Intelligent machining control method and system based on physical information neural network

Abstract

The application provides an intelligent machining control method and system based on a physical information neural network, and belongs to the field of precision machining technology, comprising the following steps: constructing a physical kernel by coupling the physical laws of machining process force, heat and vibration, combining the physical kernel with a constructed gated residual compensation network, and forming a hybrid intelligent model; establishing a digital twin with the model as a core prediction engine, and synchronously updating the digital twin by using multi-source state data; in the digital twin, using the hybrid intelligent model to perform rolling time domain forward-looking simulation based on the current synchronization state, predicting the machining results of the force, heat and vibration coupling in the future period, generating process parameter adjustment decision instructions based on the results, and delivering the instructions to a physical machining entity execution mechanism; and performing online incremental learning on the gated residual compensation network based on the machining data after the execution of the instructions, so as to realize dynamic self-adaptive optimization. The scheme realizes high-precision forward-looking prediction of the machining process and real-time active closed-loop control based on the prediction.

Description

Technical Field

[0001] This application relates to the field of precision machining technology, and in particular to an intelligent machining control method and system based on physical information neural networks. Background Technology

[0002] In the precision machining of complex aerospace components, the strong coupling and dynamic time-varying characteristics of multiple physical fields such as cutting force, heat, and vibration are the core challenges leading to machining deformation, chatter, and out-of-tolerance issues. Existing technological solutions all have fundamental flaws in addressing this challenge:

[0003] Offline simulation compensation methods based on physical mechanisms rely on a large number of idealized assumptions, making it difficult to accurately depict the real dynamic machining process. Furthermore, they are completely unable to perceive and compensate for random time-varying disturbances such as tool wear and clamping deformation that occur during machining, resulting in limited compensation effects.

[0004] The adaptive control method based on sensor data is essentially a "reactive" control method. It can only make passive adjustments after flutter or out-of-tolerance occurs, and lacks the ability to predict and intervene in advance the dynamic evolution of the processing process.

[0005] Purely data-driven processing prediction models are "black box" models. The prediction results lack physical consistency guarantees and rely heavily on a large amount of labeled data. They have weak generalization ability and insufficient engineering reliability in typical small-sample, multi-condition scenarios in the aerospace field.

[0006] In summary, existing technologies cannot achieve high-precision forward prediction of the processing and prediction-based real-time active closed-loop control. Therefore, there is an urgent need for a new method that can deeply integrate physical mechanisms and real-time data to drive an intelligent closed loop of "perception-prediction-decision-execution". Summary of the Invention

[0007] In view of this, embodiments of this application provide an intelligent machining control method and system based on physical information neural networks, which can achieve high-precision forward prediction of the machining process and real-time active closed-loop control based on prediction.

[0008] In a first aspect, embodiments of this application provide an intelligent processing control method based on a physical information neural network, comprising: The physical laws describing the coupling process of force, heat and vibration in the cutting process are constructed into a differentiable computational graph, which constitutes the physical kernel. A gated residual compensation network based on tool contact, local stiffness and mode is constructed, and the gated residual compensation network is combined with the physical kernel to form a hybrid intelligent model that deeply integrates physical mechanism and data-driven approach. Establish a digital twin with a hybrid intelligent model as the core prediction engine, collect multi-source state data of the physical processing entity in real time, and use the collected multi-source state data to synchronously update the state of the digital twin to achieve virtual-real synchronization. In the digital twin, a hybrid intelligent model is used to perform rolling time-domain look-ahead simulation based on the current synchronization state to predict the processing results caused by the coupling of force, heat and vibration in the future period, and generate risk characterization results based on the prediction results; when the risk characterization results indicate that the process parameters will be exceeded, corresponding process parameter adjustment decision instructions are generated. The process parameter adjustment decision command is sent to the execution mechanism of the physical processing entity to adjust the processing process in real time; the processing data after the decision command is executed is used as feedback to perform online incremental learning on the gated residual compensation network, realize dynamic adaptive optimization, and complete the closed-loop intelligent control of the processing process.

[0009] According to a specific implementation of an embodiment of this application, the construction of a gated residual compensation network based on tool position contact, local stiffness, and modal characteristics includes: A tool contact state encoding module is constructed. This module is used to calculate the instantaneous contact angle range, cutting-in and cutting-out states, undeformed chip thickness, local material removal rate and tool tooth phase angle at the current sampling moment based on the CNC program, tool radius, number of tool teeth, spindle speed, feed rate, axial depth of cut, radial width of cut, tool position coordinates and current material removal state, and form a tool-workpiece contact state vector. A thin-walled local stiffness mapping module is constructed. This module is used to calculate the local wall thickness, wall thickness gradient, equivalent compliance, and first to third order local natural frequencies and corresponding damping ratios of the current processing area based on the workpiece CAD model, online thickness measurement data, current local remaining wall thickness, local curvature, distance from clamping boundary, and rib adjacency relationship, forming a local stiffness description vector. A tooth frequency synchronization signal encoding module is constructed. This module is used to synchronously resample the three-dimensional cutting force signal and vibration signal according to the spindle rotation angle or the tooth passing cycle, and extract the tooth frequency fundamental wave, harmonics, chatter sideband, main frequency offset, vibration phase lag, temperature rise rate and temperature lag to form a force, heat and vibration synchronization state vector. A bounded physical parameter correction module is constructed. This module takes the tool-workpiece contact state vector, the local stiffness description vector, and the force, heat, and vibration synchronization state vector as inputs, and outputs one or more of the following corrections: tangential cutting force coefficient correction, radial cutting force coefficient correction, axial cutting force coefficient correction, cutting heat distribution coefficient correction, convective heat transfer coefficient correction, local compliance correction, modal frequency correction, and modal damping correction. Each correction is constrained within a preset physical range by a limiting function. A modal-gated residual decoding module is constructed. This module is used to receive the correction amount output by the physical parameter bounded correction module and express the residual deformation compensation amount as a weighted superposition of multiple local modal bases. The weighting coefficients of each local modal base are jointly determined by local compliance, tooth frequency energy, flutter sideband energy and temperature gradient. Each local modal base is determined by the current local residual wall thickness, clamping boundary and local geometric state. A residual amplitude output module is constructed, which is used to output the residual deformation compensation amount output by the modal gated residual decoding module.

[0010] According to a specific implementation of an embodiment of this application, the step of using a hybrid intelligent model to perform rolling time-domain look-ahead simulation based on the current synchronization state includes: The digital twin receives multi-source state data from the physical processing entity during each synchronization cycle, and constructs the current synchronization state vector X based on the multi-source state data. t The expression is: , Where, q t Given the current tool position and tool axis posture, n t Main spindle speed, v ft For the feed rate, a pt and a et These are the axial depth of cut and the radial width of cut, respectively, F t For triaxial cutting force, T t For the cutting zone temperature, a t For vibration response, h t C represents the current local remaining wall thickness. t For local equivalent compliance, Ω t Z is the set of local modal frequencies. t This indicates the current material removal status. With the current synchronization state vector X t Starting from the current time, the tool path, material removal status and candidate process parameters are used to progressively deduce the tool-workpiece contact status, undeformed chip thickness, local material removal rate, local remaining wall thickness and local stiffness changes at each discrete time. Using a hybrid intelligent model, the predicted cutting force is obtained based on the contact state between the tool and the workpiece at each discrete time in the future. The cutting heat input and temperature field diffusion are calculated based on the predicted cutting force, contact time, tool posture and local material removal rate. The predicted cutting force and temperature field diffusion are used as load inputs to the structural dynamics module to calculate the deformation response and vibration response of the workpiece in the future time domain. For the processing area of ​​thin-walled components, the hybrid intelligent model calls the correction amount output by the gated residual compensation network to perform bounded correction on the mechanical, thermal and dynamic parameters of the physical kernel, and obtain the compensated prediction results.

