Double bayesian machining condition selection method and device and storage medium

Abstract

The application relates to a double-Bayes machining condition selection method, device and storage medium, the method comprising maintaining a Bayes physical model set containing multiple physical model particles, receiving in-service machining feedback collected in an actual material removal process of dynamic engagement of a tool and a workpiece, regenerating posterior stable probability after updating particle weight through observation likelihood, constructing a bottom-connected conservative safety envelope, and transferring the same to a second-layer Bayes optimization layer as a forced constraint, screening the stable probability, the safety envelope, the engagement limit and the spindle power torque, maximizing a Bayes collection function on a feasible subset meeting the constraint, and selecting a next machining condition. The application has the advantages that a spindle speed, an axial cutting depth, a radial cutting depth and a feed per tooth combination meeting stability and machine tool safety constraints and improving a material removal rate can be selected before machining or during machining.

Description

Technical Field

[0001] This invention relates to the field of CNC machine tool cutting technology, and in particular to a method, apparatus and storage medium for selecting dual Bayesian machining conditions. Background Technology

[0002] In metal cutting, material removal rate can typically be increased by raising spindle speed, axial depth of cut, radial depth of cut, and feed per tooth, thereby improving machining productivity. However, overly aggressive cutting conditions can easily induce problems such as regenerative chatter, which in turn reduces machining efficiency and quality. To address this issue, it is necessary to select appropriate machining conditions.

[0003] Traditional methods for selecting cutting conditions mainly rely on process manual data, operator experience, trial cuts, experimental design, or exhaustive mesh search, but these methods are time-consuming and labor-intensive. Some advanced solutions introduce stability lobe diagrams or cutting time-domain simulations; however, these solutions often use offline measured static modal parameters before machining to calculate chatter boundaries, and once the parameters are selected, they cannot be dynamically calibrated and updated as the machining process progresses, making it difficult to balance chatter safety and efficiency. Summary of the Invention

[0004] The purpose of this invention is to solve the problems mentioned in the background art and to provide a method, apparatus and storage medium for selecting dual Bayesian processing conditions.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows: A method for selecting machining conditions using a dual Bayesian approach, wherein the machining conditions include spindle speed n and axial depth of cut. Radial depth of cut and feed per tooth The method includes: A first Bayesian physical model calibration layer is constructed, which includes multiple physical model particles, each of which includes uncertain modal parameters and cutting process parameters; Receive process feedback records obtained during machining processes involving dynamic engagement between the tool and the workpiece and material removal, the process feedback records including machining conditions and stability observations; For each of the aforementioned process feedback records, the observed processing results under the aforementioned processing conditions are compared with the physical stability evaluation results of each of the aforementioned physical model particles under the aforementioned processing conditions. The particle-specific likelihood is calculated, and the probability of each of the aforementioned physical model particles is updated based on the particle-specific likelihood to obtain the posterior physical model. The posterior stability probability information is reconstructed from the posterior physical model, and the bottom connected conservative safety envelope is determined based on the posterior stability probability information. By using the forced inter-layer stability interface, the posterior stability probability information and the bottom-connected conservative safety envelope are provided as forced feasibility constraints to the second Bayesian safety process optimization layer. In the second Bayesian safety process optimization layer, a candidate pool is generated and feasible subsets are formed through mandatory constraint screening. Only in the feasible subset Within the posterior safety region defined by the posterior stability probability information and the bottom-connected conservative safety envelope, a second Bayesian response surrogate model is maintained, and productivity-related or unstable responses are modeled using the second Bayesian response surrogate model. The next processing condition is selected using a Bayesian acquisition function. The selected processing conditions are applied to the processing system or simulated, and the resulting process status monitoring results are returned to the first Bayesian physical model calibration layer, the second Bayesian safety process optimization layer, or both, to close the double Bayesian loop.

[0006] In one embodiment, each of the physical model particles includes one or more tool-side modes of the spindle-tool holder-tool system and one or more workpiece-side modes of the workpiece-fixture-machine tool system, each mode including natural frequency, modal stiffness, damping ratio, modal orientation or modal participation parameters; The machining stability evaluator uses the tool's compliance with the workpiece at the cutting interface. When a position-dependent natural frequency or mode shape predictor is used, its output is used as a prior or state condition input before the processing feedback likelihood update, and the updated posterior value is allowed to deviate from the predictor output. Perform a Bayesian update on at least one of the natural frequency, modal stiffness, damping ratio, modal participation, cutting coefficient, process damping parameter, or stability bias term, and change at least one of the stability lobe peak spindle speed, lobe height, dominant mode, or meshing-related stability limit. The tool-side modal parameters are state-dependent parameters and are corrected based on actual cutting evidence related to tool wear, edge damage, built-up edge, runout, spindle speed, thermal state, cutting load, or spindle-tool holder connection state. The workpiece-side modal parameters are modified based on material removal, changes in contact position, wall thickness, remaining geometry, fixture load, or machining stage.

