micro-Doppler signal processing to perform time-frequency focusing on the acoustic emission signals and extract high-frequency transient impact components; Tensor decomposition is performed on the surface temperature field data to remove background noise from environmental thermal radiation. The extracted high-frequency transient impact components are then aligned across modes with the noise-free temperature field time series using a dynamic time warping method to generate a heterogeneous fused time series tensor.

[0008] Preferably, the process of inputting heterogeneous fused temporal tensors into the Transformer temporal feature prediction network includes: constructing a spatial topology graph of the acoustic emission sensor array, and using a graph convolutional network to extract spatial feature vectors between array nodes; Spatial feature vectors are concatenated with heterogeneous fused temporal tensors and then input into a Transformer encoder. Relative position encoding and sparse masking are introduced into the self-attention mechanism to suppress interference from irrelevant historical information in long sequences and output precursor feature sequences.

[0009] Preferably, the process of constructing a non-singular terminal sliding mode adaptive controller includes: setting the dynamic sliding mode surface boundary based on the crack risk probability, and transforming the threshold constraint of the fracture mechanics constitutive model into a penalty term for the control weight; A quantum particle swarm optimization method is introduced, with the weighted sum of the control response overshoot and steady-state error as the fitness function. Under the constraint of the penalty term, the reaching law parameters in the sliding surface are iteratively updated to complete the initial construction of the non-singular terminal sliding mode adaptive controller.

[0010] Preferably, the process of iteratively updating the reaching law parameters in the sliding surface includes: introducing Tent chaotic mapping to initialize particle positions in the quantum particle swarm optimization method to enhance the ergodicity of the initial parameter distribution; The Cauchy mutation operator is superimposed on the particle position update equation. When the fitness function value has not changed for several consecutive generations, Cauchy mutation is triggered, which drives the particle to jump out of the local minimum domain and obtain the globally optimal reaching law parameters.

[0011] Preferably, the process of outputting the crack risk probability for multiple future control cycles includes: connecting a Bayesian neural network to the back end of the Transformer encoder to model the probability distribution of the precursor feature sequence; By introducing random perturbations into the precursor feature sequence through Monte Carlo dropout sampling, the output includes a crack risk probability interval containing an upper confidence bound and a lower confidence bound, thus quantifying the uncertainty of the prediction results.

[0012] Preferably, the process of designing the sliding surface based on the crack risk probability interval includes: introducing a chance-constrained programming model from operations research, taking the probability that the upper bound of confidence does not exceed the safety threshold as the objective function, taking the physical feasible domain of grinding spindle speed, feed rate and grinding depth as the constraint condition, solving the chance-constrained programming model, and dynamically allocating the sliding surface width margin for the current control cycle.

[0013] Preferably, the process of calculating and outputting the linkage control quantity includes: constructing the grinding spindle speed, feed rate and grinding depth as three non-cooperative game players, with their respective corresponding processing efficiency and surface quality preference multi-utility functions as the payoffs; Using the dynamically allocated sliding surface width margin as a global constraint, the Nash equilibrium point of the non-cooperative game is solved, and the parameter combination corresponding to the Nash equilibrium point is used as the output of the linkage control quantity.

[0014] Preferably, the process of solving the Nash equilibrium point of a non-cooperative game includes: using a gradient aggregation strategy in federated learning to perform historical exponential smoothing aggregation of the payoff gradients of the three players during the process of solving the Nash equilibrium point in adjacent control cycles. By introducing the pheromone evaporation mechanism from biomimetic ant colony optimization, the gradient direction after historical exponential smoothing aggregation is adaptively decayed, suppressing the drastic fluctuations of the linkage control quantity in adjacent control cycles and eliminating control chattering.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention simultaneously acquires acoustic emission signals and surface temperature field data, processes them through linear frequency modulated wavelet transform and tensor decomposition, and extracts precursor feature sequences using a graph convolutional network and a converter encoder. It then utilizes a Bayesian neural network to output a crack risk probability interval with confidence intervals, achieving multi-physics coupling precursor feature extraction before crack initiation and eliminating the time lag caused by single-physical quantity threshold comparison. A chance-constrained programming model is introduced to allocate the sliding surface width margin, and the spindle speed, feed rate, and grinding depth are constructed as non-cooperative game players to solve for the Nash equilibrium point. The sliding surface width margin is used as a global constraint to output the linkage control quantity, replacing single-dimensional decoupling adjustment and avoiding control overshoot caused by single-parameter adjustment disrupting the grinding thermo-mechanical coupling balance.

