Milling process-oriented cross-fidelity time-space characteristic data fusion method
By using a cross-fidelity spatiotemporal characteristic data fusion method, combining low-fidelity simulation and high-fidelity experimental data, the problems of model accuracy and stability during milling were solved, and adaptive control of milling force and improvement of machining process stability were achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING FORESTRY UNIV
- Filing Date
- 2025-12-04
- Publication Date
- 2026-04-17
AI Technical Summary
Existing technologies struggle to balance computational efficiency of low-fidelity simulation data with accuracy of high-fidelity experimental data during milling, and lack dynamic adaptive optimization across fidelity levels, leading to instability and control difficulties in the machining process.
A cross-fidelity spatiotemporal characteristic data fusion method is adopted. A low-fidelity simulation model is constructed through the Simulink platform, and high-fidelity experimental data is combined for pre-training and sample augmentation to achieve deep fusion of multi-physics data and adaptive control of milling force. An end-to-end optimization mechanism is used for model transfer and online adjustment.
It improves the modeling accuracy and control robustness of the milling process, realizes dynamic mapping and fusion of multi-fidelity data, enhances the generalization ability and prediction accuracy of the model, and achieves real-time adaptive adjustment and dynamic safety control of milling force.
Smart Images
Figure CN121879217A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent manufacturing and complex machining process control technology, specifically to a cross-fidelity spatiotemporal characteristic data fusion method for milling processes, which can be used for fusion modeling of multi-source heterogeneous machining data, adaptive control of milling forces, and high-precision dynamic prediction. Background Technology
[0002] With the widespread application of large and complex structural components in aerospace, energy equipment, and precision mold manufacturing, their milling processes exhibit multi-scale dynamic characteristics, multi-source coupling effects, and strong nonlinear responses. In actual machining, the spatiotemporal interaction between the tool and the workpiece is complex and influenced by multiple factors such as cutting parameters, workpiece structural stiffness, tool geometry, and machine tool control characteristics. This leads to significant fluctuations in cutting force and strong vibration responses, which can easily cause damage to the machined surface, accelerated tool wear, and instability in the machine tool system.
[0003] To achieve stability control and mechanical behavior prediction in the milling process, academia and engineering typically employ two methods: simulation modeling and experimental measurement. The former, based on numerical simulation, can quickly generate a large amount of process data, but its accuracy is limited by model assumptions and simplification conditions, resulting in low-fidelity data. The latter obtains high-precision measurement results through multi-sensor experiments, but it is costly, time-consuming, and has a limited sample size, resulting in high-fidelity data.
[0004] Existing research largely focuses on the analysis and modeling of single-fidelity data, lacking effective methods for establishing a unified feature space and spatiotemporal dynamic mapping mechanism across different fidelity levels. This makes it difficult to simultaneously balance computational efficiency and prediction accuracy. Furthermore, traditional multi-source data fusion methods have limitations in handling the spatiotemporal correlation characteristics of complex milling processes, failing to achieve dynamic adaptive optimization across fidelity levels. Therefore, there is an urgent need for a cross-fidelity data fusion method that can integrate low-fidelity simulation data and high-fidelity experimental data, balancing accuracy and generalization ability, to support intelligent control and dynamic decision-making in complex milling processes. Summary of the Invention
[0005] The purpose of this invention is to overcome the problems of insufficient accuracy of low-fidelity simulation data, limited high-fidelity experimental data samples, and difficulty in achieving unified modeling of multi-source features in the existing technology. It proposes a cross-fidelity spatiotemporal characteristic data fusion method for the milling process to achieve deep fusion of multi-fidelity data and adaptive control of milling force.
[0006] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:
[0007] A cross-fidelity spatiotemporal characteristic data fusion method for milling processes, the method comprising the following steps:
[0008] S1, Low-fidelity modeling and data generation:
[0009] Taking into account the dynamic coupling relationship between tool system stiffness, workpiece structural characteristics and machine tool servo control characteristics, a coupled dynamics simulation model is built on the Simulink platform to characterize the multi-physics interaction behavior in the milling process, and a low-fidelity spatiotemporal characteristic simulation dataset of multi-edge tool edge interleaving under different milling paths is generated.