[0011] According to a specific implementation of an embodiment of this application, the risk characterization result includes: Cutting force prediction curves and force mutation risk values ​​are used to determine whether the load on the tool exceeds the preset safety threshold; Temperature field prediction results and thermal hysteresis risk values ​​are used to determine whether local thermal deformation will lead to dimensional deviations. The workpiece surface deformation cloud map and local residual wall thickness deviation map are used to determine whether there is processing deformation or overweight risk in the processing area of ​​thin-walled components; Vibration spectrum prediction results and chatter probability diagrams are used to determine whether the tool position will enter an unstable cutting zone in the future; The comprehensive processing risk index is used to uniformly characterize the quality risk and stability risk at future points in time. The expression for the comprehensive processing risk index is: , Where R(t+τ) is the comprehensive processing risk index at the future time t+τ, R u (t+τ) represents the deformation deviation risk term at the future time t+τ, R f (t+τ) represents the risk term of excessive cutting force at the future time t+τ, R T (t+τ) represents the risk term for temperature exceeding the limit or thermal hysteresis at the future time t+τ, R v (t+τ) represents the flutter risk term at the future time t+τ, R h (t+τ) represents the local residual wall thickness deviation risk term at the future time t+τ, ω u ω f ω T ω v and ω h These are the weighting coefficients for the following risk items: deformation exceeding tolerance, cutting force exceeding limit, temperature exceeding limit or thermal hysteresis, chatter, and local residual wall thickness deviation.

[0012] According to a specific implementation of an embodiment of this application, the step of generating a corresponding process parameter adjustment decision instruction when the risk characterization result indicates that the risk will exceed a preset threshold includes: When the risk characterization results indicate that the risk will exceed a preset threshold, the risk type is determined. Based on the risk type, multiple candidate process parameter adjustment schemes are generated; For each candidate process parameter adjustment scheme, the digital twin re-executes the rolling time-domain look-ahead simulation and calculates the evaluation function J corresponding to each candidate process parameter adjustment scheme.m The expression is: , Among them, R m E represents the remaining risk value after the candidate process parameter adjustment plan is implemented. m S represents the processing efficiency loss term. m Q is a constraint term for the smoothness of process parameter changes. m The constraint penalty term is applied to the machine tool, where α, β, γ, and η are R... m E m S m and Q m The corresponding weighting coefficients; Select J m The candidate process parameter adjustment scheme that is the smallest and satisfies the machine tool execution constraints is taken as the final process parameter adjustment decision instruction.

[0013] According to a specific implementation of an embodiment of this application, the step of generating multiple candidate process parameter adjustment schemes based on risk type includes: If a chatter risk is identified, a spindle speed adjustment scheme is generated based on the local modal frequency, the cutter tooth passing frequency, and the stability domain prediction results, so that the cutter tooth passing frequency avoids the current local modal sensitive frequency band. If the risk of deformation exceeding tolerance is determined, an adjustment plan will be generated first, such as reducing the feed rate, reducing the radial cutting width, or adjusting the machining allowance in the machining area of ​​the thin-walled component. If the risk of local residual wall thickness deviation is determined, an online measurement path is generated, and the adjustment scheme of local wall thickness state in the digital twin is updated after the measurement path is completed. If the risk of thermal deformation is identified, an adjustment plan is generated to reduce the material removal rate, extend the cooling interval, or adjust the toolpath sequence. If the tool is deemed to be at risk of being overloaded, an adjustment plan is generated to reduce the instantaneous cutting load or change the entry and exit methods.

[0014] According to a specific implementation of an embodiment of this application, the step of using the processed data after the execution of the decision instruction as feedback to perform online incremental learning on the gated residual compensation network to achieve dynamic adaptive optimization includes: Based on the predicted state before the execution of the decision instruction, the content of the decision instruction, the measured state after the execution of the decision instruction, and the quality inspection results, a sample of execution effect feedback is constructed. Based on the execution effect feedback samples, the force residual between the predicted cutting force and the measured cutting force, the thermal residual between the predicted temperature and the measured temperature, the vibration residual between the predicted vibration and the measured vibration, the morphology residual between the predicted deformation and the measured deformation, and the wall thickness residual between the predicted local residual wall thickness and the online thickness measurement result are calculated respectively. If any residual exceeds the corresponding threshold continuously, the online incremental learning of the gated residual compensation network is triggered. Online incremental learning employs a multi-objective loss function to update the bounded correction module of the physical parameters of the gated residual compensation network, the modal gated residual decoding module, and the residual amplitude output module. After each update, a limit check is performed on each correction value. After the update is complete, verify the validity of the update. The gated residual compensation network parameters, verified for effectiveness, are used for the next rolling time-domain look-ahead simulation and process parameter adjustment decisions, forming a dynamic adaptive optimization closed loop.

[0015] According to a specific implementation of an embodiment of this application, the expression for the multi-objective loss function of online incremental learning is: , Among them, L online L is the total loss function for online incremental learning. exe L phy L modal L thick L ctrl L bound and L mem The terms are, in order: execution effect error, physical law residual, thin-walled modal consistency, wall thickness consistency, control command smoothing, parameter bounded, and anti-forgetting. λ1, λ2, λ3, λ4, λ5, λ6, and λ7 are, in order: execution effect error, physical law residual, thin-walled modal consistency, wall thickness consistency, control command smoothing, parameter bounded, and anti-forgetting. The expression for the execution effect error term is: , in, For the Euclidean norm, F pre and F mea These are the predicted cutting force and the measured cutting force, T. pre and T mea These are the predicted temperature and the measured temperature, respectively. pre and a mea These are the predicted vibration response and the measured vibration response, respectively. pre and u meaThese represent the predicted deformation and the measured deformation, respectively, with ωF, ωT, ωa, and ωu being the first, second, third, and fourth weighting coefficients, respectively. The expression for the residual term of the physical law is: , in, C and K are the local mass matrix, damping matrix, and stiffness matrix of the current thin-walled component, respectively; To predict the deformation response; To predict deformation response The first derivative with respect to time; To predict deformation response Second derivative with respect to time; Fc is the cutting force load; ρ is the material density, c is the specific heat capacity, k is the thermal conductivity, q c For cutting heat source term; T is the predicted temperature field. To predict the first-order partial derivative of the temperature field T with respect to time t; To predict the gradient of the temperature field T in spatial coordinates; For thermal conduction and diffusion terms; The expression for the thin-walled modal consistency term is: , Where Δu is the residual deformation compensation amount, r is the selected local modal order, and ψ i (x) is the i-th local modal basis, ψ j (x) is the j-th order local modal basis, g i (t) represents the gating coefficients of the i-th mode. Let i be the amplitude of the i-th mode. This represents the contribution of the i-th local mode to the residual deformation compensation. This represents the modal reconstruction deformation caused by the superposition of r-order local modes; Denotes the i-th order local modal basis ψ i (x) and the j-th order local modal basis ψ j (x) is the inner product of the local mass matrix M of the current thin-walled member, weighted by weights. Represents ψ i Transpose of (x); The expression for the wall thickness consistency term is: , Among them, h pre To predict the local residual wall thickness, h mea Δh represents the actual local residual wall thickness obtained from online thickness measurement. pre To predict the wall thickness gradient, Δh mea This represents the measured wall thickness gradient. The expression for the control instruction smoothing term is: , Among them, u t For the current decision instruction, u t-1 For the decision instruction of the previous cycle, Δu t Let Δu be the change in the current decision instruction. t-1 This represents the change in decision instructions from the previous cycle. The expression for the bounded term with parameters is: , Where i' is the index of the physical parameter to be corrected, p is the total number of physical parameters to be corrected, and Δp i’ Let Δp be the correction amount for the i'th physical parameter to be corrected. i’,max The maximum allowable correction magnitude for the physical parameter to be corrected is given by max(·), where max(·) is the function for maximizing the value. The expression for the anti-forgetting item is: , Among them, G new For the updated compensation network, G old For the compensation network before the update, X k’ This is the k'th representative sample selected from the historical stable sample buffer, where k' is the index of the representative sample and K' is the total number of representative samples.

[0016] According to a specific implementation of an embodiment of this application, the physical kernel includes a differentiable cutting mechanics module, a heat transfer module, and a structural dynamics module. The cutting mechanics module is used to predict the cutting force based on the tool-workpiece contact state, the thickness of the undeformed chip, the tool tooth phase angle, the spindle speed, the feed rate, the axial depth of cut, and the radial width of cut. The heat transfer module is used to predict the temperature field based on the cutting force, the contact time, the material removal rate, and the cutting heat source distribution state. The structural dynamics module is used to predict the workpiece deformation and vibration response based on the cutting load, the temperature field, the current local remaining wall thickness of the workpiece, the local compliance, and the clamping boundary state.