[0007] In one embodiment, the process feedback records obtained under different radial depths of cut, feed per tooth, workpiece area, or machining conditions retain their respective actual coordinates; The particle-specific likelihood also includes a frequency likelihood, which increases the weight of particles that predict the dominant tool-side, workpiece-side, or coupled chatter frequencies and the measured chatter frequencies or related tooth-passing sidebands.

[0008] In one embodiment, when the posterior stability probability exceeds a selected stability threshold, but the process feedback record indicates flutter, a difference-sensitive likelihood is applied to reduce the weight of predicted stable particles and increase the relative weight of predicted unstable particles or particles that conform to the observed flutter frequency behavior. The posterior stability probability information and the bottom-connected conservative safety envelope are reconstructed immediately before executing / recommending another condition. The process feedback record indicating flutter is obtained during the actual material removal process, representing the meshing operation status.

[0009] In one embodiment, the posterior stable probability information is generated in multiple radial depth slices based on the shared posterior distribution of the posterior physical model. At each spindle speed, the bottom connected safety envelope independently traces upwards along the axial depth of cut from the lowest axial depth of cut grid point and stops at the first a posteriori unsafe point; For candidate radial cut depths located between storage slices, the posterior stability probability and the safe axial cut depth of the bottom-connected conservative safety envelope are both calculated by bounded interpolation between adjacent slices.

[0010] In one embodiment, the feasible subset Further requirements are made for the predicted spindle power P and predicted spindle torque T to satisfy: ; in, and The machine tool power curve and rated torque curve are related to the rotational speed; For configurable utilization factors; candidates falling outside the curve domain defined by the machine tool's rated power curve and rated torque curve are rejected; cutting force, feed rate, vibration, surface quality, temperature, tool wear, dimensional error, or feedback support confidence are also imposed as mandatory or probabilistic constraints.

[0011] In one embodiment, the second Bayesian response surrogate model is trained using analytical computation, low-cost simulation, selected time-domain simulation, measurement of processed response, or a combination thereof. The input to the second Bayesian response surrogate model includes at least one machining decision variable selected from spindle speed, axial depth of cut, radial depth of cut, and feed per tooth; The second Bayesian response surrogate model models productivity-related responses, instability responses, comprehensive scores, or physical model residuals. Where: x represents the processing parameters; r(x) represents the actual or measured process response obtained under the processing parameters; Indicates the estimated value based on the current physical parameters. The obtained physical model predicts the response; This represents the residual between the actual or measured process response and the physical model's predicted response.

[0012] The Bayesian acquisition function includes expected improvement, confidence upper bound, improvement probability, information gain, hypervolume improvement, or a combination thereof, but is evaluated only after applying the posterior stability of the first Bayesian physical model calibration layer and the bottom connectivity conservative safety envelope screening.

[0013] In one embodiment, multiple candidates in the candidate pool are evaluated by a low-cost static or quasi-static cutting force model to compute at least one of force, deformation, spindle power, spindle torque, material removal rate, or response surrogate training value. Selected candidates, initial design points, high a posteriori divergence points, boundary verification points, or shortlisted candidates are selectively evaluated through regenerative time-domain dynamic cutting simulation or actual machining operations. The resulting response is assigned fidelity or uncertainty and serves as evidence for the physical model particles of the first Bayesian physical model calibration layer, the second Bayesian response surrogate model, or both.

[0014] This invention also provides the following technical solutions: An apparatus includes one or more processors and a memory storing instructions, which, when executed by the one or more processors, cause the apparatus to perform any of the methods described above.

[0015] This invention also provides the following technical solutions: A storage medium having instructions stored thereon, which, when executed by one or more processors, cause the one or more processors to perform any of the methods described above. Effects of the invention:

[0016] This invention claims a dual Bayesian machining condition selection method, apparatus, and storage medium. It modifies the physical stability model based on feedback obtained during actual material removal machining with dynamic tool-workpiece engagement, rather than relying solely on static modal tests, offline finite element calculations, or pre-calculation simulations. Simultaneously, it captures current effective tool-side dynamic changes related to tool wear, cutting edge condition, built-up edge, runout, cutting load, thermal state, or spindle-tool holder state, as well as workpiece-side changes caused by material removal, contact position, remaining geometry, and fixture load. Transient or operational modal parameters are inferred probabilistically based on actual cutting behavior, eliminating the need for direct real-time modal testing during production. It supports real-time online, near real-time, post-cutting, periodic, and asynchronous or batch a posteriori corrections based on stored records obtained during actual material removal machining. This invention can select a combination of spindle speed, axial depth of cut, radial depth of cut, and feed per tooth that satisfies stability and machine tool safety constraints and improves material removal rate before or during machining, achieving an optimal balance between chatter safety and efficiency. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the conventional technology, the drawings used in the description of the embodiments or the conventional technology will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 The integrated two-layer Bayesian processing architecture framework of the present invention is shown in the figure: the first layer is physical stability learning, the second layer is physical response optimization, forced posterior stability interaction, selective dynamic verification, and optional measurement feedback; Figure 2 This is a framework diagram of the tool-side modal particles and workpiece-side modal particles of the present invention; Figure 3 This is a flowchart of the Bayesian mode-particle update process based on processing feedback according to the present invention. Figure 4 This is a diagram of the rear safety envelope and radial depth interpolation framework based on bottom connection of the present invention. Figure 5A This is a diagram of the second Bayesian layer of the present invention: a safe Bayesian optimization framework with posterior domain constraints. Figure 5B This is a flowchart of the multi-fidelity physical response evaluation based on posterior parameters of the present invention. Figure 6 This is a flowchart of the selective condition implementation and asynchronous model update of the present invention. Detailed Implementation

