Processing wear prediction method
By combining quantum mechanical equations and quantum machine learning, the problem of modeling accuracy in tool wear prediction under complex environments has been solved. This method achieves high-precision mapping from microscopic particle behavior to macroscopic wear state, thereby improving the accuracy and interpretability of tool wear prediction.
Patent Information
- Application Number
- CN202511068380.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-31
- Publication Date
- 2025-11-07
AI Technical Summary
Existing tool wear prediction methods have limited modeling capabilities and prediction accuracy in complex, variable, and nonlinear real-world machining environments, especially in the early wear stages and under conditions with small signals, and lack a systematic quantum machine learning modeling framework.
The quantum dynamics of the microscopic particles at the tool-workpiece interface are described using a quantum mechanical equation framework. A data-driven prediction module and a solution module with embedded physical constraints are combined. Wear prediction values are generated through a cross-scale mapping module, and quantum machine learning methods are used to improve prediction accuracy.
It achieves high-precision mapping from microscopic particle behavior to macroscopic wear state, and the prediction results have stronger interpretability and robustness, outperforming traditional models in key indicators. It is suitable for tool condition monitoring and life prediction in the manufacturing industry.
Smart Images
Figure CN120911230A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of machining monitoring and intelligent prediction, and particularly relates to a machining wear prediction method. BACKGROUND
[0002] In modern manufacturing industry, machining is one of the core processes, and the machining quality, production efficiency and stability of equipment operation are largely dependent on the performance state of the tool. Among them, tool wear not only directly affects the machining precision and workpiece surface quality, but also determines the tool life, machine downtime and the economy of the entire manufacturing process. Therefore, how to efficiently and accurately predict the tool wear state has become an important research topic in the field of intelligent manufacturing.
[0003] Existing tool wear prediction methods are mainly divided into physical modeling-based methods and data-driven machine learning methods. The former usually relies on the macro understanding of the interaction between the tool and the workpiece, such as thermal-mechanical coupling models, friction energy consumption models, etc. These models can basically describe the wear trend under ideal conditions, but their modeling ability and prediction accuracy are greatly limited when facing complex, variable and nonlinear actual machining environment. The latter, such as support vector machine, random forest, neural network, etc. data-driven model, although to some extent, breaks the dependence on prior knowledge, has good generalization ability, but its essence is statistical fitting, it is difficult to reveal the physical mechanism of wear, especially in the early wear stage and small signal working condition, the prediction effect is still not ideal.
[0004] Starting from the root of material failure, tool wear is essentially a process of migration, dissipation and structure reconstruction of micro-particles under the action of complex load (such as cutting force, friction heat). This process is controlled by micro-behavior at atomic scale or even electronic scale, and its essence follows the laws of quantum mechanics. In recent years, with the popularization of computational materials science methods such as density functional theory (DFT), academia has begun to analyze the wear mechanism from the quantum level, such as modeling the bond breaking and electron cloud distortion between atoms, studying the energy evolution mechanism at the cutting interface. These research results reveal the influence of quantum state evolution on material surface stability and wear evolution path, providing a theoretical basis for building tool wear prediction models with physical basis.
[0005] However, the high-precision solution of quantum mechanics model usually relies on a large amount of computing resources, and it is difficult to directly establish a clear correspondence with the macroscopically measurable wear parameters, which limits its engineering application. At the same time, the existing cross-scale modeling method still faces problems such as high modeling complexity and poor interpretability in the process of mapping micro-particle behavior to macro wear state.
[0006] As a new cross technology rapidly developing in recent years, quantum machine learning combines the high parallelism of quantum computing with the expressive ability of traditional machine learning, and shows unique advantages in high-dimensional quantum system modeling, feature extraction and state classification. By introducing quantum state representation, quantum kernel method and variational quantum circuit, a model with predictive ability can be built while preserving microscopic information.
[0007] However, the quantum machine learning method for tool wear prediction is still in the concept verification stage, lacking a systematic modeling system, especially in the combination of microscopic physical modeling and quantum feature selection, the integration of quantum model and actual wear data, etc. There is no mature solution.
[0008] Therefore, based on quantum mechanics theory and quantum machine learning technology, how to improve the accuracy of tool wear prediction has become a technical problem to be solved. SUMMARY
[0009] The main purpose of the present application is to provide a machining wear prediction method, which aims to improve the accuracy of tool wear prediction based on quantum mechanics theory and quantum machine learning technology.