[0017] Secondly, embodiments of this application also provide an intelligent processing control system based on a physical information neural network, used to execute the intelligent processing control method based on a physical information neural network as described in any embodiment of the first aspect, the system comprising: The hybrid intelligent modeling module is used to construct a differentiable computational graph from the physical laws describing the coupling process of force, heat and vibration in the cutting process, forming a physical kernel; it constructs a gated residual compensation network based on tool contact, local stiffness and modality, and combines the gated residual compensation network with the physical kernel to form a hybrid intelligent model that deeply integrates physical mechanism and data-driven approach. The digital twin module is used to build a digital twin with a hybrid intelligent model as the core prediction engine, collect multi-source state data of the physical processing entity in real time, and use the collected multi-source state data to synchronously update the state of the digital twin to achieve virtual-real synchronization. The forward-looking decision-making module is used in the digital twin to perform rolling time-domain forward-looking simulation based on the current synchronization state using a hybrid intelligent model. It predicts the processing results caused by the coupling of force, heat and vibration in the future period and generates risk characterization results based on the prediction results. When the risk characterization results indicate that the process parameters will be exceeded, it generates corresponding process parameter adjustment decision instructions. The closed-loop control and model optimization module is used to send process parameter adjustment decision commands to the actuators of the physical processing entity to adjust the processing process in real time; the processing data after the execution of the decision commands is used as feedback to perform online incremental learning on the gated residual compensation network to achieve dynamic adaptive optimization and complete the closed-loop intelligent control of the processing process.

[0018] Beneficial effects: The intelligent machining control method and system based on physical information neural networks in this application, through the deep integration of physical mechanisms and data-driven approaches, enables the model to maintain high-precision predictions even with limited training data, demonstrating significantly better generalization capabilities than purely data-driven methods. Digital twins combined with rolling time-domain look-ahead simulation can issue early warnings and generate adjustment instructions before machining problems occur, achieving prevention rather than remediation. A fully closed-loop intelligent control chain is formed, with each link organically connected from perception to decision-making to execution and feedback, significantly improving the automation level and robustness of the machining process. This method can be widely applied in fields with high machining quality requirements, such as aerospace thin-walled parts, precision molds, and difficult-to-machine materials, showing significant engineering application prospects and economic benefits. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of an intelligent machining control method based on a physical information neural network according to an embodiment of the present invention; Figure 2 This is a schematic diagram of an intelligent machining control system based on a physical information neural network according to an embodiment of the present invention. Detailed Implementation

[0021] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0022] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this application, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using other structures and / or functionalities besides one or more of the aspects set forth herein.

[0024] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. The illustrations only show the components related to this application and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0025] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0026] In a first aspect, embodiments of this application provide an intelligent processing control method based on a physical information neural network, referring to... Figure 1 This includes the following steps: S1. The physical laws describing the coupling process of force, heat and vibration in the cutting process are constructed into a differentiable computational graph to form a physical kernel; a gated residual compensation network based on tool contact, local stiffness and modality is constructed, and the gated residual compensation network is combined with the physical kernel to form a hybrid intelligent model that deeply integrates physical mechanism and data-driven approach. S2. Establish a digital twin with a hybrid intelligent model as the core prediction engine, collect multi-source state data of the physical processing entity in real time, and use the collected multi-source state data to synchronously update the state of the digital twin to achieve virtual-real synchronization. S3. In the digital twin, a hybrid intelligent model is used to perform rolling time-domain look-ahead simulation based on the current synchronization state to predict the processing results caused by the coupling of force, heat and vibration in the future period, and generate risk characterization results based on the prediction results; when the risk characterization results indicate that the preset threshold will be exceeded, corresponding process parameter adjustment decision instructions are generated. S4. The process parameter adjustment decision command is sent to the execution mechanism of the physical processing entity to adjust the processing process in real time; the processing data after the decision command is executed is used as feedback to perform online incremental learning on the gated residual compensation network to achieve dynamic adaptive optimization and complete the closed-loop intelligent control of the processing process.

[0027] In this embodiment, the deep integration of physical mechanisms and data-driven approaches enables the model to maintain high-precision predictions even with limited training data, demonstrating significantly better generalization capabilities than purely data-driven methods. Digital twins combined with rolling time-domain look-ahead simulation can issue early warnings and generate adjustment instructions before processing problems occur, achieving prevention rather than remediation. A fully closed-loop intelligent control chain is formed, with each link organically connected from perception to decision-making to execution and feedback, significantly improving the automation level and robustness of the processing. This method can be widely applied in fields with high processing quality requirements, such as aerospace thin-walled parts, precision molds, and difficult-to-machine materials, showing significant engineering application prospects and economic benefits.

[0028] In one embodiment, the physical kernel includes at least a differentiable cutting mechanics module, a heat transfer module, and a structural dynamics module, which are used to predict cutting forces, predict temperature fields, and predict workpiece deformation and vibration responses, respectively.

[0029] Specifically, the physical laws describing milling are first modified to be differentiable, forming a physical kernel with computable gradients. The cutting mechanics module is used to predict cutting forces based on the tool-workpiece contact state, undeformed chip thickness, tool tooth phase angle, spindle speed, feed rate, axial depth of cut, and radial width of cut; the heat transfer module is used to predict the temperature field based on cutting forces, contact time, material removal rate, and cutting heat source distribution; and the structural dynamics module is used to predict workpiece deformation and vibration response based on cutting load, temperature field, current local remaining wall thickness, local compliance, and clamping boundary state.

[0030] Due to the rapid changes in the tool-workpiece contact state, gradual thinning of local wall thickness, dynamic decrease in structural stiffness, modal frequency drift, and thermal hysteresis during the machining of thin-walled aerospace components, relying solely on the physical kernel is insufficient to accurately characterize the actual machining state. Therefore, this application constructs a gated residual compensation network based on tool contact, local stiffness, and modal response outside the physical kernel. This compensation network does not directly perform general residual fitting on the output results of the physical kernel, but rather performs bounded corrections on the cutting force parameters, heat source parameters, local compliance parameters, and modal response parameters within the physical kernel.

[0031] In one embodiment, constructing a gated residual compensation network based on tool contact, local stiffness, and modal characteristics includes: A tool contact state encoding module is constructed. This module is used to calculate the instantaneous contact angle range, cutting-in and cutting-out states, undeformed chip thickness, local material removal rate and tool tooth phase angle at the current sampling moment based on the CNC program, tool radius, number of tool teeth, spindle speed, feed rate, axial depth of cut, radial width of cut, tool position coordinates and current material removal state, and form a tool-workpiece contact state vector. A thin-walled local stiffness mapping module is constructed. This module is used to calculate the local wall thickness, wall thickness gradient, equivalent compliance, and first to third order local natural frequencies and corresponding damping ratios of the current processing area based on the workpiece CAD model, online thickness measurement data, current local remaining wall thickness, local curvature, distance from clamping boundary, and rib adjacency relationship, forming a local stiffness description vector. A tooth frequency synchronization signal encoding module is constructed. This module is used to synchronously resample the three-dimensional cutting force signal and vibration signal according to the spindle rotation angle or the tooth passing cycle, and extract the tooth frequency fundamental wave, harmonics, chatter sideband, main frequency offset, vibration phase lag, temperature rise rate and temperature lag to form a force, heat and vibration synchronization state vector. A bounded physical parameter correction module is constructed. This module takes the tool-workpiece contact state vector, the local stiffness description vector, and the force, heat, and vibration synchronization state vector as inputs, and outputs one or more of the following corrections: tangential cutting force coefficient correction, radial cutting force coefficient correction, axial cutting force coefficient correction, cutting heat distribution coefficient correction, convective heat transfer coefficient correction, local compliance correction, modal frequency correction, and modal damping correction. Each correction is constrained within a preset physical range by a limiting function to ensure that the compensation results remain physically interpretable and machine tool execution safety. A modal-gated residual decoding module is constructed. This module is used to receive the correction amount output by the physical parameter bounded correction module and express the residual deformation compensation amount as a weighted superposition of multiple local modal bases. The weighting coefficients of each local modal base are jointly determined by local compliance, tooth frequency energy, flutter sideband energy and temperature gradient. Each local modal base is determined by the current local residual wall thickness, clamping boundary and local geometric state. A residual amplitude output module is constructed, which is used to output the residual deformation compensation amount output by the modal gated residual decoding module.

