A parameter iterative algorithm valve shell investment casting precision forming method

Abstract

The present application relates to investment casting precision forming technical field, specifically disclose a kind of parameter iterative algorithm valve shell investment casting precision forming method.Method is extracted structural characteristic parameter set by three-dimensional point cloud voxelization analysis, and initial process parameter vector is generated in combination with historical process knowledge base;Response surface prediction model is established based on central composite design and finite element simulation, and optimal process parameter vector is calculated by sequential quadratic programming iteration to drive forming equipment;In the solidification process of physical casting, temperature and displacement data are collected in real time using sensor array, parameter fine tuning increment is generated by inputting response surface prediction model, online dynamic feedback control is carried out on optimal process parameter vector, until the demoulding is completed when vibration acceleration signal is lower than preset safety threshold value.The present application can effectively improve the forming size precision and organization uniformity of complex valve shell, and inhibit casting defects.

Description

Technical Field

[0001] This invention relates to the field of investment casting precision forming technology, specifically to a valve body investment casting precision forming method based on a parameter iteration algorithm. Background Technology

[0002] As a core actuator in fluid control systems, the performance of valves directly determines the operational safety of critical industrial systems. Investment casting, with its advantages of no parting line and high dimensional accuracy, has become the mainstream forming process for valve bodies. With the development of supercritical energy and deep-sea exploration technology, valve bodies are evolving towards variable wall thickness, complex flow channels, and high pressure resistance. The complex structure leads to strong nonlinear characteristics in the filling and solidification process of molten metal. Specifically, thin-walled areas are prone to cold shut defects due to excessively rapid cooling, thick areas are prone to shrinkage defects due to insufficient feeding, and locations with drastic changes in wall thickness are prone to thermal stress and thermal cracking defects due to excessive temperature gradients.

[0003] Current investment casting process parameters largely rely on the experience of engineers for setting, and are adjusted through repeated trial pours. These process parameters include pouring temperature, pouring speed, cooling gas flow rate, cooling time, vibration frequency, and amplitude. This trial-and-error approach, when applied to various valve body types, reveals problems such as long optimization cycles and high scrap rates.

[0004] Existing technologies, such as patent document CN121373374A, disclose a steam dewaxing and multi-directional uniform cooling process, but lack adaptive adjustment algorithms for cooling gas flow rate and time; patent document CN113290234A discloses a fixed-parameter water cooling and vibration device, but faces the bottleneck of not being able to intelligently adapt parameters to complex flow channels. The aforementioned technical solutions all remain at the static open-loop control level, lacking a low-level optimization mechanism based on physical feature extraction and multi-physics coupling iteration, and failing to construct a closed-loop data system covering the entire lifecycle from filling to demolding for state perception and dynamic parameter correction, resulting in the inability to overcome the process bottleneck in the molding yield of highly complex valve bodies. Summary of the Invention

[0005] The purpose of this invention is to provide a precision forming method for valve body investment casting based on a parameter iteration algorithm, which solves the technical problem of low yield of complex valve body forming caused by the reliance on open-loop static experience settings in existing investment casting processes.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A parameter iterative algorithm-based method for precision forming of valve body by investment casting includes the following steps: Step S1: Use a 3D scanning device to acquire 3D point cloud data of the physical valve body, perform geometric analysis on the 3D point cloud data to extract a set of structural feature parameters, and store the set of structural feature parameters in the valve body structural feature parameter library. Step S2: Based on the set of structural feature parameters in the valve body structure feature parameter library, and combined with the historical process knowledge base, perform similarity matching calculation to generate an initial process parameter vector; Step S3: Using the initial process parameter vector as the input node, perform finite element simulation based on the central composite design method, extract the objective function response set, and fit to generate a response surface prediction model; wherein, the boundary conditions of the finite element simulation are set based on the boundary voxel set obtained in step S1. Step S4: Invoke the response surface prediction model and perform sequential quadratic programming iteration guided by the objective function response set to calculate and output the optimal process parameter vector; Step S5: Drive the melting equipment, casting equipment, cooling system and vibration demolding system according to the optimal process parameter vector to generate a physical casting in the forming process state; Step S6: A sensor array is used to collect real-time temperature and displacement data of the physical casting in the forming process. The real-time temperature and displacement data are input into the response surface prediction model to calculate the parameter fine-tuning increment. The parameter fine-tuning increment is superimposed on the optimal process parameter vector for dynamic feedback control until the vibration acceleration signal collected by the sensor is lower than the preset vibration acceleration safety threshold, at which point demolding is determined to be complete. After demolding, the set of structural feature parameters of qualified valve bodies in this batch and the optimal process parameter vector are written back to the valve body structural feature parameter library and the historical process knowledge base.

[0007] Furthermore, the step S1 of performing geometric analysis on the 3D point cloud data to extract the set of structural feature parameters further includes: The 3D point cloud data is converted into a 3D voxel mesh using an octree spatial partitioning algorithm; Calculate the minimum Euclidean distance from the center point of any voxel in the three-dimensional voxel grid to the physical boundary of the three-dimensional point cloud data, and multiply the minimum Euclidean distance by two to generate the wall thickness parameter in the corresponding spatial coordinates.

[0008] Furthermore, the step of extracting the set of structural feature parameters in step S1 further includes: Based on the region growing algorithm, the hollow flow channel connected domains within the three-dimensional voxel mesh are identified, and the topological skeleton of the hollow flow channel connected domains is extracted as flow channel parameters. A binarization filtering extraction operation below a preset wall thickness safety threshold is performed on the wall thickness parameter, and the area ratio and spatial coordinate domain set of the region below the preset wall thickness safety threshold are extracted as the thin-wall parameter; The wall thickness parameter, the flow channel parameter, and the thin wall parameter are assembled together to form the set of structural feature parameters.