[0019] For clarity, the "double Bayesian" in this invention refers to two functionally distinct probabilistic mechanisms. The first Bayesian mechanism updates the machining stability physical model, which includes uncertain modal parameters and cutting process parameters. The second Bayesian mechanism performs safe process optimization by learning or correcting productivity-related responses, comprehensive scores, or one or more constraint responses, and selecting the next condition through a Bayesian acquisition function. Forced constraints can be evaluated probabilistically, deterministically, or using a hybrid rule approach, and can be used to screen the candidate pool before maximizing the acquisition function. Simply updating the stability probability map and selecting stability boundary test points using expected improvement or utility rules does not constitute the complete double Bayesian architecture of this invention.

[0020] The two Bayesian mechanisms are connected through a mandatory inter-layer stability interface, rather than operating as independent or adjacent modules. The first Bayesian physical model calibration layer provides its reconstructed posterior stability probability information and bottom-connected conservative safety envelope as mandatory feasibility constraints to the second Bayesian safety process optimization layer. The second layer models and optimizes productivity-related or instability-related responses only within the posterior safety domain provided by the first layer. After simulation or processing under selected conditions, process state monitoring returns the obtained stable state and available response evidence to the first, second, or both layers, thus closing the coupled dual Bayesian loop.

[0021] To avoid ambiguity, a model that only learns the mapping from tool geometry, remaining material, or predetermined toolpath coordinates to workpiece natural frequencies or mode shapes is not itself the first Bayesian mechanism of this disclosure. This model can serve as an initial, conditional, or auxiliary predictor of workpiece-side modal states, but it cannot replace the machining feedback likelihood update of the physical model set.

[0022] The first Bayesian mechanism is defined as follows: It compares the observed machining results under actual process conditions with multiple physically evaluable particles, calculates the particle-specific likelihood, updates the particle probabilities, and reconstructs the posterior stability information from the reweighted physical set. Therefore, the learned state is not limited to the workpiece's natural frequencies or mode shapes, but can also include tool-side dynamics, workpiece-side dynamics, cutting coefficients, process damping, modal coupling, and stability bias.

[0023] The second Bayesian mechanism also differs functionally from the position-modal predictor. It operates after the posterior stable information is generated, learning or correcting productivity-related responses, unstable responses, physical model residuals, or comprehensive scores for safe candidate selection. Therefore, the first physical posterior and the second response agent have different training evidence, predictive power, and loop closure effects.

[0024] The flutter stability solver and any time-domain dynamic model serve as auxiliary computational components. This invention is not limited to specific stability lobe equations, frequency-domain representations, time-domain motion equations, cutting force laws, numerical solvers, discretization methods, integration methods, or software implementations. Any technically applicable and sufficiently accurate analytical, numerical, empirical, data-driven, commercial, or hybrid model can be used, provided it provides the stability or dynamic response information required for the disclosed Bayesian update and safety optimization process. The concept of this invention lies in the probabilistic physics model update, posterior stability generation, safety screening, and Bayesian condition selection process, rather than considering the underlying flutter stability model or time-domain model in isolation.

[0025] This invention provides a dual Bayesian method for selecting machining conditions, including spindle speed n and axial depth of cut. Radial depth of cut and feed per tooth The methods include: The first Bayesian physics model calibration layer includes: (a) Maintenance includes The first Bayesian physics model set of i particles, each particle i including a physics parameter vector And the previous probability weights, physical parameter vector It includes at least one tool-side or workpiece-side modal parameter with physical significance, and optionally includes at least one cutting process parameter.

[0026] When a position-dependent workpiece dynamics predictor is used, it is used to initialize or conditionalize the set, but does not prevent subsequent corrections through machining feedback likelihood.

[0027] (b) Receive processing feedback observations under actual processing conditions The actual processing conditions are expressed as follows: ; And processing feedback observations when available With the workpiece area or processing state Related.

[0028] Among them, processing feedback observation It includes at least one in-process observation obtained during actual material removal machining when the rotating tool and the workpiece are dynamically engaged at the cutting interface, and is not based solely on static non-cutting modal tests, finite element calculations, or offline simulations.

[0029] Among them, the machining observations can be provided to the first Bayesian physical model calibration layer immediately or in near real time, can be provided after the cutting conditions or machining segment is completed, can be provided periodically after multiple observations are acquired, or can be provided subsequently as a stored sequence of observations acquired during actual material removal machining, so that the meshing cutting acquisition state rather than the Bayesian processing delay defines the observation as actual machining feedback.

[0030] (c) For each particle i, the machining stability evaluator is used to evaluate it under actual machining conditions and in the workpiece area or machining state to obtain the particle-specific critical axial depth of cut or equivalent particle-specific stability response, as well as the particle-specific predicted stability probability. .