[0010] S2, High-fidelity data acquisition:
[0011] Through dynamic milling experiments with varying depth of cut, high-fidelity test data on the multi-physics field, including cutting force, vibration, and temperature, are obtained regarding the influence of cutting parameters on the structural characteristics of the workpiece and its load-bearing behavior.
[0012] S3, Low-fidelity pre-training:
[0013] With high-fidelity data disabled, low-fidelity simulation data is used for pre-training to establish an initial milling force adaptive control model.
[0014] S4, Sample Augmentation and Fusion Optimization:
[0015] The sample is expanded by dynamic adjustment and enhancement strategy of low-fidelity simulation data, and multi-physics high-fidelity experimental data is introduced. Under the model credibility constraint, the cross-fidelity features are deeply fused, and the initial milling force adaptive control model is jointly optimized and trained end-to-end to obtain the cross-fidelity milling force adaptive control model.
[0016] S5, Model Transfer and Application:
[0017] The trained cross-fidelity milling force adaptive control model is transferred to the online machining site to perform spatiotemporal adaptive adjustment and precise control of the milling force.
[0018] Step S1 further includes:
[0019] Set different milling paths and cutting parameter set ,in, For feed rate, This represents the axial cutting depth. Radial cutting width, Main spindle speed The helix angle of the cutting tool;
[0020] The entire cutting process is divided into multiple operations, each of which consists of several steps. The cutting segment within each step is further discretized into several spatial nodes to introduce spatial distribution feature data. At the same time, on the time axis t, the timing information of key simulation nodes, including tool entry, exit and transition states, is recorded to introduce time feature data.
[0021] By combining spatial distribution feature data and temporal feature data, a low-fidelity spatiotemporal characteristic simulation dataset of multi-bladed tools under different working conditions is formed. , Indicates nodal displacement. Indicates vibration response, This indicates the instantaneous change in cutting force.
[0022] Step S2 further includes:
[0023] A multi-source sensor acquisition system, including displacement sensors, vibration accelerometers, and triaxial force sensors, is deployed on the CNC machining center; among them, the displacement sensors are used to measure the displacement signal of the workpiece. The vibration accelerometer is used to collect the vibration response signal of the spindle end face. A triaxial force sensor is used to measure the cutting force signal during the milling process. ;
[0024] By designing dynamic milling experiments with varying depth of cut and feed rate, high-fidelity multiphysics data on the interaction between the tool and workpiece were collected using a multi-source sensor system under multiple working conditions. This yielded a high-fidelity spatiotemporal dataset of the influence of cutting parameters on the workpiece's structural characteristics and its loading behavior. .
[0025] Furthermore, without introducing high-fidelity data, low-fidelity data is used to pre-train the initial fuzzy logic controller to establish an adaptive control model for the initial milling force.
[0026] Furthermore, in step S4, the dynamic adjustment and enhancement strategy for the low-fidelity simulation data includes randomly perturbing the cutting parameters, dynamically adjusting the tool feed trajectory, and the simulation resolution.
[0027] Furthermore, in step S4, the process of sample augmentation through dynamic adjustment and enhancement strategies of low-fidelity simulation data includes:
[0028] Let the original input parameter vector be... ;in, For feed rate, This represents the axial cutting depth. Radial cutting width, Main spindle speed The helix angle of the cutting tool;
[0029] The dynamic perturbation strategy is defined as follows:
[0030] ;
[0031] in, Perturbation sample The i-th perturbation quantity. The update rule for the dynamic covariance matrix associated with training iteration step t is as follows:
[0032] ;
[0033] in, As a smoothing factor, Let d represent the standard deviation of the perturbation in the j-th input dimension, and d be the number of input dimensions.