[0009] Furthermore, the step of generating the initial process parameter vector in step S2 further includes: Extract the historical feature dimension vector stored in the historical process knowledge base; The distance matching value between the set of structural feature parameters and each of the historical feature dimension vectors is calculated using a weighted Euclidean distance metric. Extract the six values ​​bound to the target historical process record with the smallest distance matching value: pouring temperature, pouring speed, cooling gas flow rate, cooling time, vibration frequency, and amplitude. Map and assign these six values ​​to generate the initial process parameter vector.

[0010] Furthermore, the step of extracting the objective function response set in step S3 further includes: Under each experimental matrix node generated by the central composite design, the thermohydrodynamic solver is invoked to calculate the spatiotemporal physical field of the molten metal filling and solidification process; The coordinates of the boundary points of the thermal stress deformation field are extracted from the spatiotemporal physical field. The coordinates of the boundary points of the thermal stress deformation field are then subjected to a three-dimensional Boolean difference operation with the three-dimensional point cloud data. The output size deviation is used as the first element of the objective function response set.

[0011] Furthermore, the step of extracting the objective function response set in step S3 further includes: The solidification latent heat release rate span is extracted from the spatiotemporal physical field, and the grain size standard deviation distribution rate is calculated based on the non-uniform nucleation model. The uniformity of the microstructure is then output as the second element of the objective function response set. The dimensional deviation and tissue uniformity are assembled into a tensor to form the objective function response set.

[0012] Furthermore, the regression equation for fitting and generating the response surface prediction model in step S3 is:

[0013] In the formula, The objective function response set represents the predicted output, including two output dimensions: size deviation and fabric uniformity. , This represents the 6-dimensional process input variables, which correspond to pouring temperature, pouring speed, cooling gas flow rate, cooling time, vibration frequency, and amplitude, respectively. This represents the intercept constant term in the regression model; This represents the partial regression coefficients of the first-order linear main effects; This represents the partial regression coefficient for a second-order pure quadratic effect; This represents the partial regression coefficient of the two-parameter cross-coupling effect; i and j represent the process parameter dimension indices, and satisfy... .

[0014] Furthermore, step S4, which involves performing sequential quadratic programming iterations guided by the objective function response set, includes: The elements in the response set of the objective function are weighted and integrated to construct a comprehensive objective cost function. In each internal iteration loop, the comprehensive objective cost function is expanded by Taylor second order to construct an approximate Hessian matrix of the Lagrange function to solve the quadratic programming subproblem, and an optimization search step size direction vector is generated to update the parameter vector nodes.

[0015] Furthermore, the step of calculating and outputting the optimal process parameter vector in step S4 includes: when the absolute value of the difference change of the comprehensive objective cost function is strictly lower than the preset convergence threshold, stopping the iterative calculation, and forcibly outputting the parameter vector node corresponding to the stopping time as the optimal process parameter vector.

[0016] Furthermore, step S5, which involves driving each piece of equipment based on the optimal process parameter vector, includes: The values ​​of pouring temperature and pouring speed in the optimal process parameter vector are analyzed and converted into electrical drive signals to control the smelting equipment and the pouring equipment; The cooling gas flow rate and cooling time values ​​in the optimal process parameter vector are analyzed, and the cooling system is driven to perform forced heat exchange. The vibration frequency and amplitude values ​​in the optimal process parameter vector are analyzed and the vibration demolding system is driven to perform mechanical peeling.

[0017] Furthermore, the step of driving the cooling system to perform forced heat exchange further includes: The cooling system is constructed as a spatially addressed microchannel array cooling mechanism controlled by multiple sets of independent electromagnetic proportional servo valves. Based on the numerical distribution of wall thickness parameters in the set of structural feature parameters and the ratio of the maximum to minimum wall thickness of the casting, the external ceramic shell surface of the physical casting is divided into N spatially independent cooling zones using the K-means clustering algorithm. N is a positive integer adaptively determined by the K-means clustering algorithm, and the value of N ranges from 3 to 6. The K-means clustering iteration terminates when the change in cluster centers is less than 0.1 mm. The single cooling gas flow rate value in the optimal process parameter vector is reduced in dimension and expanded into an N-dimensional initial cooling vector, which independently drives the electromagnetic proportional servo valves corresponding to the N independent cooling zones.

[0018] Furthermore, the step of generating the real-time temperature data further includes: The spatial partial derivatives of the discrete matrix of the two-dimensional temperature field on the surface are solved in Cartesian coordinates to extract and generate a real-time spatial thermal gradient tensor including the normal temperature gradient and the tangential temperature gradient. The real-time spatial thermal gradient tensor is compared with the critical phase transition thermal stress gradient threshold preset in the spatiotemporal physical field simulation stage, and the local chilling coordinate set exceeding the threshold is extracted.

[0019] Furthermore, the specific execution steps of the dynamic feedback control include: The nonlinear model predictive controller receives the local chill coordinate set as additional hard constraints; it deconstructs the cooling gas flow rate increment in the parameter fine-tuning increment into an N-dimensional spatial correction control vector that maps one-to-one with the N spatially independent cooling zones, and replaces the cooling gas flow rate scalar in the parameter fine-tuning increment with the N-dimensional spatial correction control vector to form a corrected optimal process parameter vector; in the iterative solution of the minimizing model predictive control quadratic cost function, a dynamic decay penalty weight is applied to the vector elements in the N-dimensional spatial correction control vector corresponding to the local chill coordinate set; the dynamic decay penalty weight is adaptively selected based on the multiple by which the local thermal gradient exceeds a threshold, and the value range is [missing value]. .