[0031] in, This represents the probability that the response to the processing feedback record k is stable under particle i. The output can be expressed as: ; ; Among them, the particle-specific response represents the current effective dynamic state of the tool-workpiece under cutting engagement, and is allowed to differ from the state obtained before machining, in static tests, or in simulations.

[0032] (d) Calculate the observation likelihood for each particle i ,in, For particle i, the target or label is a stable object observed in practice. And when available, for auxiliary evidence vectors ( The probability mass, probability density, or other nonnegative likelihood measure (including flutter frequency, fast Fourier transform evidence, force, vibration, power, surface, or time-domain evidence) assigned to it. , and Given the condition, the likelihood can be expressed as: ; And according to the following formula, through unnormalized weights and normalized posterior weights Update particle weights: ; The weight update infers the current effective dynamic state using actual meshing cutting observations, eliminating the need for direct real-time measurement of transient modal parameters. It also allows for correction of tool-side dynamics related to wear, edge condition, thermal state, runout, cutting load, or toolholder connection state, as well as workpiece-side dynamics related to material removal, contact position, remaining geometry, or fixture load. The first Bayesian physics model calibration layer sequentially processes machining feedback observations upon observation reception. Processing can be performed after the corresponding cutting conditions or machining segment is completed; it can be performed periodically after a selected number of observations; or it can be performed asynchronously according to the stored sequence. The stored sequence can be processed sequentially according to the acquisition order, or processed through combined batch likelihood.

[0033] (e) The posterior stable probability of the candidate processing condition x is reconstructed based on the updated physical model particles, instead of simply fitting the boundary output model. The posterior stable probability is: ; (f) For each spindle speed n, radial depth of cut, feed per tooth, and selectable machining state, construct a bottom-connected allowable axial depth of cut set A_conn. The axial depth of cut is only included in the set if every intermediate axial depth of cut from the lowest considered axial depth of cut to a certain axial depth of cut satisfies the posterior stability criterion, thereby excluding disconnected high-depth safety islands. The set can be represented as:

[0034] And by subtracting the axial depth of cut safety margin The bottom-connected conservative safety envelope is obtained according to the following formula: ; The interface for enforced inter-layer stability includes: (g) The reconstructed posterior stable probability The graph and bottom connected conservative safety envelope are transferred from the first Bayesian physical model calibration layer to the second Bayesian safety process optimization layer as a forced stability constraint; The second Bayesian safety process optimization layer differs functionally from the first Bayesian physical model set and any optional location-related modal parameter predictor.

[0035] The second Bayesian safety process optimization layer includes: (h) Generate candidate pool And form a feasible subset according to the following formula. : ; Wherein, only if the candidate satisfies at least , Radial engagement limit, spindle power and spindle torque limits related to the following speeds, and any additional mandatory constraints for x. hour, The value is 1: ; (i) Maintain a second Bayesian response surrogate model, different from the first Bayesian physical model set, for prediction. The candidate's productivity-related response, instability-related physical response, physical model residuals or comprehensive score, and prediction uncertainty.

[0036] (j) only Maximize the Bayesian acquisition function to select the next processing condition; the acquisition function can be selectively multiplied by a soft feasibility probability weight. ; (k) Output the selected next processing condition for recommendation, simulation, operator approval or processing execution.

[0037] Monitoring and closed-loop feedback connections include: (l) After simulation or processing implementation of the selected next processing condition, process status data is collected, and the corresponding response under the actual processing condition is classified into at least stable, early warning, slight chatter, chatter, boundary, sensor failure or model failure.

[0038] Among them, the process status data generated during the actual material removal process can be streamed to the monitoring and Bayesian layer, or recorded for delayed transmission and processing.

[0039] (m) At the corresponding actual processing coordinates, the classification response and available sensors, controllers or simulation evidence are back to the first Bayesian physical model calibration layer to update the physical model particle posterior, back to the second Bayesian safety process optimization layer to update the second Bayesian response surrogate model, or back to both layers at the same time.

[0040] The feedback can be executed immediately, after the corresponding cutting or machining segment, periodically, or asynchronously based on the stored actual machining records.

[0041] (n) When the classification response contradicts the stability prediction, the posterior stable probability graph and the bottom connected conservative safety envelope are reconstructed through the first Bayesian physical model calibration layer, and the updated graph and envelope are transmitted through the forced interlayer stability interface before selecting subsequent processing conditions.

[0042] Each particle contains one or more tool-side modes from the spindle-tool holder-tool system and one or more workpiece-side modes from the workpiece-fixture-machine tool system. Each mode includes natural frequency, modal stiffness, damping ratio, modal orientation, or modal participation parameters. The machining stability evaluator uses the tool-to-workpiece compliance at the cutting interface.

[0043] When a position-dependent natural frequency or mode shape predictor is used, its output is used as a prior or state condition input before the processing feedback likelihood update, and the updated posterior value is allowed to deviate from the predictor output.

[0044] Furthermore, performing a Bayesian update on at least one of the natural frequency, modal stiffness, damping ratio, modal participation, cutting coefficient, process damping parameter, or stability bias term will change at least one of the stability lobe peak spindle speed, lobe height, dominant mode, or meshing-related stability limit.

