Tool structure data prediction method, device, equipment and storage medium

Abstract

The present application relates to the technical field of tool structure data prediction, and particularly relates to a tool structure data prediction method, device, equipment and storage medium, the method firstly constructs a data sample library for training, performs a simulation operation based on the data sample library to construct a training sample set, performs multi-round iteration training on an initial index prediction model by using the training sample set, calls a target index prediction model to respectively predict machining key indexes of each group of tool structure data to be predicted, obtains an index data group corresponding to each group of tool structure data to be predicted, obtains an initial tool structure data group based on multiple groups of tool structure data to be predicted, adopts a genetic algorithm, and performs multi-round iteration update on the initial tool structure data group based on the multiple groups of tool structure data to be predicted and multiple groups of index data, and when a preset iteration stop condition is met, an optimized tool structure data group is obtained, aiming at accurately predicting key physical quantities of tool structure.

Description

Technical Field

[0001] This invention relates to the field of tool structure data prediction technology, and in particular to a tool structure data prediction method, apparatus, device and storage medium. Background Technology

[0002] Fused silica, with its superior performance, plays a crucial role in aerospace, laser electronics, and the manufacturing of high-generation inkjet-printed OLEDs (organic light-emitting diodes). Key components made from fused silica, such as printhead microplates and mask substrates, place stringent demands on surface accuracy, integrity, and hole precision in ultra-precision machining. In ultra-precision grinding, subsurface damage directly impacts component performance and reliability, affecting display pixel accuracy and production yield, thus becoming a core bottleneck restricting processing quality.

[0003] To meet the demands for high precision and low defects, in-situ laser-assisted ultra-precision machining technology and negative rake angle tools have been applied. The former can improve the machining performance of fused silica in the ductile region and reduce cracks, while the latter can suppress brittle fracture and promote plastic removal by enhancing hydrostatic pressure. The tool rake angle, clearance angle, and machining temperature are key parameters affecting the machining effect. Reasonable matching of the three can optimize the material removal mode, but excessively high temperatures can easily cause thermal damage. The clearance angle design needs to balance cutting stability and tool strength.

[0004] Current research still faces many limitations. The removal mechanism in fused silica processing is not fully understood, and microscopic changes are difficult to observe directly. Traditional finite element methods are prone to mesh distortion when simulating large deformations and crack propagation, and traditional prediction methods such as response surface models are unable to capture the complex relationships of highly nonlinear systems. Existing tool optimization patents are mostly aimed at conventional machining and are not fully adapted to the micro- and nano-scale characteristics and brittle-plastic mechanisms of ultra-precision machining, resulting in insufficient versatility and accuracy. Summary of the Invention

[0005] In order to overcome the shortcomings of the prior art, the present invention aims to provide a tool structure data prediction method, device, equipment and storage medium, which aims to accurately predict the key physical quantities of the tool structure, realize the efficient optimization of tool structure and machining parameters, and improve machining quality and efficiency.

[0006] The first aspect of this invention provides a method for predicting tool structure data, comprising: constructing a data sample library for training; performing simulation operations based on the data sample library to construct a training sample set; constructing an initial index prediction model; performing multiple rounds of iterative training on the initial index prediction model using the training sample set to obtain a target index prediction model; acquiring multiple sets of tool structure data to be predicted; calling the target index prediction model to predict key machining indicators for each set of tool structure data to be predicted, thereby obtaining an index data set corresponding to each set of tool structure data to be predicted; acquiring an initial tool structure data set based on the multiple sets of tool structure data to be predicted; employing a genetic algorithm to perform multiple rounds of iterative updates on the initial tool structure data set based on the multiple sets of tool structure data to be predicted and the multiple sets of index data sets; and when a preset iteration stopping condition is met, using the updated tool structure data set as an optimized tool structure data set.

[0007] Optionally, in a first implementation of the first aspect of the present invention, the training sample set includes a training set and a test set, and the data sample library includes various tool processing parameters; the step of performing simulation operations based on the data sample library to construct the training sample set includes: using a full permutation and combination algorithm to perform full permutation and combination of the various tool processing parameters to obtain various input working condition combinations; performing simulation operations based on each working condition combination to obtain a simulation index data set corresponding to each working condition combination; integrating the various tool processing parameters and the simulation index data set corresponding to each working condition combination to obtain the training sample set; and dividing the training sample set into a training set and a test set based on a preset partitioning ratio.

[0008] Optionally, in a second implementation of the first aspect of the present invention, the initial indicator prediction model includes an initial Actor network and an initial Critic network; the step of performing multiple rounds of iterative training on the initial indicator prediction model using the training sample set to obtain the target indicator prediction model includes: constructing an initial experience replay pool based on the training sample set, and obtaining current state samples and next state samples from the initial experience replay pool; using the initial Actor network to predict the action of the current state sample to obtain the current action; obtaining a predefined reward function, calculating the reward value of the current action based on the reward function to obtain the current reward value; storing the current state sample, the current action, the current reward value, and the next state sample in the initial experience replay pool to obtain an expanded experience replay pool; and performing multiple rounds of iterative update processing on the initial Actor network and the initial Critic network based on the expanded experience replay pool to obtain the target indicator prediction model.

[0009] Optionally, in a third implementation of the first aspect of the present invention, the step of performing multiple rounds of iterative updates on the initial Actor network and the initial Critic network based on the expanded experience replay pool to obtain the target indicator prediction model includes: obtaining the current state sample, the current action, the current reward value, and the next state sample based on the expanded experience replay pool; using the initial Actor network to predict the action of the next state sample to obtain the next action; using the initial Critic network to perform value evaluation processing based on the current state sample, the current action, the next state sample, and the next action to obtain the current state value and the next state value; calculating the error based on the current reward value, the current state value, and the next state value to obtain the TD error; and performing multiple rounds of iterative updates on the initial Actor network and the initial Critic network based on the TD error to obtain the target indicator prediction model.

