A control method and equipment for a rotary flipping fixture in a CNC milling machining center

Abstract

This invention discloses a control method and equipment for a rotary flipping fixture suitable for CNC milling machining centers, relating to the field of CNC machining technology. This invention optimizes clamping force through multi-source sensing and digital twin technology to prevent workpiece deformation; integrates thermal compensation and error propagation chain models to accurately correct the flipping angle and suppress error accumulation; and utilizes digital twin pre-verification and multi-sensor closed-loop monitoring to achieve real-time comparison, traceability, and adaptive adjustment of the machining process. It solves the problems of precision deviation and batch consistency issues caused by unstable clamping, thermal deformation, and error propagation in multi-face milling of high-precision parts, significantly improving the precision, reliability, and process adaptability of complex part machining.

Description

Technical Field

[0001] This invention relates to the field of CNC machining technology, specifically a control method and equipment for a rotary flipping fixture suitable for CNC milling machining centers. Background Technology

[0002] High-precision machining of commercial vehicle components is a crucial link in the automotive manufacturing industry, directly affecting the overall safety and operational reliability of vehicles. This is especially true for multi-faceted milling of core components such as engines and transmission systems, which requires ensuring extremely high standards of dimensional and positional accuracy on each machined surface. Otherwise, subsequent assembly and overall vehicle performance will be affected. Current machining methods largely rely on traditional fixtures and fixed programs. When dealing with workpieces of different materials and complex shapes, the clamping force is often difficult to adjust flexibly, leading to localized stress concentrations in easily deformable materials during machining. Furthermore, positioning during rotation and flipping relies on mechanical limits and manual calibration, and angular deviations can easily accumulate over multiple flips, forming cumulative errors. These defects make it difficult to maintain machining consistency, especially evident in mass production.

[0003] In multi-face machining, insufficient clamping force adaptability directly affects the stability of the workpiece's posture. Different materials, such as aluminum alloys and magnesium alloys, exhibit significantly different sensitivities to pressure. Excessive clamping force may cause workpiece deformation, while insufficient force cannot withstand cutting forces leading to displacement. Instability in the workpiece's posture further amplifies the difficulty of rotational positioning. High-precision rotation requires precise control of the flipping angle, but due to thermal deformation caused by temperature changes and minor fluctuations in servo drive, the actual angle easily deviates from the expected value. This is especially true when continuous multi-face flipping is required; a small deviation in the previous positioning will be passed on to the next, resulting in error accumulation. For example, when machining a transmission housing that requires sequential milling of six faces, if the first flipping angle has a deviation of 0.005°, subsequent flips will gradually increase the positional deviation of the opposite faces, eventually leading to out-of-tolerance hole positions and preventing precise mating with adjacent parts. Summary of the Invention

[0004] The purpose of this invention is to provide a control method and equipment for a rotary flipping fixture suitable for CNC milling machining centers, which solves the problems of workpiece deformation, positioning error accumulation and batch consistency in the multi-face machining of high-precision parts.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] This application provides a control method for a rotary flipping fixture suitable for CNC milling machining centers, including the following steps:

[0007] S1. Collect workpiece material difference data and initial shape parameters in real time through sensors, and match the corresponding elastic modulus and stress threshold using a preset material database to obtain the initial value of targeted clamping force.

[0008] S2. Based on the obtained initial value of clamping force, apply an adaptive control algorithm to dynamically adjust the pressure distribution of the hydraulic fixture, simulate the effect of cutting force on the workpiece before processing, and determine an optimized clamping scheme to avoid deformation.

[0009] S3. Extract workpiece posture stability index from the determined optimized clamping scheme, obtain real-time displacement and vibration data fed back by multi-axis sensors, and determine if the displacement exceeds the preset threshold to activate the feedback loop and obtain stable posture data.

[0010] S4. Based on the obtained stable attitude data, the flipping angle is pre-calibrated using a rotary servo system, and the ambient temperature sensor input is integrated through the thermal deformation compensation module to obtain accurate rotation path planning;

[0011] S5. Extract the angle deviation vector from the precise rotation path plan, apply the error compensation algorithm to analyze the impact of the residual deviation from the previous flip on the current face, and determine the cumulative error correction coefficient.

[0012] S6. Based on the determined cumulative error correction coefficient, obtain the position coordinate adjustment value of each face in the multi-face milling sequence. If the coordinate deviation exceeds the accuracy standard, trigger iterative optimization. After continuous simulation verification, obtain the final machining path.

[0013] S7. Integrate the data of each machining surface from the obtained final machining path, and use a closed-loop monitoring system to compare the actual milling results with the expected model in real time to obtain a complete error accumulation control record.

[0014] A control device for a rotary flipping fixture in a CNC milling machining center, used to implement a control method for a rotary flipping fixture in a CNC milling machining center, comprising:

[0015] The perception and preliminary judgment module scans the workpiece using lidar and hyperspectral imager, constructs a 3D model containing material and geometric features using edge computing, matches material mechanical parameters and generates dynamic stress thresholds using deep neural networks and dynamic knowledge graphs, and quickly calculates the initial scheme of safe clamping force by calling intelligent rule base and machine learning model.

[0016] The twin and optimization module loads the model, simulates cutting loads, and calculates dynamic stress and deformation through a high-fidelity digital twin system to identify machining risk areas; it integrates an adaptive model predictive controller to perform multi-objective iterative optimization of hydraulic clamp pressure distribution in the twin environment, generates an optimized clamping scheme, and uses feedforward-feedback composite control to achieve dynamic fine-tuning during machining;

[0017] The monitoring and attitude stabilization module collects six-degree-of-freedom motion and vibration data of the workpiece in real time through a monitoring network consisting of a multi-axis inertial measurement unit, a laser displacement sensor, and a vibration accelerometer. Based on the built-in stability index system, it performs online comparison and anomaly diagnosis, and calls the real-time controller and reinforcement learning optimizer to dynamically adjust the clamping parameters and output stable attitude data.

[0018] The planning and compensation module filters and reduces noise from the stable attitude data, calculates the thermally induced angle offset in real time by combining environmental and workpiece surface temperature data, plans a smooth S-shaped rotation trajectory and maps it to a servo axis command sequence through inverse kinematics, and then performs collision detection and look-ahead control optimization to output a precise rotation path.

[0019] The modeling and error module constructs an error propagation state space model based on historical angle deviation data, uses fuzzy clustering algorithm to intelligently stratify and extract key terms of cumulative error, and adopts a neural network-driven adaptive compensator combined with multiple linear regression method to dynamically calculate and fit the cumulative error correction coefficient.

[0020] The dynamic compensation module adaptively adjusts the multi-face milling coordinate sequence based on the correction coefficient through a graph network-based dynamic error compensation model. When the coordinate deviation exceeds the tolerance, the reinforcement learning optimization engine is activated to iteratively generate new coordinates. The stability, surface quality and interference risk of the path are evaluated by Monte Carlo simulation and vibration spectrum analysis using a high-fidelity simulation platform. After feedback optimization, the final machining path is output.

[0021] The monitoring and evolution module generates a structured surface-process related dataset through a data fusion engine. It integrates multi-sensor synchronous acquisition and real-time comparison systems to locate processing deviations. It uses an error tracing model based on graph neural networks and a prediction model that fuses support vector machines and time series data for trend early warning. It also records the entire process data into the process knowledge base to achieve adaptive iterative updates of system parameters and strategies.

[0022] The beneficial effects of this invention are as follows:

[0023] By employing multi-source sensing and digital twin simulation, the problem of workpiece deformation and instability caused by insufficient clamping force adaptability is solved. LiDAR and hyperspectral imager are used to collect workpiece geometry and material data in real time, which are then input into a high-fidelity digital twin system for transient finite element analysis. This simulates the dynamic influence of cutting forces, identifies deformation risk areas, and uses an adaptive model predictive control algorithm to iteratively optimize the optimal clamping force distribution scheme. This achieves a shift from fixed clamping to material-specific force application, effectively suppressing stress concentration, plastic deformation, or machining displacement caused by improper clamping, and ensuring the basic posture stability of the workpiece throughout the machining process.

[0024] By solving the problems of rotational positioning deviation and error accumulation in multi-face flipping through real-time thermal deformation compensation and error transmission chain modeling, the system integrates environmental and workpiece surface temperature data to compensate for angular offset caused by thermal deformation in real time, and establishes the error transmission law based on the state space model. The system identification and neural network algorithm are used to quantitatively analyze the impact of previous residual deviations on the current face and calculate the cumulative error correction coefficient. This achieves a leap from single mechanical positioning to global error prediction and compensation, effectively cutting off the error transmission chain in continuous flipping, controlling the positioning accuracy of each machined face within an extremely high standard, and avoiding assembly and fit problems such as hole position deviation caused by error accumulation.

[0025] This approach addresses the challenge of maintaining consistent machining processes in mass production by employing digital twin-based path pre-verification and closed-loop monitoring throughout the entire process. The method involves performing multi-stage continuous cutting simulation and Monte Carlo analysis on a digital twin platform using the error-compensated coordinate sequence to pre-verify path stability. During actual machining, multiple sensors synchronously collect data and compare it in real-time with the expected model to quickly pinpoint deviations. Furthermore, graph neural networks are used for deviation tracing and trend prediction, and structured records are fed back to the process knowledge base. This upgrade from open-loop execution to a fully closed-loop intelligent control system encompassing prediction, execution, monitoring, and optimization enables real-time deviation correction in single machining operations and continuous optimization of process parameters in mass production, thereby ensuring and continuously improving the consistency and reliability of high-precision parts mass production. Attached Figure Description

[0026] To better understand and implement this application, the technical solution is described in detail below with reference to the accompanying drawings.