[0032] Furthermore, the modal-gated residual decoding module restricts the residual deformation compensation to the local modal basis of the thin-walled component, and its output satisfies: ; Where N is the total modal order, and Δu(x,t+τ) is the position of the workpiece surface at the future time t+τ. The residual deformation compensation amount at the location, ψ i (x) is the i-th local modal basis, ψ i (x) is determined by the current local remaining wall thickness, clamping boundary, and local geometry, a i (t+τ) represents the predicted amplitude of the i-th mode, g i (t) represents the gating coefficient of the i-th mode, g i (t) is determined by the local wall thickness, equivalent flexibility, tooth frequency energy, flutter sideband energy, and temperature gradient. Through the above settings, the residual deformation field output by the gated residual compensation network is limited to the range of force-induced deformation, thermal deformation, and vibration deformation modes that may actually occur in aerospace thin-walled components, thus avoiding arbitrary residual compensation without physical constraints.

[0033] Furthermore, the hybrid intelligent model is trained using a multi-objective loss function. This multi-objective loss function includes cutting force prediction error terms, temperature prediction error terms, deformation prediction error terms, online thickness measurement error terms, structural dynamics residual terms, heat conduction residual terms, modal consistency constraint terms, and parameter bounded constraint terms. Through this training method, the hybrid intelligent model not only approximates historical machining data and measured feedback data, but also ensures that the model output satisfies the constraints of cutting mechanics, heat transfer, and structural dynamics.

[0034] In one embodiment, step S2 specifically includes: On an edge server or industrial computer, create a virtual environment that includes machine tool geometry, workpiece CAD model, tool model, fixture model, material removal status, and online thickness measurement status. Integrate the hybrid intelligent model formed in step S1 into the virtual environment as the core prediction engine, enabling the digital twin to have dynamic prediction capabilities. A multi-source sensor network is deployed on the physical machining entity, including a smart tool holder or force measuring platform for acquiring three-dimensional cutting forces and torques, an infrared thermal imager or thermocouple for acquiring temperature fields, an accelerometer for acquiring vibration responses, a grating ruler or machine tool CNC system for acquiring position status, and an online measuring device for acquiring local residual wall thickness or local topography. The acquired data is aggregated to an edge computing gateway via fieldbus or industrial Ethernet. After filtering, outlier removal, timestamp synchronization, and coordinate unification of the multi-source data, the gateway drives the digital twin to update its current state at fixed intervals.

[0035] The synchronized state of a digital twin includes at least the current tool position coordinates, tool axis posture, spindle speed, feed rate, axial depth of cut, radial width of cut, three-dimensional cutting forces, spindle power, cutting zone temperature, vibration response, online thickness measurement data, current local remaining wall thickness, local equivalent compliance, local modal frequency, and material removal status. Through these settings, a complete digital mirror of the machining process is achieved, providing a synchronized state foundation for subsequent forward simulation and proactive decision-making.

[0036] In one embodiment, the rolling time-domain look-ahead simulation includes: generating a sequence of tool-workpiece contact states in the future rolling time domain based on the current tool position, tool axis posture, material removal state, current local remaining wall thickness, local equivalent compliance, and local modal frequency; and predicting the cutting force, temperature field, deformation response, and vibration response in the future time domain based on the contact state sequence. Risk characterization results include at least one of the following: cutting force prediction curve, temperature field prediction results, workpiece surface deformation cloud map, local remaining wall thickness deviation map, vibration spectrum prediction results, chatter probability map, and comprehensive machining risk index. The process parameter adjustment decision instruction generation process includes: performing look-ahead simulations for multiple candidate process parameter adjustment schemes, calculating evaluation functions based on residual risk values, processing efficiency losses, parameter change smoothness, and machine tool execution constraints, and selecting the candidate scheme with the smallest evaluation function that satisfies the machine tool execution constraints as the final decision instruction.

[0037] In one embodiment, the use of a hybrid intelligent model to perform rolling time-domain look-ahead simulation based on the current synchronization state includes: S31. Construct the current synchronization state vector. The digital twin receives multi-source state data from the physical processing entity during each synchronization cycle and constructs the current synchronization state vector X based on this multi-source state data. t The expression is: , Where, q t Given the current tool position and tool axis posture, n t Main spindle speed, v ft For the feed rate, apt and a et These are the axial depth of cut and the radial width of cut, respectively, F t For triaxial cutting force, T t For the cutting zone temperature, a t For vibration response, h t C represents the current local remaining wall thickness. t For local equivalent compliance, Ω t Z is the set of local modal frequencies. t This indicates the current material removal status. Specifically, the multi-source status data includes the current tool position coordinates, tool axis posture, spindle speed, feed rate, axial depth of cut, radial width of cut, three-dimensional cutting force, spindle power, cutting zone temperature, vibration signal, online thickness measurement data, workpiece area with removed material, current local remaining wall thickness, and clamping boundary status; S32. Generate a sequence of candidate processing states for the future rolling time domain. Using the current synchronization state vector X... t Starting from the current time, the tool path, material removal status, and candidate process parameters are used to progressively deduce the tool-workpiece contact status, undeformed chip thickness, local material removal rate, local remaining wall thickness, and local stiffness changes at each discrete time. Specifically, the length H of the look-ahead time window can be adaptively set according to the type of processing area. When the processing area is a thick-walled or highly rigid area, H takes a shorter time window; when the processing area is a thin-walled, weakly rigid area or an area nearing the final wall thickness, H takes a longer time window. S33. Perform force-thermal-vibration coupled look-ahead prediction. Using a hybrid intelligent model, the predicted cutting force is obtained based on the tool-workpiece contact state at each discrete time point in the future. The cutting heat input and temperature field diffusion are calculated based on the predicted cutting force, contact time, tool posture, and local material removal rate. The predicted cutting force and temperature field diffusion are input as loads to the structural dynamics module to calculate the deformation response and vibration response of the workpiece in the future time domain. For the machining area of ​​thin-walled components, the hybrid intelligent model calls the correction amount output by the gated residual compensation network to perform bounded correction on the mechanical, thermal, and dynamic parameters of the physical kernel, obtaining the compensated prediction results (including cutting force, temperature, deformation, and vibration prediction results). Specifically, the correction amount output by the gated residual compensation network includes the correction amount of tangential cutting force coefficient, radial cutting force coefficient, axial cutting force coefficient, cutting heat distribution coefficient, local compliance, modal frequency, and modal damping.

[0038] In one embodiment, step S3 includes: S34. Generate prediction results, which include multiple risk characterization results in the future time domain.

[0039] Furthermore, the risk characterization results include: Cutting force prediction curves and force mutation risk values ​​are used to determine whether the load on the tool exceeds the preset safety threshold; Temperature field prediction results and thermal hysteresis risk values ​​are used to determine whether local thermal deformation will lead to dimensional deviations. The workpiece surface deformation cloud map and local residual wall thickness deviation map are used to determine whether there is processing deformation or overweight risk in the processing area of ​​thin-walled components; Vibration spectrum prediction results and chatter probability diagrams are used to determine whether the tool position will enter an unstable cutting zone in the future; The comprehensive processing risk index is used to uniformly characterize the quality risk and stability risk at future points in time. The expression for the comprehensive processing risk index is: , Where R(t+τ) is the comprehensive processing risk index at the future time t+τ, R u (t+τ) represents the deformation deviation risk term at the future time t+τ, R f (t+τ) represents the risk term of excessive cutting force at the future time t+τ, R T (t+τ) represents the risk term for temperature exceeding the limit or thermal hysteresis at the future time t+τ, R v (t+τ) represents the flutter risk term at the future time t+τ, R h (t+τ) represents the local residual wall thickness deviation risk term at the future time t+τ, ω u ω f ω T ω v and ω h These are the weighting coefficients for the following risk items: deformation exceeding tolerance, cutting force exceeding limit, temperature exceeding limit or thermal hysteresis, chatter, and local residual wall thickness deviation.