[0010] Optionally, in a fourth implementation of the first aspect of the present invention, the step of performing multiple rounds of iterative updates on the initial Actor network and the initial Critic network based on the TD error to obtain the target index prediction model includes: using a backpropagation algorithm to perform multiple rounds of iterative updates on the initial Critic network based on the TD error to obtain an optimized Critic network; using a policy gradient algorithm to perform multiple rounds of iterative updates on the initial Actor network based on the TD error to obtain an optimized Actor network; and combining the optimized Critic network and the optimized Actor network to form the target index prediction model.

[0011] Optionally, in a fifth implementation of the first aspect of the present invention, the genetic algorithm includes a roulette wheel algorithm, a crossover algorithm, and a mutation algorithm; the step of using the genetic algorithm to perform multiple rounds of iterative updates on the initial tool structure data set based on multiple sets of the tool structure data set to be predicted and multiple sets of the index data set, and when a preset iteration stopping condition is met, using the updated tool structure data set as the optimized tool structure data set, includes: using the roulette wheel algorithm to perform roulette wheel selection processing on multiple sets of the tool structure data set to be predicted based on multiple sets of the index data set to obtain multiple sets of parent tool structure data sets; using the crossover algorithm to perform gene exchange processing on multiple sets of the parent tool structure data sets to obtain multiple sets of crossover tool structure data sets; using the mutation algorithm to perform gene fine-tuning processing on multiple sets of the crossover tool structure data sets to obtain multiple sets of mutated tool structure data sets; and performing multiple rounds of iterative updates on the initial tool structure data set based on multiple sets of mutated tool structure data sets, and when a preset iteration stopping condition is met, using the updated tool structure data set as the optimized tool structure data set.

[0012] Optionally, in a sixth implementation of the first aspect of the present invention, the step of employing the roulette wheel selection algorithm to perform roulette wheel selection processing on multiple sets of tool structure data sets to be predicted based on multiple sets of index data sets to obtain multiple sets of parent tool structure data sets includes: obtaining a predefined fitness function; calculating the fitness of multiple sets of index data sets based on the fitness function to obtain fitness values ​​corresponding to the multiple sets of tool structure data sets to be predicted; employing the roulette wheel selection algorithm to calculate roulette wheel selection probabilities based on multiple fitness values ​​to obtain roulette wheel selection probabilities corresponding to the multiple sets of tool structure data sets to be predicted; obtaining a preset selection rule; and performing roulette wheel selection processing on multiple sets of tool structure data sets to be predicted based on the selection rule and multiple roulette wheel selection probabilities to obtain multiple sets of parent tool structure data sets.

[0013] A second aspect of the present invention provides a tool structure data prediction device, comprising: a simulation module for constructing a data sample library for training and performing simulation operations based on the data sample library to construct a training sample set; a training model for constructing an initial index prediction model and performing multiple rounds of iterative training on the initial index prediction model using the training sample set to obtain a target index prediction model; a prediction module for acquiring multiple sets of tool structure data to be predicted, calling the target index prediction model to predict key machining indicators for each set of tool structure data to be predicted, and obtaining an index data set corresponding to each set of tool structure data to be predicted; a data acquisition module for acquiring an initial tool structure data set based on the multiple sets of tool structure data to be predicted; and an optimization module for using a genetic algorithm to perform multiple rounds of iterative updates on the initial tool structure data set based on the multiple sets of tool structure data to be predicted and the multiple sets of index data sets, and when a preset iteration stopping condition is met, using the updated tool structure data set as the optimized tool structure data set.

[0014] A third aspect of the present invention provides a tool structure data prediction device, the tool structure data prediction device comprising: a memory and at least one processor, the memory storing instructions; the at least one processor calling the instructions in the memory to cause the tool structure data prediction device to perform the various steps of the tool structure data prediction method described in any of the preceding claims.

[0015] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the steps of the tool structure data prediction method described in any of the preceding claims.

[0016] In the technical solution of this invention, firstly, a data sample library for training is constructed. Simulation operations are performed based on the data sample library to construct a training sample set. Then, an initial index prediction model is constructed. The initial index prediction model is iteratively trained multiple times using the training sample set to obtain a target index prediction model. Next, multiple sets of tool structure data to be predicted are obtained. The target index prediction model is then used to predict key machining indicators for each set of tool structure data, resulting in an index data set corresponding to each set of tool structure data. An initial tool structure data set is obtained based on these multiple sets of tool structure data. Finally, a genetic algorithm is used to predict the initial tool structure data based on these multiple sets of tool structure data. The initial tool structure data set is iteratively updated multiple times by a set of index data sets and multiple sets of index data sets. When the preset iteration stopping condition is met, the updated tool structure data set is used as the optimized tool structure data set. The aim is to accurately predict the key physical quantities of the tool structure and achieve efficient optimization of tool structure and machining parameters to improve machining quality and efficiency. This improves the machining consistency, surface accuracy and surface quality of fused silica components, effectively extends tool life and reduces energy consumption. It can fully meet the extreme requirements of inkjet printing equipment for key quartz components in terms of micro-hole accuracy, positioning stability and wear resistance. It also provides a new way for large-scale, low-loss precision manufacturing of high value-added optical glass components. Attached Figure Description

[0017] Figure 1 A flowchart illustrating the tool structure data prediction method provided in this embodiment of the invention;

[0018] Figure 2 This is a schematic diagram of the tool structure data prediction device provided in an embodiment of the present invention;