[0020] Furthermore, the step of acquiring data using a sensor array in step S6 includes: using an infrared thermal imaging sensor array to capture the two-dimensional temperature field discrete matrix of the surface of the physical casting in the forming process state, and generating the real-time temperature data; using a non-contact laser interferometric displacement sensor to detect the instantaneous linear displacement offset of the physical casting in the forming process state, and generating the real-time displacement data.

[0021] In step S6, the minimization model predictive control quadratic cost function of the dynamic feedback control is:

[0022] In the formula, This represents the dimensionless value of the quadratic cost function used to predict the control model. This represents a prediction step size range constant, with a value range of 15-30; This represents a constant controlling the step size range, with a value range of 3-8. This indicates the prediction made by the response surface methodology at time k. The state quantity formed at any moment; This represents the zero-deviation benchmark target value for defect-free forming; This represents the incremental sequence of process parameter fine-tuning; The positive definite diagonal weight matrix representing the penalty for deviation in forming state is: matrix; The parameter adjustment amplitude suppresses the positive definite diagonal weight matrix, which is... The matrix; k represents the real-time control sampling time; j represents the iteration step index.

[0023] The above prediction step size and control step size The value of is crucial to the performance of the controller.

[0024] This embodiment sets the sensor sampling time window. Based on Fourier thermal conductivity ( The solidification time constant at the maximum wall thickness of 185 mm for solid valve shells was estimated. Combined with a thermoelastic mechanics model, the dominant frequency period of thermal disturbance at the flange-thin-wall interface was statistically measured, yielding a dominant frequency period range of 20 s to 40 s for this type of valve shell. To ensure the accuracy of the prediction time domain... It can fully cover at least one thermal disturbance cycle to effectively predict the long-term effects of thermal disturbances, while avoiding the accumulation of model mismatch errors caused by the prediction time domain exceeding the effective matching interval of the model. Based on the Nyquist sampling theorem and system identification results, The constraints are 15 to 30. Simultaneously, based on the controllability criterion for discrete systems, to ensure the control degrees of freedom... It can perform a sufficient number of correction actions under the excitation of the maximum rate of change of the measured thermal gradient (approximately 3℃ / cm / s), and the output will not oscillate at high frequencies due to excessively long control step size (i.e., it satisfies the stability constraints of discrete systems). ),Will The constraints are 3 to 8. Offline simulation verification shows that under this value combination, the model predicts the quadratic cost function. The convergence speed is compared to A 40% increase compared to The computational burden is reduced by 60%, and the dynamic overshoot is suppressed to within 5%, thus achieving the optimal balance between control accuracy and computational efficiency.

[0025] Furthermore, steps S3 and S4 are replaced by the following steps: The initial process parameter vector and the set of structural feature parameters are input into a deep learning prediction proxy model based on a long short-term memory network architecture to generate a prediction target set. The deep learning prediction proxy model includes 3 LSTM layers and 2 fully connected layers. The deep learning prediction agent model is invoked, and a collection function including the expectation improvement criterion is constructed based on the Gaussian process to perform Bayesian global optimization. The coordinates of the new parameter node that maximizes the absolute value of the comprehensive utility improvement are searched as the optimal process parameter vector. The maximum number of exploration steps for the Bayesian optimization is 200 steps.

[0026] Compared with the prior art, the present invention has the following beneficial effects: This invention utilizes 3D point cloud voxelization analysis to precisely extract wall thickness, flow channel, and thin-wall parameters from a solid valve shell, transforming complex geometric constraints into a quantifiable and transferable set of structural feature parameters. This provides a rigid input benchmark for subsequent simulation and optimization. Based on this, a response surface prediction model is constructed using a central composite design and thermodynamics solver. Furthermore, through sequential quadratic programming iterations, the optimal process parameter vector for dimensional deviation and microstructure uniformity is found on a continuous surrogate surface. This achieves a high-precision mathematical mapping from structural features to process parameters, significantly improving the scientific rigor and efficiency of parameter optimization.

[0027] Furthermore, this invention employs a nonlinear model predictive control architecture, using real-time temperature data collected by an infrared thermal imaging sensor array and real-time displacement data collected by a non-contact laser interferometric displacement sensor as feedback. During solidification, the optimal process parameter vector is incrementally adjusted online, forming active compensation and closed-loop suppression for transient environmental thermal disturbances, effectively overcoming the formation of intergranular defects such as shrinkage porosity, thermal cracking, and distortion. In the subordinate scheme, the cooling system is reconstructed into a spatially addressed microchannel array cooling mechanism using the K-means clustering algorithm, and a dynamic attenuation penalty weight with spatiotemporal dual dimensions is applied based on the real-time spatial thermal gradient tensor. This achieves targeted intervention of the local cooling rate of the valve shell with extreme wall thickness differences, further improving the consistency of molding density and dimensional accuracy. Attached Figure Description

[0028] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0029] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0030] Figure 2 This is a flowchart of the dynamic feedback control process for spatial addressing cooling matrix clustering and real-time spatial thermal gradient tensor constraints in this invention.

[0031] Figure 3This is a flowchart illustrating the construction of the deep learning prediction agent model and the Bayesian global optimization alternative path of this invention.

[0032] Figure 4 This is one of the operation interface diagrams of the system during the implementation of the present invention.

[0033] Figure 5 This is the second diagram of the system's operation interface during the implementation of this invention.