[0016] 2. This invention uses a Venn diagram to generate a random polygonal grain topology and combines it with the extended finite element method to establish a fracture mechanics constitutive model. The dynamic thermo-mechanical coupling distribution is used as a boundary condition, improving the physical accuracy of crack propagation threshold mapping under different grain sizes. In the parameter optimization of the non-singular terminal sliding mode adaptive controller, tent chaotic mapping initialization and Cauchy mutation operators are introduced to prevent the reaching law parameters from getting trapped in local minima. In solving the Nash equilibrium point, federated learning gradient aggregation combined with an ant colony pheromone evaporation mechanism is used to adaptively decay the historical gain gradient, smoothing the variation amplitude of the linkage control quantity in adjacent control cycles and reducing the chattering phenomenon inherent in terminal sliding mode control. Attached Figure Description

[0017] Figure 1 This is the overall flowchart of the multi-dimensional anti-cracking adaptive control for alumina ceramic grinding according to the present invention; Figure 2 This is a flowchart illustrating the construction process of the fracture mechanics constitutive model of the present invention. Figure 3 This is a flowchart of the multi-source data preprocessing and feature extraction process of the present invention; Figure 4 This is a flowchart of the crack risk probability prediction process of the present invention; Figure 5 This is a flowchart illustrating the construction process of the non-singular terminal sliding mode controller of the present invention. Figure 6 This is a flowchart of the linkage control quantity solution and chatter suppression of the present invention. Detailed Implementation

[0018] As a preferred embodiment, please attach Figure 1 To be continued Figure 3The industrial control unit pre-establishes a fracture mechanics constitutive model for the alumina ceramic grinding process. This model quantifies the mapping relationship between grinding force, grinding temperature, material grain size, and crack initiation and propagation thresholds. Specifically, the grinding force includes normal and tangential grinding forces; the grinding temperature is the peak temperature of the grinding contact zone; the material grain size is the average grain diameter of the alumina ceramic to be processed; and the crack initiation and propagation thresholds include the critical stress intensity factor threshold for crack initiation and the critical energy release rate threshold for crack propagation.

[0019] During the alumina ceramic grinding process, the acoustic emission sensing unit and the infrared thermal imaging unit synchronously acquire acoustic emission signals and surface temperature field data during the grinding process, and transmit the acquired raw data to the industrial control unit in real time. The industrial control unit preprocesses the received acoustic emission signals and surface temperature field data to generate a heterogeneous fused temporal tensor that meets the input format requirements. This heterogeneous fused temporal tensor is then input into a pre-trained Transformer temporal feature prediction network. The Transformer temporal feature prediction network extracts the crack initiation precursor feature sequence corresponding to the grinding process and outputs the crack risk probability for the next N control cycles, where N is a positive integer greater than or equal to 2, and the duration of the control cycle is in the range of 1 ms. 10ms, which can be adaptively adjusted according to the processing accuracy requirements.

[0020] The industrial control unit uses the crack risk probability and the crack initiation and propagation threshold constraint output by the fracture mechanics constitutive model as the sliding surface design parameters to construct a non-singular terminal sliding mode adaptive controller. Based on the input sliding surface design parameters, the non-singular terminal sliding mode adaptive controller calculates and outputs the linkage control quantities of the grinding spindle speed, feed rate and grinding depth, and simultaneously generates correction instructions for the grinding wheel dressing parameters, including dressing feed, dressing depth and dressing linear speed.