[0019] Figure 3 This is a schematic diagram of the tool structure data prediction device provided in an embodiment of the present invention. Detailed Implementation

[0020] This invention provides a method, apparatus, device, and storage medium for predicting tool structure data. In this invention, the terms "first," "second," "third," "fourth," etc. (if present)," in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0021] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the tool structure data prediction method in this invention includes:

[0022] 101. Construct a data sample library for training, and perform simulation operations based on the data sample library to construct a training sample set;

[0023] In this embodiment, a data sample library for training is constructed, using the structural parameters of quartz glass and diamond cutting tools, along with cutting temperature, as core variables. Multiple sets of values ​​for the tool rake angle, tool clearance angle, and cutting temperature are determined as the data sample library. When performing simulation operations based on this sample library, a corresponding cutting simulation model is built in Abaqus software using the Smooth Particle Dynamics (SPH) method. The workpiece adopts the SPH particle model, and the model parameters can be set to dimensions... The total number of particles was set to 600,000, and the cutting tool was configured as a diamond tool with a Lagrange mesh. The model parameters were strictly matched to the physical properties of quartz glass and diamond, such as density, Poisson's ratio, and specific heat capacity. A training sample set was constructed based on simulation operations using the SPH method. This accurately captures the nonlinear behaviors such as large deformation and crack propagation during the cutting of hard and brittle materials. Compared with the traditional finite element method, it effectively avoids the mesh mismatch problem, making the output physical quantities in the sample set more consistent with the mechanical and material removal characteristics of actual machining.

[0024] 102. Construct an initial indicator prediction model, and use the training sample set to perform multiple rounds of iterative training on the initial indicator prediction model to obtain a target indicator prediction model;

[0025] In this embodiment, the initial index prediction model is constructed using the Actor-Critic deep reinforcement learning framework. The Actor network uses a neural network as its core structure, taking into account machining parameters such as the tool's rake angle, clearance angle, and cutting temperature, and outputting key machining physical quantities such as the shearing angle, shearing plane length, and maximum hydrostatic pressure. It is responsible for generating continuous prediction actions. The Critic network is also built on a neural network, taking into account the current machining parameters and the predicted physical quantities output by the Actor, and outputting the value assessment value corresponding to the state-action combination. This enables the quantitative judgment of the prediction result quality. Together, they constitute the core architecture of the initial index prediction model. During the multiple rounds of iterative training of the initial index prediction model, the model parameters are continuously adjusted and optimized based on the input parameters (tool rake angle, clearance angle, cutting temperature) and corresponding machining physical quantity data (maximum hydrostatic pressure, number of high-pressure particles, shear angle of the cutting area, and length of the shear plane) in the training sample set. This allows the model to gradually learn the correlation between input and output. When the model prediction accuracy meets the preset threshold or reaches the training number, the target index prediction model can be obtained. The target index prediction model can accurately adapt to the complex nonlinear relationship between the tool and process parameters and key machining physical quantities in the ultra-precision machining of fused silica components. It can quickly output the prediction results of machining physical quantities corresponding to different parameter combinations, providing a direct and reliable quantitative basis for the intelligent optimization of the tool structure. This effectively improves the surface accuracy and surface integrity of ultra-precision machining of fused silica components.

[0026] 103. Obtain multiple sets of tool structure data groups to be predicted, and call the target index prediction model to predict the key processing indicators for each set of tool structure data groups to be predicted, so as to obtain the index data group corresponding to each set of tool structure data groups to be predicted.

[0027] In this embodiment, the acquired multiple sets of tool structure data to be predicted specifically refer to the set of tool rake angle, clearance angle, and cutting temperature parameters covering different combinations of values. The parameter value range closely matches the actual process requirements of precision machining of molten stone. By reasonably covering the potential optimization range of key parameters, a dataset to be predicted is formed. When calling the trained target index prediction model, the rake angle, clearance angle, and cutting temperature parameters in each set of tool structure data to be predicted are used as model input. Relying on the correlation between parameters and machining indicators learned by the model, the model quickly outputs the corresponding key machining indicator results, i.e., the indicator data set corresponding one-to-one with each set of input parameter combinations. Each indicator data set includes maximum hydrostatic pressure, number of high-pressure particles, shear angle of the cutting area, and shear plane length. The target index prediction model can get rid of the complex calculation process and long time consumption of traditional numerical simulation, and complete the prediction of machining indicators for multiple parameter combinations in an efficient and accurate manner, significantly reducing the time and computational costs of parameter evaluation. Meanwhile, the predicted index data set can directly quantify the machining effect corresponding to different tool structure parameters, providing a comprehensive and intuitive decision-making basis for subsequent multi-objective optimization of tool structure, helping to quickly select the optimal tool structure parameter combination that meets the needs of ultra-precision machining, thereby ensuring the machining surface accuracy and surface integrity of fused silica parts, and improving the optimization efficiency and reliability of machining process.

[0028] 104. Obtain an initial tool structure data set based on multiple sets of the tool structure data sets to be predicted;

[0029] In this embodiment, the initial tool structure data set is obtained based on multiple sets of tool structure data to be predicted. Essentially, within the reasonable range of tool rake angle, clearance angle, and cutting temperature values ​​already defined in the data set to be predicted, a specific number of parameter combinations are selected by random sampling to form the initial tool structure data set. This provides a diverse search starting point for subsequent tool structure optimization, helps the optimization algorithm to fully explore the potential optimal solution in the parameter space, improves the efficiency of the optimization process and the global optimization capability, and lays a balanced and reliable foundation for finally selecting the optimal tool structure parameter combination that meets the requirements of ultra-precision machining.