[0045] The tool-side modal parameters are state-dependent parameters and can be modified in response to actual cutting evidence related to tool wear, edge damage, built-up edge, runout, spindle speed, thermal state, cutting load, or spindle-tool holder connection state.

[0046] Workpiece-side modal parameters can be modified in response to material removal, changes in contact position, wall thickness, remaining geometry, fixture load, or machining stage.

[0047] Feedback records obtained under different radial depths of cut, feed per tooth, workpiece area, or machining conditions retain their respective actual coordinates. Stable observations have a soft target close to 1, chatter observations have a soft target close to 0, and early warning or slight chatter observations have an intermediate target.

[0048] Particle likelihood Optionally, it includes frequency likelihood, which increases the weight of the predicted dominant tool-side, workpiece-side, or coupled chatter frequencies to the measured chatter frequencies or the particles corresponding to the tooth-passing sideband.

[0049] When the predicted stability probability before the update Exceeding the selected stability threshold, but processing feedback observation This demonstrates that when flutter occurs, a difference-sensitive likelihood is applied: the weights of predicted stable particles are reduced, the relative weights of predicted unstable particles or particles that conform to the observed flutter frequency behavior are increased, and the posterior stability map and conservative safety envelope are reconstructed immediately before implementing / recommending another condition. This indicates that flutter observations are acquired during actual material removal processing, thus representing the meshing operation state, rather than just the simulation or static test state.

[0050] A posterior map is generated based on a shared posterior physical model distributed across multiple radial depth slices. At each spindle speed, the bottom-connected safety envelope independently traces upwards along the axial depth and stops at the first posterior unsafe point. For candidate radial depths located between storage slices, the posterior stability probability is... and safe axial depth of cut All values ​​are obtained through bounded interpolation between adjacent slices, rather than treating observations of different radial depths as occurring at the same common radial depth.

[0051] feasible subset Further requirements are made for the predicted spindle power and predicted spindle torque to meet the following conditions: ; in, and The rated curve of the machine tool is related to the rotational speed; This is the configurable utilization factor.

[0052] Candidates falling outside the input machine tool curve domain are rejected rather than accepted by extrapolation; cutting force, feed rate, vibration, surface quality, temperature, tool wear, dimensional error, or feedback support confidence are also imposed as mandatory or probabilistic constraints.

[0053] The second Bayesian response surrogate model is trained using analytical computation, low-cost simulation, selected time-domain simulation, measurement of machining response, or a combination thereof. Its inputs include at least one machining decision variable selected from spindle speed, axial depth of cut, radial depth of cut, and feed per tooth; its modeling objects are productivity-related responses, instability responses, comprehensive scores, or physical model residuals. Instead of replacing it for calculating posterior stability The physical particle state; where x represents the processing parameters; r(x) represents the actual or measured process response obtained under the processing parameters; Indicates the estimated value based on the current physical parameters. The obtained physical model predicts the response; This represents the residual between the actual or measured process response and the physical model's predicted response.

[0054] Bayesian acquisition functions include expectation improvement, confidence upper bound, improvement probability, information gain, hypervolume improvement, or combinations thereof, but are evaluated only after applying the first-level posterior stability and safety envelope screening.

[0055] Multiple generated candidates are evaluated by a low-cost static or quasi-static cutting force model to compute at least one of the following: force, deformation, spindle power, spindle torque, material removal rate, or response surrogate training value.

[0056] Selected candidates, initial design points, high a posteriori divergence points, boundary verification points, or shortlisted candidates are selectively evaluated through regenerated time-domain dynamic cutting simulation or actual machining operations, while other generated candidates do not need to undergo this high-fidelity operation.

[0057] The resulting response is assigned fidelity or uncertainty and serves as evidence for the first set of Bayesian physical models, the second set of Bayesian response surrogate models, or both.

[0058] The processing stability evaluator, static / quasi-static model, and time-domain simulation can be implemented by any proven analytical, numerical, empirical, commercial, reduced-order, finite element, frequency domain, time domain, data-driven, or hybrid model, as long as it can provide the required stability or dynamic response output, without being limited to a specific stability equation, static force formula, regenerative force formula, or time integration method.

[0059] Simulation responses can supplement, initialize, or sort the model, but cannot replace at least one actual material removal feedback observation required for the first-level physical posterior correction.

[0060] The actual material removal feedback observations can be provided online / near real-time, after a single cutting / machining segment, periodically, or from a stored sequence of recorded cutting results; delayed processing of stored observations will not convert them into simulation evidence.

[0061] In summary, the method of the present invention is defined by four clearly connected parts: (1) updating the first Bayesian physical model calibration layer with physical meaning modal parameters or cutting process parameters; (2) transmitting the forced interlayer stability interface of the reconstructed posterior stability probability map and the bottom connected conservative safety envelope; (3) selecting the second Bayesian safety process optimization layer with high productivity processing conditions only within the transmitted posterior safety domain; and (4) returning the feedback to the process status monitoring of the first, second or both layers.