[0021] The industrial control unit synchronously updates the sensing and acquisition results of acoustic emission signals and surface temperature field data, as well as the crack risk probability prediction results output by the Transformer temporal feature prediction network, according to a fixed control cycle. Based on the updated parameters, the control parameters of the non-singular terminal sliding mode adaptive controller are iteratively corrected to achieve multi-dimensional crack prevention adaptive closed-loop control in the alumina ceramic grinding process.

[0022] As a preferred embodiment, the industrial control unit incorporates thermo-elastic-plastic mechanics finite element simulation to obtain the dynamic coupling distribution of the stress and temperature fields in the grinding zone. Specifically, a macroscopic finite element model of alumina ceramic grinding is constructed, and the thermophysical and mechanical properties of the alumina ceramic to be processed are set, including elastic modulus, Poisson's ratio, thermal conductivity, specific heat capacity, coefficient of thermal expansion, yield strength, and fracture toughness. The abrasive grain distribution model and motion boundary conditions of the grinding wheel are set. Using a sequential coupling method, the temperature field distribution during the grinding process is first solved through transient heat conduction analysis. Then, the temperature field results are input as a body load to the static structure analysis module to solve the stress field distribution in the grinding zone. Finally, the dynamic coupling distribution of the stress and temperature fields in the grinding zone is obtained. The mathematical expression for this dynamic coupling distribution is: in, Let t be the coordinates within the grinding zone. The stress tensor components at that point This represents the temperature value at the corresponding location and time. Let be the normal grinding force at time t. Let be the tangential grinding force at time t. Let be the grinding spindle speed at time t. Let be the feed rate at time t. Let t be the grinding depth at time t.

[0023] The industrial control unit, taking into account the anisotropic characteristics of alumina ceramic microcrystals, uses a Vinio diagram to generate a random polygonal grain topology. Specifically, based on the average grain size and grain size distribution range of the alumina ceramic to be processed, random seed points conforming to a normal distribution are generated in a two-dimensional plane. A Vinio diagram is generated based on these random seed points, with each Vinio unit corresponding to an alumina ceramic grain, and the common edge of adjacent Vinio units corresponding to grain boundaries. Each grain unit is assigned elastic modulus and fracture toughness parameters conforming to the anisotropy of alumina ceramic single crystals, thus completing the construction of the random polygonal grain topology.

[0024] The industrial control unit inputs the aforementioned dynamically coupled distribution as boundary conditions into the random polygonal grain topology, and uses the extended finite element method to simulate the crack propagation path along grain boundaries or transgranular structures. Specifically, an enrichment term is introduced into the displacement approximation function of the extended finite element method. This enrichment term includes the step function of the crack surface and the asymptotic displacement function of the crack tip. The expression for the displacement approximation function is: in, This is the spatial coordinate vector of the calculation point in the two-dimensional finite element model. Let I be the shape function of node I; Let I be the normal displacement degrees of freedom for node I. Let be the step function of the crack surface. α represents the enriched degrees of freedom corresponding to the step function, and α is the term index variable of the asymptotic displacement function at the crack tip. Let be the asymptotic displacement function at the crack tip. Let N be the set of all nodes in the model, where N represents the enriched degrees of freedom at the crack tip.

[0025] By using extended finite element method (FEA) solutions, the critical stress intensity factor for crack initiation and the critical energy release rate for crack propagation under different boundary conditions are obtained. The industrial control unit traverses different combinations of grain size and grinding parameters, including multiple orthogonal combinations of grinding spindle speed, feed rate, and grinding depth. Through multivariate nonlinear fitting, the mapping relationship between grinding force, grinding temperature, material grain size, and crack initiation and propagation threshold is generated, thus completing the construction of the fracture mechanics constitutive model.