[0027] Figure 1 A flowchart illustrating a control method for a rotary flipping fixture suitable for a CNC milling machining center, provided in Embodiment 1 of this application;

[0028] Figure 2 This is a flowchart illustrating step S4 of a control method for a rotary flipping fixture for a CNC milling machining center, as provided in Embodiment 1 of this application.

[0029] Figure 3 This is a flowchart illustrating step S5 of a control method for a rotary flipping fixture for a CNC milling machining center, as provided in Embodiment 1 of this application.

[0030] Figure 4 This is a schematic diagram of the structure of a control device for a rotary flipping fixture suitable for a CNC milling machining center, provided in Embodiment 2 of this application. Detailed Implementation

[0031] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, exemplary embodiments will be described in detail below, examples of which are illustrated in the accompanying drawings. In the following description, when referring to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0032] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used herein are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0033] The following detailed description of the specific implementation methods, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided in detail.

[0034] Example 1, please refer to Figures 1-3 This embodiment provides a control method for a rotary flipping fixture suitable for CNC milling machining centers, including the following steps:

[0035] S1. Real-time data on workpiece material differences and initial shape parameters are collected by sensors. The corresponding elastic modulus and stress threshold are matched with a preset material database to obtain the initial value of the targeted clamping force.

[0036] Further, step S1 specifically includes:

[0037] By scanning in conjunction with LiDAR and hyperspectral imager, the three-dimensional point cloud and spectral information of the workpiece surface are collected in real time. The edge computing unit is used for data fusion and noise reduction to construct a three-dimensional feature model of the workpiece containing geometric contours, material composition and surface roughness, and extract key shape parameters and material feature vectors.

[0038] The extracted material features are input into a pre-trained deep neural network model, combined with a dynamically updated material knowledge graph database, to match and output the elastic modulus distribution, yield strength curve and strain hardening parameters that match the workpiece in real time, and generate a dynamic stress threshold field that changes with the geometric shape.

[0039] By inputting geometric features and material parameters into an intelligent rule base built on historical processing data, the system retrieves the best historical cases through a similarity matching algorithm. Combined with preset calculation rules and machine learning models, it quickly calculates the initial value and distribution scheme of the safety clamping force that meets deformation control requirements.

[0040] Furthermore, a three-dimensional feature model of the workpiece, including geometric contours, material composition, and surface roughness, is constructed. Specifically, this includes: acquiring three-dimensional point cloud data of the workpiece surface through LiDAR scanning, and simultaneously acquiring spectral information of the workpiece surface in the 400-1000nm band using a hyperspectral imager; in the edge computing unit, spatial alignment of point cloud coordinates and pixel coordinates is first achieved through a joint standard method based on a calibration board, and multi-view point cloud registration is completed using an improved ICP (Iterative Closest Point) algorithm to form a complete three-dimensional point cloud framework. After dimensionality reduction using principal component analysis, the hyperspectral data is compared with a standard material spectral library using a spectral angle matching algorithm to identify the material composition corresponding to each pixel. In the data fusion stage, a feature-level fusion strategy is adopted: the spectral feature vector of the nearest neighbor pixel is matched for each spatial point in the 3D point cloud, forming a multi-attribute point cloud data structure in which each spatial point contains 3D coordinates (X,Y,Z), reflection intensity (I), and a 128-dimensional spectral feature vector. Subsequently, a statistical outlier removal filter and a bilateral filtering algorithm are used to eliminate point cloud noise and high-frequency noise in the spectral data, respectively. Based on the fused multi-attribute point cloud data, an improved Poisson surface reconstruction algorithm is used to generate a continuous surface model through octree spatial partitioning and implicit function fitting. During the reconstruction process, the spectral feature vector of each vertex is retained as an additional attribute to form a digital 3D feature model containing geometric topology, material composition distribution, and surface roughness parameters calculated from the variance of the local point cloud normal vector. The model is ultimately stored in a multi-layer data structure, where the basic geometry layer stores the triangular mesh topology, and the attribute layer stores the material feature vector and roughness parameters corresponding to each vertex, providing multi-dimensional input with both spatial accuracy and physical properties for subsequent intelligent matching.

[0041] The process involves generating a dynamic stress threshold field that varies with geometric shape. Specifically, this includes: collecting and structurally storing complete case data from successful past machining operations (including workpiece features, clamping parameters, and machining results); formalizing domain expert experience into executable production rules; inputting theoretical constraints from material handbooks, mechanical formulas, and safety specifications; managing all rules and cases through a unified framework; and continuously verifying and optimizing confidence levels and applicability using historical data to form a hybrid knowledge system supporting intelligent reasoning and self-learning. A pre-trained deep neural network model focuses on accurately mapping the relationship between spectra and mechanical parameters, based on large-scale spectral-mechanical properties. The model can be paired with a dataset and uses a dual-branch fusion architecture to process the spatial-spectral features of hyperspectral images and the spectral subgraph of the knowledge graph based on material composition. The mapping relationship is initially established through end-to-end supervised learning, and then self-supervised contrastive learning is introduced to enhance the ability to identify subtle differences in the spectrum. Online incremental learning is achieved by combining the validation data accumulated in actual production, so that the model can continuously evolve and accurately and robustly infer key mechanical properties such as elastic modulus and yield strength from the surface spectrum of the workpiece. Finally, based on the above model and knowledge system, combined with the material feature vector of the workpiece and three-dimensional geometric information, a dynamic stress threshold field that changes with the geometric shape is generated.

[0042] The optimal historical cases are retrieved through a similarity matching algorithm, combined with preset calculation rules and a machine learning model. Specifically, this involves: using a weighted similarity matching algorithm based on Mahalanobis distance to retrieve several optimal historical cases with the highest overall similarity to the current workpiece and their corresponding successful clamping solutions from the historical case library; then, a parallel collaborative decision-making process is initiated: on the one hand, a preset calculation rule library based on material mechanics and contact mechanics is invoked to theoretically scale and correct the case solutions according to the size ratio and material differences between the current workpiece and similar cases; on the other hand, the feature vectors and case solutions are input into a trained ensemble machine learning model to mine complex nonlinear relationships through a data-driven approach to output optimization strategies; simultaneously, the above candidate solutions are rapidly verified through finite element simulation in a lightweight digital twin environment; finally, the performance of the output results of each path is dynamically evaluated through a meta-reinforcement learning module, and the conclusions of rule correction, data prediction, and simulation verification are intelligently integrated to generate a highly robust initial value and distribution scheme of safe clamping force that combines the reliability of historical experience, the rigor of physical laws, and the intelligent adaptability of data.

[0043] The intelligent rule base is constructed by integrating historical processing data, expert experience, and industry standards. First, it collects and structures complete case data from past successful processing, including workpiece characteristics, clamping parameters, and processing results. Simultaneously, it formalizes the experience and knowledge of domain experts into executable production rules and incorporates theoretical constraints from material handbooks, mechanical formulas, and safety specifications. All rules and cases are managed through a unified framework, and their confidence and applicability are continuously validated and optimized using historical data, forming a hybrid knowledge system that supports intelligent reasoning and self-learning.

[0044] The pre-trained deep neural network model is based on a large-scale spectral-mechanical property paired dataset and adopts a dual-branch fusion architecture to process the spatial-spectral features of hyperspectral images and the spectral subgraph of the knowledge graph constructed based on material composition, respectively. The mapping relationship between spectrum and mechanical parameters is initially established through end-to-end supervised learning. Then, self-supervised contrastive learning is introduced to enhance the ability to identify subtle differences in the spectrum. In addition, online incremental learning is achieved by combining the validation data accumulated in actual production, so that the model can continuously evolve and accurately and robustly realize intelligent inference from the workpiece surface spectrum to key mechanical properties such as elastic modulus and yield strength.

[0045] Specifically, by integrating multi-source sensing technologies such as lidar and hyperspectral imaging, and combining intelligent decision-making with deep neural networks and historical knowledge bases, the core problems in the background technology, such as insufficient adaptability of clamping force and inability to flexibly adjust for workpieces of different materials and complex shapes, caused by reliance on traditional fixed fixtures and programs, are solved. This effectively avoids workpiece deformation and stress concentration caused by improper clamping force, laying a stable and reliable foundation for subsequent high-precision machining.

[0046] S2. Based on the obtained initial clamping force value, the pressure distribution of the hydraulic clamp is dynamically adjusted by applying an adaptive control algorithm. Before machining, the influence of cutting force on the workpiece is simulated to determine an optimized clamping scheme to avoid deformation.

[0047] Furthermore, step S2 specifically includes:

[0048] The initial value of clamping force, the three-dimensional geometric model of the workpiece, and the material property parameters are input into the high-fidelity digital twin system. The simulated time-varying cutting load is applied according to the actual machining path and process parameters. The dynamic stress field and cumulative deformation distribution of the workpiece under the current clamping state are calculated through transient finite element analysis, and the plastic deformation risk area and structural instability critical point in the entire machining process are identified.