[0062] Please see Figure 1 Module 101 receives the nominal frequency response function / modal data and meshing machining feedback; Module 110 represents physical modal particles and soft stability likelihood; Module 120 outputs posterior modal parameters and posterior stability probability plots; Module 125 provides a bottom-connected safety envelope and a forced posterior stability interface. Module 130 generates a continuous candidate pool; Module 140 performs forced posterior lobes and machine tool constraint feasibility screening; Module 150 evaluates static or quasi-static physical responses and updates the Gaussian process surrogate model for force, power, torque, or related responses; Module 160 performs safety-screened material removal rate acquisition and forms a diversified shortlist. Module 170 performs regeneration dynamic verification and acceptance judgment for selected candidates. Module 180 represents optional machining implementation and acquisition in machining measurement feedback. The measurement feedback from Module 180 returns to the first-layer physical posterior and can further update the second-layer response or surrogate model. The first Bayesian layer reweights the physically evaluable modes and cutting process particles using feedback obtained during actual meshing material removal machining. The resulting posterior stable probability map and bottom connected safety envelope are passed to module 140 as a forced posterior safety domain. The second layer utilizes machine tool constraints and a low-cost static or quasi-static response model to screen a large candidate pool, performs safety-screened material removal rate acquisition, and sends only a shortlist of candidates to module 170 for selective regeneration time-domain verification and acceptance judgment. Accepted candidates can be implemented in module 180. When measurement processing feedback is obtained, this feedback is returned to the first Bayesian physical model calibration layer to update the posterior, and can further update the second-layer response surrogate model.

[0063] Please see Figure 2Modules 201, 202, and 203 provide tool-side modal priors, workpiece-side modal priors, and cutting process priors, respectively; module 210 performs bounded particle sampling and physical validity screening; modules 221 to 223 represent the physical model particles; module 230 calculates the relative cutting interface compliance; and module 240 performs particle-specific stability lobes or time-domain prediction. The tool-side modal priors, workpiece-side modal priors, and cutting process uncertainties are jointly sampled into bounded physical model particles. Each particle represents a physically consistent combination of tool dynamics, workpiece dynamics, cutting coefficients, process damping, and optional stability biases. The relative cutting interface compliance is calculated for each particle, and particle-specific stability lobes or time-domain responses are generated accordingly.

[0064] Please see Figure 3 The nominal modal prior 301 and particles 310 generate particle-specific stability lobes 320. Feedback logs 302 are evaluated under actual machining conditions, and optional evidence 303 is provided for likelihood calculation; module 330 predicts particle-specific stability probabilities; module 340 forms soft tags and optional frequency or sensor likelihoods; module 350 updates and normalizes particle weights, and optionally resamples; module 360 ​​outputs the posterior mean or maximum posterior modal parameters and intervals. Feedback logs are evaluated under their actual spindle speed, axial depth of cut, radial depth of cut, feed per tooth, and optional machining conditions. Each particle predicts the stability probability under these conditions. Observed stability, warnings, minor chatter, or chatter results, together with optional frequencies, forces, vibrations, or other evidence, determine the particle-specific likelihood. Weights are normalized and resampled as needed to generate posterior modal parameters and uncertainty intervals.

[0065] Please see Figure 4 A bottom connectivity scan 410 is performed on the posterior graph 401 of the stored radial cut depths; module 420 excludes isolated high-cut-depth safety islands; module 430 applies a safety margin; module 440 processes candidate radial cut depths between storage slices through interpolation; and module 450 outputs the resulting conservative safety envelope to the optimizer. For each stored radial cut-depth slice, the posterior safety boundary traces upwards from the lowest axial cut depth and terminates at the first unsafe point. This bottom connectivity rule excludes isolated high-cut-depth probability islands. The configurable axial cut-depth margin is then subtracted, and bounded interpolation is used to evaluate candidate radial cut depths between storage slices before providing the resulting conservative envelope to the optimizer.

[0066] Please see Figure 5AThe second Bayesian optimization layer starts with candidate condition 510, applies posterior lobe safety screening 520 and machine tool or meshing feasibility checks 530, calculates the low-cost physical response in module 540, updates the Gaussian process response surrogate model in module 550, ranks feasible candidates through safe Bayesian acquisition in module 560, selects the next candidate and updates the Gaussian process in module 570, forms a diversified safety margin shortlist in module 580, and sends the selected candidates to time-domain / fast Fourier transform verification in module 590. In a preferred embodiment, the stability region used in module 520 is provided by the experimentally updated posterior lobe, rather than by the independent stability Gaussian process. Candidate conditions are first screened by posterior lobe stability requirements and machine tool or meshing constraints. The low-cost physical response is then calculated and used to update the second-layer Gaussian process response surrogate model. The safe-screened Bayesian acquisition only ranks feasible candidates, updates the candidate search, and forms a diversified shortlist. The selected shortlist candidates are then passed to a higher-fidelity verification stage.