[0034] For each perturbation sample Calculate its corresponding low-fidelity output response:
[0035] ;
[0036] in, An initial milling force adaptive control model is generated; a diverse set of samples is produced:
[0037] ;
[0038] A sample selection mechanism based on target similarity is introduced, and the similarity function of the generated samples is calculated as follows:
[0039] ;
[0040] Wherein represents The i-th generated sample is located in the L-th layer feature vector of the surrogate model. This represents the feature vector of the high-fidelity experimental data at the Lth layer. This is the bandwidth parameter.
[0041] When similarity If the generated augmented samples are not found, they are discarded; otherwise, they are retained. This is a preset threshold.
[0042] Furthermore, in step S4, the process of performing end-to-end joint optimization training on the initial milling force adaptive control model includes the following steps:
[0043] In the cross-fidelity feature mapping stage, let the model prediction output be... The confidence interval for low-fidelity data is set as follows: ;in, This indicates that the model is at the input. The prediction uncertainty is as follows. The confidence coefficient is used to determine the model's performance. When the model receives new online monitoring data streams, it automatically triggers parameter fine-tuning based on the real-time calculated confidence interval, enabling the model to perform end-to-end closed-loop correction using high-fidelity data without retraining. The specific correction process includes:
[0044] Model control block parameter set is processed in an end-to-end manner. Joint training and dynamic optimization are performed, and the objective function is defined as follows:
[0045] ;
[0046] The model control block parameter set includes adjustment coefficient, proportional coefficient, and control range. The multi-objective comprehensive loss function includes a prediction error term, a confidence penalty term, and a fidelity consistency term:
[0047] ;
[0048] In the formula, This represents the prediction error term. For high-fidelity, realistic output To predict high-fidelity output for the model; Indicates a credibility penalty item; Indicates the fidelity consistency term. For high-fidelity, realistic output The uncertainty of the model's prediction of the input 𝑥 is denoted by 𝛽, which is an adjustment coefficient used to control the intensity of the penalty. An adaptive mapping operator for high- and low-fidelity feature spaces; weight parameters , , Determined by Bayesian optimization or cross-validation;
[0049] High-fidelity and low-fidelity data are alternately input into the model. Low-fidelity simulation data provides wide-area distribution characteristics to maintain the structural stability of the model, while high-fidelity data is fed back to the model control block in real time to correct the control gradient. The control gradient update rule is defined as follows:
[0050] ;
[0051] in, and These represent the loss terms corresponding to high-fidelity and low-fidelity data, respectively; weighting coefficients. Determined by the confidence function of the online monitoring data:
[0052] ;
[0053] in, and These represent the mean uncertainties for the low-fidelity and high-fidelity data domains, respectively. To adjust the sensitivity coefficient.
[0054] Step S5 further includes:
[0055] Deploy the trained cross-fidelity milling force adaptive control model to edge computing nodes;
[0056] Real-time acquisition of vibration and force signals from the machining site; prediction and spatiotemporal adjustment of milling force through model; online feedback control of cutting force fluctuations.
[0057] When the system response is detected to deviate from the prediction range, the model automatically updates local parameters to perform online adaptive adjustment.
[0058] The present invention provides a cross-fidelity spatiotemporal characteristic data fusion method for milling processes. By introducing a cross-fidelity data fusion mechanism, it balances the scalability of low-fidelity models with the accuracy of high-fidelity experiments, significantly improving the modeling accuracy and control robustness of complex milling processes. Compared with existing technologies, the present invention has the following beneficial effects:
[0059] (1) It realizes the dynamic mapping and fusion of multi-fidelity data in a unified feature space, which improves the generalization ability of the model;
[0060] (2) The reliability and prediction accuracy of the model under different working conditions were improved by the end-to-end joint optimization mechanism;
[0061] (3) It can realize real-time adaptive adjustment of milling force and dynamic safety control of the machining process, and has high engineering application value. Attached Figure Description
[0062] Figure 1 This is a schematic diagram of the overall process of the cross-fidelity spatiotemporal characteristic data fusion method for milling process according to the present invention.