[0049] Based on the deformation prediction results, an adaptive model predictive control algorithm is adopted, with minimizing the maximum deformation of the workpiece as the main control objective. Combined with clamping energy consumption optimization and vibration suppression auxiliary objectives, the pressure output value of each hydraulic clamping unit is dynamically adjusted. After each adjustment, the cutting and clamping coupling simulation is re-performed in the digital twin environment, and the workpiece deformation is evaluated to determine whether it meets the preset form and position tolerance threshold. This forms a closed-loop iterative optimization process until a clamping force distribution optimization scheme that meets multiple constraints is output.

[0050] The optimized clamping force distribution scheme is converted into pressure setting commands for each clamping unit and executed by a high-precision hydraulic servo system. During the machining process, based on the real-time data fed back by the strain sensor and optical deformation monitoring device integrated into the clamping system, the actual deformation is compared with the simulation prediction results online. The clamping force is dynamically fine-tuned through a feedforward and feedback composite control strategy to ensure that the workpiece is continuously in a deformation-controlled state throughout the entire process, ultimately achieving intelligent clamping control that avoids deformation.

[0051] Furthermore, the dynamic stress field and cumulative deformation distribution of the workpiece under the current clamping state are calculated through transient finite element analysis. Specifically, this includes: in the digital twin system, adaptive mesh generation is performed based on the three-dimensional geometric model of the workpiece; for areas with large curvature, abrupt changes in wall thickness, or fine features (such as hole edges and fillets), a finer mesh is used to ensure accurate representation of geometric details; for flat or gently stressed areas, a coarser mesh is used to balance computational accuracy and efficiency; material properties are spatially differentiated based on hyperspectral identification results; and each element is associated with its corresponding material constitutive model (such as elastoplastic model, creep model, etc.) to reflect the non-uniformity of material distribution.

[0052] The analysis employs a combined explicit and implicit transient solution strategy. Within each simulation time step: First, based on the process planning path, the built-in cutting force prediction model is invoked to calculate the dynamic cutting load (including tangential force, radial force, and axial force) corresponding to the current tool position, and this load is applied to the corresponding position of the workpiece as nodal force or surface force. Simultaneously, based on the clamping force distribution scheme in the current optimization iteration, the pressure of the hydraulic clamp is applied to the clamp-workpiece contact surface through a contact algorithm. The nonlinear equilibrium equation is solved using the complete Newton-Raphson iterative method to calculate the stress increment and strain increment at this time step, and the stress field and deformation field of the workpiece are updated.

[0053] Through continuous time-step cumulative calculations, the system can output dynamic stress cloud maps and cumulative deformation evolution animations of the workpiece throughout the entire processing process. Based on this, it automatically identifies two types of key risks: one is the plastic deformation risk area, which is the area where the equivalent stress continuously exceeds the yield strength of the material, and marks its location, range, and degree of stress exceeding the limit; the other is the structural instability critical point, which is the critical moment and location at which the workpiece may suddenly warp, vibrate, or resonate by using eigenvalue buckling analysis or monitoring a sudden increase in local deformation rate, providing a clear suppression target for subsequent clamping force optimization.

[0054] An adaptive model predictive control algorithm is employed, with minimizing the maximum workpiece deformation as the primary control objective. This is combined with clamping energy consumption optimization and vibration suppression as auxiliary objectives to dynamically adjust the pressure output values ​​of each hydraulic clamping unit. Specifically, this involves constructing a finite-time domain rolling optimization problem within each control cycle, using the current clamping state and the predicted cutting load sequence as input. A sequential quadratic programming method is used to solve online a multi-objective function that minimizes the maximum workpiece deformation within the predicted time domain, while simultaneously integrating the total energy consumption of the hydraulic system and the root mean square of vibration acceleration at key points. This dynamically calculates the optimal pressure adjustment sequence for each hydraulic actuator. The algorithm concretizes the vibration suppression objective as a constraint on the amplitude of the dominant resonant frequency in the machining spectrum, incorporating it as a penalty term into the optimization function, thereby simultaneously disrupting the resonance condition when adjusting the pressure distribution. After each optimization, the system performs a one-step look-ahead simulation in a digital twin environment to verify performance and feeds the results back to the next cycle, forming an adaptive closed loop of prediction-optimization-verification-rolling, achieving real-time, multi-objective coordinated control of clamping pressure.

[0055] The clamping force is dynamically fine-tuned through a combined feedforward and feedback control strategy. Specifically, the feedforward prediction and compensation path is as follows: Based on the pre-stored complete machining path and process parameters, the dynamic load change trend caused by cutting force, spindle centrifugal force, etc. is predicted in advance at each moment. Based on this prediction and digital twin model, a set of ideal clamping force compensation is pre-calculated and sent to the hydraulic servo system as the basic setting value. The goal is to actively counteract the influence of interference before it occurs, thereby improving the system's response speed and stability.

[0056] Feedback Real-time Correction Path: The strain changes of key support points of the fixture are monitored in real time by a micro strain sensor array integrated into the fixture body, and the micron-level displacement of key features of the workpiece is captured in real time by a laser displacement meter or vision measurement system installed at the workstation. The real-time collected strain and displacement data are compared with the predicted value of the digital twin system at the current moment in milliseconds to generate a deviation signal.

[0057] Composite and Dynamic Adjustment: The deviation signal is input to the fuzzy PID controller, which weights and fuses the feedforward compensation output with the feedback correction amount calculated in real time based on the deviation. The fusion weight is dynamically adjusted according to the machining stage (e.g., feedforward is more relied upon in roughing and high cutting force stages; feedback is more relied upon in finishing and high precision stages). The final generated composite control command drives the hydraulic cylinder through a high-bandwidth servo valve, achieving microsecond-level dynamic fine-tuning of the clamping force. At the same time, the controller fine-tunes its own control parameters online based on the statistical characteristics of the deviation (e.g., mean, variance), and can feed back significant and recurring deviation patterns to the digital twin system for updating and correcting its simulation model.

[0058] Specifically, by performing pre-simulation and in-process closed-loop control in a digital twin environment, the problem of workpiece deformation caused by the fixed program of traditional fixtures and the inability to flexibly adjust the clamping force according to the dynamic machining load was solved in the background technology; it effectively resisted cutting interference and ensured that the workpiece remained in a stable and controlled geometric posture throughout the entire process, laying a solid foundation for subsequent high-precision milling.

[0059] S3. Extract workpiece posture stability index from the determined optimized clamping scheme, obtain real-time displacement and vibration data fed back by multi-axis sensors, and determine if the displacement exceeds the preset threshold to activate the feedback loop and obtain stable posture data.

[0060] Furthermore, step S3 specifically includes:

[0061] The key attitude stability index system is extracted from the determined optimized clamping scheme, including but not limited to the displacement tolerance range of the workpiece in the machining coordinate system, the vibration amplitude threshold of each degree of freedom, the safe range of resonance frequency, and the dynamic stiffness coefficient; the stability index system is correlated with the geometric features, material properties and process parameters of the workpiece to determine the stability monitoring benchmark.

[0062] A monitoring network consisting of a multi-axis inertial measurement unit, a laser displacement sensor, and a vibration accelerometer is used to collect real-time data on the six degrees of freedom displacement, vibration spectrum, and micro-deformation of the workpiece during the processing. The real-time data is compared online with the stability index system. If any index continues to exceed the threshold, an abnormal state is triggered and a feedback control signal containing the deviation type, degree, and adjustment priority is generated.

[0063] Based on the feedback control signal, the control model in the optimized clamping scheme is invoked to generate targeted clamping parameter adjustment instructions (including pressure redistribution of each hydraulic actuator, fine adjustment of support point position, and application of active damping); after the adjustment is executed, sensor data is re-acquired for stability evaluation; if the requirements are still not met, a reinforcement learning strategy is initiated for iterative optimization until all stability indicators continue to converge within the allowable range.

[0064] Once the workpiece's posture reaches a stable state through closed-loop adjustment, a complete stable posture dataset containing steady-state displacement coordinates, vibration residual energy spectrum, dynamic stiffness coefficient, and final clamping parameters is collected and output. This dataset, along with the adjustment process record, is synchronously stored in the process knowledge base to form a traceable and reusable posture stability control case, providing an adaptive prediction and control benchmark for subsequent processing.

[0065] Furthermore, the stability index system is modeled in association with the workpiece's geometric features, material properties, and process parameters. Specifically, this includes: establishing a stability association model with the workpiece's structural topology, material distribution characteristics, and machining load spectrum as core parameters; for geometric features, the system extracts key dimensional ratios (such as length-to-diameter ratio and thickness ratio), center of mass position, inertia tensor, and coordinates of locally weak stiffness areas; for material properties, it imports the spatially differentiated elastic modulus distribution field and damping characteristic parameters obtained in step S1; for process parameters, it loads the cutting force time history curve, spindle speed spectrum, and tool-workpiece contact state sequence for the current process; based on these inputs, the workpiece is calculated under specific clamping conditions using a multi-physics coupled modal analysis module. The model identifies the first six natural frequencies, mode shapes, and modal damping ratios, and maps these dynamic characteristics to a stability index system. For example, the dynamic stiffness coefficient is defined as the response function of key points on the workpiece at a specific excitation frequency. The safe range of the resonance frequency is set as ±15% bandwidth of each natural frequency. The vibration amplitude threshold is differentiated according to the material fatigue limit and machining accuracy requirements by frequency band. The displacement tolerance range is determined by combining the workpiece's form and position tolerances with the theoretical elastic deformation caused by the cutting force. This correlation model is calibrated by comparing historical machining data with experimental test results and has online update capabilities. It can dynamically fine-tune various thresholds according to actual working conditions, forming a personalized stability monitoring benchmark that is precisely matched to specific workpieces and processes.