[0030] 105. Using a genetic algorithm, the initial tool structure data set is iteratively updated multiple times based on multiple sets of the tool structure data set to be predicted and multiple sets of the index data set. When the preset iteration stopping condition is met, the updated tool structure data set is used as the optimized tool structure data set.

[0031] In this embodiment, when using a genetic algorithm to optimize tool structure parameters, multiple sets of tool structure data to be predicted are used as the basis for parameter search, and corresponding index data sets are used as the evaluation criteria. Multiple rounds of iterative updates are performed on the randomly selected initial tool structure data sets. During the iteration process, the genetic algorithm evaluates and filters the suitability of each tool structure parameter combination based on the processing effect reflected by the index data sets, until a preset iteration stopping condition is met. At this point, the finally updated tool structure data set is determined as the optimized tool structure data set. The genetic algorithm has a powerful global optimization capability, capable of efficiently exploring the optimal solution within a wide parameter space based on the quantitative feedback of key processing indicators, effectively avoiding local optima problems. Simultaneously, the optimization process closely integrates the process feasibility range of the data sets to be predicted with the quantitative results of the processing effect of the index data sets, ensuring that the final optimized tool structure data set accurately matches the quality requirements of molten stone precision machining, ensuring better material removal during processing, thereby improving the surface accuracy and surface integrity of the parts, and providing reliable support for the efficient optimization of the processing technology.

[0032] In this embodiment of the invention, the training sample set includes a training set and a test set, and the data sample library includes various tool processing parameters. The step of performing simulation operations based on the data sample library to construct the training sample set includes: using a full permutation and combination algorithm to perform full permutation and combination of the various tool processing parameters to obtain various input condition combinations; performing simulation operations based on each of the working condition combinations to obtain simulation index data sets corresponding to each working condition combination; integrating the various tool processing parameters and the simulation index data sets corresponding to each working condition combination to obtain the training sample set; and dividing the training sample set into a training set and a test set based on a preset partitioning ratio.

[0033] In this embodiment, the data sample library covers various tool machining parameters, specifically the tool rake angle, clearance angle, and cutting temperature. A full permutation combination algorithm is used to perform full permutation combinations on these parameters, resulting in various input condition combinations. For example, when the tool rake angle includes four values: 0°, -15°, -35°, and -65°; the clearance angle includes four values: 2°, 5°, 10°, and 15°; and the cutting temperature includes seven gradient values ​​between 25°C and 600°C, the full permutation combination algorithm can be used to combine these various tool machining parameters to obtain 112 different input condition combinations. Each combination fully covers the value dimensions of the tool rake angle, clearance angle, and cutting temperature, and fits the feasible range of the precision machining process for molten stone. Then, simulation operations were performed for each input working condition combination. Specifically, a corresponding cutting simulation model was constructed in Abaqus software using the Smooth Particle Dynamics (SPH) method. The workpiece used a 240×40×4μm³ SPH particle model with a total of 600,000 particles. The tool was configured as a diamond tool with a Lagrange mesh. The model parameters strictly matched the physical properties of quartz glass and diamond, such as density and Poisson's ratio. Through simulation calculations, machining characteristic data for each working condition could be obtained. For example, for the working condition combination of a rake angle of -15° and a temperature of 25°C, the simulation showed that the material was removed in a brittle mode, producing discontinuous fragments, corresponding to shearing... The simulation yields results such as the shear angle and shear plane length. For the working condition combination with a rake angle of 15° and a temperature of 600°C, the simulation shows that the material exhibits plastic flow characteristics and continuous fragmentation, and the corresponding shear angle and other data are also obtained. For the working condition combination where the rake angle gradually changes from 0° to -65°, the simulation shows that the shear plane length continuously increases and the maximum hydrostatic pressure gradually increases. These data together constitute the simulation index data set corresponding to each working condition combination. That is, the simulation calculation obtains the shear angle, shear plane length, maximum hydrostatic pressure, and the number of particles with hydrostatic tensile stress less than -200MPa under each working condition, forming the simulation index data set corresponding to that working condition combination. Subsequently, various tool processing parameters and the corresponding simulation index data sets are integrated to form a structurally complete training sample set. Then, based on a preset partitioning ratio (such as 9:1), the training sample set is decomposed into a training set for model parameter learning and a test set for verifying the model's generalization ability.The application of the full permutation and combination algorithm ensures that the input working condition combination can fully cover the potential value range of tool machining parameters, avoiding insufficient representativeness of the sample set due to the lack of local parameter dimensions. The simulation operation based on the smooth particle dynamics method can accurately capture the nonlinear mechanical behavior in the cutting process of hard and brittle materials, making the simulation index data set more consistent with the material removal and stress distribution characteristics of actual machining, and providing high-quality labeled data for the training sample set. The reasonable division of the training set and the test set not only ensures that the model can fully learn the correlation between parameters and indices through the training set, but also effectively verify the generalization performance of the model with the help of the test set, avoiding the model from falling into overfitting, and laying a reliable data foundation for the training of the subsequent index prediction model.

[0034] In this embodiment of the invention, the initial indicator prediction model includes an initial Actor network and an initial Critic network. The step of performing multiple rounds of iterative training on the initial indicator prediction model using the training sample set to obtain a target indicator prediction model includes: constructing an initial experience replay pool based on the training sample set, and obtaining current state samples and next state samples from the initial experience replay pool; using the initial Actor network to predict the action of the current state sample to obtain the current action; obtaining a predefined reward function, and calculating the reward value of the current action based on the reward function to obtain the current reward value; storing the current state sample, the current action, the current reward value, and the next state sample in the initial experience replay pool to obtain an expanded experience replay pool; and performing multiple rounds of iterative update processing on the initial Actor network and the initial Critic network based on the expanded experience replay pool to obtain the target indicator prediction model.