[0026] In a preferred embodiment, the industrial control unit utilizes linear frequency-modulated wavelet transform from radar micro-Doppler signal processing to perform time-frequency focusing on the acoustic emission signal and extract high-frequency transient impact components. Specifically, for the original acoustic emission signal... Perform a linear frequency modulated wavelet transform, with the kernel function being the linear frequency modulated wavelet basis function. The transform expression is: in, The linear frequency modulated wavelet transform coefficients of the acoustic emission signal are denoted as . for The amplitude of the original acoustic emission signal at the corresponding moment; For the time-domain integral variable of the linear frequency modulated wavelet transform, it represents the time shift during the wavelet transform process and corresponds to the time sampling time of the original acoustic emission signal; For the mother wavelet function, The conjugate of the mother wavelet function, For frequency variables, For time variables, This is a linear frequency modulation coefficient used to adjust the time-frequency focusing performance.

[0027] By using linear frequency modulated wavelet transform, the original acoustic emission signal is converted into a time-frequency domain distribution. Based on the frequency range of the acoustic emission signal corresponding to the crack initiation of alumina ceramics, a bandpass filter interval is set to extract the high-frequency transient impact component in the corresponding time-frequency domain, thereby filtering out low-frequency mechanical vibration noise and high-frequency electrical noise.

[0028] The industrial control unit performs tensor decomposition on the surface temperature field data to remove ambient thermal radiation background noise. Specifically, it constructs a third-order tensor from M consecutive frames of surface temperature field data. Where H is the height of the temperature field image, W is the width of the temperature field image, and M is the number of consecutive frames; a high-order orthogonal iterative algorithm is used to perform Tucker decomposition on the third-order tensor, and the decomposition expression is: in, For the core tensor, , , These are the factor matrices corresponding to the height, width, and time dimensions, respectively. The residual tensor is used as the low-rank component of the corresponding time dimension. The low-rank component is taken as the background noise of environmental thermal radiation and extracted from the original tensor to obtain the time series sequence of the temperature field after removing the background noise.

[0029] The industrial control unit utilizes a dynamic time warping method to perform cross-modal time alignment between the extracted high-frequency transient impact component and the time series sequence of the temperature field after noise stripping, generating a heterogeneous fused time series tensor. Specifically, the time series sequence of the high-frequency transient impact component is used as the reference sequence. The temperature field time series is used as the sequence to be aligned. Construct an L×K distance matrix The distance matrix elements The optimal normalized path of the distance matrix is ​​found through dynamic programming, which minimizes the cumulative distance. The recursive formula for the cumulative distance is: in, To reach the location The minimum cumulative distance. Based on the optimal regularization path, the time dimensions of the two heterogeneous time series are aligned. The aligned high-frequency transient impact component is then concatenated with the temperature field time series along the feature dimension, generating a dimension of... The heterogeneous fusion temporal tensor, where T is the aligned temporal length and F is the feature dimension.

[0030] In a preferred embodiment, the industrial control unit constructs a spatial topology graph of the acoustic emission sensor array and extracts spatial feature vectors between array nodes using a graph convolutional network. Specifically, each sensor in the acoustic emission sensor array is treated as a graph node, and an adjacency matrix is ​​constructed based on the actual installation location of the sensor. Adjacency matrix elements The value is: when the spatial distance between node i and node j is less than a preset threshold. ,otherwise Construct the degree matrix ,in A two-layer graph convolutional network is used to extract spatial features from the acoustic emission signals of the sensor array. The layer propagation formula for graph convolution is: in, The adjacency matrix for adding self-loops, where I is the identity matrix. for The corresponding degree matrix, Let l be the input feature matrix of the l-th layer. Let L be the trainable weight matrix of the l-th layer. It is a non-linear activation function. Let be the output feature matrix of the l-th layer.

[0031] Spatial feature vectors between nodes of the acoustic emission sensor array are extracted through forward propagation of a two-layer graph convolutional network. The dimension of the spatial feature vectors matches the feature dimension of the heterogeneous fusion temporal tensor.