[0066] Based on the feedback control signal, the control model in the optimized clamping scheme is invoked to generate targeted clamping parameter adjustment instructions. Specifically, upon receiving a feedback control signal containing the type, degree, and priority of deviation, the high-fidelity digital twin control model established in step S2 is first invoked as the simulation test environment. For displacement deviation issues, the system uses a stiffness reconstruction algorithm to calculate how to change the overall stiffness distribution of the workpiece-clamping system by adjusting the pressure ratio of each hydraulic actuator to suppress excessive deformation in a specific direction. For abnormal vibration, the active damping injection module is activated. This module calculates the dynamic damping force sequence that needs to be applied at specific support points based on the vibration spectrum characteristics and the principle of antiphase superposition. The specific adjustment command generation adopts a hierarchical decision-making architecture: the bottom fast response layer generates a preliminary adjustment scheme based on a preset rule base (such as "increase the pressure difference of the axially symmetrical support point if the axial displacement is too large"); the middle optimization layer performs multi-step look-ahead simulation of the preliminary scheme in a digital twin environment based on a model predictive control framework, in order to make all stability indicators return to the threshold in the shortest adjustment time; if convergence is still not achieved after several iterations, the top decision layer activates a reinforcement learning agent, which explores pressure and support point combinations different from conventional strategies and evaluates their long-term stability benefits in a digital twin environment, thereby finding a better adjustment strategy; the final generated adjustment command is a complete set of executable instructions containing the precise pressure setpoint of each hydraulic unit, pressure change gradient, support point position fine adjustment (for movable support units with servo drive), and active damping application parameters. The system will drive the actuator to make precise adjustments based on this and immediately enter the next monitoring-evaluation cycle.

[0067] Specifically, by constructing a real-time monitoring and adaptive adjustment closed loop, the problem of unstable working posture caused by dynamic changes in cutting force in the background technology is solved; active maintenance and dynamic control of workpiece posture are realized, effectively suppressing machining errors caused by vibration or displacement, and providing reliable state guarantee for subsequent high-precision flipping and milling.

[0068] S4. Based on the obtained stable attitude data, the flipping angle is pre-calibrated using a rotary servo system, and the ambient temperature sensor input is integrated through the thermal deformation compensation module to obtain a precise rotation path plan.

[0069] Furthermore, step S4 specifically includes:

[0070] S41. Based on stable attitude data, Kalman filtering and wavelet denoising algorithms are used to remove sensor noise and environmental interference to obtain a high-confidence clean attitude dataset. Based on the geometric relationship between the clean attitude dataset and the target machining surface of the workpiece, the theoretical reference rotation angle required for accurate flipping is calculated as the initial input for path planning.

[0071] S42. Integrate ambient temperature sensor and infrared temperature measurement data of workpiece surface, and input them into thermal deformation compensation module; The thermal deformation compensation module calculates the angular offset caused by temperature gradient in real time according to the material thermal expansion coefficient, structural heat transfer model and historical temperature rise-deformation mapping relationship; The angular offset is dynamically superimposed with the reference rotation angle to generate the compensated target flipping angle.

[0072] S43. Based on the compensated target angle and purification posture data, combined with the dynamic constraints of the rotary servo system, a smooth S-shaped acceleration and deceleration rotation trajectory is planned, and the angle, angular velocity and angular acceleration sequence of each axis is calculated; the rotation trajectory parameters are mapped to the position and velocity command sequence of each axis of the servo motor through the inverse kinematics model, forming an executable preliminary rotation path.

[0073] S44. Collision interference detection and vibration suppression optimization are performed on the initial rotation path. The continuity of the command sequence is optimized by using a look-ahead control algorithm. A precise rotation path plan containing position, velocity, acceleration and temperature compensation parameters is generated and output to the servo execution unit to achieve high-precision, low-vibration flipping control.

[0074] Furthermore, Kalman filtering and wavelet denoising algorithms are employed to remove sensor noise and environmental interference, resulting in a high-confidence purified attitude dataset. Specifically, this involves: receiving raw attitude data streams (containing six degrees of freedom displacement, acceleration, and angular velocity) from multi-axis sensors; performing preliminary filtering using an improved Kalman filter; this filter uses the rigid body motion equation of the workpiece as the state equation and sensor observations as measurements, dynamically estimating the optimal attitude value by calculating and updating the state covariance matrix in real time, effectively suppressing random white noise; subsequently, wavelet packet decomposition is performed on the filtered data: the db4 wavelet basis function is selected to perform a 5-level decomposition of the signal, obtaining sub-band signals of different frequency bands; the coefficients of each sub-band are processed using an improved threshold function (such as a soft-hard threshold tradeoff function), focusing on removing high-frequency noise components caused by environmental vibration and electromagnetic interference; finally, wavelet reconstruction is performed to obtain a high-confidence attitude dataset with dual time-frequency domain purification; this dataset not only contains smoothed attitude parameters but also retains the true dynamic change trend, providing reliable input for subsequent accurate calculations.

[0075] The system calculates the angular offset caused by the temperature gradient in real time. Specifically, the thermal deformation compensation module receives two types of temperature inputs in real time: one is data from a multi-point thermocouple array deployed in the processing environment, used to construct the ambient temperature field; the other is the temperature distribution map of the workpiece surface obtained by scanning with an infrared thermal imager. First, a three-dimensional transient heat transfer finite element model is used, combined with material thermal properties (thermal conductivity, specific heat capacity), to calculate the three-dimensional temperature gradient field inside the workpiece in real time. Then, based on the workpiece's geometric model and the material's thermal expansion coefficient, a thermo-structural coupling analysis is used to calculate the non-uniform thermal expansion caused by this temperature gradient field. For the angular offset, the system focuses on the thermal deformation of key geometric features on the workpiece used for flipping and positioning (such as the reference plane and the axis of the positioning hole). By solving for the directional changes of these features in space caused by thermal expansion, a spatial geometric transformation model (such as Rodrigues' rotation formula) is used to calculate the angular offset vector relative to the initial calibration state. The entire process runs online at millisecond intervals to ensure the real-time nature of the compensation values.

[0076] The process of generating the compensated target flipping angle specifically includes: fusing the reference rotation angle (geometrically calculated from the cleanup posture data) and the real-time angle offset in a unified machine tool coordinate system, and performing calculations based on the spatial posture transformation matrix. First, the reference rotation angle is converted into a rotation matrix; simultaneously, the angle offset (including small-angle rotations around multiple axes) is converted into another small rotation matrix; then, the two rotation matrices are multiplied sequentially (the compensation is applied before or after the reference rotation based on the process logic) to obtain a comprehensive rotation matrix; finally, the compensated target flipping angle for path planning is extracted from the comprehensive rotation matrix using a matrix inverse algorithm (e.g., converting to quaternions and then to Euler angles). This angle integrates the dual effects of workpiece stable posture and thermal deformation, and is the direct target value for driving the rotary servo system to achieve precise flipping.

[0077] The rotation trajectory parameters are mapped to position and velocity command sequences for each axis of the servo motors using an inverse kinematics model, forming an executable preliminary rotation path. Specifically, this involves: planning a smooth S-shaped velocity curve that follows the dynamic constraints of the rotary servo system (such as maximum angular velocity, maximum angular acceleration, and maximum jerk) based on the compensated target flipping angle; this curve describes the angle of the main rotation axis as a function of time; and obtaining the angular velocity and angular acceleration sequences through differentiation based on this function. Subsequently, this one-dimensional main axis rotation sequence is input into the inverse kinematics model describing the mechanical structure of the rotary table. This model establishes a mapping relationship from the target angle of the main axis to the required rotation amounts of each servo motor axis based on the actual transmission chain of the table (e.g., including two rotation axes, A and C). For a two-axis rotary table, the inverse kinematics process typically involves spherical trigonometry to determine the required rotation angles of the A and C axes, thereby collaboratively achieving the spatial orientation of the main axis. Finally, the output consists of independent, synchronous position-time and velocity-time sequence commands for each servo motor axis, constituting a preliminary rotation path that the control actuator can directly execute.

[0078] The continuity of the command sequence is optimized using a look-ahead control algorithm. Specifically, after generating the initial motor shaft position sequence, a multi-step look-ahead control algorithm is introduced for smooth optimization. This involves first forward-reading dozens to hundreds of subsequent path points, analyzing the second-order difference (acceleration) and third-order difference (jerk) of the position commands. When a sudden change in acceleration or jerk is detected, speed look-ahead planning is initiated. Without changing the final position and total time, transitional B-spline curve segments are inserted or the speed curve is locally Bezier-modified to make the changes in acceleration and jerk smooth and continuous. This process is equivalent to online smoothing of the original path. The optimized command sequence has continuous and differentiable speed and acceleration curves, fundamentally eliminating the servo system start-stop shock, increased tracking error, and structural resonance that may be caused by discontinuous commands, ensuring the stability and high precision of the rotation process.