[0034] Figure 6 This is the third diagram of the system's operation interface during the implementation of this invention. Detailed Implementation

[0035] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0036] The following is in conjunction with the appendix Figures 1-6 The embodiments of the present invention will be described in detail below.

[0037] Example 1: This example discloses a method for precision forming of valve body by investment casting using a parameter iteration algorithm, including the following steps: Step S1: Use a 3D scanning device to acquire 3D point cloud data of the physical valve body, perform geometric analysis on the 3D point cloud data to extract a set of structural feature parameters, and input the set of structural feature parameters into the valve body structural feature parameter library.

[0038] This step transforms the complex boundary constraints of the valve housing into digital tensors that can be invoked by the algorithm. The process begins with the activation of a 3D optical scanning device. A non-contact optical scanner equipped with binocular vision and a blue structured light projector is used, with the structured light projector wavelength set to 450nm. A standard metal solid valve housing is placed on a high-precision single-axis rotary table, and a phase-shifted fringe coding pattern is projected onto the surface of the solid valve housing by the structured light projector. A binocular vision camera synchronously captures the modulation deformation fringes on the surface of the solid valve housing at a sampling frame rate of 8 frames per second. Based on the fundamental principles of triangulation and epipolar geometric constraints, the absolute spatial coordinates of the solid surface are calculated, generating local point cloud blocks from multiple viewpoints.

[0039] An iterative nearest-neighbor registration and fusion operation based on a fast point feature histogram descriptor is performed on local point cloud blocks to generate raw 3D point cloud data including the global outer contour of the entity and the internal flow channel topology. After removing outliers, the standard set of no-noise points is defined as the 3D point cloud data. 3D point cloud data The physical coordinates of any random point in space are marked as follows: .

[0040] In order to transform 3D point cloud data without topological attributes To convert these into geometric constraint parameters that guide solidification phase transitions, it is necessary to process the 3D point cloud data. By performing voxelization and dimensionality reduction analysis, the three core variables affecting the heat transfer topology were accurately separated.

[0041] Voxelization meshing action, setting a bounding box to fully cover the 3D point cloud data The extreme coordinate range is determined by using octree space partitioning logic to uniformly divide the bounding box into segments with side lengths of [missing information]. A cube-shaped voxel unit. Voxel side length. The thickness is adaptively selected based on the casting's outline dimensions. For small castings with an outline dimension less than 500mm, a thickness of 0.5-1.0mm is used; for large castings with an outline dimension greater than or equal to 500mm, a thickness of 1.0-2.0mm is used. This embodiment is for large valve bodies. The value is set to 1.0 mm. This will contain 3D point cloud data. The voxel units that the projection falls into are labeled as the boundary voxel set. Boundary voxel set The physical flow-blocking surface and heat dissipation surface of the valve body are defined.

[0042] Extracting wall thickness parameters In the solidification thermodynamics of castings, the normal wall thickness directly determines the location of local hot spots and the solidification delay time. The extraction command is applied to the boundary voxel set. An arbitrary solid central voxel element inside a closed surface, with coordinates denoted as . Search its projection onto the boundary voxel set The nearest Euclidean distance on the wall thickness. The mathematical definition equation is expressed as:

[0043] In the formula, Representing coordinates The scalar wall thickness parameter at the location, in mm; This represents the set of boundary voxels that constitute the inner and outer surfaces; Represents the boundary voxel set The Middle Spatial coordinates of discrete points; This indicates the minimum value operation. The scalar wall thickness calculation results corresponding to the central voxel elements of all internal solids are arranged according to the coordinate sequence to form a wall thickness parameter matrix. .

[0044] Extract flow channel parameters The flow channel topology determines the Reynolds number distribution characteristics and the risk of air entrapment during molten metal filling. This is achieved through analysis of three-dimensional point cloud data. The enclosed hollow region undergoes a 3D thinning morphological erosion operation, stripping away boundary voxels down to the central axis skeleton, which is reduced to a single pixel width. The spatial 3D coordinate sequence of the central axis skeleton and the gradient vector of cross-sectional area change along the tangent direction are extracted. The skeleton coordinate sequence and cross-sectional gradient vector are then integrated into a one-dimensional array, which serves as the flow channel parameters. This one-dimensional array can be directly used as the feature dimension input for the subsequent historical process knowledge base matching module and finite element simulation model.

[0045] Extracting thin-wall parameters Thin-walled regions are high-risk coordinate domains that can cause cooling at the molten metal flow front, leading to cold shut defects. The system pre-determines a fixed empirical constant as a safe threshold for wall thickness. This threshold is determined based on the critical thin-wall thickness of the valve body material; for heat-resistant steel materials, The value range is 12~15mm, and this embodiment has a fixed setting. Iterate through the wall thickness parameter matrix output by the previous calculations. Execution logic judgment formula:

[0046] In the formula, This represents the extracted thin-wall parameter tensor; The three-dimensional spatial coordinates of the center voxel of the casting solid; This represents the wall thickness parameter at the corresponding spatial coordinate point, in mm. This represents the preset wall thickness safety threshold, set to 15.0 mm. Simultaneously, the area integral percentage of this point set in the global surface area is extracted as a scalar.

[0047] Thus far, the 3D point cloud data Wall thickness parameters output by dimensionality reduction analysis Flow channel parameters and thin-wall parameters Strictly packaged and bound, the combination is defined as a set of structural feature parameters. The system executes the data storage instruction, transferring the set of structural feature parameters. Write the valve housing structural feature parameter library to the corresponding index entry in the solid-state drive. (Structural feature parameter set) It is the thermodynamic intrinsic cause of internal defects in metal casting and must be used as a rigid input for subsequent parameter mapping steps.