[0063] Figure 2 This is a schematic diagram of the Simulink coupled dynamics model of the tool-part-machine tool control system of the present invention.
[0064] Figure 3 This is a schematic diagram of the spatiotemporal characteristic data output of the tool-workpiece cutting process according to the present invention.
[0065] Figure 4 This is a schematic diagram of the high-fidelity dynamic milling test and data acquisition system based on multi-source sensors of the present invention.
[0066] Figure 5 This is a schematic diagram of the pre-training and end-to-end joint optimization process of the cross-fidelity model of the present invention. Detailed Implementation
[0067] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.
[0068] This invention discloses a cross-fidelity spatiotemporal characteristic data fusion method for milling processes, the method comprising the following steps:
[0069] S1, Low-fidelity modeling and data generation:
[0070] Taking into account the dynamic coupling relationship between tool system stiffness, workpiece structural characteristics and machine tool servo control characteristics, a coupled dynamics simulation model is built on the Simulink platform to characterize the multi-physics interaction behavior in the milling process, and a low-fidelity spatiotemporal characteristic simulation dataset of multi-edge tool edge interleaving under different milling paths is generated.
[0071] S2, High-fidelity data acquisition:
[0072] Through dynamic milling experiments with varying depth of cut, high-fidelity test data on the multi-physics field, including cutting force, vibration, and temperature, are obtained regarding the influence of cutting parameters on the structural characteristics of the workpiece and its load-bearing behavior.
[0073] S3, Low-fidelity pre-training:
[0074] With high-fidelity data disabled, low-fidelity simulation data is used for pre-training to establish an initial milling force adaptive control model.
[0075] S4, Sample Augmentation and Fusion Optimization:
[0076] The sample is expanded by dynamic adjustment and enhancement strategy of low-fidelity simulation data, and multi-physics high-fidelity experimental data is introduced. Under the model credibility constraint, the cross-fidelity features are deeply fused, and the initial milling force adaptive control model is jointly optimized and trained end-to-end to obtain the cross-fidelity milling force adaptive control model.
[0077] S5, Model Transfer and Application:
[0078] The trained cross-fidelity milling force adaptive control model is transferred to the online machining site, and the milling force is spatiotemporally adapted. The invention will be further explained below with reference to specific embodiments.
[0079] Example 1: Overall Process of Cross-Fidelity Spatiotemporal Characteristic Data Fusion
[0080] like Figure 1 As shown, the cross-fidelity spatiotemporal characteristic data fusion method of the present invention includes the following steps:
[0081] 1. Low-fidelity modeling and simulation data generation
[0082] a. Establish a coupled dynamic model of the tool-part-machine tool control system based on the Simulink platform. ,like Figure 2 As shown in the figure, this model comprehensively considers the dynamic coupling relationship between the tool system stiffness, workpiece structural characteristics, and machine tool servo control characteristics.
[0083] b. The established model includes components such as tool geometry, workpiece structural features, milling process, and simulation nodes, used to describe the cutting interaction between the tool and workpiece. After simulation, it can output spatiotemporal characteristic data of the tool-workpiece cutting process, such as... Figure 3 As shown.
[0084] c. In the simulation, different milling paths are set. and cutting parameter set ,in, For feed rate, This represents the axial cutting depth. Radial cutting width, Main spindle speed The helix angle of the cutting tool.
[0085] The entire cutting process is divided into multiple operations, each consisting of several steps. The cutting segment (i.e., the cutting trajectory) within each step is further discretized into several spatial nodes, thus introducing spatial distribution characteristic data. Simultaneously, on the time axis t, the model records the temporal information of key simulation nodes such as tool entry, exit, and transition states, thus introducing temporal characteristic data.