[0035] In this embodiment, the initial index prediction model is composed of an initial Actor network and an initial Critic network, both of which use neural networks as their core structure. When conducting multiple rounds of iterative training using the training sample set, an initial experience replay pool is first constructed based on the training sample set. This pool stores experience samples containing states, actions, rewards, and the next state. The tool rake angle, clearance angle, and cutting temperature in the training sample set are the state samples, from which the current state sample and the next state sample can be obtained. Subsequently, the initial Actor network is used to predict the action of the current state sample. Here, the action is a continuous prediction of key machining indicators such as maximum hydrostatic pressure, number of high-pressure particles, shear angle in the cutting area, and shear plane length, thus obtaining the current action. Next, a predefined reward function is obtained. This function is calculated based on the negative mean square error between the predicted and actual values. Based on this, the actual machining indicators corresponding to the current state sample and the current action are calculated to obtain the current reward value. The calculation expression of the reward function is:

[0036] ,

[0037] in, This indicates the current action, specifically the predicted values ​​for key machining parameters such as maximum hydrostatic pressure, number of high-pressure particles, shear angle in the cutting zone, and length of the shear plane. This represents the actual value corresponding to the current action. This indicates the dimension of the output variable, corresponding to the number of key processing indicators that need to be predicted (such as maximum hydrostatic pressure, number of high-pressure particles, etc.). Indicates representative t The reward value at any given moment.

[0038] Subsequently, the current state sample, current action, current reward value, and next state sample are stored in the initial experience replay pool to form an expanded experience replay pool, enabling continuous accumulation and updating of experience samples. Then, based on the expanded experience replay pool, multiple rounds of iterative updates are performed on the initial Actor network and the initial Critic network until the network converges, ultimately obtaining the target index prediction model. This model achieves continuous and accurate prediction of key machining indicators through the Actor network and continuously corrects the prediction strategy through the value evaluation of the Critic network. This allows the target index prediction model to fully capture the complex nonlinear relationship between tool parameters and machining indicators, enhancing the model's generalization ability and prediction accuracy, and providing reliable quantitative support for subsequent intelligent optimization of tool structures.

[0039] In this embodiment of the invention, the step of performing multiple rounds of iterative updates on the initial Actor network and the initial Critic network based on the expanded experience replay pool to obtain the target indicator prediction model includes: obtaining the current state sample, the current action, the current reward value, and the next state sample based on the expanded experience replay pool; using the initial Actor network to predict the action of the next state sample to obtain the next action; using the initial Critic network to perform value evaluation processing based on the current state sample, the current action, the next state sample, and the next action to obtain the current state value and the next state value; calculating the error based on the current reward value, the current state value, and the next state value to obtain the TD (Temporal Difference) error; and performing multiple rounds of iterative updates on the initial Actor network and the initial Critic network based on the TD error to obtain the target indicator prediction model.

[0040] In this embodiment, when iterating the initial Actor network and the initial Critic network multiple times based on the expanded experience replay pool, the current state sample, current action, current reward value, and next state sample are first obtained from the expanded experience replay pool. The current state sample is a combination of the tool rake angle, clearance angle, and cutting temperature. The current action is the prediction result of the machining index for this state by the initial Actor network, and the current reward value is calculated using a predefined reward function. Subsequently, the initial Actor network is used to predict the action of the next state sample to obtain the next action, i.e., continuous prediction of the machining index corresponding to the next state sample. Simultaneously, the initial Critic network is used to perform value evaluation processing, using the current state sample... With the current action As input, we obtain the current state value. The following is a sample of the state. With the next action As input, we obtain the value of the next state. Next, based on the current reward value and the current state value... and the value of the next state Error calculation is performed to obtain the TD error. The specific calculation expression is as follows:

[0041] ,

[0042] in, Indicates TD error, This represents the current reward value. This represents the discount factor, typically set to 0.99, used to ensure algorithm convergence. Indicates the value of the next state. Indicates the value of the current state. express The current state sample at time t. express The current action at any given moment. express The next state sample at time t. express The next action at any given moment.

[0043] Finally, based on the TD error, the initial Actor network and the initial Critic network are iteratively updated multiple times until the network converges, ultimately obtaining the target index prediction model. The introduction of the TD error enables precise quantification of state value deviation, allowing the evaluation capability of the initial Critic network to continuously iterate and providing reliable feedback for the policy optimization of the initial Actor network.

[0044] In this embodiment of the invention, the step of performing multiple rounds of iterative updates on the initial Actor network and the initial Critic network based on the TD error to obtain the target index prediction model includes: using a backpropagation algorithm to perform multiple rounds of iterative updates on the initial Critic network based on the TD error to obtain an optimized Critic network; using a policy gradient algorithm to perform multiple rounds of iterative updates on the initial Actor network based on the TD error to obtain an optimized Actor network; and combining the optimized Critic network and the optimized Actor network to form the target index prediction model.

[0045] In this embodiment, when performing multiple iterations to update the initial Actor network and the initial Critic network based on the TD error, the initial Critic network is first updated using the backpropagation algorithm. Through backpropagation, the TD error is propagated layer by layer from the network output layer to the input layer. The weights and thresholds of the initial Critic network are adjusted according to the error magnitude to minimize the bias in value assessment. After multiple iterations until the network converges, an optimized Critic network with accurate value assessment capabilities is obtained. For the initial Actor network, its function is to generate predicted actions for processing indicators based on state samples. It is updated using the policy gradient algorithm, and the specific calculation expression is as follows:

[0046] ,

[0047] in, This represents the Actor network parameters, i.e., the weights and thresholds of the neural network. The left arrow (←) represents the assignment operator, used to update the parameters. Indicates the learning rate. The gradient operator is defined as the object of gradient calculation, which is the parameters of the Actor network. , This represents the prediction strategy of the Actor network. The algorithm uses the TD error output by the optimized Critic network as a feedback signal. By adjusting the network parameters, it optimizes the prediction strategy, making the prediction results of the initial Actor network closer to the actual values ​​of the machining indicators. After multiple iterations, an optimized Actor network with accurate prediction capabilities is obtained. Finally, the optimized Critic network and the optimized Actor network are combined to form a target indicator prediction model. The optimized Actor network receives state samples such as the tool rake angle, clearance angle, and cutting temperature, and outputs prediction results for machining indicators such as maximum hydrostatic pressure and shear angle. The optimized Critic network continuously evaluates the value of the output actions of the optimized Actor network, providing a basis for fine-tuning its subsequent prediction strategy. Together, they achieve accurate prediction of machining indicators.