[0048] Step S2: Based on the set of structural feature parameters in the valve body structure feature parameter library, and combined with the historical process knowledge base, perform similarity matching calculation to generate an initial process parameter vector.

[0049] This step provides a high-confidence starting point for parameters, avoiding computational divergence caused by blindly searching globally. This step receives the set of structural feature parameters output from step S1. As the starting point for calculation.

[0050] The system's underlying database contains a historical process knowledge base. Historical Craftsmanship Knowledge Base It pre-stores data copies of hundreds of thousands of successful casting cases from various dynasties. Historical process knowledge base. Each index record in the index is marked as an index. It consists of two parts: The first half is the historical feature dimension vector. It is also made up of historical barriers. Historical Flow Historical thin wall Composition; the latter half consists of six static control quantities corresponding to the physical casting process at that time: historical pouring temperature. Historical casting speed Historical cooling gas flow rate Historical cooldown time Historical vibration frequency Historical amplitude .

[0051] The matching algorithm loads the set of structural feature parameters of the current input. traversing the historical technology knowledge base All historical feature dimension vectors Find the weighted Euclidean distance between the two. The weighted Euclidean distance control formula is defined as follows:

[0052] In the formula, Indicates the current valve entity characteristics and the first The weighted Euclidean distance between historical feature records, which is dimensionless; , , This represents the preset weighting coefficients, which satisfy... In this embodiment, the value is fixed. The weighting of defects in valve casting wall thickness is set based on experience. Denotes the Frobenius norm; The Fréchet distance-based flow channel topology difference mapping function, used to quantify the morphological deviation between two spatial curves, is defined as follows:

[0053] in and For two parameterized space curves, Let be a parametric transformation function whose range is continuous and non-decreasing in the interval [0,1]. These are the curve parameters. During discretization, each flow channel curve is uniformly sampled, with the number of sampling points ranging from 20 to 200. The discretization step length and voxel edge length are used. Maintain consistency. Represents the L2 norm; Indicates the index number of historical process records.

[0054] The optimization processor obtains the weighted Euclidean distance. Specific index symbol for reaching the minimum value At this point, the target historical process record is locked. The six static control variables bound to the latter half of this record are extracted. These six values ​​serve as the benchmark for system optimization, are mapped and assigned by the system, and the formal output is defined as the initial process parameter vector. Initial process parameter vector The internal data deconstruction is expressed as:

[0055] In the formula, This represents the initial pouring temperature scalar value, in °C. This represents the initial pouring speed as a scalar value, in kg / s. This represents the initial cooling gas flow rate as a scalar, in L / min. This represents the initial cooldown time scalar, in seconds. This represents the initial vibration frequency scalar, with the unit being Hz. The initial amplitude scalar is represented in mm; the superscript T indicates the matrix transpose operation.

[0056] The aforementioned generated initial process parameter vector It serves as a hub connecting static historical experience with dynamic mathematical optimization, transmitting data losslessly to the next computational node.

[0057] Step S3: Using the initial process parameter vector as the input node, perform finite element simulation based on the central composite design method, extract the objective function response set, and fit to generate a response surface prediction model.

[0058] This step is used to construct a continuous mathematical mapping proxy space between physical fields and static control parameters, transforming the cost of physical trial and error into low-cost computational iteration.

[0059] The initial process parameter vector received in step S2 As the origin of basic design.

[0060] Start the central composite design logic matrix generator. For the initial process parameter vector... It includes 6 independent input parameters, namely Each positive and negative perturbation boundary is set to generate an experimental matrix with a spatial star-shaped distribution. .

[0061] This embodiment employs a standard six-factor central composite design, including cubic points, There are 77 experimental nodes, including one axial point and one center point. The positive and negative disturbance ratios of the parameters are selected according to the casting accuracy requirements. The value is 10%-20% for normal working conditions and 8%-12% for high-precision casting conditions. In this embodiment, it is 15%.

[0062] For the experimental matrix Each parameter combination node in the process calls the finite element thermodynamics solver to perform virtual casting simulation calculations. The boundary conditions of the 3D model for the finite element simulation are strictly imported from the boundary voxel set including the inner and outer surfaces generated in step S1. .

[0063] The simulation parameters are set as follows: The transient heat conduction time step is 0.1–1.0 s, the fluid filling time step is 0.05–0.5 s, and the solution convergence threshold is set to... Mesh size and voxel side length Linked matching. The finite element thermodynamics solver solves the three-dimensional viscous incompressible Navier-Stokes equations and Fourier transient thermal conduction control equations of molten metal within discrete time steps, calculating and outputting the spatiotemporal physical fields of the entire process from molten metal filling to solidification and cooling. This includes the temperature time history and stress-strain distribution on the three-dimensional mesh nodes.

[0064] From the spacetime physical field The coordinates of the boundary points of the thermal stress deformation field induced by contraction resistance and non-uniform temperature gradient are extracted. A global rigid registration method is used to connect the boundary point coordinates of the thermal stress deformation field with the lossless 3D point cloud data transferred from step S1. Alignment is performed, with a registration error threshold of ≤0.05mm. After registration, 3D Boolean subtraction and spatial registration operations are performed to extract the normal offset distance at each corresponding spatial node. The absolute value of the average volume deviation is then calculated by integration and defined as a scalar form of dimensional deviation. Dimensional deviation It serves as the first element of the objective function response set.