[0079] Specifically, by integrating real-time thermal deformation compensation and high-precision path planning, the problem of rotational positioning deviation caused by temperature changes and servo fluctuations in the background technology has been solved; the upgrade from relying on fixed mechanical limits to dynamic temperature compensation and intelligent planning has been realized, effectively eliminating the impact of thermal deformation and unstable mechanical response on angular accuracy, ensuring the accuracy and repeatability of each flipping angle, and curbing the transmission and accumulation of errors in multi-faceted processing from key links.

[0080] S5. Extract the angle deviation vector from the precise rotation path planning, apply the error compensation algorithm to analyze the impact of the residual deviation from the previous flip on the current face, and determine the cumulative error correction coefficient.

[0081] Furthermore, step S5 specifically includes:

[0082] S51. Extract angle deviation data from the precise rotation path planning and preprocess it using Kalman filtering. Based on the preprocessed data, establish a state space model describing the error transmission law between multiple processes. Solve the coupling effect matrix of the previous flip residual deviation on the current processing surface through the system identification algorithm, thereby quantifying the specific influence value of historical error sources on the current surface.

[0083] S52. By integrating real-time sensor data and historical error records, assess the spatial distribution and evolution trend of the current cumulative error. Use fuzzy clustering algorithm to intelligently stratify the cumulative error into different error sets such as systematic deviation, random fluctuation, clamping residue, and thermal deformation. Based on the amplitude, time-varying characteristics, and sensitivity to the final processing accuracy of each type of error, extract the key error items that have a significant impact and use them as priority targets for compensation calculation.

[0084] S53. For the extracted key error terms, a neural network-driven adaptive compensator, combined with a preset error compensation mechanism, calculates an adaptive compensation value for each term and integrates them to form a global error adjustment strategy. Based on the error adjustment records after compensation according to the error adjustment strategy, a multiple linear regression model is used to analyze the mapping relationship between the cumulative error and the parameter to be corrected, dynamically fitting and calculating the cumulative error correction coefficient suitable for the current working condition. This coefficient achieves root-cause suppression of error by mapping the compensation amount back to the source deviation of path planning.

[0085] Furthermore, a state-space model describing the error propagation law between multiple processes is established. Specifically, this includes: based on the preprocessed historical angle deviation sequence, a recursive least squares system identification method is used to construct the error propagation state-space model. The state vector x_k of this model is defined as the error state of the key geometric features of the workpiece after the k-th process (such as the normal deviation of the positioning surface, the axial deviation of the datum hole, etc.). The input vector u_k contains the known error sources of the k-th process (such as clamping repeat positioning error, residual thermal deformation from the previous process), and the output vector... y_k represents the actual measurement deviation of the current process. By solving for model parameters of the form x_{k+1}=A*x_k+B*u_k+w_k, y_k=C*x_k+v_k, where A, B, and C are matrices to be identified, and w_k and v_k represent process and measurement noise, the contribution of preceding errors (included in state x_k) to the output y_k of the current process is clearly quantified. During the identification process, the system automatically adjusts the model order to ensure that the model's prediction error is minimized in the typical process chain of multi-face milling of automotive parts. This model can clearly reveal the transmission path of phenomena such as how initial clamping misalignment leads to accumulated deviations in hole positions during subsequent flipping.

[0086] Fuzzy clustering algorithm is used to intelligently stratify cumulative errors, classifying them into error sets of different natures such as systematic deviations, random fluctuations, clamping residues, and thermally induced deformation. Specifically, this involves: extracting multi-dimensional feature vectors of the cumulative errors in the current process, including: spatial distribution (error values ​​at each measurement point), temporal stability (fluctuations between consecutive processing batches), spectral characteristics (whether it has periodicity), and correlation with operating conditions (such as whether it is strongly correlated with temperature and load); then, fuzzy C-means clustering algorithm is used to design a membership function with error amplitude stability, spatial correlation, and operating condition sensitivity as coordinate axes; and all error numbers are then... The data points (derived from different measurement points and processing batches) are iteratively divided into four preset clusters: systematic deviation (high amplitude, high stability, high spatial correlation), random fluctuation (low amplitude, low stability, low correlation), clamping residue (high amplitude, medium stability, strongly correlated with fixture layout), and thermally induced deformation (amplitude exhibits time-varying / temperature-varying patterns, medium spatial correlation). Based on the cluster center characteristics of each cluster and the contribution weight of the error to the final contour / position accuracy, the key error items with the greatest impact are automatically selected (usually systematic deviation and clamping residue have the highest priority), providing precise targets for subsequent compensation.

[0087] An adaptive compensator driven by a neural network, combined with a pre-defined error compensation mechanism, calculates an adaptive compensation value for each error. Specifically, for each extracted key error term, a dedicated lightweight feedforward neural network is deployed. For example, for residual clamping errors, the network's input layer includes the clamping position coordinates, clamping force sequence, and residual error vector from the previous process. For thermally induced deformation, the inputs are the workpiece's initial temperature, the ambient temperature rise curve, and the processing thermal load. The network learns online to establish a nonlinear mapping from these inputs to the required angle or position compensation amount. The compensation mechanism is pre-defined: for example, for systematic deviations, the compensation strategy is direct reverse offset; for time-varying errors, the strategy is feedforward predictive compensation. The neural network accurately calculates the compensation amount under this strategy. The system continuously fine-tunes the network weights using the deviation between the latest processing results and the predicted compensation effect, achieving adaptive evolution of the compensator and ensuring that it can provide corresponding accurate compensation values ​​for different error sources when processing complex workpieces such as oil tank flanges.

[0088] The mapping relationship between cumulative error and the parameters to be corrected is analyzed using a multiple linear regression model. The cumulative error correction coefficient, adapted to the current working conditions, is dynamically fitted and calculated. Specifically, this involves using a matrix of historical cumulative error observations over a period (e.g., 10 consecutive workpieces) as the dependent variable Y, and the matrix of various error compensations calculated and actually executed by a neural network compensator during the same period as the independent variable X. Through regularized multiple linear regression, the coefficient matrix β in the form Y=X*β is solved, where β is the cumulative error correction coefficient. Its physical meaning is: how many units of compensation need to be applied at the source of the error to eliminate a unit of a certain type of cumulative error. The calculation process is dynamic: after each new workpiece is processed, new data is added to the regression dataset, and the coefficient β is refitted, so that the correction coefficient can reflect the influence of slowly changing factors such as equipment status and tool wear in real time. The final correction coefficient will be directly used to update the rotation path planning of subsequent workpieces in stage S4 or the coordinate adjustment values ​​in stage S6, achieving root-cause suppression and closed-loop control of similar errors.

[0089] Specifically, by establishing an error propagation model and an intelligent traceability and compensation mechanism, the core problem of the gradual accumulation of angular deviations caused by multiple flipping in the background technology has been solved. This has realized the transformation from single-link delayed correction to global forward compensation of the transmission chain, which can effectively cut off the accumulation path of errors between multiple processes. It ensures that the relative positional accuracy between the surfaces of multi-faceted workpieces such as transmission housings is always controlled during continuous flipping processing, thereby fundamentally avoiding assembly deviations caused by error superposition.

[0090] S6. Based on the determined cumulative error correction coefficient, obtain the position coordinate adjustment value of each face in the multi-face milling sequence. If the coordinate deviation exceeds the accuracy standard, iterative optimization is triggered. After continuous simulation verification, the final machining path is obtained.

[0091] Furthermore, step S6 specifically includes:

[0092] Based on the cumulative error correction coefficient, the initial position coordinates of each face in the multi-face milling sequence are adaptively adjusted through a pre-constructed dynamic error compensation model to generate coordinate correction values ​​for each face. The dynamic error compensation model combines the geometric topology of the workpiece with the error propagation chain and uses machine learning prediction algorithms to estimate the adjusted coordinate offset in real time, forming a preliminary coordinate compensation dataset.

[0093] The calculated coordinate adjustment values ​​are compared with the preset accuracy range in real time. If the position deviation of any face exceeds the threshold, the multi-objective iterative optimization process is started. Based on the reinforcement learning optimization module, combined with multiple constraints such as machining stability, cutting force distribution and tool life, a new temporary coordinate data sequence is dynamically generated and the compensation strategy is updated in real time until the coordinates of each face meet the accuracy standard.

[0094] The adjusted coordinate sequence was loaded into a high-fidelity digital twin platform to perform multi-process continuous cutting simulation. Monte Carlo simulation and vibration spectrum analysis were used to evaluate the stability of the machining path under dynamic load, the consistency of surface quality, and the risk of tool-workpiece interference.

[0095] If the simulation finds that local vibration exceeds the standard or path interference, it will be fed back to the optimization module for secondary adjustment. After continuous simulation verification, the final machining path dataset containing the precise coordinates of each surface, tool path, cutting parameters and compensation records will be output.