[0048] In this embodiment of the invention, the genetic algorithm includes a roulette wheel algorithm, a crossover algorithm, and a mutation algorithm. The step of using the genetic algorithm to iteratively update an initial tool structure data set based on multiple sets of data for the tool structure to be predicted and multiple sets of data for the index, and using the updated tool structure data set as the optimized tool structure data set when a preset iteration stopping condition is met, includes: using the roulette wheel algorithm to perform roulette wheel selection on multiple sets of data for the tool structure to be predicted based on multiple sets of data for the index, obtaining multiple sets of parent tool structure data sets; using the crossover algorithm to perform gene exchange on multiple sets of parent tool structure data sets, obtaining multiple sets of crossover tool structure data sets; using the mutation algorithm to perform gene fine-tuning on multiple sets of crossover tool structure data sets, obtaining multiple sets of mutated tool structure data sets; and using the multiple sets of mutated tool structure data sets to iteratively update the initial tool structure data set through multiple rounds of iteration, and using the updated tool structure data set as the optimized tool structure data set when a preset iteration stopping condition is met.

[0049] In this embodiment, when using a genetic algorithm to iteratively update the initial tool structure data set in multiple rounds, the first step is to process multiple sets of tool structure data sets to be predicted and their corresponding index data sets using a roulette wheel algorithm. Specifically, the parent generation is selected based on the fitness value corresponding to each set of tool structure data sets to be predicted. Individuals with higher fitness values ​​have a greater probability of being selected. A random number between 0 and 1 is generated. If this random number is greater than the probability of an individual being selected, that individual is retained for the next iteration. Then, a crossover operation algorithm is used to perform gene exchange processing on the parent tool structure data sets. Here, the genes correspond to parameters such as tool rake angle, clearance angle, and cutting temperature. By selecting and exchanging the corresponding parameter dimensions of different parent data sets, the parameter combinations are recombined, resulting in multiple sets of cross-cutting tool structure data sets, thereby expanding the diversity of parameter search. Next, a mutation operation algorithm is used to perform gene fine-tuning processing on the cross-cutting tool structure data sets, specifically fine-tuning parameters (such as tool rake angle). By introducing small-amplitude random parameter changes, the algorithm avoids getting trapped in local optima. After this processing, multiple sets of mutated tool structure data sets are obtained. The specific calculation expression is as follows:

[0050] ,

[0051] in, Indicates the relationship with the first The iteration, the The rake angle value corresponding to the group of tool structure data. Indicates the relationship with the first The iteration, the... The rake angle value corresponding to the group of tool structure data. This represents a random parameter, usually a normal distribution, ranging from -1 to 1.

[0052] Subsequently, the initial tool structure data set is iteratively updated multiple times based on multiple sets of mutated tool structure data sets. Each iteration repeats the roulette wheel selection, crossover, and mutation process until the preset iteration stopping condition is met, i.e., the maximum number of iterations is reached, or the difference in fitness values ​​between two consecutive iterations is less than a threshold. At this point, the updated tool structure data set is determined as the optimized tool structure data set. The roulette wheel algorithm relies on the fitness function to achieve directional selection of high-quality parameter combinations, ensuring that the iteration process converges towards better machining results. The crossover operation algorithm expands the diversity of parameter combinations and improves the coverage of the parameter search space. The mutation operation algorithm introduces new parameter mutations through random fine-tuning, effectively avoiding the local optimum problem. The three work together to enable the genetic algorithm to efficiently explore the global optimum within the tool parameter space. The final optimized tool structure data set can accurately adapt to the process requirements of molten stone precision machining, significantly improving the surface quality and process stability.

[0053] In this embodiment of the invention, the step of employing the roulette wheel selection algorithm to perform roulette wheel selection processing on multiple sets of tool structure data sets to be predicted based on multiple sets of index data sets to obtain multiple sets of parent tool structure data sets includes: obtaining a predefined fitness function; calculating the fitness of multiple sets of index data sets based on the fitness function to obtain fitness values ​​corresponding to the multiple sets of tool structure data sets to be predicted; using the roulette wheel selection algorithm to calculate roulette wheel selection probabilities based on multiple fitness values ​​to obtain roulette wheel selection probabilities corresponding to the multiple sets of tool structure data sets to be predicted; obtaining a preset selection rule; and performing roulette wheel selection processing on the multiple sets of tool structure data sets to be predicted based on the selection rule and multiple roulette wheel selection probabilities to obtain multiple sets of parent tool structure data sets.