[0086] Therefore, the synchronous modeling and output of spatiotemporal characteristics were achieved during the dynamic simulation process, resulting in a low-fidelity spatiotemporal characteristic simulation dataset of multi-bladed tools under different working conditions. Including nodal displacements Vibration response and instantaneous cutting force changes This low-fidelity dataset provides foundational samples and spatiotemporal correlation features for subsequent cross-fidelity fusion modeling.
[0087] 2. High-fidelity dynamic milling test and data acquisition
[0088] a. To verify the effectiveness of the low-fidelity simulation model and construct cross-fidelity data mapping relationships, a multi-source sensor acquisition system is deployed on the CNC machining center, such as... Figure 4 As shown.
[0089] b. Among them, the displacement sensor is used to measure the displacement signal of the workpiece. The vibration accelerometer is used to collect the vibration response signal of the spindle end face. The triaxial force sensor is used to measure the cutting force signal during the milling process. .
[0090] c. By designing dynamic milling experiments with varying depth of cut and feed rate, the system acquires high-fidelity multiphysics data on tool-workpiece interaction under multiple working conditions, thereby obtaining a high-fidelity spatiotemporal dataset of the influence of cutting parameters on workpiece structural features and its loading behavior. It is used to guide the feature correction, credibility assessment and cross-fidelity fusion training of pre-trained simulation models.
[0091] 3. Low-fidelity data pre-training and sample augmentation
[0092] a. Without introducing high-fidelity data, pre-train the initial fuzzy logic controller using low-fidelity data to establish the basic control structure for adaptive milling force control, such as... Figure 5 As shown.
[0093] b. Employ a dynamic perturbation strategy to sample the input parameters in a distributed manner, and expand the diversity of training samples through data augmentation.
[0094] Let the original input parameter vector be...
[0095]
[0096] in, For feed rate, This represents the axial cutting depth. Radial cutting width, Main spindle speed The helix angle of the cutting tool.
[0097] The dynamic perturbation strategy is defined as follows:
[0098]
[0099] in, Let be the dynamic covariance matrix associated with training iteration step t, used to control the adaptive adjustment of perturbation strength during training. The dynamic update rule for the covariance is:
[0100]
[0101] in, As a smoothing factor, Let d represent the standard deviation of the perturbation of the j-th input dimension, and d be the number of input dimensions.
[0102] For each perturbation sample Calculate its corresponding low-fidelity output response:
[0103]
[0104] in, This is a low-fidelity simulation model.
[0105] Through the above process, while maintaining the physical rationality of the input parameters, a diverse set of samples is generated:
[0106]
[0107] This sample set covers a larger area of the original parameter space in terms of statistical distribution, which can effectively improve the robustness and generalization performance of the model under complex conditions.
[0108] To further suppress the bias caused by excessive perturbation, a sample selection mechanism based on target similarity is introduced, and the similarity function is defined as:
[0109]
[0110] when Time (of which) (If the threshold is set), the sample is discarded, thus ensuring that the augmented data remains consistent with the original operating conditions at the physical level.
[0111] c. Low-fidelity data is set to a dynamically adjusted state in the model to continuously optimize the representation capability of the low-fidelity feature space.
[0112] 4. High-fidelity data fusion and end-to-end joint optimization
[0113] a. Introduce high-fidelity experimental data into model training, and establish an adaptive fusion between high-fidelity and low-fidelity features through a confidence interval constraint mechanism, such as... Figure 5 As shown.
[0114] Let the low-fidelity sample set be The high-fidelity sample set is ;in, For the input parameter vector, and These represent low-fidelity and high-fidelity output responses (such as cutting force, vibration, or temperature characteristics), respectively.
[0115] (1) Data fusion with confidence interval constraints
[0116] In the cross-fidelity feature mapping stage, this invention introduces data fusion based on confidence interval constraints to balance the influence weights between high-fidelity and low-fidelity data. Let the model prediction output be... Its confidence interval is defined as:
[0117]
[0118] in, This indicates that the model is at the input. The prediction uncertainty is as follows. is the confidence coefficient.