[0010] In order to achieve the above purpose, the present application provides a machining wear prediction method, comprising the following steps: The quantum state dynamics of the micro-particles at the tool-workpiece contact interface is described by a quantum mechanics equation framework, wherein the framework is constructed based on the Schrodinger equation with dissipation term; In response to the input machining parameters and cutting force data, the total energy of the machining process is output by the data-driven prediction module, wherein the machining parameters include cutting speed, feed rate, tool rake angle, cutting depth and cutting width; When the total energy output exceeds the set threshold, the quantum mechanics equation is solved by the solving module embedded with physical constraints to obtain the wave function probability distribution of the micro-particles; Based on the wave function probability distribution, the particle transition behavior is calculated and the macro tool wear is associated by the cross-scale mapping module, thereby generating the wear prediction value.
[0011] In an embodiment of the present application, the potential function in the quantum mechanics equation framework is simplified as a constant, wherein the quantum state description is triggered by real-time acquisition of three-directional cutting force data when the cutting process starts.
[0012] In an embodiment of the present application, the data-driven prediction module adopts a neural network architecture, wherein in response to the machining parameter input, the total energy is output within a set time after receiving the sensor data.
[0013] In an embodiment of the present application, the embedded physical constraint solving module is a physical information neural network, wherein the neural network outputs wave function real part and imaginary part through input space coordinates.
[0014] In an embodiment of the present application, the particle transition behavior calculation includes solving transition probability formula through perturbation theory:
[0015] wherein, denotes transition probability from state to . denotes reduced Planck constant; denotes perturbation Hamiltonian matrix element; denotes angular frequency difference; denotes time.
[0016] In an embodiment of the present application, the neural network architecture adopts adaptive optimizer to update parameters during training, wherein learning rate decays with training rounds.
[0017] In an embodiment of the present application, the data-driven prediction module receives standardized feature vectors at the input layer, wherein the feature standardization module is connected with the data acquisition sensor through a bus.
[0018] In an embodiment of the present application, the physical information neural network contains three layers of hidden layers, wherein each layer of neurons adopts hyperbolic tangent activation function.
[0019] In an embodiment of the present application, the macro tool wear amount is calculated through a wear rate formula:
[0020] wherein, denotes volume loss amount of tool material per unit time, denotes total number of quantum states with energy less than . denotes wear area volume; denotes transition rate.
[0021] In an embodiment of the present application, the total number of quantum states is calculated based on normalized state density:
[0022] wherein, denotes effective mass of particle; denotes particle energy; denotes reduced Planck constant.
[0023] Compared with the prior art, the present application has the following beneficial effects: 1. A cross-scale modeling method combining quantum mechanics and machine learning is proposed, establishing a mapping mechanism from microscopic particle dissipation behavior to macroscopic wear amount, breaking through the limitations of traditional models lacking physical interpretation.
[0024] 2. The present application innovatively combines physical information neural network (PINN) with quantum mechanics equations, embedding Schrodinger equation constraints in the neural network training process, and realizes the physically interpretable solution of particle wave function.
[0025] 3. The present application introduces verifiable physical parameters (such as particle transition probability, state density, etc.) as intermediate features, making the prediction results more interpretable. The results show that the model is superior to the comparative method in terms of mean absolute percentage error (MAPE), confidence interval coverage (PICP) and other key indicators. Compared with traditional data-driven models, the method shows better results in key performance indicators such as prediction error and confidence interval coverage, and has stronger physical consistency and robustness.
[0026] 4. The method uses sensing data and tool material characteristics in the cutting process to construct a prediction model that can be transferred between numerical simulation and industrial field. Thanks to the low demand for a small amount of calibration data and the adaptive learning ability of the algorithm to physical laws, the technology has high universality and reliability, and is expected to be applied to manufacturing tool condition monitoring and life prediction, which is of great significance to improve processing quality and reduce production cost. BRIEF DESCRIPTION OF DRAWINGS
[0027] The present application will be described in detail below with specific embodiments and drawings, in which: Figure 1 The flow structure diagram of the first embodiment of the present application is shown in the figure; Figure 2 The particle dissipation diagram in the cutting process is shown in the figure; Figure 3 The energy prediction neural network structure diagram is shown in the figure; Figure 4 Quantum machine learning method for solving Schrodinger equation. DETAILED DESCRIPTION
[0028] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be described in detail below with the drawings and examples. It should be understood that the following specific examples are only used to explain the present application and do not limit the present application.