[0040] Specifically, the weighting coefficient ω u ω f ω T ω v and ω h It can adaptively adjust according to the type of processing area. When the processing area is a thin-walled final forming area, it can improve R. u (t+τ) and R h The weight of (t+τ); when the energy of the frequency band increases due to flutter in the processing area, the weight of R is increased. v The weight of (t+τ); when the temperature rise rate or temperature gradient exceeds the threshold, increase R.T The weight of (t+τ).

[0041] In one embodiment, generating a corresponding process parameter adjustment decision instruction when the risk characterization result indicates that the risk will exceed a preset threshold includes: S35. Determine the risk type based on the prediction results. When the risk characterization results indicate that the risk will exceed a preset threshold, determine the risk type. S36. Generate process parameter adjustment decision instructions. Based on the risk type, generate multiple candidate process parameter adjustment schemes; for each candidate scheme, the digital twin re-executes the rolling time-domain look-ahead simulation and calculates the evaluation function J corresponding to each scheme. m The expression is: , Among them, R m E represents the remaining risk value after the candidate process parameter adjustment plan is implemented. m S represents the processing efficiency loss term. m Q is a constraint term for the smoothness of process parameter changes. m The constraint penalty term is applied to the machine tool, where α, β, γ, and η are R... m E m S m and Q m The corresponding weighting coefficients; Select J m The candidate process parameter adjustment scheme that is the smallest and satisfies the machine tool execution constraints is taken as the final process parameter adjustment decision instruction.

[0042] Specifically, when the comprehensive machining risk index R(t+τ) at any discrete moment in the future time domain exceeds a preset risk threshold, or when any single risk exceeds the corresponding safety threshold, the system determines that the current process parameters have potential instability or out-of-tolerance risks. If the chatter probability exceeds a preset threshold or the energy of the chatter sideband continues to rise, it is determined to be a chatter risk; if the predicted deformation exceeds the form and position tolerance threshold, it is determined to be a deformation out-of-tolerance risk; if the predicted value of the local remaining wall thickness deviates from the target wall thickness and exceeds a preset threshold, it is determined to be a local remaining wall thickness deviation risk; if the temperature field prediction result shows that the local temperature gradient continues to rise, it is determined to be a thermal deformation risk; if the peak value of the cutting force or the force mutation rate exceeds the threshold, it is determined to be a tool load risk.

[0043] Specifically, the candidate process parameter adjustment schemes include at least one of the following: spindle speed adjustment, feed rate adjustment, axial depth of cut adjustment, radial width of cut adjustment, material removal rate adjustment, cooling interval adjustment, tool path local smoothing correction, thin-walled area layering allowance redistribution, online measurement path generation, pause machining for retesting, cutting load adjustment, or cutting in / out method adjustment.

[0044] In one embodiment, generating multiple candidate process parameter adjustment schemes based on risk type includes: If a chatter risk is identified, a spindle speed adjustment scheme is generated based on the local modal frequency, the cutter tooth passing frequency, and the stability domain prediction results, so that the cutter tooth passing frequency avoids the current local modal sensitive frequency band. If the risk of deformation exceeding tolerance is determined, an adjustment plan will be generated first, such as reducing the feed rate, reducing the radial cutting width, or adjusting the machining allowance in the machining area of ​​the thin-walled component (redistribution of the layered allowance in the thin-walled area). If the risk of local residual wall thickness deviation is determined, an online measurement path is generated, and the adjustment scheme of local wall thickness state in the digital twin is updated after the measurement path is completed. If the risk of thermal deformation is identified, an adjustment plan is generated to reduce the material removal rate, extend the cooling interval, or adjust the toolpath sequence (local smoothing correction of the toolpath). If the tool is deemed to be at risk of being overloaded, an adjustment plan is generated to reduce the instantaneous cutting load or change the entry and exit methods.

[0045] With the above settings, step S3 of this application is not only based on the current measurement error for post-event compensation, but also performs rolling prediction of future tool position, future material removal state, future local stiffness change and future force, heat and vibration coupling response in the digital twin, and screens process parameter adjustment schemes based on the predicted risks, thereby generating executable active control commands before the aerospace thin-walled components experience deformation deviation, wall thickness deviation or flutter instability.

[0046] In one embodiment, the actuator that issues process parameter adjustment decision instructions to the physical processing entity includes: S41. The final process parameter adjustment decision command is sent to the CNC system, spindle driver, feed axis servo driver, online measurement device, or cooling execution unit via a standardized communication interface. The decision command includes at least one of the following: spindle speed adjustment, feed rate adjustment, radial width adjustment, axial depth of cut adjustment, toolpath local correction, online measurement path adjustment, or cooling interval adjustment. Upon receiving the decision command, the CNC system performs parameter adjustments according to preset acceleration, jerk, and trajectory continuity constraints to avoid new vibrations or impacts caused by sudden changes in control values.

[0047] In one embodiment, the step of using the processed data after the execution of the decision instruction as feedback to perform online incremental learning on the gated residual compensation network to achieve dynamic adaptive optimization includes: S42. Construct execution effect feedback samples. Based on the predicted state before the execution of the decision instruction, the content of the decision instruction, the measured state after the execution of the decision instruction, and the quality inspection results, construct execution effect feedback samples; S43. Calculate the execution effect error. Based on the execution effect feedback samples, calculate the force residual between the predicted cutting force and the measured cutting force, the thermal residual between the predicted temperature and the measured temperature, the vibration residual between the predicted vibration and the measured vibration, the morphology residual between the predicted deformation and the measured deformation, and the wall thickness residual between the predicted local residual wall thickness and the online thickness measurement result. If any residual continuously exceeds the corresponding threshold, online incremental learning of the gated residual compensation network is triggered; if all the above residuals are below the corresponding threshold, the gated residual compensation network is not triggered to update, and the execution effect feedback sample is only stored in the stable sample cache. S44. Online incremental learning uses a multi-objective loss function to update the bounded correction module of the physical parameters of the gated residual compensation network, the modal gated residual decoding module, and the residual amplitude output module. After each update, a limit check is performed on each correction value. S45. After the update is completed, verify the validity of the update; S46. The gated residual compensation network parameters that have passed the validity verification are used for the next rolling time-domain look-ahead simulation and process parameter adjustment decision, forming a dynamic adaptive optimization closed loop.

[0048] Specifically, step S42 also includes: within a preset feedback time window after the decision command is executed, collecting data on three-dimensional cutting force, spindle power, cutting zone temperature, vibration signal, online thickness measurement data, tool position coordinates, actual feed rate, actual spindle speed, and dimensional deviation data obtained from online or offline detection. Execution effect feedback sample D t The expression is: , Among them, X t Let u be the current synchronization state vector (the synchronization state vector before executing the decision instruction). t For the current decision instruction, Y t +Δtpre represents the future state predicted by the hybrid intelligent model before the execution of the decision instruction, and Yt+Δtmea represents the measured state collected after the execution of the decision instruction.

[0049] Furthermore, the expression for the multi-objective loss function of the online incremental learning is as follows: , Among them, L online L is the total loss function for online incremental learning. exe L phy L modal L thick L ctrl L bound and L memThe terms are, in order: execution effect error, physical law residual, thin-walled modal consistency, wall thickness consistency, control command smoothing, parameter bounded, and anti-forgetting. λ1, λ2, λ3, λ4, λ5, λ6, and λ7 are the non-negative weight coefficients corresponding to the execution effect error, physical law residual, thin-walled modal consistency, wall thickness consistency, control command smoothing, parameter bounded, and anti-forgetting terms, respectively.

[0050] The execution effect error term, used to constrain the predicted result after the execution of the decision instruction to approximate the measured result, is expressed as follows: , in, For the Euclidean norm, F pre and F mea These are the predicted cutting force and the measured cutting force, T. pre and T mea These are the predicted temperature and the measured temperature, respectively. pre and a mea These are the predicted vibration response and the measured vibration response, respectively. pre and u mea These represent the predicted deformation and the measured deformation, respectively, with ωF, ωT, ωa, and ωu being the first, second, third, and fourth weighting coefficients, respectively.