[0065] At the same time, from the spacetime physical field Extract the span of the latent heat release rate during solidification. Track the boundary voxel set. The solidification time difference between different coordinate nodes inside the liquidus line and the solidus line is considered when the temperature drops. Substituting this solidification time difference into the heterogeneous nucleation prediction equation, the global grain size standard deviation distribution is calculated. This heterogeneous nucleation prediction equation is based on classical nucleation theory, and its expression is:

[0066] In the formula, Nucleation density; For potential nucleation particle density, it can be taken as [value missing] for heat-resistant steel. ; The critical nucleation work can be obtained by consulting or calculating based on the material composition. Boltzmann constant, standard value ; The absolute temperature is used. Combining the solidification time difference and cooling rate, the grain size at each node is calculated, and its standard deviation distribution is statistically analyzed. The output of the grain size standard deviation distribution is defined as the microstructure uniformity in scalar form. Tissue uniformity It serves as the second element of the objective function response set.

[0067] dimensional deviation With tissue uniformity Assembled into tensor form, defined as the objective function response set At this point, the objective function response set With the experimental matrix This forms a discrete mapping pair between input and output.

[0068] A second-order polynomial regression algorithm is introduced, using the parameter nodes of the external input as independent variables to extract the output objective function response set. As the dependent variable, a continuous analytical form of the response surface prediction model is fitted. By using least squares estimation to determine all regression coefficients, a complete and usable predictive model can be obtained. (Response surface prediction model) The core fitting equation is expressed as:

[0069] In the formula, This represents the set of objective function responses to the predicted output; and The input consists of six independent parameter variables, namely, pouring temperature, pouring speed, cooling gas flow rate, cooling time, vibration frequency, and amplitude. This represents the intercept constant term in the regression model; This represents the partial regression coefficients of the first-order linear main effects; This represents the regression coefficient of the second-order pure quadratic nonlinear effect; It represents the partial regression coefficient of the cross-coupling interaction effect between two different input parameters.

[0070] Response surface prediction model and the extracted objective function response set Both are used together as lossless output and are transmitted forward to step S4.

[0071] Step S4: Call the response surface prediction model, perform sequential quadratic programming iteration guided by the objective function response set, and calculate and output the optimal process parameter vector.

[0072] The controller simultaneously loads the response surface prediction model output from step S3. The initial process parameter vector established in step S2 The initial process parameter vector This serves as the initial guess for the zero point in the iterative computation. To unify the search direction of the bi-objective optimization, the objective function response set... Internal dimensional deviations With tissue uniformity We perform weighted integration to construct a single comprehensive objective cost function. :

[0073] In the formula, This represents the value of the comprehensive objective cost function, which is dimensionless. This represents the predicted dimensional deviation, in mm. This represents the dimensional deviation normalization factor, with a value of 2.0 mm, corresponding to the maximum allowable dimensional deviation of castings in the industry. This represents the predicted value for tissue evenness, and is dimensionless. This represents the tissue homogeneity normalization factor, with a value of 1.0, corresponding to the baseline extreme value of tissue homogeneity. and Let be the weighting coefficient, satisfying In this embodiment, the value is fixed. Prioritize ensuring dimensional accuracy.

[0074] Start the sequential quadratic programming solver engine. In each internal iteration loop... In the middle, the solver engine is at the current process parameter vector node At this point, for the comprehensive objective cost function Perform a second-order Taylor expansion to construct an approximate Hessian matrix for the Lagrange function. Solve the quadratic programming subproblem to generate the optimization search step size direction vector. A line search strategy is used to determine the step size factor. The formula for updating the process parameter vector node is:

[0075] In the formula, and They represent the first and The process parameter vector at the next iteration is a 6-dimensional column vector; Indicates the first The step size factor of the iteration is a scalar; Indicates the first The search direction vector for each iteration is a 6-dimensional column vector. The upper limit of the iterations is set to 500. If convergence is not achieved after exceeding this limit, the process is forcibly terminated and the current optimal solution is returned.

[0076] When the combined objective cost function is used in three consecutive iterations The absolute value of the difference change is strictly lower than the preset convergence threshold. When the convergence threshold is reached, the iteration stops. The general range of values ​​is This embodiment has a fixed setting. To balance optimization accuracy and computational cost, the parameter vector nodes corresponding to the stopping moment are extracted and determined as the optimal process parameter vector. .

[0077] Optimal process parameter vector The mathematical structure corresponds precisely to:

[0078] In the formula, It is a one-dimensional column vector; This represents the absolute value of the optimal pouring temperature, in °C. This represents the absolute value of the optimal pouring speed, in kg / s. This represents the absolute amount of the optimal cooling gas flow rate, in L / min. This represents the absolute value of the optimal cooling time, in seconds. This represents the absolute value of the optimal vibration frequency, in Hz. This represents the absolute value of the optimal amplitude, in mm. Optimal process parameter vector. It will completely jump out of the digital domain and enter the physical domain, becoming the sole input source for instruction execution step S5.

[0079] Step S5: Drive the melting equipment, casting equipment, cooling system, and vibration demolding system according to the optimal process parameter vector to generate a physical casting in the forming process state. This is used to forcibly map the optimal parameters calculated in mathematical space to the servo control input voltage of heavy industrial machinery.

[0080] The equipment ranges for each process parameter are as follows: casting temperature range 1400-1650℃, casting speed range 5-15kg / s, cooling gas flow rate range 0~500L / min, vibration frequency range 20~80Hz, and amplitude range 0.5~3.0mm.

[0081] Receive the optimal process parameter vector output in step S4 .

[0082] Central programmable logic controller (PLC) parses the optimal process parameter vector The absolute value of the optimal pouring temperature The output power control pulse width modulation duty cycle signal is sent to the intermediate frequency induction melting furnace. The temperature of the molten metal inside the intermediate frequency induction melting furnace strictly approaches and stabilizes at the optimal pouring temperature. The allowable error window is .