[0067] Please see Figure 5B Module 510 provides candidate machining conditions and posterior physical parameters to the multi-fidelity physical response evaluator. Modules 541, 542, 543, and 544 represent static chip thickness calculation, static cutting force calculation, use of posterior stiffness or flexibility, and generation of static force, deformation, power, and torque outputs, respectively. Modules 550 and 560 utilize low-cost responses for Gaussian process response learning and secure acquisition sorting; modules 570 and 580 form a selected or shortlist of candidate sets. The shortlist is first provided to module 592. Modules 592, 593, 594, and 597 / 598 then sequentially calculate regenerated chip thickness, dynamic cutting force, posterior modal response, numerical integration, and time-domain or frequency-domain stability evidence. Dynamic evidence or measured machining response can return a first Bayesian physical posterior, a second Bayesian response surrogate model, or both. Static or quasi-static paths evaluate a large number of candidates through static undeformed chip thickness, cutting force, posterior stiffness or flexibility, and force, deformation, power, and torque calculations. These outputs support Gaussian process response learning and secure acquisition. Modules 570 / 580 provide a shortlist to module 592, rather than directly to the final evidence module. The selected candidates are sequentially processed by module 592 for regenerated chip thickness calculation, module 593 for dynamic cutting force calculation, module 594 for posterior modal response, and modules 597 / 598 for numerical integration and time / frequency domain evidence calculation. The resulting evidence can update the physical posterior, the response surrogate model, or both.

[0068] Please see Figure 6The initial information module 601 provides tool-side and workpiece-side modal parameters, tool / workpiece / material information, and machine tool constraints or machining requirements. Module 602 generates Bayesian modal particle posterior and stability lobes; module 603 updates the Gaussian process response surrogate model using time-domain prediction and / or verification results; module 604 performs constrained Bayesian optimization; and module 605 provides the next candidate machining condition. The decision module 606 determines whether to accept the candidate condition. Accepted conditions are applied through module 607, while module 608 continues with the current machining conditions. The machining system 609, sensors 610, data acquisition system 611, and process status and stability monitoring / classification module 612 provide asynchronous feedback to update the physical posterior, stability lobes, response surrogate model, or optimization status. Initial tool, workpiece, material, modal, and machine tool constraint information is used for posterior stability lobe generation, response surrogate model updates, and constrained Bayesian optimization. Candidate conditions are accepted only when the degree of improvement, safety, magnitude of change, and permissible process boundary requirements are all met. Accepted conditions are applied at approved boundaries; otherwise, the current validated conditions continue. Sensor and controller data are acquired during processing and can be processed synchronously or asynchronously to update the first-layer posterior, stability lobes, second-layer response surrogate model, or optimization state.

[0069] The present invention also provides a processing optimization apparatus, including one or more processors and a memory storing instructions, wherein when the instructions are executed by the one or more processors, the apparatus performs the method described above.

[0070] The present invention also provides a non-transitory computer-readable storage medium having instructions stored thereon, which, when executed by one or more processors, cause the one or more processors to perform the method described in any one of the above.

[0071] This invention explicitly uses actual material removal feedback obtained during the dynamic engagement of the tool and workpiece; therefore, static modal test data, finite metadata, or offline simulation data alone cannot meet the required physical posterior correction. The necessary technical source for the feedback is its acquisition during the actual engagement material removal process; it does not require simultaneous first Bayesian calculation and acquisition, and covers real-time, post-cutting, periodic, and stored record processing.

[0072] For clarity, an optional geometric position-modal predictor can provide prior values ​​of workpiece modalities, but the required first Bayesian mechanism must subsequently use evidence of the machining results for likelihood-based correction and reconstruction of posterior stability information; the required second Bayesian mechanism is an independent response optimization process that operates within the reconstructed posterior safety domain.

Claims

1. A method for selecting processing conditions using a dual Bayesian approach, characterized in that, The machining conditions include spindle speed n and axial depth of cut. Radial depth of cut and feed per tooth The method includes: A first Bayesian physical model calibration layer is constructed, which includes multiple physical model particles, each of which includes uncertain modal parameters and cutting process parameters; Receive process feedback records obtained during machining processes involving dynamic engagement between the tool and the workpiece and material removal, the process feedback records including machining conditions and stability observations; For each of the aforementioned process feedback records, the observed processing results under the aforementioned processing conditions are compared with the physical stability evaluation results of each of the aforementioned physical model particles under the aforementioned processing conditions. The particle-specific likelihood is calculated, and the probability of each of the aforementioned physical model particles is updated based on the particle-specific likelihood to obtain the posterior physical model. The posterior stability probability information is reconstructed from the posterior physical model, and the bottom connected conservative safety envelope is determined based on the posterior stability probability information. By using the forced inter-layer stability interface, the posterior stability probability information and the bottom-connected conservative safety envelope are provided as forced feasibility constraints to the second Bayesian safety process optimization layer. In the second Bayesian safety process optimization layer, a candidate pool is generated and feasible subsets are formed through mandatory constraint screening. Only in the feasible subset Within the posterior safety region defined by the posterior stability probability information and the bottom-connected conservative safety envelope, a second Bayesian response surrogate model is maintained, and productivity-related or unstable responses are modeled using the second Bayesian response surrogate model. The next processing condition is selected using a Bayesian acquisition function. The selected processing conditions are applied to the processing system or simulated, and the resulting process status monitoring results are returned to the first Bayesian physical model calibration layer, the second Bayesian safety process optimization layer, or both, to close the double Bayesian loop.