[0051] The physical law residual term is used to ensure that the output of the updated gated residual compensation network still satisfies the force-thermal-vibration coupling physical relationship. Its expression is: , in, C and K are the local mass matrix, damping matrix, and stiffness matrix of the current thin-walled component, respectively; To predict the deformation response; To predict deformation response The first derivative with respect to time; To predict deformation response Second derivative with respect to time; Fc is the cutting force load; ρ is the material density, c is the specific heat capacity, k is the thermal conductivity, q c For cutting heat source term; T is the predicted temperature field. The first partial derivative of the temperature field T with respect to time t is used to predict the rate of change of the temperature in the current processing area with time. To predict the gradient of the temperature field T in spatial coordinates, we represent the rate of change of the temperature in the current processing area along the spatial direction; This is the heat conduction diffusion term, used to characterize the diffusion process of the temperature field in space under the influence of the thermal conductivity k.

[0052] The thin-walled modal consistency term is used to constrain the residual deformation compensation amount to expand along the local modal basis of the current thin-walled region, preventing the gated residual compensation network from generating arbitrary residuals that do not conform to the true deformation shape of the thin-walled component. Its expression is: , Where Δu is the residual deformation compensation amount, and r is the selected local modal order; ψ i (x) is the i-th order local modal basis, ψ i (x) is used to characterize the spatial deformation morphology of the current thin-walled region in the i-th order mode; ψ j (x) is the j-th order local modal basis; g i (t) represents the gating coefficient of the i-th mode, g i (t) is used to characterize the degree of participation of the i-th mode in the residual deformation compensation under the current processing state; Let i be the amplitude of the i-th mode. Used to characterize the deformation amplitude corresponding to the i-th order mode at future time; This represents the contribution of the i-th local mode to the residual deformation compensation. This represents the modal reconstruction deformation caused by the superposition of r-order local modes; Denotes the i-th order local modal basis ψ i (x) and the j-th local modal basis ψ j (x) is the inner product of the local mass matrix M of the current thin-walled member, weighted by weights. Represents ψ i The transpose of (x); when i ≠ j, This is used to characterize the orthogonality deviation between different orders of local modes. Ideally, different orders of local modes satisfy a mass orthogonality relation, i.e. =0; therefore, This is used to constrain different local modal bases to remain independent of each other, and to avoid the compensation network generating unreasonable residual deformations caused by the superposition of non-orthogonal modes.

[0053] The wall thickness consistency term is used to ensure that the predicted remaining wall thickness after constraint compensation is consistent with the online thickness measurement result. Its expression is: , Among them, h pre To predict the local residual wall thickness, h mea Δh represents the actual local residual wall thickness obtained from online thickness measurement. pre To predict the wall thickness gradient, Δh mea This represents the measured wall thickness gradient. The control instruction smoothing term is used to prevent abrupt changes in the control variables generated after online learning. Its expression is: , Among them, ut For the current decision instruction, u t-1 For the decision instruction of the previous cycle, Δu t Let Δu be the change in the current decision instruction. t-1 This represents the change in decision instructions from the previous cycle.

[0054] The bounded parameter is used to constrain the cutting force coefficient corrections (tangential cutting force coefficient correction, radial cutting force coefficient correction, and axial cutting force coefficient correction), cutting heat distribution coefficient correction, local compliance correction, modal frequency correction, and modal damping correction of the gated residual compensation network output to be within a preset physical range. Its expression is: , Where i' is the index of the physical parameter to be corrected, p is the total number of physical parameters to be corrected, and Δp i’ Let Δp be the correction amount for the i'th physical parameter to be corrected. i’,max The maximum allowable correction magnitude for the physical parameter to be corrected is given by , and max(·) is the function for taking the maximum value.

[0055] The anti-forgetting term is used to constrain the prediction performance of the gated residual compensation network after online updates on historical typical working condition samples to prevent significant degradation. Its expression is: , Among them, G new For the updated compensation network, G old For the compensation network before the update, X k’ This is the k'th representative sample selected from the historical stable sample buffer, where k' is the index of the representative sample and K' is the total number of representative samples.

[0056] In specific implementation, λ1, λ2, λ3, λ4, λ5, λ6, and λ7 can be preset or adaptively adjusted according to the processing area type, risk type, and online feedback residual status. When the current processing area is the thin-walled final forming area, λ3 and λ4 are increased to strengthen the thin-walled modal consistency and wall thickness consistency constraints. When the prediction result is characterized as chatter risk, λ2 and λ3 are increased to strengthen the structural dynamics residual and thin-walled modal consistency constraints. When the change in control command is close to the machine tool execution constraints, λ5 and λ6 are increased to limit the abrupt change in control quantity and the out-of-bounds movement of physical parameter correction. When the number of online feedback samples is small or the working condition changes frequently, λ7 is increased to suppress the degradation of historical working condition prediction performance caused by online updates.

[0057] Preferably, λ1, λ2, λ3, λ4, λ5, λ6 and λ7 are all greater than or equal to 0, and can be normalized so that λ1+λ2+λ3+λ4+λ5+λ6+λ7=1; or a fixed weight combination can be used according to different processing stages to ensure the stability and repeatability of the online incremental learning process.

[0058] In one embodiment, after each update, a limit check is performed on each correction amount, including: After triggering online incremental learning, the system only updates the bounded correction module for physical parameters, the modal gated residual decoding module, and the residual amplitude output module in the gated residual compensation network, without updating the physical kernel parameters and the backbone feature extraction layer. Online updates use small batches of feedback samples, and after each update round, a limit check is performed on the correction amounts for cutting force coefficient, cutting heat distribution coefficient, local compliance, modal frequency, and modal damping. If any correction amount exceeds the preset physical range, or if the updated model prediction causes the control command to exceed the machine tool execution constraints, the update round is canceled, and the system reverts to the previous stable model parameters.

[0059] In one embodiment, after the update is completed, an update validity verification is performed, including: The system uses the updated hybrid intelligent model in the digital twin to replay and verify the same feedback time window, and compares the prediction error, physical law residual, and control command smoothness before and after the update. Only when the execution effect error term decreases, and the physical law residual term, parameter bounded term, and control command smoothness term all meet the preset thresholds, are the updated gated residual compensation network parameters confirmed as valid parameters; otherwise, the update result is not adopted.

[0060] In one embodiment, the gated residual compensation network parameters that have passed validity verification are used for the next rolling time-domain look-ahead simulation and process parameter adjustment decisions, forming a dynamic adaptive optimization closed loop, including: Verified and effective gated residual compensation network parameters are used for subsequent rolling time-domain lookahead simulations and process parameter adjustment decisions. Simultaneously, the feedback samples of this execution effect, risk type, decision instructions, execution effect error, network update amount, and update validity results are stored in an online learning sample library. Subsequently, when machining areas, wall thickness states, local modal frequencies, or tool wear states are similar, the system prioritizes calling the gated residual compensation network parameters corresponding to similar samples as initial values ​​to improve the model's adaptability in continuous and batch machining of aerospace thin-walled components.

[0061] Through the above settings, a complete closed loop of perception—prediction—decision—execution—feedback—safety update is achieved. This not only allows for proactive adjustment of machining parameters based on prediction results, but also enables constrained updates to the gating residual compensation network based on the execution effect of each decision command. While ensuring physical consistency and execution safety, this improves the model's adaptability to changes in current tool wear, local stiffness variations, thermal drift, and clamping status.