[0054] In this embodiment, during the roulette wheel selection process, a predefined fitness function is first obtained. This function is used to quantify the machining adaptability corresponding to the tool structure data set to be predicted, assuming the objective is to maximize the maximum hydrostatic pressure. and maximizing the length of the shear plane Minimize the shear angle and minimizing the number of particles under high hydrostatic tensile stress The fitness function can then be defined as:

[0055] ,

[0056] in, , Indicates the relationship with the first The iteration, the... Tool structure data set Indicates the relationship with the first The iteration, the... The rake angle value corresponding to the group of tool structure data. Indicates the relationship with the first The iteration, the... The back angle value corresponding to the group of tool structure data. Indicates the relationship with the first The iteration, the... The cutting temperature values ​​corresponding to the tool structure data group. Indicates and The corresponding fitness value, Indicates the maximum hydrostatic pressure. Indicates the length of the shearing plane. Indicates the shear angle. This indicates the number of particles under high static water tensile stress.

[0057] Subsequently, based on the fitness function, the parameters of the index data group corresponding to each group of tool structure data to be predicted are substituted into the calculation to obtain the fitness value corresponding to each group of tool structure data to be predicted. The magnitude of this value directly reflects the quality of the machining effect of the corresponding parameter combination.

[0058] Next, the roulette wheel algorithm is used to calculate the selection probability. The specific calculation expression is as follows:

[0059] ,

[0060] in, Indicates the relationship with the first The iteration, the... Tool structure data set Indicates and The corresponding roulette wheel selection probability, Indicates and The corresponding fitness value, This represents the population size, i.e., the number of data sets of the tool structure to be predicted, which is generally set to 50. The formula determines the probability of each data set of the tool structure to be predicted being selected based on the proportion of fitness values ​​of each group. The higher the fitness value of a group, the greater the probability of selection.

[0061] Next, a preset selection rule is obtained, typically by generating random numbers between 0 and 1. The selection probabilities of all the tool structure data sets to be predicted are accumulated in order to form a continuous interval. The tool structure data set to be predicted corresponding to the interval in which the random number falls is selected. This operation is repeated until the required number of parent tool structure data sets are obtained. The fitness function realizes the quantitative evaluation of the machining effect, making the quality of the tool structure data sets to be predicted comparable. The calculation of the roulette wheel selection probability uses a probabilistic mechanism to selectively filter parameter combinations with higher fitness, ensuring that the subsequent iteration process converges to a better machining effect. At the same time, the characteristics of probabilistic selection retain a certain degree of population diversity, avoiding the algorithm from converging to a local optimum too early, and providing a high-quality and diverse parent foundation for subsequent crossover and mutation operations.

[0062] The above describes the tool structure data prediction method in the embodiments of the present invention. The following describes the tool structure data prediction device in the embodiments of the present invention. Please refer to [link / reference]. Figure 2 One embodiment of the tool structure data prediction device in this invention includes:

[0063] Simulation module 201: used to construct a data sample library for training, and to perform simulation operations based on the data sample library to construct a training sample set;

[0064] Training model 202: used to construct an initial indicator prediction model, and to perform multiple rounds of iterative training on the initial indicator prediction model using the training sample set to obtain a target indicator prediction model;

[0065] Prediction module 203: used to acquire multiple sets of tool structure data groups to be predicted, call the target index prediction model to predict the key processing indicators for each set of tool structure data groups to be predicted, and obtain the index data group corresponding to each set of tool structure data groups to be predicted.

[0066] Data acquisition module 204: used to acquire an initial tool structure data set based on multiple sets of the tool structure data sets to be predicted;

[0067] Optimization module 205: Used to perform multiple rounds of iterative updates on the initial tool structure data set based on multiple sets of the tool structure data set to be predicted and multiple sets of the index data set using a genetic algorithm. When the preset iteration stop condition is met, the updated tool structure data set is used as the optimized tool structure data set.

[0068] Based on the same ideas as the methods in the above embodiments, the apparatus provided in this application can implement the methods in the above embodiments.

[0069] above Figure 2The tool structure data prediction device in this embodiment of the invention is described in detail from the perspective of modular functional entities. The tool structure data prediction device in this embodiment of the invention is described in detail below from the perspective of hardware processing.

[0070] Figure 3 This is a schematic diagram of a tool structure data prediction device 300 provided in an embodiment of the present invention. The tool structure data prediction device 300 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors) and a memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) storing application programs 333 or data 332. The memory 320 and storage media 330 can be temporary or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the tool structure data prediction device 300. Furthermore, the processor 310 may be configured to communicate with the storage media 330 and execute the series of instruction operations in the storage media 330 on the tool structure data prediction device 300 to implement the steps of the tool structure data prediction method provided in the above-described method embodiments.

[0071] The tool structure data prediction device 300 may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input / output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 3 The illustrated tool structure data prediction device structure does not constitute a limitation on the tool structure data prediction device, and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0072] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of the tool structure data prediction method.

[0073] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system, device, or unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0074] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0075] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting tool structure data, characterized in that, include: Construct a data sample library for training, and perform simulation operations based on the data sample library to construct a training sample set; An initial indicator prediction model is constructed, which includes an initial Actor network and an initial Critic network; an initial experience replay pool is constructed based on the training sample set, and current state samples and next state samples are obtained from the initial experience replay pool; the initial Actor network is used to predict the action of the current state sample to obtain the current action; a predefined reward function is obtained, and the reward value of the current action is calculated based on the reward function to obtain the current reward value; The current state sample, the current action, the current reward value, and the next state sample are stored in the initial experience replay pool to obtain the expanded experience replay pool; Based on the expanded experience replay pool, the current state sample, the current action, the current reward value, and the next state sample are obtained; the initial Actor network is used to predict the action of the next state sample to obtain the next action; the initial Critic network is used to perform value evaluation processing based on the current state sample, the current action, the next state sample, and the next action to obtain the current state value and the next state value. The TD error is obtained by calculating the error based on the current reward value, the current state value, and the next state value. Based on the TD error, the initial Actor network and the initial Critic network are iteratively updated multiple times to obtain the target index prediction model. Multiple sets of tool structure data to be predicted are obtained, and the target index prediction model is called to predict the key machining indexes for each set of tool structure data to be predicted, so as to obtain the index data set corresponding to each set of tool structure data to be predicted. An initial tool structure data set is obtained based on multiple sets of the tool structure data sets to be predicted. A roulette wheel selection algorithm is used to select multiple sets of tool structure data sets to be predicted based on multiple sets of index data sets, resulting in multiple sets of parent tool structure data sets. A crossover operation algorithm is used to perform gene exchange processing on the multiple sets of parent tool structure data sets, resulting in multiple sets of cross tool structure data sets. A mutation operation algorithm is used to perform gene fine-tuning processing on the multiple sets of cross tool structure data sets, resulting in multiple sets of mutated tool structure data sets. Based on the multiple sets of mutated tool structure data sets, the initial tool structure data set is iteratively updated multiple times. When a preset iteration stopping condition is met, the updated tool structure data set is used as the optimized tool structure data set.