[0032] The industrial control unit concatenates spatial feature vectors and heterogeneous fused temporal tensors along the feature dimension, then inputs them into a Transformer encoder. Relative position encoding and sparse masking are introduced into the self-attention mechanism to suppress interference from irrelevant historical information in long sequences, outputting a precursor feature sequence. Specifically, the Transformer encoder comprises N sequentially stacked encoding layers. Each encoding layer includes a multi-head self-attention module and a feedforward neural network module. The outputs of both modules are connected to a layer normalization module and a residual connection structure.

[0033] In the multi-head self-attention module, the relative position encoding is introduced as follows: A relative positional deviation term between the query vector and the key vector is added to the attention score calculation. The formula for calculating the attention score is: Where Q is the query matrix, K is the key matrix, and V is the value matrix. Let P be the dimension of the key vector, and P be the relative position encoding matrix.

[0034] The sparse mask is a lower triangular sparse mask. In the mask matrix, only the elements of the current time and the historical time of a preset length are kept as 1, and the rest are 0. This is used to mask irrelevant historical information that exceeds the preset length in the long sequence, reducing computational complexity while suppressing historical interference. After multi-layer encoding processing by the Transformer encoder, the output is the crack initiation precursor feature sequence corresponding to the grinding process.

[0035] The industrial control unit connects a Bayesian neural network to the back end of the Transformer encoder to model the probability distribution of the precursor feature sequence. By introducing random perturbations into the precursor feature sequence through Monte Carlo dropout sampling, it outputs a crack risk probability interval containing upper and lower confidence bounds, quantifying the uncertainty of the prediction result. Specifically, the Bayesian neural network consists of two fully connected layers, each followed by a dropout layer with a dropout rate set to 0.1. 0.3; During the model inference phase, the dropout layer remains active, and S Monte Carlo dropout samplings are performed on the precursor feature sequence to obtain S sets of predicted crack risk probabilities. Based on the S sets of predicted values, the mean and standard deviation are calculated, and the mean of the crack risk probability is... with standard deviation The calculation formula is:

[0036] in, The crack risk probability prediction value is the output of the discarded sample in the s-th Monte Carlo iteration. Based on a preset confidence level, the upper and lower confidence bounds of the crack risk probability are calculated to generate a crack risk probability interval. The upper confidence bound... and the lower bound of confidence The calculation formula is:

[0037] in, The quantiles of the standard normal distribution correspond to the preset confidence level.

[0038] As a preferred embodiment, refer to the appendix. Figure 4 and attached Figure 5 The industrial control unit sets a dynamic sliding surface boundary based on the crack risk probability, transforming the threshold constraint of the fracture mechanics constitutive model into a penalty term for the control weights. Specifically, a non-singular terminal sliding surface is constructed, and the expression for the sliding surface is: in, The system state error is the difference between the current crack risk probability and the safe crack risk probability corresponding to the crack initiation and propagation threshold output by the fracture mechanics constitutive model. The sliding surface coefficient is positive. The coefficients are power coefficients and satisfy the following conditions: This ensures that the sliding surface is non-singular at the point where the system state error is 0, thus avoiding the singularity problem of traditional terminal sliding surfaces.

[0039] The dynamic sliding surface boundary is set based on the upper confidence bound of the crack risk probability. When the upper confidence bound of the crack risk probability increases, the boundary range of the sliding surface is reduced to improve the control response speed. The crack initiation and propagation threshold constraint output by the fracture mechanics constitutive model is transformed into a penalty term for the control weight. The penalty term is used to limit the output range of the control quantity to avoid the control quantity from exceeding the threshold constraint and causing crack initiation.

[0040] The industrial control unit incorporates a quantum particle swarm optimization method, using the weighted sum of control response overshoot and steady-state error as the fitness function. Under penalty constraints, iteratively updates the reaching law parameters in the sliding surface, completing the initial construction of the non-singular terminal sliding mode adaptive controller. Specifically, the reaching law adopts a fast power-law reaching law, expressed as: in, , For the parameters to be optimized in the approach law, The coefficients are power coefficients and satisfy the following conditions: .