[0062] Secondly, embodiments of this application also provide an intelligent machining control system based on a physical information neural network, used to execute the intelligent machining control method based on a physical information neural network as described in any embodiment of the first aspect, referring to... Figure 2 The system includes: The hybrid intelligent modeling module is used to construct a differentiable computational graph from the physical laws describing the coupling process of force, heat and vibration in the cutting process, forming a physical kernel; it constructs a gated residual compensation network based on tool contact, local stiffness and modality, and combines the gated residual compensation network with the physical kernel to form a hybrid intelligent model that deeply integrates physical mechanism and data-driven approach. The digital twin module is used to build a digital twin with a hybrid intelligent model as the core prediction engine, collect multi-source state data of the physical processing entity in real time, and use the collected multi-source state data to synchronously update the state of the digital twin to achieve virtual-real synchronization. The forward-looking decision-making module is used in the digital twin to perform rolling time-domain forward-looking simulation based on the current synchronization state using a hybrid intelligent model. It predicts the processing results caused by the coupling of force, heat and vibration in the future period and generates risk characterization results based on the prediction results. When the risk characterization results indicate that the process parameters will be exceeded, it generates corresponding process parameter adjustment decision instructions. The closed-loop control and model optimization module is used to send process parameter adjustment decision commands to the actuators of the physical processing entity to adjust the processing process in real time; the processing data after the execution of the decision commands is used as feedback to perform online incremental learning on the gated residual compensation network to achieve dynamic adaptive optimization and complete the closed-loop intelligent control of the processing process.

[0063] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A smart processing control method based on a physical information neural network, characterized in that, include: The physical laws describing the coupling process of force, heat and vibration in the cutting process are constructed into a differentiable computational graph, which constitutes the physical kernel. A gated residual compensation network based on tool contact, local stiffness and mode is constructed, and the gated residual compensation network is combined with the physical kernel to form a hybrid intelligent model that deeply integrates physical mechanism and data-driven approach. Establish a digital twin with a hybrid intelligent model as the core prediction engine, collect multi-source state data of the physical processing entity in real time, and use the collected multi-source state data to synchronously update the state of the digital twin to achieve virtual-real synchronization. In the digital twin, a hybrid intelligent model is used to perform rolling time-domain look-ahead simulation based on the current synchronization state to predict the processing results caused by the coupling of force, heat and vibration in the future period, and generate risk characterization results based on the prediction results. When the risk characterization results indicate that the preset threshold will be exceeded, a corresponding process parameter adjustment decision instruction will be generated. The process parameter adjustment decision command is sent to the execution mechanism of the physical processing entity to adjust the processing process in real time; the processing data after the decision command is executed is used as feedback to perform online incremental learning on the gated residual compensation network, realize dynamic adaptive optimization, and complete the closed-loop intelligent control of the processing process.

2. The intelligent processing control method based on physical information neural network according to claim 1, characterized in that, The construction of the gated residual compensation network based on tool position contact, local stiffness, and modal characteristics includes: A tool contact state encoding module is constructed. This module is used to calculate the instantaneous contact angle range, cutting-in and cutting-out states, undeformed chip thickness, local material removal rate and tool tooth phase angle at the current sampling moment based on the CNC program, tool radius, number of tool teeth, spindle speed, feed rate, axial depth of cut, radial width of cut, tool position coordinates and current material removal state, and form a tool-workpiece contact state vector. A thin-walled local stiffness mapping module is constructed. This module is used to calculate the local wall thickness, wall thickness gradient, equivalent compliance, and first to third order local natural frequencies and corresponding damping ratios of the current processing area based on the workpiece CAD model, online thickness measurement data, current local remaining wall thickness, local curvature, distance from clamping boundary, and rib adjacency relationship, forming a local stiffness description vector. A tooth frequency synchronization signal encoding module is constructed. This module is used to synchronously resample the three-dimensional cutting force signal and vibration signal according to the spindle rotation angle or the tooth passing cycle, and extract the tooth frequency fundamental wave, harmonics, chatter sideband, main frequency offset, vibration phase lag, temperature rise rate and temperature lag to form a force, heat and vibration synchronization state vector. A bounded physical parameter correction module is constructed. This module takes the tool-workpiece contact state vector, the local stiffness description vector, and the force, heat, and vibration synchronization state vector as inputs, and outputs one or more of the following corrections: tangential cutting force coefficient correction, radial cutting force coefficient correction, axial cutting force coefficient correction, cutting heat distribution coefficient correction, convective heat transfer coefficient correction, local compliance correction, modal frequency correction, and modal damping correction. Each correction is constrained within a preset physical range by a limiting function. A modal-gated residual decoding module is constructed. This module is used to receive the correction amount output by the physical parameter bounded correction module and express the residual deformation compensation amount as a weighted superposition of multiple local modal bases. The weighting coefficients of each local modal base are jointly determined by local compliance, tooth frequency energy, flutter sideband energy and temperature gradient. Each local modal base is determined by the current local residual wall thickness, clamping boundary and local geometric state. A residual amplitude output module is constructed, which is used to output the residual deformation compensation amount output by the modal gated residual decoding module.

3. The intelligent processing control method based on physical information neural network according to claim 2, characterized in that, The method of using a hybrid intelligent model to perform rolling time-domain look-ahead simulation based on the current synchronization state includes: The digital twin receives multi-source state data from the physical processing entity during each synchronization cycle, and constructs the current synchronization state vector X based on the multi-source state data. t The expression is: ， Where, q t Given the current tool position and tool axis posture, n t Main spindle speed, v ft For the feed rate, a pt and a et These are the axial depth of cut and the radial width of cut, respectively. t For triaxial cutting force, T t For the cutting zone temperature, a t For vibration response, h t C represents the current local remaining wall thickness. t For local equivalent compliance, Ω t Z is the set of local modal frequencies. t This indicates the current material removal status. With the current synchronization state vector X t Starting from the current time, the tool path, material removal status and candidate process parameters are used to progressively deduce the tool-workpiece contact status, undeformed chip thickness, local material removal rate, local remaining wall thickness and local stiffness changes at each discrete time. Using a hybrid intelligent model, the predicted cutting force is obtained based on the contact state between the tool and the workpiece at each discrete time in the future. The cutting heat input and temperature field diffusion are calculated based on the predicted cutting force, contact time, tool posture and local material removal rate. The predicted cutting force and temperature field diffusion are used as load inputs to the structural dynamics module to calculate the deformation response and vibration response of the workpiece in the future time domain. For the processing area of ​​thin-walled components, the hybrid intelligent model calls the correction amount output by the gated residual compensation network to perform bounded correction on the mechanical, thermal and dynamic parameters of the physical kernel, and obtain the compensated prediction results.

4. The intelligent processing control method based on physical information neural network according to claim 3, characterized in that, The risk characterization results include: Cutting force prediction curves and force mutation risk values ​​are used to determine whether the load on the tool exceeds the preset safety threshold; Temperature field prediction results and thermal hysteresis risk values ​​are used to determine whether local thermal deformation will lead to dimensional deviations. The workpiece surface deformation cloud map and local remaining wall thickness deviation map are used to determine whether there is processing deformation or overweight risk in the processing area of ​​thin-walled components; Vibration spectrum prediction results and chatter probability diagrams are used to determine whether the tool position will enter an unstable cutting zone in the future; The comprehensive processing risk index is used to uniformly characterize the quality risk and stability risk at future points in time. The expression for the comprehensive processing risk index is: ， Where R(t+τ) is the comprehensive processing risk index at the future time t+τ, R u (t+τ) represents the deformation deviation risk term at the future time t+τ, R f (t+τ) represents the risk term of excessive cutting force at the future time t+τ, R T (t+τ) represents the risk term for temperature exceeding the limit or thermal hysteresis at the future time t+τ, R v (t+τ) represents the flutter risk term at the future time t+τ, R h (t+τ) represents the local residual wall thickness deviation risk term at the future time t+τ, ω u ω f ω T ω v and ω h These are the weighting coefficients for the following risk items: deformation exceeding tolerance, cutting force exceeding limit, temperature exceeding limit or thermal hysteresis, chatter, and local residual wall thickness deviation.

5. The intelligent processing control method based on physical information neural network according to claim 4, characterized in that, When the risk characterization result indicates that the risk will exceed a preset threshold, a corresponding process parameter adjustment decision instruction is generated, including: When the risk characterization results indicate that the risk will exceed a preset threshold, the risk type is determined. Based on the risk type, multiple candidate process parameter adjustment schemes are generated; For each candidate process parameter adjustment scheme, the digital twin re-executes the rolling time-domain look-ahead simulation and calculates the evaluation function J corresponding to each candidate process parameter adjustment scheme. m The expression is: ， Among them, R m E represents the remaining risk value after the candidate process parameter adjustment plan is implemented. m S represents the processing efficiency loss term. m Q is a constraint term for the smoothness of process parameter changes. m The constraint penalty term is applied to the machine tool, where α, β, γ, and η are R... m E m S m and Q m The corresponding weighting coefficients; Select J m The candidate process parameter adjustment scheme that is the smallest and satisfies the machine tool execution constraints is taken as the final process parameter adjustment decision instruction.