[0096] Furthermore, the pre-built dynamic error compensation model adaptively adjusts the initial position coordinates of each face in the multi-face milling sequence, generating coordinate correction values ​​for each face. Specifically, this includes: constructing a directed weighted graph network describing the propagation of errors between processes based on the workpiece's assembly feature tree and machining process dependency graph; the nodes in the graph represent the theoretical coordinates of each machining face, and the edge weights are dynamically determined by the cumulative error correction coefficients determined by S5 and the geometric constraints between processes (such as perpendicularity and coaxiality); when the initial coordinate sequence is input, the model first propagates forward along the machining sequence, through the graph... The integral network calculates the expected offset of each node due to the accumulation of errors in the preceding process; then, backpropagation compensation is performed: starting from the last process, the calculated offset is back-allocated to its predecessor node (machined surface) according to the correction coefficient and geometric constraints, forming a six-degree-of-freedom coordinate compensation vector for each surface; the whole process is completed online at the millisecond level, and the output coordinate correction value includes not only the position translation, but also the normal vector rotation adjustment required to compensate for the accumulated deviation of the angle, ensuring that the position and orientation of each surface are synchronously and accurately compensated when machining multi-faceted workpieces such as automobile engine cylinder blocks.

[0097] The adjusted coordinate offset is estimated in real time using machine learning prediction algorithms. Specifically, an online learning gradient boosting decision tree (GBDT) ensemble model is used for real-time offset prediction. The input feature vector of this model includes: the previous reference coordinates, various cumulative error values ​​provided by S5, the geometric features of the surface (such as area, curvature, and normal), the material removal rate, and the actual machining error feedback of the previous surface. After each compensation, the model immediately performs a very short micro-cutting process simulation in a digital twin sandbox environment based on the updated coordinates, predicting secondary offsets such as tool deformation and thermal drift that may occur during cutting at this coordinate. The offset predicted by this micro-simulation is fused with the macro-compensation amount based on the graph network to output the final estimated coordinate offset. This GBDT model continuously compares the predicted offset with the subsequent actual measurement deviation, and uses online gradient descent to update its decision tree structure and weights in real time, thereby continuously approximating the true offset pattern under complex working conditions.

[0098] Based on the reinforcement learning optimization module, combined with multiple constraints such as machining stability, cutting force distribution and tool life, a new temporary coordinate data sequence is dynamically generated. Specifically, the coordinate adjustment problem is constructed as a partially observable Markov decision process; the agent's state is the current coordinate sequence, real-time process state (cutting force, vibration) and predicted accuracy deviation; the action is the fine adjustment of the coordinates of each surface; the reward function is a negative multi-objective weighted sum: R=-(w1*accuracy deviation+w2*vibration energy+w3*cutting force fluctuation+w4*tool life consumption). The agent learns using a proximal policy optimization algorithm. In each iteration, the agent explores multiple temporary coordinate adjustment trajectories in parallel within a digital twin environment. For each trajectory, the system calls the built-in fast process simulator to evaluate its machining stability (through simulated vibration), cutting force distribution (through a prediction model), and tool life consumption (through empirical formulas) within seconds. The agent updates its policy network based on these evaluation results, generating the next batch of better-performing temporary coordinate sequences. This process is repeated until a Pareto front solution is found that achieves an optimal balance between vibration, cutting force stability, and tool wear while ensuring accuracy. This solution is then output as a new temporary coordinate data sequence.

[0099] Monte Carlo simulation and vibration spectrum analysis were used to evaluate the stability of the machining path under dynamic loads, surface quality consistency, and tool-workpiece interference risk. Specifically, the evaluation process first constructed a parameter space including material property fluctuations (e.g., yield strength ±3% normal distribution), tool wear state (exponential distribution based on Taylor formula), and clamping stiffness changes (±10% uniform distribution). Latin hypercube sampling was used to generate 300 parameter combinations, and explicit dynamic simulations were executed in parallel on a digital twin platform. For vibration spectrum analysis, acceleration time-history data at key points of the spindle and workpiece were extracted, and short-time Fourier transforms were performed to generate time-frequency diagrams. Vibration frequency bands with energy exceeding 4 m / s² and duration exceeding 50 ms were automatically identified and compared with the machine tool structure modal library. The system identifies the frequency of chatter risk. Surface quality assessment reconstructs the microscopic surface morphology by recording the relative vibration displacement between the tool and workpiece, and calculates the 95% confidence interval of the arithmetic mean deviation Ra of each machined surface profile. If the upper limit of the interval exceeds the process requirement value (e.g., Ra 1.6 μm), the quality is considered unstable. Interference risk detection employs a hierarchical bounding box-based continuous collision detection algorithm to calculate the minimum signed distance between the tool envelope and the workpiece / fixture in real time with a resolution of 0.005 mm. Risk segments with a distance ≤ 0.1 mm are marked, and the tool pose at the time of occurrence is recorded. Finally, a quantitative report is generated by integrating all simulation samples, clearly marking areas with a vibration exceedance probability > 5%, a surface quality compliance rate < 99%, or an interference risk > 0, providing accurate defect location and data support for path optimization.

[0100] Specifically, by using path-level global compensation and pre-verification, the problems of coordinate-level residual deviation and batch consistency were solved. The complete process from local parameter adjustment to global path optimization generated high-precision and high-reliability final machining instructions, ensuring that each workpiece can be processed according to the uniformly verified optimal path in mass production, thereby achieving and maintaining a high degree of machining consistency.

[0101] S7. Integrate the data of each machining surface from the obtained final machining path, and use a closed-loop monitoring system to compare the actual milling results with the expected model in real time to obtain a complete error accumulation control record.

[0102] The actual milling result refers to the physical machining data collected in real time by multiple sensors (such as displacement sensors, vibration sensors, laser measuring instruments, etc.) during the machining process, including the actual machining coordinates, dynamic vibration spectrum, cutting force fluctuations, and online measured workpiece surface morphology and dimensional information; the expected model refers to the final machining path data, which includes the theoretical position coordinates of each machining surface, the planned tool path, cutting parameters, and the corresponding workpiece design CAD model.

[0103] Furthermore, step S7 specifically includes:

[0104] Based on the final processing path, the geometric features, process parameters and historical compensation data of each processing surface are extracted. Spatiotemporal alignment and feature extraction are performed through a multi-source data fusion engine to generate a structured surface-process association dataset containing spatial topological relationships, processing time sequence and error labels. A dynamic mapping is established with the expected model in the digital twin system.

[0105] The system acquires vibration, displacement, cutting force and temperature data in real time through a multi-sensor synchronous acquisition system. It then performs multimodal feature fusion analysis by combining structured datasets to dynamically identify abnormal fluctuations in milling results. Based on an adaptive threshold adjustment strategy, when key indicators are detected to exceed the dynamic tolerance range, a real-time comparison engine is activated to perform high-precision matching between actual machining data and expected models to locate the process position, type and intensity of deviations.

[0106] An error propagation modeling method based on graph neural networks is adopted to conduct multi-process source tracing analysis on the identified deviation data and construct a spatiotemporal evolution map of error accumulation. Based on the spatiotemporal evolution map, support vector machine and time series prediction model are integrated to make rolling predictions on the error development trend and output the adjustment direction of future processes and risk warning signals.

[0107] By integrating real-time monitoring data, deviation analysis results, and predictive information, a structured error control record is automatically generated through an intelligent recording engine and simultaneously archived to the process knowledge base. Based on the record feedback, the parameter configuration and early warning mechanism of the closed-loop monitoring system are dynamically optimized to achieve adaptive iteration of error control strategies and continuous accumulation of process knowledge, forming an error accumulation control system that can be verified in a closed loop and evolve autonomously.

[0108] Furthermore, a multi-source data fusion engine is used for spatiotemporal alignment and feature extraction to generate a structured surface-process association dataset containing spatial topological relationships, processing sequence, and error labels. Specifically, this includes: extracting geometric features, process parameters, and historical compensation data of each processing surface from the final processing path; using the interpolation clock of the CNC system as a reference, aligning theoretical data, process parameters, and subsequent sensor acquisition plans under a unified spatiotemporal reference through timestamp synchronization and spatial coordinate transformation; for each processing surface, the system extracts features such as its center coordinates, normal vector, and boundary contour, and combines them with processing sequence and geometric adjacency to construct an initial graph structure with processing surfaces as nodes and spatiotemporal relationships as edges; in this process, the system performs feature-level fusion of theoretical data, process settings, and online measurement results, adding an error label field to each node, and finally generating a structured surface-process association map containing spatial topology, processing sequence, and error labels. This map is directly mapped to the expected model in the digital twin system, providing a queryable and inferable data foundation for real-time comparison and traceability.

[0109] An error propagation modeling method based on graph neural networks is adopted to perform multi-process source tracing analysis on the identified deviation data and construct a spatiotemporal evolution map of error accumulation. Specifically, when a deviation is identified in the current process (node ​​K in the map), the deviation is used as an initial signal and input into a pre-trained spatiotemporal graph convolutional network. This network captures the error propagation caused by geometric adjacency relationships (such as shared edges and coplanarity) through spatial convolutional layers, and traces the residual effects of previous processes along the processing sequence through temporal convolutional layers. The network calculates the contribution weight of each previous node in the map to the current deviation through multi-layer message passing and attention mechanisms, thereby automatically outlining the main error propagation paths (such as clamping deviation at node A, cumulative error at node C, and dimensional deviation at node K). The deviation event, contribution weight, propagation path, and working condition snapshot are added as a source tracing record to the global error spatiotemporal evolution map. This map is continuously enriched with each processing batch, gradually revealing the generation, propagation, and accumulation patterns of errors under specific workpiece structures and process routes.