[0041] The expression for the fitness function is: in, To control the overshoot of the response, To control the steady-state error of the response, , These are weighting coefficients. P is the penalty coefficient, which is a penalty term generated based on the threshold constraint of the fracture mechanics constitutive model. When the control quantity exceeds the threshold constraint, P takes a positive maximum value; otherwise, P takes 0.

[0042] In quantum particle swarm optimization, particle position updates are based on the probabilistic properties of quantum potential wells, and the particle position update equation is:

[0043] in, Let be the position of the i-th particle at time t. Let be the optimal position for the i-th particle. This represents the global optimal position of the particle swarm. is the contraction / expansion coefficient, and u is a random number between 0 and 1.

[0044] The industrial control unit incorporates Tent chaotic mapping into the quantum particle swarm optimization method to initialize particle positions, enhancing the ergodicity of the initial parameter distribution. The expression for the Tent chaotic mapping is: in, For the chaotic variable in the nth iteration, the initial position of the particle swarm is initialized based on the chaotic sequence generated by the Tent chaotic map, ensuring that the initial particles are uniformly distributed in the parameter search space.

[0045] The industrial control unit superimposes a Cauchy mutation operator into the particle position update equation. When the fitness function value remains unchanged for several consecutive generations, Cauchy mutation is triggered, driving the particle out of the local minimum region and obtaining the globally optimal reaching law parameters. The expression for the Cauchy mutation operator is: in, The position of the mutated particle. Let be a random number derived from the standard Cauchy distribution. The probability density function of the standard Cauchy distribution is: By leveraging the wide tail characteristic of Cauchy mutation, the global search capability of the particle swarm is enhanced, avoiding parameter optimization from getting trapped in local minima. Ultimately, the globally optimal reaching law parameters are obtained, completing the initialization construction of the non-singular terminal sliding mode adaptive controller.

[0046] As a preferred embodiment, refer to the appendix. Figure 6 The industrial control unit incorporates a chance-constrained programming model from operations research. The objective function is to satisfy a set condition with the probability that the upper bound of the confidence threshold does not exceed it. The constraints are the physical feasible regions of the grinding spindle speed, feed rate, and grinding depth. The chance-constrained programming model is solved to dynamically allocate the sliding surface width margin for the current control cycle. Specifically, the expression for the chance-constrained programming model is: in, This is the width margin of the sliding surface. This represents the upper bound of the crack risk probability confidence level under the corresponding sliding surface width margin. This is the safe threshold for the probability of crack risk. The preset confidence probability, , These are the minimum and maximum physical limits for the grinding spindle speed, respectively. , These are the minimum and maximum physical limits for the feed rate, respectively. , These are the minimum and maximum physical limits for grinding depth, respectively.

[0047] By solving the above chance-constrained programming model using a convex optimization algorithm, the optimal sliding surface width margin under the current control cycle is obtained, thus completing the dynamic design of the sliding surface.

[0048] The industrial control unit constructs the grinding spindle speed, feed rate, and grinding depth as three non-cooperative game players, each with their respective processing efficiency and surface quality preference utility functions as payoffs. Using dynamically allocated sliding surface width margin as a global constraint, it solves for the Nash equilibrium point of the non-cooperative game, and outputs the parameter combination corresponding to the Nash equilibrium point as the linkage control quantity. Specifically, the three game players are: Player 1 corresponds to the grinding spindle speed... The feed rate corresponding to player 2 in the game. Game player 3 corresponds to grinding depth Each player's strategy space is the physical feasible region of the corresponding parameters, and the global constraint is the crack risk probability constraint corresponding to the sliding surface width margin.

[0049] Each player's multi-utility function, i.e., payoff function, taking into account both processing efficiency and surface quality, is expressed as follows: in, Let i be the strategy variable of player i. Let i be the combination of strategy variables for the other two players besides player i. The material removal rate is an indicator. Surface roughness index , Let be the preference weight coefficients of player i for the two indicators.