6. The intelligent processing control method based on physical information neural network according to claim 5, characterized in that, The process of generating multiple candidate process parameter adjustment schemes based on risk type includes: If a chatter risk is identified, a spindle speed adjustment scheme is generated based on the local modal frequency, the cutter tooth passing frequency, and the stability domain prediction results, so that the cutter tooth passing frequency avoids the current local modal sensitive frequency band. If the risk of deformation exceeding tolerance is determined, an adjustment plan will be generated first, such as reducing the feed rate, reducing the radial cutting width, or adjusting the machining allowance in the machining area of ​​the thin-walled component. If the risk of local residual wall thickness deviation is determined, an online measurement path is generated, and the adjustment scheme of local wall thickness state in the digital twin is updated after the measurement path is completed. If the risk of thermal deformation is identified, an adjustment plan is generated to reduce the material removal rate, extend the cooling interval, or adjust the toolpath sequence. If the tool is deemed to be at risk of being overloaded, an adjustment plan is generated to reduce the instantaneous cutting load or change the entry and exit methods.

7. The intelligent processing control method based on physical information neural network according to claim 3, characterized in that, The step of using the processed data after the execution of the decision instruction as feedback to perform online incremental learning on the gated residual compensation network to achieve dynamic adaptive optimization includes: Based on the predicted state before the execution of the decision instruction, the content of the decision instruction, the measured state after the execution of the decision instruction, and the quality inspection results, a sample of execution effect feedback is constructed. Based on the execution effect feedback samples, the force residual between the predicted cutting force and the measured cutting force, the thermal residual between the predicted temperature and the measured temperature, the vibration residual between the predicted vibration and the measured vibration, the morphology residual between the predicted deformation and the measured deformation, and the wall thickness residual between the predicted local residual wall thickness and the online thickness measurement result are calculated respectively. If any residual exceeds the corresponding threshold continuously, the online incremental learning of the gated residual compensation network is triggered. Online incremental learning employs a multi-objective loss function to update the bounded correction module of the physical parameters of the gated residual compensation network, the modal gated residual decoding module, and the residual amplitude output module. After each update, a limit check is performed on each correction value. After the update is complete, verify the validity of the update. The gated residual compensation network parameters, verified for effectiveness, are used for the next rolling time-domain look-ahead simulation and process parameter adjustment decisions, forming a dynamic adaptive optimization closed loop.

8. The intelligent processing control method based on physical information neural network according to claim 7, characterized in that, The expression for the multi-objective loss function of the online incremental learning is: ， Among them, L online L is the total loss function for online incremental learning. exe L phy L modal L thick L ctrl L bound and L mem The terms are, in order: execution effect error, physical law residual, thin-walled modal consistency, wall thickness consistency, control command smoothing, parameter bounded, and anti-forgetting. λ1, λ2, λ3, λ4, λ5, λ6, and λ7 are, in order: execution effect error, physical law residual, thin-walled modal consistency, wall thickness consistency, control command smoothing, parameter bounded, and anti-forgetting. The expression for the execution effect error term is: ， in, For the Euclidean norm, F pre and F mea These are the predicted cutting force and the measured cutting force, T, respectively. pre and T mea These are the predicted temperature and the measured temperature, respectively. pre and a mea These are the predicted vibration response and the measured vibration response, respectively. pre and u mea These represent the predicted deformation and the measured deformation, respectively, with ωF, ωT, ωa, and ωu being the first, second, third, and fourth weighting coefficients, respectively. The expression for the residual term of the physical law is: ， in, C and K are the local mass matrix, damping matrix, and stiffness matrix of the current thin-walled component, respectively; To predict the deformation response; To predict deformation response The first derivative with respect to time; To predict deformation response Second derivative with respect to time; Fc is the cutting force load; ρ is the material density, c is the specific heat capacity, k is the thermal conductivity, q c For cutting heat source term; T is the predicted temperature field. To predict the first-order partial derivative of the temperature field T with respect to time t; To predict the gradient of the temperature field T in spatial coordinates; For thermal conduction and diffusion terms; The expression for the thin-walled modal conformance term is: ， Where Δu is the residual deformation compensation amount, r is the selected local modal order, and ψ i (x) is the i-th order local modal basis, ψ j (x) is the j-th order local modal basis, g i (t) represents the gating coefficients of the i-th mode. The amplitude of the i-th mode This represents the contribution of the i-th local mode to the residual deformation compensation. This represents the modal reconstruction deformation caused by the superposition of r-order local modes; Denotes the i-th order local modal basis ψ i (x) and the j-th local modal basis ψ j (x) is the inner product of the local mass matrix M of the current thin-walled member, weighted by weights. Represents ψ i The transpose of (x); The expression for the wall thickness consistency term is: ， Among them, h pre To predict the local residual wall thickness, h mea Δh represents the actual local residual wall thickness obtained from online thickness measurement. pre To predict the wall thickness gradient, Δh mea This represents the measured wall thickness gradient. The expression for the control instruction smoothing term is: ， Among them, u t For the current decision instruction, u t-1 For the decision instruction of the previous cycle, Δu t Let Δu be the change in the current decision instruction. t-1 This represents the change in decision instructions from the previous cycle; The expression for the bounded term with parameters is: ， Where i' is the index of the physical parameter to be corrected, p is the total number of physical parameters to be corrected, and Δp i’ Let Δp be the correction amount for the i'th physical parameter to be corrected. i’,max The maximum allowable correction magnitude for the physical parameter to be corrected is given by max(·), where max(·) is the function for maximizing the value. The expression for the anti-forgetting item is: ， Among them, G new For the updated compensation network, G old For the compensation network before the update, X k’ This is the k'th representative sample selected from the historical stable sample buffer, where k' is the index of the representative sample and K' is the total number of representative samples.

9. The intelligent processing control method based on physical information neural network according to claim 1, characterized in that, The physical kernel includes a differentiable cutting mechanics module, a heat transfer module, and a structural dynamics module. The cutting mechanics module is used to predict the cutting force based on the tool-workpiece contact state, undeformed chip thickness, tool tooth phase angle, spindle speed, feed rate, axial depth of cut, and radial width of cut. The heat transfer module is used to predict the temperature field based on the cutting force, contact time, material removal rate, and cutting heat source distribution state. The structural dynamics module is used to predict the workpiece deformation and vibration response based on the cutting load, temperature field, current local remaining wall thickness of the workpiece, local compliance, and clamping boundary state.

10. An intelligent machining control system based on a physical information neural network, used to execute the intelligent machining control method based on a physical information neural network as described in any one of claims 1-9, characterized in that, The system includes: The hybrid intelligent modeling module is used to construct a differentiable computational graph from the physical laws describing the coupling process of force, heat and vibration in the cutting process, forming a physical kernel; it constructs a gated residual compensation network based on tool contact, local stiffness and modality, and combines the gated residual compensation network with the physical kernel to form a hybrid intelligent model that deeply integrates physical mechanism and data-driven approach. The digital twin module is used to build a digital twin with a hybrid intelligent model as the core prediction engine, collect multi-source state data of the physical processing entity in real time, and use the collected multi-source state data to synchronously update the state of the digital twin to achieve virtual-real synchronization. The forward-looking decision-making module is used in the digital twin to perform rolling time-domain forward-looking simulation based on the current synchronization state using a hybrid intelligent model. It predicts the processing results caused by the coupling of force, heat and vibration in the future period and generates risk characterization results based on the prediction results. When the risk characterization results indicate that the process parameters will be exceeded, it generates corresponding process parameter adjustment decision instructions. The closed-loop control and model optimization module is used to send process parameter adjustment decision commands to the actuators of the physical processing entity to adjust the processing process in real time; the processing data after the execution of the decision commands is used as feedback to perform online incremental learning on the gated residual compensation network to achieve dynamic adaptive optimization and complete the closed-loop intelligent control of the processing process.

Images