[0110] This system integrates Support Vector Machine (SVM) and time series prediction models to perform rolling predictions of error development trends, outputting adjustment directions and risk warning signals for future processes. Specifically, based on a constructed spatiotemporal evolution map, the system performs two types of predictions in parallel: First, it extracts static and dynamic feature vectors from the current node and operating conditions, inputs them into a multi-class SVM, which quickly identifies the most likely error evolution pattern (such as rapid divergence or oscillating convergence) based on a historical pattern library and outputs a qualitative risk level. Second, it extracts time series data of similar errors from the historical map, inputs them into a Long Short-Term Memory (LSTM) network for quantitative extrapolation, and predicts the specific numerical change trend within the next few process steps. The qualitative discrimination results of SVM and the quantitative prediction curve of LSTM are intelligently weighted and fused. If the fusion result indicates a high risk level and the predicted value will exceed the threshold, a structured signal containing a specific adjustment direction (such as increasing the fixture pressure by X% before the next process to compensate for the decrease in stiffness) and a warning level is automatically generated, providing a forward-looking decision-making basis for subsequent processing and realizing closed-loop control from passive response to proactive prevention.

[0111] Specifically, by constructing a closed-loop monitoring and intelligent evolution system for the entire process, the ultimate problem of unsustainable consistency in mass production caused by the unknowability of the processing process, the difficulty in tracing deviations, and the difficulty in accumulating process knowledge has been solved. It not only realizes the transformation from post-processing inspection to real-time perception and intervention during the process, but also automatically generates knowledge records from all process data and feeds them back to optimize system parameters, forming an intelligent closed loop that can continuously learn from historical experience and autonomously evolve process strategies, thereby dynamically maintaining and continuously improving processing accuracy and stability in long-term mass production.

[0112] Example 2, please refer to Figure 4 This embodiment provides a control device for a rotary flipping fixture suitable for CNC milling machining centers, used to implement a control method for a rotary flipping fixture suitable for CNC milling machining centers, including:

[0113] The perception and preliminary judgment module scans the workpiece using LiDAR and hyperspectral imager, constructs a 3D model containing material and geometric features using edge computing, and combines deep neural networks and dynamic knowledge graphs to match material mechanical parameters and generate dynamic stress thresholds. Then, it calls on the intelligent rule base and machine learning model to quickly calculate the initial scheme of safe clamping force. Its hardware consists of a LiDAR scanning head and hyperspectral imaging lens installed above the workstation, an edge computing server, a material database server, and an AI inference unit in the control cabinet.

[0114] The twin and optimization module loads the model, simulates cutting loads, and calculates dynamic stress and deformation through a high-fidelity digital twin system to identify machining risk areas. It integrates an adaptive model predictive controller to perform multi-objective iterative optimization of hydraulic fixture pressure distribution in the twin environment, generate an optimized clamping scheme, and use feedforward-feedback composite control to achieve dynamic fine-tuning during machining. Its core is a high-performance simulation workstation, which receives input through a data interface unit and outputs optimization commands via a dedicated motion control card.

[0115] The monitoring and attitude stabilization module collects six-degree-of-freedom motion and vibration data of the workpiece in real time through a monitoring network consisting of a multi-axis inertial measurement unit, a laser displacement sensor, and a vibration accelerometer. Based on the built-in stability index system, it performs online comparison and anomaly diagnosis, and calls the real-time controller and reinforcement learning optimizer to dynamically adjust the clamping parameters and output stable attitude data. Its physical components include a miniature six-axis inertial measurement unit embedded in the fixture, a distributed laser displacement sensor group, and a triaxial vibration accelerometer. All sensors are connected to a multi-channel synchronous data acquisition instrument with integrated controller via a high-speed fieldbus.

[0116] The planning and compensation module filters and reduces noise from stable attitude data, calculates thermally induced angular offset in real time by combining environmental and workpiece surface temperature data, plans a smooth S-shaped rotation trajectory and maps it to a servo axis command sequence through inverse kinematics, and then performs collision detection and look-ahead control optimization to output a precise rotation path. Its hardware includes a multi-point environmental temperature sensor array, an infrared thermal imager and an embedded filtering and calculation board. The path planning algorithm runs on a dedicated controller for the rotation axis and communicates with the CNC system through a fiber optic bus.

[0117] The modeling and error module constructs an error propagation state space model based on historical angle deviation data. It uses fuzzy clustering algorithm to intelligently stratify and extract key terms of cumulative error, and adopts a neural network-driven adaptive compensator combined with multiple linear regression method to dynamically calculate and fit the cumulative error correction coefficient. Its functions are mainly implemented by software algorithms running on a data analysis server, which acquires measurement data through industrial Ethernet.

[0118] The dynamic compensation module adaptively adjusts the multi-face milling coordinate sequence based on the correction coefficient using a graph network-based dynamic error compensation model. When the coordinate deviation exceeds the tolerance, a reinforcement learning optimization engine is activated to iteratively generate new coordinates. The module also uses a high-fidelity simulation platform to evaluate the stability, surface quality, and interference risk of the machining path through Monte Carlo simulation and vibration spectrum analysis. After feedback optimization, the final machining path is output. This module software is integrated into the process planning computer and uses a simulation accelerator card to improve computational efficiency. The optimization data is sent to the CNC system through a secure communication protocol.

[0119] The monitoring and evolution module generates a structured surface-process correlation dataset through a data fusion engine. It integrates a multi-sensor synchronous acquisition and real-time comparison system to locate processing deviations. It uses an error tracing model based on graph neural networks and a prediction model that fuses support vector machines and time series data for trend early warning. It also records the entire process data to the process knowledge base to achieve adaptive iterative updates of system parameters and strategies. Its hardware foundation is an enterprise-level process data server, supplemented by workshop edge gateway devices and monitoring and alarm terminals, which form a closed loop through factory network interconnection.

[0120] This includes a central control unit (using an industrial PC or high-performance PLC) as the main controller, a hydraulic execution subsystem for executing clamping commands (including a servo hydraulic pump station, distributed valve island, hydraulic cylinders and pressure sensors), a rotary execution subsystem for executing rotary motions (including a high-precision rotary table, dual-axis servo drives and high-torque servo motors), and a high-speed communication network using time-sensitive networking or deterministic Ethernet to ensure the synchronization and real-time performance of commands and data throughout the system.

[0121] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A control method for a rotary flipping fixture suitable for CNC milling machining centers, characterized in that: Includes the following steps: S1. Collect workpiece material difference data and initial shape parameters in real time through sensors, and match the corresponding elastic modulus and stress threshold using a preset material database to obtain the initial value of targeted clamping force. S2. Based on the obtained initial value of clamping force, apply an adaptive control algorithm to dynamically adjust the pressure distribution of the hydraulic fixture, simulate the effect of cutting force on the workpiece before processing, and determine an optimized clamping scheme to avoid deformation. S3. Extract workpiece posture stability index from the determined optimized clamping scheme, obtain real-time displacement and vibration data fed back by multi-axis sensors, and determine if the displacement exceeds the preset threshold to activate the feedback loop and obtain stable posture data. S4. Based on the obtained stable attitude data, the flipping angle is pre-calibrated using a rotary servo system, and the ambient temperature sensor input is integrated through the thermal deformation compensation module to obtain accurate rotation path planning; S5. Extract the angle deviation vector from the precise rotation path plan, apply the error compensation algorithm to analyze the impact of the residual deviation from the previous flip on the current face, and determine the cumulative error correction coefficient. S6. Based on the determined cumulative error correction coefficient, obtain the position coordinate adjustment value of each face in the multi-face milling sequence. If the coordinate deviation exceeds the accuracy standard, trigger iterative optimization. After continuous simulation verification, obtain the final machining path. S7. Integrate the data of each machining surface from the obtained final machining path, and use a closed-loop monitoring system to compare the actual milling results with the expected model in real time to obtain a complete error accumulation control record.

2. The control method for a rotary flipping fixture suitable for CNC milling machining centers according to claim 1, characterized in that: Step S1 specifically includes: By using LiDAR and hyperspectral imager to scan in tandem, the three-dimensional point cloud and spectral information of the workpiece surface are collected in real time. The edge computing unit is used for data fusion and noise reduction to construct a three-dimensional feature model of the workpiece containing geometric contours, material composition and surface roughness, and extract key shape parameters and material feature vectors. The extracted material features are input into a pre-trained deep neural network model, combined with a dynamically updated material knowledge graph database, to match and output the elastic modulus distribution, yield strength curve and strain hardening parameters that match the workpiece in real time, and generate a dynamic stress threshold field that changes with the geometric shape. By inputting geometric features and material parameters into an intelligent rule base built on historical processing data, the system retrieves the best historical cases through a similarity matching algorithm. Combined with preset calculation rules and machine learning models, it quickly calculates the initial value of the safety clamping force and its distribution scheme that meet the deformation control requirements.

3. The control method for a rotary flipping fixture suitable for CNC milling machining centers according to claim 1, characterized in that: Step S2 specifically includes: The initial value of clamping force, the three-dimensional geometric model of the workpiece, and the material property parameters are input into the high-fidelity digital twin system. The simulated time-varying cutting load is applied according to the actual machining path and process parameters. The dynamic stress field and cumulative deformation distribution of the workpiece under the current clamping state are calculated through transient finite element analysis, and the plastic deformation risk area and structural instability critical point in the entire machining process are identified. Based on the deformation prediction results, an adaptive model predictive control algorithm is adopted, with minimizing the maximum deformation of the workpiece as the main control objective. Combined with clamping energy consumption optimization and vibration suppression auxiliary objectives, the pressure output value of each hydraulic clamping unit is dynamically adjusted, and closed-loop iterative optimization is performed in a digital twin environment until the output of a clamping force distribution optimization scheme that meets multiple constraints is obtained. The optimized clamping force distribution scheme is converted into pressure setting instructions for each clamping unit and executed by a high-precision hydraulic servo system. During the machining process, the clamping force is dynamically fine-tuned through a feedforward and feedback composite control strategy.