[0050] The Nash equilibrium point is a strategy combination that satisfies the following conditions. For any player i, any strategy variable All satisfy This means that each player's payoff is optimal when the strategies of the other players are fixed.

[0051] The industrial control unit employs a gradient aggregation strategy from federated learning. During the Nash equilibrium point solution process in adjacent control cycles, it performs historical exponential smoothing aggregation of the payoff gradients of the three players. It also introduces a pheromone evaporation mechanism from biomimetic ant colony optimization to adaptively decay the gradient direction after historical exponential smoothing aggregation, suppressing drastic fluctuations in the coordinated control quantities between adjacent control cycles and eliminating control chattering. Specifically, the expression for historical exponential smoothing aggregation is: in, Let i be the aggregate gradient of player i in the t-th control cycle. Let be the real-time gradient of the payoff function of player i in the t-th control cycle. This is a smoothing coefficient, with a value ranging from 0 to 1. Let be the aggregate gradient of player i in the (t-1)th control cycle.

[0052] The adaptive decay expression for the pheromone volatilization mechanism is as follows: in, The aggregate gradient after decay. is the pheromone evaporation coefficient, ranging from 0 to 1, and k is the number of iterations in which the crack risk probability does not change significantly within a continuous control period.

[0053] The Nash equilibrium point is iteratively solved based on the decayed aggregate gradient, which smooths the gradient change amplitude of adjacent control cycles, suppresses the drastic jump of the linkage control quantity, eliminates the control chattering phenomenon inherent in the non-singular terminal sliding mode adaptive controller, and finally obtains stable linkage control quantity and grinding wheel dressing parameter correction instructions, which are output to the servo drive unit and grinding wheel dressing unit for execution.

Claims

1. A multi-dimensional crack prevention adaptive control method for grinding alumina ceramics, characterized in that, include: A fracture mechanics constitutive model for the grinding process of alumina ceramics was established to quantify the mapping relationship between grinding force, grinding temperature, material grain size and crack initiation and propagation threshold. Acoustic emission signals and surface temperature field data during the grinding process are collected synchronously, and after preprocessing, they are input into the Transformer temporal feature prediction network to extract the crack initiation precursor feature sequence and output the crack risk probability for multiple future control cycles. Using the crack risk probability and the threshold constraint of the fracture mechanics constitutive model as the sliding surface design parameters, a non-singular terminal sliding mode adaptive controller is constructed to solve and output the linkage control quantities of grinding spindle speed, feed rate and grinding depth, and simultaneously correct the grinding wheel dressing parameters. The sensing data and prediction results are updated according to a fixed control cycle, and the control parameters are iteratively corrected.

2. The multi-dimensional crack prevention adaptive control method for alumina ceramic grinding according to claim 1, characterized in that, The process of establishing a constitutive model for fracture mechanics includes: Thermo-elastic-plastic mechanics finite element simulation is introduced to obtain the dynamic coupling distribution of stress field and temperature field in the grinding zone; combined with the anisotropic characteristics of alumina ceramic micrograins, a random polygonal grain topology is generated using a Vinyson diagram. The dynamic coupling distribution is input as a boundary condition into the random polygonal grain topology. The extended finite element method is used to simulate the crack propagation path along grain boundaries or transgranular paths. By traversing different grain sizes and grinding parameter combinations, the mapping relationship is fitted and generated.

3. The multi-dimensional crack prevention adaptive control method for grinding alumina ceramics according to claim 1, characterized in that, The preprocessing of acoustic emission signals and surface temperature field data includes: By drawing on the linear frequency modulated wavelet transform in radar micro-Doppler signal processing, time-frequency focusing is performed on the acoustic emission signal to extract the high-frequency transient impact component; Tensor decomposition is performed on the surface temperature field data to remove background noise from environmental thermal radiation. The extracted high-frequency transient impact components are then aligned across modes with the noise-free temperature field time series using a dynamic time warping method to generate a heterogeneous fused time series tensor.