[0119] The credibility constraint is characterized in the following form:
[0120]
[0121] That is, when the prediction error exceeds the confidence interval, the system applies a penalty, thereby constraining the stability and reliability of the model in the high-fidelity domain.
[0122] b. Design of multi-objective loss function
[0123] To achieve dynamic alignment and deep fusion of multi-fidelity feature spaces, a multi-objective loss function consisting of three parts was constructed:
[0124]
[0125] Wherein, the prediction error term used to minimize the difference between the model prediction and the actual output is: Credibility penalty item Defined as above, it is used to maintain the consistency of model confidence intervals under high-fidelity data constraints; the fidelity consistency term is... .in This is an adaptive mapping operator for high-fidelity and low-fidelity feature spaces, used to ensure the compatibility of two types of data in the latent feature space. Weight parameters. , , The ratio of contributions among prediction accuracy, reliability stability, and fidelity consistency is determined by Bayesian optimization or cross-validation.
[0126] c. End-to-end joint optimization and dynamic feature transfer.
[0127] Model control block parameter set is processed in an end-to-end manner. Joint training and dynamic optimization are performed to achieve feature fusion and dynamic transfer between the simulation domain and the online monitoring domain. The optimization objective function is defined as follows:
[0128]
[0129] in, It is a multi-objective comprehensive loss function that includes a prediction error term, a confidence penalty term, and a fidelity consistency term, and is used to simultaneously constrain the prediction accuracy, stability, and cross-fidelity feature consistency of the model.
[0130] During training and online operation, the system uses alternating inputs of high-fidelity and low-fidelity data. Low-fidelity simulation data provides wide-area distribution characteristics to maintain the structural stability of the model, while high-fidelity data obtained from online monitoring is fed back to the model control block in real time to correct the control gradient, enabling the control logic to be updated online from simulation optimization to actual processing.
[0131] The gradient update rule is defined as follows:
[0132]
[0133] in, and These represent the loss terms for high-fidelity (online monitoring) and low-fidelity (simulation) data, respectively. Weighting coefficients. The confidence function of the line monitoring data determines the proportion of high-fidelity and low-fidelity data in gradient updates.
[0134]
[0135] in, and These represent the mean uncertainties for the low-fidelity and high-fidelity data domains, respectively. To adjust the sensitivity coefficient, balanced learning and dynamic feature transfer between the two domains can be achieved.
[0136] When the model receives a new online monitoring data stream, the system automatically triggers parameter fine-tuning based on the real-time calculated confidence interval, enabling the pre-trained simulation model to achieve end-to-end closed-loop correction using high-fidelity data without retraining.
[0137] 5. Cross-fidelity model transfer and online adaptive control
[0138] a. Deploy the trained cross-fidelity milling force adaptive control model to edge computing nodes.
[0139] b. Real-time acquisition of vibration and force signals from the machining site, and prediction and spatiotemporal adjustment of milling force through modeling, to achieve online feedback control of cutting force fluctuations.
[0140] c. When the system response is detected to deviate from the prediction range, the model automatically updates local parameters to achieve online adaptive adjustment.
[0141] Example 2: Credibility Constraints and Dynamic Fusion Mechanism
[0142] To improve the robustness of the model in fusing multiple fidelity data, this invention introduces a credibility constraint mechanism.
[0143] a. During the joint optimization process, set confidence intervals for low-fidelity data;
[0144] b. When the number of high-fidelity samples is small, their contribution to the loss function can be controlled by confidence weight;
[0145] c. Bayesian optimization is used to dynamically adjust the fusion weights of high-fidelity and low-fidelity features, so as to achieve a smooth transition and steady-state convergence of the model between different fidelities.