[0029] As Figures 1 to 4 shown, in order to achieve the above purpose, the present application proposes a machining wear prediction method, including the following steps: describing quantum state dynamics of micro-particles at the tool-workpiece contact interface through a quantum mechanics equation framework, wherein the framework is built based on a Schrödinger equation with a dissipation term; outputting, by a data-driven prediction module, a total energy of the machining process in response to input machining parameters and cutting force data, wherein the machining parameters include a cutting speed, a feed rate, a tool rake angle, a cutting depth, and a cutting width; when the total energy output exceeds a set threshold, solving the quantum mechanics equation through a physics constraint embedded solving module to obtain a wave function probability distribution of the micro-particles; based on the wave function probability distribution, calculating, by a cross-scale mapping module, particle transition behavior and relating a macroscopic tool wear amount to generate a wear prediction value.
[0030] In an embodiment of the present application, a potential energy function in the quantum mechanics equation framework is simplified as a constant, wherein quantum state description is triggered by real-time acquisition of three-directional cutting force data when the cutting process starts.
[0031] In an embodiment of the present application, the data-driven prediction module adopts a neural network architecture, wherein the total energy is output within a set time after sensor data is received in response to input machining parameters.
[0032] In an embodiment of the present application, the physics constraint embedded solving module is a physics information neural network, wherein the neural network outputs a real part and an imaginary part of the wave function through input spatial coordinates.
[0033] In an embodiment of the present application, the particle transition behavior calculation includes solving a transition probability formula through perturbation theory:
[0034] wherein, denotes a transition probability from a state to ; denotes a reduced Planck constant; denotes a perturbation Hamiltonian matrix element; denotes an angular frequency difference; denotes time.
[0035] In an embodiment of the present application, the neural network architecture adopts an adaptive optimizer to update parameters during training, wherein a learning rate decays with training rounds.
[0036] In an embodiment of the present application, the data-driven prediction module receives a standardized feature vector at an input layer, wherein a feature standardization module is connected to a data acquisition sensor through a bus.
[0037] In an embodiment of the present application, the physical information neural network comprises three layers of hidden layers, wherein each layer of neurons adopts a hyperbolic tangent activation function.
[0038] In an embodiment of the present application, the macro tool wear amount is calculated by a wear rate formula:
[0039] wherein, represents the volume loss amount of tool material per unit time, represents the total number of quantum states with energy less than ; represents the volume of the wear area; represents the transition rate.
[0040] In an embodiment of the present application, the total number of quantum states is calculated based on a normalized state density:
[0041] wherein, represents the effective mass of the particle; represents the energy of the particle; represents the reduced Planck constant.
[0042] The detailed description is as follows: First, the Schrodinger quantum dynamics equation at the contact between the tool and the workpiece is established: According to the wave-particle duality of real particles, it is assumed that the angular momentum of the worn tool particles is , and the energy is , then the wave vector and the angular frequency of the corresponding wave are respectively ; (1) wherein is called the reduced Planck constant, which is a basic constant in quantum mechanics, and directly relates the angular frequency , the angular momentum and the quantum phenomenon; Unlike the macro Newtonian mechanics in classical mechanics, which describes the physical phenomena at the friction and wear between the tool and the workpiece, from a micro level, the motion of the tool particles is described by the wave function to describe the motion state of the micro particles, i.e. the quantum state, which gives the probability amplitude of the particle at time, at ; (2) For non-free tool particles in such problems, gives the probability amplitude of the particle at At any moment, particles are... The probability density of occurrence at a given location It is the wave function at the initial time ( = 0) and the origin position ( The amplitude value of (= 0); It represents the imaginary unit.
[0043] Differentiating both sides of equation (2) with respect to time and space respectively, we get: (3) in, This represents the time partial derivative operator in the non-relativistic case. Furthermore, since the result of energy acting on the wave function is different from that obtained using the operator... The result of the action on the wave function is the same, and the result of the action of momentum on the wave function is the same as that of the action of the operator. The result is the same when applied to the wave function, so equation (3) becomes: (4) That is, the wave function described when no potential function is applied. Schrödinger equation.