4. The control method for a rotary flipping fixture suitable for CNC milling machining centers according to claim 1, characterized in that: Step S3 specifically includes: From the determined optimized clamping scheme, an attitude stability index system including displacement tolerance range, vibration amplitude threshold, resonant frequency safety range and dynamic stiffness coefficient is extracted, and modeled in association with workpiece characteristics and process parameters to determine the monitoring benchmark. The six-degree-of-freedom displacement, vibration spectrum and micro-deformation data of the workpiece during the processing are collected in real time by a multi-axis sensor monitoring network and compared online with the stability index system. If the index exceeds the threshold, an abnormal state is triggered and a feedback control signal is generated. Based on the feedback control signal, the control model in the optimized clamping scheme is invoked to generate targeted clamping parameter adjustment instructions. Iterative optimization is then performed through a reinforcement learning strategy until all stability indicators continuously converge within the allowable range, and a complete stable posture dataset is output.

5. A control method for a rotary flipping fixture suitable for CNC milling machining centers according to claim 4, characterized in that: Step S4 specifically includes: S41. Based on stable attitude data, Kalman filtering and wavelet denoising algorithms are used to remove sensor noise and environmental interference to obtain a high-confidence clean attitude dataset; based on the geometric relationship between the clean attitude dataset and the target machining surface of the workpiece, the theoretical reference rotation angle is calculated. S42. Integrate ambient temperature sensor and workpiece surface infrared temperature measurement data, input into thermal deformation compensation module, calculate the angular offset caused by temperature gradient in real time, dynamically superimpose the angular offset with the reference rotation angle, and generate the compensated target flipping angle. S43. Based on the compensated target angle and purification posture data, combined with the dynamic constraints of the rotary servo system, a smooth S-shaped acceleration and deceleration rotation trajectory is planned, and the rotation trajectory parameters are mapped to generate the position and speed command sequence of each axis of the servo motor through the inverse kinematics model, forming a preliminary rotation path. S44. Collision interference detection and vibration suppression optimization are performed on the initial rotation path, and the continuity of the command sequence is optimized by using a look-ahead control algorithm; a precise rotation path plan containing position, velocity, acceleration and temperature compensation parameters is generated.

6. The control method for a rotary flipping fixture suitable for CNC milling machining centers according to claim 1, characterized in that: Step S5 specifically includes: S51. Extract angle deviation data from the precise rotation path planning, and preprocess it using Kalman filtering. Establish a state-space model describing the error transmission law between multiple processes. Solve the coupling effect matrix of the previous flip residual deviation on the current machining surface through the system identification algorithm to quantify the specific influence value of historical error sources. S52. By integrating real-time sensor data and historical error records, assess the spatial distribution and evolution trend of the current cumulative error, use fuzzy clustering algorithm to intelligently stratify the cumulative error, and extract key error items with significant impact as priority targets for compensation calculation. S53. For the extracted key error terms, an adaptive compensator driven by a neural network is used to calculate the adaptive compensation value and integrate them to form a global error adjustment strategy. Based on the error adjustment record after compensation, the mapping relationship between the cumulative error and the parameter to be corrected is analyzed through a multiple linear regression model. The cumulative error correction coefficient that is suitable for the current working condition is dynamically fitted and calculated.

7. A control method for a rotary flipping fixture suitable for CNC milling machining centers according to claim 6, characterized in that: Establish a state-space model to describe the error propagation law between multiple processes. Specifically, based on the preprocessed historical angle deviation sequence, a recursive least squares system identification method is used to construct an error propagation state-space model based on the historical angle deviation sequence. The contribution of the preceding error to the current process is quantified by solving the model parameters, and the model order is automatically adjusted during the identification process to minimize the prediction error.

8. A control method for a rotary flipping fixture suitable for CNC milling machining centers according to claim 1, characterized in that: Step S6 specifically includes: Based on the cumulative error correction coefficient, the initial position coordinates of each face in the multi-face milling sequence are adaptively adjusted through a pre-constructed dynamic error compensation model to generate coordinate correction values ​​for each face. The dynamic error compensation model combines the geometric topology of the workpiece with the error propagation chain and uses machine learning prediction algorithms to estimate the adjusted coordinate offset in real time. The calculated coordinate adjustment value is compared with the preset accuracy range in real time. If the position deviation exceeds the threshold, the multi-objective iterative optimization process is started. Based on the reinforcement learning optimization module, combined with multiple constraints such as machining stability, cutting force distribution and tool life, a new temporary coordinate data sequence is dynamically generated. The adjusted coordinate sequence is loaded into the digital twin platform to perform multi-process continuous cutting simulation. The stability of the machining path under dynamic load, surface quality consistency and tool-workpiece interference risk are evaluated by Monte Carlo simulation and vibration spectrum analysis. If it is not qualified, it is fed back for optimization until it is verified to be qualified and the final machining path dataset is output.

9. A control method for a rotary flipping fixture suitable for CNC milling machining centers according to claim 1, characterized in that: Step S7 specifically includes: Based on the final processing path, the geometric features, process parameters and historical compensation data of each processing surface are extracted. Spatiotemporal alignment and feature extraction are performed through a multi-source data fusion engine to generate a structured surface-process association dataset containing spatial topological relationships, processing time sequence and error labels. A dynamic mapping is established with the expected model in the digital twin system. The system acquires vibration, displacement, cutting force and temperature data in real time through a multi-sensor synchronous acquisition system, dynamically identifies abnormal fluctuations in milling results, and activates a real-time comparison engine when key indicators exceed tolerances. This engine performs high-precision matching between actual machining data and the expected model to pinpoint the process location, type and intensity of the deviation. An error propagation modeling method based on graph neural networks is used to perform multi-process source tracing analysis on the identified deviation data, construct a spatiotemporal evolution map of error accumulation, and integrate support vector machine and time series prediction model to make rolling predictions on error development trends, output adjustment direction and risk warning signals; By integrating real-time monitoring data, deviation analysis results, and prediction information, structured error control records are automatically generated and synchronously archived into the process knowledge base. Based on the feedback from these records, the parameter configuration and early warning mechanism of the closed-loop monitoring system are dynamically optimized, enabling adaptive iteration of error control strategies and continuous accumulation of process knowledge.

10. A control device for a rotary flipping fixture of a CNC milling machining center, applied to the control method for a rotary flipping fixture of a CNC milling machining center as described in any one of claims 1-9, characterized in that: include: The perception and preliminary judgment module scans the workpiece using lidar and hyperspectral imager, constructs a 3D model containing material and geometric features using edge computing, matches material mechanical parameters and generates dynamic stress thresholds using deep neural networks and dynamic knowledge graphs, and quickly calculates the initial scheme of safe clamping force by calling intelligent rule base and machine learning model. The twin and optimization module loads the model, simulates cutting loads, and calculates dynamic stress and deformation through a high-fidelity digital twin system to identify machining risk areas; it integrates an adaptive model predictive controller to perform multi-objective iterative optimization of hydraulic clamp pressure distribution in the twin environment, generates an optimized clamping scheme, and uses feedforward-feedback composite control to achieve dynamic fine-tuning during machining; The monitoring and attitude stabilization module collects six-degree-of-freedom motion and vibration data of the workpiece in real time through a monitoring network consisting of a multi-axis inertial measurement unit, a laser displacement sensor, and a vibration accelerometer. Based on the built-in stability index system, it performs online comparison and anomaly diagnosis, and calls the real-time controller and reinforcement learning optimizer to dynamically adjust the clamping parameters and output stable attitude data. The planning and compensation module filters and reduces noise from the stable attitude data, calculates the thermally induced angle offset in real time by combining environmental and workpiece surface temperature data, plans a smooth S-shaped rotation trajectory and maps it into a servo axis command sequence through inverse kinematics, performs collision detection and look-ahead control optimization, and outputs a precise rotation path. The modeling and error module constructs an error propagation state space model based on historical angle deviation data, uses fuzzy clustering algorithm to intelligently stratify and extract key terms of cumulative error, and adopts a neural network-driven adaptive compensator combined with multiple linear regression method to dynamically calculate and fit the cumulative error correction coefficient. The dynamic compensation module adaptively adjusts the multi-face milling coordinate sequence based on the correction coefficient through a graph network-based dynamic error compensation model. When the coordinate deviation exceeds the tolerance, the reinforcement learning optimization engine is activated to iteratively generate new coordinates. The stability, surface quality and interference risk of the path are evaluated by Monte Carlo simulation and vibration spectrum analysis using a high-fidelity simulation platform. After feedback optimization, the final machining path is output. The monitoring and evolution module generates a structured surface-process related dataset through a data fusion engine. It integrates multi-sensor synchronous acquisition and real-time comparison systems to locate processing deviations. It uses an error tracing model based on graph neural networks and a prediction model that fuses support vector machines and time series data for trend early warning. It also records the entire process data into the process knowledge base to achieve adaptive iterative updates of system parameters and strategies.

Images