[0146] Example 3: Application Verification in Online Processing Scenarios
[0147] In precision milling tests of aerospace structural components, the cross-fidelity model described in this invention is deployed on an edge computing terminal:
[0148] a. Communicates with the machine tool control system in real time via industrial bus to collect spindle power signals;
[0149] b. The model updates at a frequency of 10 Hz to predict milling force and correct control parameters, thereby achieving synchronous optimization of feed rate and spindle speed;
[0150] c. Experimental results show that, after adopting the method of the present invention, the milling force fluctuation amplitude is reduced by about 25%, the workpiece surface roughness is improved by about 18%, and the system response stability is significantly improved for adjustment and precise control.
[0151] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.
[0152] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.
Claims
1. A cross-fidelity spatiotemporal property data fusion method for a milling process, characterized in that, The method includes the following steps: S1, Low-fidelity modeling and data generation: Taking into account the dynamic coupling relationship between tool system stiffness, workpiece structural characteristics and machine tool servo control characteristics, a coupled dynamics simulation model is built on the Simulink platform to characterize the multi-physics interaction behavior in the milling process, and a low-fidelity spatiotemporal characteristic simulation dataset of multi-edge tool edge interleaving under different milling paths is generated. S2, High-fidelity data acquisition: Through dynamic milling experiments with varying depth of cut, high-fidelity test data on the multi-physics field, including cutting force, vibration, and temperature, are obtained regarding the influence of cutting parameters on the structural characteristics of the workpiece and its load-bearing behavior. S3, Low-fidelity pre-training: With high-fidelity data disabled, low-fidelity simulation data is used for pre-training to establish an initial milling force adaptive control model. S4, Sample Augmentation and Fusion Optimization: The sample is expanded by dynamic adjustment and enhancement strategy of low-fidelity simulation data, and multi-physics high-fidelity experimental data is introduced. Under the model credibility constraint, the cross-fidelity features are deeply fused, and the initial milling force adaptive control model is jointly optimized and trained end-to-end to obtain the cross-fidelity milling force adaptive control model. S5, Model Transfer and Application: The trained cross-fidelity milling force adaptive control model is transferred to the online machining site to perform spatiotemporal adaptive adjustment and precise control of the milling force.
2. The milling process oriented cross-fidelity spatio-temporal characteristics data fusion method of claim 1, wherein, Step S1 further includes: Setting different milling paths and sets of cutting parameters wherein, is the feed speed, is the axial depth of cut, is the radial depth of cut, is the spindle speed, is the tool helix angle; The entire cutting process is divided into multiple operations, each of which consists of several steps. The cutting segment within each step is further discretized into several spatial nodes to introduce spatial distribution feature data. At the same time, on the time axis t, the timing information of key simulation nodes, including tool entry, exit and transition states, is recorded to introduce time feature data. By combining spatial distribution feature data and temporal feature data, a low-fidelity spatiotemporal characteristic simulation dataset of multi-bladed tools under different working conditions is formed. , Indicates nodal displacement. Indicates vibration response, This indicates the instantaneous change in cutting force.
3. The cross-fidelity spatiotemporal characteristic data fusion method for milling processes according to claim 1, characterized in that, Step S2 further includes: A multi-source sensor acquisition system, including displacement sensors, vibration accelerometers, and triaxial force sensors, is deployed on the CNC machining center; among them, the displacement sensors are used to measure the displacement signal of the workpiece. The vibration accelerometer is used to collect the vibration response signal of the spindle end face. A triaxial force sensor is used to measure the cutting force signal during the milling process. ; By designing dynamic milling experiments with varying depth of cut and feed rate, high-fidelity multiphysics data on the interaction between the tool and workpiece were collected using a multi-source sensor system under multiple working conditions. This yielded a high-fidelity spatiotemporal dataset of the influence of cutting parameters on the workpiece's structural characteristics and its loading behavior. .
4. The cross-fidelity spatiotemporal characteristic data fusion method according to claim 1, characterized in that, In step S3, without introducing high-fidelity data, low-fidelity data is used to pre-train the initial fuzzy logic controller to establish an initial milling force adaptive control model.