[0044] Suppose a single-particle system The function is This describes the total energy of the system, where Since it is a potential function, when a potential function is applied, The equation can be expressed as: (5) In quantum mechanics, the Schrödinger equation is a known mathematical framework used to describe the evolution of the state (wave function) of a quantum system over time or space. After establishing the Schrödinger quantum dynamics equation, the wave function is generally solved based on the known Schrödinger equation to further obtain the probability density of the positions of the particles dissipated by each tool in quantum space.
[0045] Combine machine learning methods to obtain the total energy correlation relationship during the processing: Based on experimentally measured machining parameters, including cutting speed Feed rate Rake angle of the cutting tool Depth of cut Cutting width According to the cutting speed and workpiece diameter Calculate spindle speed : The components of the cutting force in three directions are measured using a three-dimensional force gauge, including the main cutting force along the cutting speed direction. Feed force along the feed direction and the back force perpendicular to the cutting plane The total cutting force is the vector sum of the three directional components In addition, the cutting power is calculated from the main cutting force and the cutting speed The machining time is calculated according to the workpiece cutting length and the feed amount , and the total energy of the machining process is calculated according to the physical method The calculated energy value is used as the label of the supervised learning.
[0046] Based on the measured and calculated machining parameters and other data, a measured cutting force data set of the machining process is established, and the measured machining parameters, cutting force data and data obtained by physical method are uniformly combined and divided into data sets in time sequence, and divided into training set and test set in the ratio of 8:2.
[0047] For the measured data, abnormal values need to be cleaned, and method is used to identify and eliminate abnormal samples exceeding 3 times the standard deviation, such as negative cutting force values caused by sensor failure. Subsequently, the features are standardized to eliminate dimensional differences, and the features , , , , , etc. are normalized.
[0048] The core challenge of tool machining energy prediction is to capture the complex nonlinear relationship between cutting speed, feed rate and energy. For this task, neural networks achieve high-precision modeling through multiple nonlinear transformations, with the following specific design: (1) Input layer design: The input layer receives the standardized feature vector, and the dimension is determined by the experimental parameters and derived features. The original parameters include cutting speed , feed rate , tool rake angle , cutting depth , cutting width , sensor data includes three cutting forces , , , and derived cutting power and synthetic cutting force , so the input features are 10 dimensions. Standardization uses method to ensure that each feature has a mean of 0 and a variance of 1.
[0049] (2) Hidden layer architecture: Based on the empirical rule, the number of hidden layers and nodes expands with the increase of data volume. For the large amount of data collected by the tool (about 10000-100000), three hidden layers (256-128-64 nodes) are adopted, each layer is followed by activation function: where is the input feature vector, and are the weights and biases.
[0050] Subsequently, the regularization design is carried out, and is added after each layer to prevent overfitting, and an L2 regularization term is added in the loss function to limit the weight amplitude.
[0051] (3) Output layer: The output layer adopts single-node linear activation to predict the energy value , and in order to ensure that the prediction conforms to the physical law, a self-defined loss function is fused with mean square error (MSE) and negative energy penalty term to force the model output .
[0052] In the model training process, the adaptive moment estimation optimizer is used for parameter update, and the parameter settings are , , the initial learning rate is set to 0.001, and an exponential decay strategy is given for training rounds. After each training cycle, the learning rate is decayed by a factor of 0.96, effectively balancing the rapid convergence in the early training period and the fine tuning in the later period. To enhance the generalization ability of the model, a composite loss function with three constraints is constructed: the mean square error (MSE) is used to measure the deviation between the predicted value and the true label, the L2 regularization term (coefficient ) is used to constrain the network weight norm to prevent overfitting, and the negative energy penalty term (coefficient ) is introduced to impose a quadratic loss on the output with negative value, ensuring that the energy prediction strictly follows the physical conservation law. To prevent overtraining, an early stopping monitoring mechanism is set, which automatically terminates training when the validation set loss does not decrease for 10 consecutive training cycles, and rolls back to the historical optimal weight state.
[0053] In the model verification stage, a multi-dimensional evaluation system is established. First, the power conservation constraint is used to verify the physical reasonableness, and the theoretical cutting power and dynamically compared with the model's implicit power prediction value, when the deviation exceeds 2 times the standard deviation of the measured power, the model recalibration process is triggered; Monte Carlo Dropout method is used to quantify the prediction uncertainty, keeping 20% of neurons randomly inactive during the test phase, and calculating the mean value of the prediction value through 50 forward propagations and variance , when the root mean square error of the training set exceeds , it is determined as a low confidence prediction. After testing the test set, the model still maintains , low prediction accuracy and negative energy output ratio within , meeting the prediction accuracy requirement.