[0083] The central programmable logic controller analyzes the absolute value of the optimal pouring speed. The digital quantity is converted into the control voltage for the servo hydraulic pump. The servo hydraulic pump drives the vacuum casting tube, preheating it to the optimal pouring temperature. The liquid alloy is injected into the inner cavity of the ceramic mold at a constant mass flow rate, which is exactly equal to the absolute value of the optimal pouring rate. .

[0084] After the molten metal completes its full-load filling process, the system immediately transitions to the non-uniform directional solidification intervention stage. The central programmable logic controller (PLC) analyzes the optimal absolute flow rate of the cooling gas. Absolute quantity of optimal cooling time The space-addressable microchannel array cooling mechanism, arranged around the outer boundary of the ceramic housing, is activated. The electromagnetic proportional servo valve operates based on the optimal absolute flow rate of the cooling gas. The flow rate of nitrogen gas entering the microchannel is adjusted. The high-speed nitrogen gas flow induces forced convection cooling boundary conditions in the backflow region and high-curvature surface of the ceramic shell. The duration of this cooling behavior is strictly equal to the absolute value of the optimal cooling time. .

[0085] Once the temperature is below the solidus and the crystal framework is established, the system enters the stress release phase after the shell is removed. The central programmable logic controller (PLC) analyzes the absolute value of the optimal vibrational frequency. With the absolute value of the optimal amplitude The piezoelectric ceramic vibration table receives an alternating sinusoidal electrical signal synthesized from these two parameters, which excites the ceramic shell and applies shear mechanical work to the solidified shell interface, forcing the ceramic shell to brittlely fracture and peel off.

[0086] After the aforementioned four control processes are completed, the digital parameters originally stored in memory are fully instantiated. At this point, the output physical entity is no longer a set of data, but a three-dimensional heavy-duty component with a temperature distribution field and an internal thermal strain distribution field, which is defined as a physical casting in the forming process state. Physical castings in the forming process. It is a continuously changing spatiotemporal entity that constitutes the physical target detected by the sensor array in step S6.

[0087] Step S6: A sensor array is used to collect real-time temperature and displacement data of the physical casting during the forming process. This data is then input into a response surface prediction model to calculate parameter fine-tuning increments. These increments are then superimposed onto the optimal process parameter vector for dynamic feedback control until the vibration acceleration signal collected by the sensors falls below a preset safety threshold, at which point demolding is considered complete. This step is used to counteract transient random environmental thermal disturbances in the physical casting workshop that cannot be perfectly modeled in step S3, such as ambient temperature drift and slight differences in metal batch composition, preventing the accumulation of open-loop execution deviations that could lead to thermal cracking and scrap.

[0088] Step S6 is a closed-loop control function that executes continuously on the time axis, forcibly exiting only when demolding is completely finished. The sensor physical sampling time window is defined as follows: Its general value range is 0.5~2s, and it is selected according to the solidification rate of the casting. In this embodiment, it is fixed. The execution cycle of dynamic feedback control is an integer multiple of this sampling time window.

[0089] At any physical sampling time k, the infrared thermal imaging sensor array (temperature measurement range 200-1600℃) is aligned with the physical casting in the forming process. The surface blackbody radiation energy is captured, and after solving the inverse Planck radiation law, the discrete matrix of the two-dimensional surface temperature field is output, which is defined as the real-time temperature data. .

[0090] Simultaneously, a non-contact laser interferometric displacement sensor (displacement range 0-5mm, resolution ≤0.01mm) is applied to the physical casting in the forming process. The probe laser beam emitted from the endpoint of the easily distorted cantilever beam is analyzed as an instantaneous linear displacement shift, which is defined as real-time displacement data. Real-time temperature data With real-time displacement data Together, they characterize the true solidification deviation state of the physical casting at time k.

[0091] The data link then points back to the response surface prediction model that was losslessly saved in step S3. At this stage, the system is equipped with a nonlinear model predictive controller. The controller receives real-time temperature data. With real-time displacement data As initial state constraints, the response surface prediction model is used. As an internal feedforward rolling prediction function, the predictive controller solves for the future. Within each prediction step, the model predicts the control quadratic cost function. The optimal control sequence when the minimum value is reached:

[0092] In the formula, This represents the dimensionless value of the quadratic cost function used to predict the control model. This represents the prediction step size range constant. Based on the aforementioned thermal disturbance period coverage constraint and model mismatch error suppression requirements, its value range is limited to 15~30. In this embodiment, the measured thermal disturbance dominant frequency for this nuclear power plant valve is 22s. ; This represents the control step size range constant, based on the controllability criterion of discrete systems and The stability constraint is limited to a value range of 3 to 8. In this embodiment, the value is set to 8 while taking into account the real-time performance of the computation. ; This represents the result predicted by the response surface prediction model at time k. The objective function state variables at time t; This represents the zero-deviation target value for defect-free ideal forming, which is fixed at 0. This represents the sequence of incremental adjustments to the vector of process parameters that need to be applied. Let the positive definite diagonal weight matrix of the predicted state penalty be denoted as . Matrix, values ​​taken in this embodiment ; The control action amplitude suppression positive definite diagonal weight matrix is, for example, Matrix, values ​​in this embodiment ; Denotes the square of the weighted Euclidean norm, i.e. .