2. The method according to claim 1, characterized in that, Each of the physical model particles includes one or more tool-side modes of the spindle-tool holder-tool system and one or more workpiece-side modes of the workpiece-fixture-machine tool system. Each mode includes natural frequency, modal stiffness, damping ratio, modal orientation or modal participation parameters. The machining stability evaluator uses the tool's compliance with the workpiece at the cutting interface. When a position-dependent natural frequency or mode shape predictor is used, its output is used as a prior or state condition input before the processing feedback likelihood update, and the updated posterior value is allowed to deviate from the predictor output. Perform a Bayesian update on at least one of the natural frequency, modal stiffness, damping ratio, modal participation, cutting coefficient, process damping parameter, or stability bias term, and change at least one of the stability lobe peak spindle speed, lobe height, dominant mode, or meshing-related stability limit. The tool-side modal parameters are state-dependent parameters and are corrected based on actual cutting evidence related to tool wear, edge damage, built-up edge, runout, spindle speed, thermal state, cutting load, or spindle-tool holder connection state. The workpiece-side modal parameters are modified based on material removal, changes in contact position, wall thickness, remaining geometry, fixture load, or machining stage.

3. The method according to claim 1, characterized in that, The process feedback records obtained under different radial depths of cut, feed per tooth, workpiece area, or machining conditions retain their respective actual coordinates. The particle-specific likelihood also includes a frequency likelihood, which increases the weight of particles that predict the dominant tool-side, workpiece-side, or coupled chatter frequencies and the measured chatter frequencies or related tooth-passing sidebands.

4. The method according to claim 1, characterized in that, When the posterior stability probability exceeds the selected stability threshold, but the process feedback record indicates flutter, apply difference-sensitive likelihood, reduce the weight of predicted stable particles, increase the relative weight of predicted unstable particles or particles that conform to the observed flutter frequency behavior, and immediately reconstruct the posterior stability probability information and the bottom connected conservative safety envelope before executing / recommending another condition. The process feedback record indicating flutter is obtained during the actual material removal process, representing the meshing operation status.

5. The method according to claim 1, characterized in that, Based on the shared posterior distribution of the posterior physical model, the posterior stable probability information is generated in multiple radial depth slices; At each spindle speed, the bottom connected safety envelope independently traces upwards along the axial depth of cut from the lowest axial depth of cut grid point and stops at the first a posteriori unsafe point; For candidate radial cut depths located between storage slices, the posterior stability probability and the safe axial cut depth of the bottom-connected conservative safety envelope are both calculated by bounded interpolation between adjacent slices.

6. The method according to claim 1, characterized in that, The feasible subset Further requirements are made for the predicted spindle power P and predicted spindle torque T to satisfy: ； in, and The machine tool power curve and rated torque curve are related to the rotational speed; For configurable utilization factors; candidates falling outside the curve domain defined by the machine tool's rated power curve and rated torque curve are rejected; cutting force, feed rate, vibration, surface quality, temperature, tool wear, dimensional error, or feedback support confidence are also imposed as mandatory or probabilistic constraints.

7. The method according to claim 1, characterized in that, The second Bayesian response surrogate model is trained using analytical computation, low-cost simulation, selected time-domain simulation, measurement of processing response, or a combination thereof; The input to the second Bayesian response surrogate model includes at least one machining decision variable selected from spindle speed, axial depth of cut, radial depth of cut, and feed per tooth; The second Bayesian response surrogate model models productivity-related responses, instability responses, comprehensive scores, or physical model residuals. Where x represents the processing parameters; r(x) represents the actual or measured process response obtained under the processing parameters; Indicates the estimated value based on the current physical parameters. The obtained physical model predicts the response; This represents the residual between the actual or measured process response and the physical model's predicted response; The Bayesian acquisition function includes expected improvement, confidence upper bound, improvement probability, information gain, hypervolume improvement, or a combination thereof, but is evaluated only after applying the posterior stability of the first Bayesian physical model calibration layer and the bottom connectivity conservative safety envelope screening.

8. The method according to claim 1, characterized in that, Multiple candidates in the candidate pool are evaluated by a low-cost static or quasi-static cutting force model to compute at least one of force, deformation, spindle power, spindle torque, material removal rate, or response surrogate training value. Selected candidates, initial design points, high a posteriori divergence points, boundary verification points, or shortlisted candidates are selectively evaluated through regenerative time-domain dynamic cutting simulation or actual machining operations. The resulting response is assigned fidelity or uncertainty and serves as evidence for the physical model particles of the first Bayesian physical model calibration layer, the second Bayesian response surrogate model, or both.

9. An apparatus comprising one or more processors and a memory storing instructions, characterized in that, When the instructions are executed by the one or more processors, the apparatus performs the method of any one of claims 1-8.

10. A storage medium storing instructions thereon, characterized in that, When the instructions are executed by one or more processors, the one or more processors perform the method of any one of claims 1-8.

Images