[0054] 3. Solve the Schrödinger equation using the physical information guided neural network PINN: It is known that the tool wear particles in the cutting process obey quantum dissipation behavior, and its dynamics is described by the time-dependent stationary Schrödinger equation: where the total energy of the system is predicted by the above neural network method, and the potential energy function is simplified as a constant, Generate training data in the solution domain, divided into internal points and boundary points, internal points are randomly and uniformly sampled in the tool wear area (x-direction cutting depth [0, L]), used to calculate the Schrödinger equation residual loss; the boundary points are taken from both boundaries, and the wave function is forced to satisfy the boundary conditions.
[0055] Then design a physical information guided neural network, build a fully connected neural network, input coordinate x, output real and imaginary parts of the wave function, design one node in the input layer, three layers in the hidden layer, 64 neurons in each layer, use tanh as the activation function, and two nodes in the output layer to output the real and imaginary parts respectively.
[0056] Encode physical laws as loss functions, here embodied as the Schrödinger equation and its boundary conditions, to constrain the output of the neural network. Use Pytorch to implement the training loop, and finally get the numerical solution of the wave function The value does not directly represent the physical quantity of the particle, but through its modulus square represents the probability density of the particle at position x, and its peak corresponds to the position where the tool surface wear particles are most likely to gather. Further, we connect the macroscopic tool wear condition through the wave function of the microscopic particle, so as to obtain the most intuitive and accurate wear change.
[0057] 4. Connect the microscopic wave function with the macroscopic wear: By simulating the micro-particle behavior of the tool material through machine learning, the state distribution (wave function) of these particles at the quantum level can be calculated, which essentially reflects the probability of particles appearing at different positions. On this basis, combined with the quantum transition theory, i.e. the law of state switching of particles under the action of external energy (such as cutting pressure), the probability and rate of these particles leaving their original positions are analyzed.
[0058] Further, by statistically analyzing the particle density (state density) that may occur on the tool surface, the probability of a single particle leaving is converted into the collective number of particles leaving in a unit of time.
[0059] Finally, by experimentally calibrating the contribution of a single particle leaving to material loss, the cumulative effect of micro-events is converted into macroscopically observable wear rate and total amount.
[0060] In short, this process directly links the nanoscale particle dynamics to the millimeter-scale engineering wear loss through a three-step logic of "quantum behavior prediction → group probability statistics → experimental parameter calibration".
[0061] The specific calculation process is as follows: According to the simplified time-dependent Schrödinger equation and the physical information guided machine learning method to solve the Schrödinger equation, given the quantum state of the tool particle in the quantum space, i.e. the wave function of the particle; The total energy of the system is represented by and this can be represented as where does not contain time, so the Schrödinger equation is transformed into the following form: (6) From the converted time-dependent Schrödinger equation, the particle transition probability and transition rate are solved. Obviously, any quantum state of the particle can be expanded into a generalized series: (7) Substituting equation (6) into equation (7) gives: (8) Given the relationship: ; , Take the inner product of both sides of the above equation with to get: (9) Further, the representation in the representation is obtained Equation: (10) where , ; and since the general cannot be solved accurately, when , a perturbation solution is introduced, let , substitute into the equation to get: (11) Compare the same power of to get: (12) The zero-order approximate solution can be obtained from the first formula in (12): ( is a constant); Assume , the system is in the th eigenstate, that is ; The first-order approximate solution can be obtained from the second formula in (12): (13) Thus we can get: .
[0062] Therefore, when , the system is in the th eigenstate, that is, the transition probability , denoted by , so (14) The transition rate is the transition probability per unit time: (15) During the wear process, the transition of electrons from bound states (such as valence bands) to delocalized states (such as conduction bands) will weaken the interatomic bonding force. The state density quantifies the number of quantum states within a specific energy range that can participate in such destructive transitions, directly affecting the probability of ion detachment. Its meaning represents the density of quantum states that can exist in a unit volume within a unit energy interval, and its general form is: (16) where is the energy-momentum relationship of the particle, , is the effective mass of the electron, is the wave vector, denotes the Dirac function.