5. The cross-fidelity spatiotemporal characteristic data fusion method for milling processes according to claim 1, characterized in that, In step S4, the dynamic adjustment and enhancement strategy for the low-fidelity simulation data includes random disturbance of cutting parameters, dynamic adjustment of tool feed trajectory, and simulation resolution.
6. The cross-fidelity spatiotemporal characteristic data fusion method for milling processes according to claim 1, characterized in that, Step S4, the process of sample augmentation using dynamic adjustment and enhancement strategies for low-fidelity simulation data, includes: Let the original input parameter vector be... ;in, For feed rate, This represents the axial cutting depth. Radial cutting width, Main spindle speed The helix angle of the cutting tool; The dynamic perturbation strategy is defined as follows: ； in, Perturbation sample The i-th perturbation quantity. The update rule for the dynamic covariance matrix associated with training iteration step t is as follows: ； in, As a smoothing factor, Let d represent the standard deviation of the perturbation in the j-th input dimension, and d be the number of input dimensions. For each perturbation sample Calculate its corresponding low-fidelity output response: ； in, An initial milling force adaptive control model is generated; a diverse set of samples is produced: ； A sample selection mechanism based on target similarity is introduced, and the similarity function of the generated samples is calculated as follows: ； Wherein represents The i-th generated sample is located in the L-th layer feature vector of the surrogate model. This represents the feature vector of the high-fidelity experimental data at the Lth layer. This is the bandwidth parameter.
7. When similarity If the generated augmented samples are not found, they are discarded; otherwise, they are retained. This is a preset threshold.
8. The cross-fidelity spatiotemporal characteristic data fusion method for milling processes according to claim 1, characterized in that, Step S4, the process of performing end-to-end joint optimization training on the initial milling force adaptive control model, includes the following steps: In the cross-fidelity feature mapping stage, let the model prediction output be... The confidence interval for low-fidelity data is set as follows: ;in, This indicates that the model is at the input The prediction uncertainty is as follows. The confidence coefficient is used to determine the model's performance. When the model receives new online monitoring data streams, it automatically triggers parameter fine-tuning based on the real-time calculated confidence interval, enabling the model to perform end-to-end closed-loop correction using high-fidelity data without retraining. The specific correction process includes: Model control block parameter set is processed in an end-to-end manner. Joint training and dynamic optimization are performed, and the objective function is defined as follows: ； The model control block parameter set includes adjustment coefficients, proportional coefficients, and control intervals. The multi-objective comprehensive loss function includes a prediction error term, a confidence penalty term, and a fidelity consistency term: ； In the formula, This represents the prediction error term. For high-fidelity, realistic output To predict high-fidelity output for the model; Indicates a credibility penalty item; Indicates the fidelity consistency term. For high-fidelity, realistic output The uncertainty of the model's prediction of the input 𝑥 is denoted by 𝛽, which is an adjustment coefficient used to control the intensity of the penalty. An adaptive mapping operator for high- and low-fidelity feature spaces; weight parameters , , Determined by Bayesian optimization or cross-validation; High-fidelity and low-fidelity data are alternately input into the model. Low-fidelity simulation data provides wide-area distribution characteristics to maintain the structural stability of the model, while high-fidelity data is fed back to the model control block in real time to correct the control gradient. The control gradient update rule is defined as follows: ； in, and These represent the loss terms corresponding to high-fidelity and low-fidelity data, respectively; weighting coefficients. Determined by the confidence function of the online monitoring data: ； in, and These represent the mean uncertainties for the low-fidelity and high-fidelity data domains, respectively. To adjust the sensitivity coefficient.
9. The cross-fidelity spatiotemporal characteristic data fusion method for milling processes according to claim 1, characterized in that, Step S5 further includes: Deploy the trained cross-fidelity milling force adaptive control model to edge computing nodes; Real-time acquisition of vibration and force signals from the machining site; prediction and spatiotemporal adjustment of milling force through model; online feedback control of cutting force fluctuations. When the system response is detected to deviate from the prediction range, the model automatically updates local parameters to perform online adaptive adjustment.