4. The multi-dimensional crack prevention adaptive control method for alumina ceramic grinding according to claim 3, characterized in that, The process of inputting heterogeneous fused temporal tensors into the Transformer temporal feature prediction network includes: Construct a spatial topology graph of the acoustic emission sensor array, and use a graph convolutional network to extract spatial feature vectors between array nodes; Spatial feature vectors are concatenated with heterogeneous fused temporal tensors and then input into a Transformer encoder. Relative position encoding and sparse masking are introduced into the self-attention mechanism to suppress interference from irrelevant historical information in long sequences and output precursor feature sequences.

5. The multi-dimensional crack prevention adaptive control method for grinding alumina ceramics according to claim 1, characterized in that, The process of constructing a non-singular terminal sliding mode adaptive controller includes: Based on the crack risk probability, the dynamic sliding surface boundary is set, and the threshold constraint of the fracture mechanics constitutive model is transformed into a penalty term for the control weight. The quantum particle swarm optimization method is introduced, and the fitness function is the weighted sum of the control response overshoot and steady-state error. Under the penalty term constraint, the reaching law parameters in the sliding surface are iteratively updated to complete the initial construction of the non-singular terminal sliding mode adaptive controller.

6. The multi-dimensional crack prevention adaptive control method for grinding alumina ceramics according to claim 5, characterized in that, The process of iteratively updating the reaching law parameters in the sliding surface includes: In the quantum particle swarm optimization method, Tent chaotic mapping is introduced to initialize particle positions, thereby enhancing the ergodicity of the initial parameter distribution. The Cauchy mutation operator is superimposed on the particle position update equation. When the fitness function value has not changed for several consecutive generations, Cauchy mutation is triggered, which drives the particle to jump out of the local minimum domain and obtain the globally optimal reaching law parameters.

7. The multi-dimensional crack prevention adaptive control method for grinding alumina ceramics according to claim 4, characterized in that, The process of outputting the crack risk probability for multiple future control cycles includes: A Bayesian neural network is connected to the back end of the Transformer encoder to model the probability distribution of the precursor feature sequence. By introducing random perturbations into the precursor feature sequence through Monte Carlo dropout sampling, the output includes a crack risk probability interval containing an upper confidence bound and a lower confidence bound, thus quantifying the uncertainty of the prediction results.

8. The multi-dimensional crack prevention adaptive control method for grinding alumina ceramics according to claim 7, characterized in that, The process of designing a sliding mode surface based on the crack risk probability range includes: We introduce the opportunity-constrained programming model from operations research, with the objective function being the probability that the upper bound of confidence does not exceed the safety threshold and the physical feasible region of grinding spindle speed, feed rate and grinding depth as constraints. We solve the opportunity-constrained programming model and dynamically allocate the sliding surface width margin for the current control cycle.

9. The multi-dimensional crack prevention adaptive control method for grinding alumina ceramics according to claim 8, characterized in that, The process of calculating and outputting the linkage control quantity includes: The grinding spindle speed, feed rate, and grinding depth are constructed as three non-cooperative game players, with their respective processing efficiency and surface quality preference utility functions as their payoffs. Using the dynamically allocated sliding surface width margin as a global constraint, the Nash equilibrium point of the non-cooperative game is solved, and the parameter combination corresponding to the Nash equilibrium point is used as the output of the linkage control quantity.

10. The multi-dimensional crack prevention adaptive control method for grinding alumina ceramics according to claim 9, characterized in that, The process of finding the Nash equilibrium point in a non-cooperative game includes: The gradient aggregation strategy in federated learning is adopted to perform historical exponential smoothing aggregation of the payoff gradients of the three players during the process of solving the Nash equilibrium point in adjacent control cycles. A pheromone evaporation mechanism from biomimetic ant colony optimization is introduced to adaptively decay the gradient direction after historical exponential smoothing aggregation.

Images