2. The tool structure data prediction method according to claim 1, characterized in that, The training sample set includes a training set and a test set, and the data sample library includes various tool machining parameters; the step of performing simulation operations based on the data sample library to construct the training sample set includes: A full permutation combination algorithm is used to perform full permutation combinations on various tool machining parameters to obtain various input working condition combinations; Simulation operations are performed based on each of the aforementioned working condition combinations to obtain a set of simulation index data corresponding to each working condition combination. By integrating various tool processing parameters and simulation index data sets corresponding to each working condition combination, a training sample set is obtained; Based on a preset partitioning ratio, the training sample set is divided into a training set and a test set.

3. The tool structure data prediction method according to claim 1, characterized in that, The process of performing multiple rounds of iterative updates on the initial Actor network and the initial Critic network based on the TD error to obtain the target index prediction model includes: The backpropagation algorithm is used to perform multiple rounds of iterative updates on the initial Critic network based on the TD error in order to obtain an optimized Critic network. The policy gradient algorithm is used to perform multiple rounds of iterative updates on the initial Actor network based on the TD error to obtain an optimized Actor network. The optimized Critic network and the optimized Actor network are combined to form the target indicator prediction model.

4. The tool structure data prediction method according to claim 1, characterized in that, The method employs the roulette wheel selection algorithm, performing roulette wheel selection on multiple sets of the tool structure data sets to be predicted based on multiple sets of the index data sets, to obtain multiple sets of parent tool structure data sets, including: Obtain the predefined fitness function; Based on the fitness function, the fitness of multiple sets of index data groups is calculated to obtain fitness values ​​corresponding to multiple sets of tool structure data groups to be predicted. Using the aforementioned roulette wheel selection algorithm, the roulette wheel selection probability is calculated based on multiple fitness values ​​to obtain the roulette wheel selection probability corresponding to multiple sets of the tool structure data sets to be predicted; A preset selection rule is obtained, and a roulette wheel selection process is performed on multiple sets of the tool structure data sets to be predicted based on the selection rule and multiple roulette wheel selection probabilities to obtain multiple sets of the parent tool structure data sets.

5. A tool structure data prediction device, characterized in that, include: Simulation module: used to build a data sample library for training, and to perform simulation operations based on the data sample library to build a training sample set; Training Model: Used to construct an initial metric prediction model, which includes an initial Actor network and an initial Critic network; constructing an initial experience replay pool based on the training sample set, and obtaining current state samples and next state samples from the initial experience replay pool; using the initial Actor network to predict the action of the current state sample to obtain the current action; obtaining a predefined reward function, and calculating the reward value of the current action based on the reward function to obtain the current reward value; The current state sample, the current action, the current reward value, and the next state sample are stored in the initial experience replay pool to obtain the expanded experience replay pool; Based on the expanded experience replay pool, the current state sample, the current action, the current reward value, and the next state sample are obtained; the initial Actor network is used to predict the action of the next state sample to obtain the next action; the initial Critic network is used to perform value evaluation processing based on the current state sample, the current action, the next state sample, and the next action to obtain the current state value and the next state value. The TD error is obtained by calculating the error based on the current reward value, the current state value, and the next state value. Based on the TD error, the initial Actor network and the initial Critic network are iteratively updated multiple times to obtain the target index prediction model. Prediction module: used to acquire multiple sets of tool structure data to be predicted, call the target index prediction model to predict the key machining indicators for each set of tool structure data to be predicted, and obtain the index data set corresponding to each set of tool structure data to be predicted. Data acquisition module: used to acquire an initial tool structure data set based on multiple sets of the tool structure data sets to be predicted; The optimization module employs a roulette wheel selection algorithm to select multiple sets of the tool structure data sets to be predicted based on multiple sets of the aforementioned index data sets, resulting in multiple sets of parent tool structure data sets. It then uses a crossover operation algorithm to perform gene exchange processing on the multiple sets of parent tool structure data sets, resulting in multiple sets of cross-cutting tool structure data sets. Finally, it uses a mutation operation algorithm to fine-tune the genes of the multiple sets of cross-cutting tool structure data sets, resulting in multiple sets of mutated tool structure data sets. Based on these mutated tool structure data sets, it iterates and updates the initial tool structure data set multiple times. When a preset iteration stopping condition is met, the updated tool structure data set is used as the optimized tool structure data set.

6. A tool structure data prediction device, characterized in that, The tool structure data prediction device includes: a memory and at least one processor, wherein the memory stores instructions; At least one of the processors invokes the instructions in the memory to cause the tool structure data prediction device to perform the steps of the tool structure data prediction method as claimed in any one of claims 1-4.

7. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the various steps of the tool structure data prediction method as described in any one of claims 1-4.

Images