[0063] In three-dimensional space, the total number of electronic states with energy less than (17) Taking the derivative of the energy gives the state density: (18) where represents the volume; represents the effective mass of the electron.
[0064] Further normalization to the state density per unit volume: (19) The final approximate estimate of the rate of tool wear is: (20) Thus, the tool wear amount on a macroscopic level can be obtained according to this formula.
[0065] Compared with the prior art, the present application has the following beneficial effects: 1. A cross-scale modeling method combining quantum mechanics and machine learning is proposed, and a mapping mechanism from the microscopic particle dissipation behavior to the macroscopic wear amount is established, breaking through the limitation of the traditional model lacking physical interpretation.
[0066] 2. The present application innovatively combines the physical information neural network (PINN) with the quantum mechanics equation, and embeds the Schrödinger equation constraint in the neural network training process, realizing the physically interpretable solution of the particle wave function.
[0067] 3. The present application introduces verifiable physical parameters (such as particle transition probability, state density, etc.) as intermediate features, making the prediction results have stronger interpretability. The results show that the model is superior to the comparative method in terms of mean absolute percentage error (MAPE), confidence interval coverage rate (PICP) and other key indicators. Compared with traditional data-driven models, the method shows better results in key performance indicators such as prediction error and confidence interval coverage, and has stronger physical consistency and robustness.
[0068] 4. The method uses sensing data and tool material characteristics in the cutting process to construct a prediction model that can be transferred between numerical simulation and industrial field. Thanks to the low requirement of the algorithm for a small amount of calibration data and the adaptive learning ability of physical laws, the technology has high universality and reliability, and is expected to be applied to manufacturing tool state monitoring and life prediction, which is of great significance to improve processing quality and reduce production cost.
[0069] Microstate samples of the tool-workpiece interface are generated in a simulated environment. First, an initial quantum state model of the tool-workpiece interface is established based on the machining process parameters (such as cutting speed, feed rate, tool geometry parameters, etc.) and the microstructure characteristics of the tool material. The time-domain Schrödinger equation is solved by numerical methods (such as finite difference method or molecular dynamics simulation) to obtain the wave functions of the micro-particles at different time steps and the corresponding particle probability density distributions To enrich the diversity of the samples, known analytical solutions or defined functions (such as harmonic oscillator ground state wave functions) can be used to generate training samples under different parameter conditions.
[0070] The parameters in the simulation and learning process are set. This includes the potential function form in the Schrödinger equation, the time step and the size of the spatial discrete grid; the structure parameters of the PINN network (such as the number of hidden layers, the number of neurons in each layer, the type of activation function, etc.); the hyperparameters of the training process (such as learning rate, number of iterations, batch size, etc.); and the input and output dimensions of the mapping model. In this embodiment, the potential of the Schrödinger equation can use the approximate potential well model, the time step is set to Δt = 0.01 seconds, and the spatial grid interval is Δx = 0.001 nanometers; the PINN network contains three hidden layers, each with 64 neurons, and the activation function uses tanh; the training iteration is 5000 times, and the initial learning rate is set to 0.001.
[0071] The quantum mechanics equation is embedded in the neural network, and the physical information guided neural network (PINN) is used to solve the Schrödinger equation. Starting from the initial wave function generated from the sample, the PINN network satisfies the time-dependent Schrödinger equation and the boundary conditions at each space-time point through the loss function, and obtains the real and imaginary parts of the wave function. Figure 2 The dissipation and transition behavior of the micro-particles inside the tool material during the cutting process is illustrated: as the cutting process proceeds, the particles continuously transition from low energy levels to high energy levels and escape from the bound state. Through the PINN solving process, the particle probability distribution and transition rate of each simulation sample can be obtained.
[0072] Then the training set and the validation set are divided, and the above generated sample set is divided into training set and validation set. The data is randomly divided in the ratio of 8:2. In the training set, the PINN network and the macro mapping model are trained using multiple samples, and the samples in the validation set are used to evaluate the prediction ability of the model and prevent overfitting. To enhance the generalization of the model, the network can also be trained multiple times using cross-validation or Monte Carlo Dropout method, and the average prediction performance is recorded.
[0073] After the training, the new machining conditions are used to predict the wear. After inputting the new cutting parameters and process conditions, first, the trained PINN network is used to solve the particle wave function distribution under this condition, and the transition probability and state density are calculated. Then, the key quantum state characteristics (such as transition rate, state density, etc.) are input into the mapping model to obtain the macroscopic wear. Figure 3 The energy prediction neural network structure shown is used to calculate the instantaneous power and total energy in the cutting process, and these physical quantities are combined with the machine learning model as supervised information to improve the prediction accuracy. The final output is the predicted value of tool wear and its uncertainty range.
[0074] Error evaluation: Quantitative evaluation of prediction results to test the effectiveness of the method. Common error indicators include root mean square error (RMSE), mean absolute percentage error (MAPE), etc. By comparing the true wear amount of the validation set samples with the model predicted value, various error indicators are calculated. In order to evaluate the uncertainty prediction performance, indicators such as confidence interval coverage rate (PICP) can be used to measure the reliability of the prediction interval. If the error exceeds the preset threshold, the network structure or simulation parameters can be adjusted for retraining.
[0075] Finally, a reasonable simulation data scheme is used to verify the prediction ability of the model. Specifically, based on theoretical physical formulas and numerical solutions, training and validation data sets can be generated. For example, the solution of the time-dependent Schrödinger equation under given initial potential well conditions has an analytical expression or high-precision numerical solution. A series of wave function solutions can be generated under different boundary conditions and material parameters. Then, according to the wave function, the particle transition rate and state density are calculated, and they are mapped to the macroscopic wear through a pre-set function.
[0076] The above only describes the preferred embodiments of the present application, and does not limit the patent scope of the present application. Any equivalent structural transformation based on the inventive concept of the present application, or direct / indirect application in other related technical fields is included in the patent protection scope of the present application.
Claims
1. A method of machining wear prediction, characterized by, The method comprises the following steps: describing quantum state dynamics of micro-particles at the tool-workpiece contact interface via a quantum mechanics equation framework, wherein the framework is built based on a Schrödinger equation containing a dissipation term; outputting total energy of the machining process via a data-driven prediction module in response to input machining parameters and cutting force data, wherein the machining parameters include cutting speed, feed rate, tool rake angle, cutting depth, and cutting width; when the total energy output exceeds a set threshold, solving the quantum mechanics equation via a physics constraint embedded solving module to obtain a wave function probability distribution of micro-particles; based on the wave function probability distribution, calculating particle transition behavior and correlating macroscopic tool wear via a cross-scale mapping module to generate a wear prediction value.
2. The machining wear prediction method of claim 1, wherein, the potential energy function in the quantum mechanics equation framework is simplified as a constant, wherein quantum state description is triggered by real-time acquisition of three-directional cutting force data when the machining process starts.
3. The machining wear prediction method of claim 1 wherein, the data-driven prediction module adopts a neural network architecture, wherein total energy is output within a set time after sensor data is received in response to machining parameter input.
4. The process wear prediction method of claim 1, wherein, the physics constraint embedded solving module is a physics information neural network, wherein the neural network outputs real and imaginary parts of the wave function via input spatial coordinates.
5. The machining wear prediction method of claim 2, wherein, the particle transition behavior calculation includes solving a transition probability formula via perturbation theory: wherein denotes the transition probability from state to ; denotes the reduced Planck constant; denotes the perturbation Hamiltonian matrix element; denotes the angular frequency difference; denotes the time.
6. The machining wear prediction method of claim 3, wherein, the neural network architecture adopts an adaptive optimizer to update parameters during training, wherein the learning rate decays with training rounds.
7. The machining wear prediction method of claim 3 wherein, the data-driven prediction module receives a standardized feature vector at the input layer, wherein a feature standardization module is connected to a data acquisition sensor via a bus.
8. The process wear prediction method of claim 4, wherein, the physics information neural network contains three layers of hidden layers, wherein each layer of neurons adopts a hyperbolic tangent activation function.
9. The process wear prediction method of claim 5, wherein, the macroscopic tool wear is calculated via a wear rate formula: wherein, represents the volume loss amount of the tool material per unit time, represents the total number of quantum states having an energy less than represents the volume of the wear region; represents the transition rate. 10. The process wear prediction method of claim 9, wherein, the total number of quantum states is calculated based on a normalized state density: wherein, denotes the effective mass of the particle; denotes the energy of the particle; denotes the reduced Planck constant.