[0093] After the solver completes its calculations, it extracts the first element of the optimal control increment sequence, defines it as the parameter fine-tuning increment ΔM(k), and outputs it. The parameter fine-tuning increment ΔM(k) is immediately fed forward and superimposed onto the optimal process parameter vector output in step S4 within the current execution cycle. Above, synthesize new real-time drive signals. :

[0094] In the formula, This represents the final physical correction command amount actually issued to the servo hydraulic pump, gas servo proportional valve, and piezoelectric vibration table at the current moment; This represents the global baseline control vector passed from step S4; This represents a small parameter correction amount to suppress the current thermal disturbance compensation output.

[0095] When the final physical correction command is executed by the servo mechanism, the physical casting is in the forming process state. This will produce a corrected deformation of the physical field. At this point, time has progressed to... During the cycle, the sensor array collects new real-time temperature and displacement data again, and the data stream loops back into the controller, forming a continuously operating data-closed defense line. The control cycle terminates when the vibration acceleration signal detected by the sensor array falls below a preset safety threshold (this threshold is a vibration acceleration safety threshold, ranging from 0.01 to 0.05g, indicating that the shell has been completely peeled off).

[0096] After demolding, the structural characteristic parameters of the qualified valve bodies in this batch will be collected. With the optimal process parameter vector As a new record entry, it is automatically written back to the valve body structure feature parameter library and the historical process knowledge base to complete the knowledge iteration.

[0097] Abnormal operating condition handling rules: To ensure stable operation in industrial settings, the system has the following pre-defined abnormal handling logic: When the sensor signal is interrupted, the control system continues to operate by fine-tuning the parameters of the previous effective cycle and triggering an alarm. When the sequential quadratic programming or Bayesian optimization iterative calculation diverges or exceeds the maximum number of iterations, it automatically reverts to the initial process parameter vector or the qualified solution of the previous iteration. When the number of missing valid points in the 3D point cloud exceeds 15% of the total number of original complete point clouds, the system prompts to re-execute the 3D scanning operation. If the missing percentage does not exceed 15%, the local completion algorithm is started to interpolate and estimate the missing areas.

[0098] Example 2: In practical applications, to meet the casting environment of valve bodies with extremely complex flow channels that are non-convex and highly nonlinear, this example provides an alternative path based on Example 1. This example shares the basic physical execution boundaries of steps S1, S2, and S5 with Example 1, but performs algorithm reconstruction at the model abstraction layer of steps S3 and S4 and the online feedback mode of step S6, broadening the optimization dimension while ensuring that the data transmission closed loop remains unchanged.

[0099] Alternative step S3: Using the initial process parameter vector as input nodes, reconstruct the physical mapping based on the long short-term memory network to generate a deep learning predictive agent model.

[0100] Receive the initial process parameter vector transmitted without loss in step S2. and the set of structural feature parameters extracted in step S1 Instead of relying on the central composite design and finite element simulation, it directly accesses and traverses the historical process knowledge base. All prior time series data records of successes and failures.

[0101] A deep learning predictive agent model based on a long short-term memory (LSTM) architecture is constructed. The deep learning predictive agent model consists of an input layer, five hidden feedforward neural layers, and a single output layer. Of the five hidden layers, the first three are LSTM layers with 128, 64, and 32 neurons respectively, and the activation function is the hyperbolic tangent function; the last two are fully connected layers with 16 and 8 neurons respectively, and the activation function is the linear rectified function. The structural feature parameters are then set... Flattened into a one-dimensional feature vector according to row-major rules, and compared with the initial process parameter vector. Perform serial concatenation; the resulting vector has a dimension of [missing value]. After flattening, the vector dimensions are validated to remove any outlier vectors. This concatenated vector then serves as the input node vector for the deep learning predictive proxy model.

[0102] Historical Crafts Knowledge Base The absolute dimensional deviation values ​​and the microstructure grain variation coefficients obtained from the destructive testing after physical forming are recorded as absolute target values ​​for supervised learning.

[0103] The training process is as follows: Historical data was randomly divided into training and test sets in an 8:2 ratio. A temporal backpropagation gradient descent algorithm with a batch size of 32 was used to update the gating weight matrix of neurons within the model. The specific implementation method is as follows: using the mean squared error between the predicted values ​​and the true values ​​in the training set as the loss function, the gradient is solved layer by layer in reverse, and the network weights and bias parameters are adaptively updated until the loss function converges. The model training uses a fixed learning rate of 0.001 and a maximum number of iterations of 1000 steps, with a loss function convergence threshold. The mathematical update logic of the forget gate in a deep learning predictive agent model is expressed as follows:

[0104] In the formula, This represents the output tensor of the forgetting gated state at time step t, with a value ranging from 0 to 1. The Sigmoid non-linear activation function is represented by the formula: ; This represents the trainable weight matrix tensor of the forget gate, initialized as a random normal distribution matrix; express Time-varying hidden state propagation in a neural network; This represents the concatenated joint state tensor input at time t; This represents the linear bias constant vector of the forget gate, initialized to zero. This represents a vector concatenation operation, where the concatenated vector has the dimension of the hidden state plus the dimension of the input.

[0105] When the mean squared error loss function value calculated by backpropagation is lower than the preset network convergence threshold, all trainable weight matrix tensors are frozen. At this point, the solidified deep learning predictive surrogate model is rigorously defined and output as... In this network architecture, the input joint state tensor will propagate forward, and the direct output is defined as the predicted target set. It also includes two elements: scalar size deviation and tissue uniformity. Deep learning predictive surrogate model. Together with the prediction target set Proceed to the next execution step.

[0106] Alternative step S4: Invoke the deep learning prediction agent model, perform Bayesian global optimization guided by the prediction target set, and calculate the output optimal process parameter vector.

[0107] Bayesian optimization uses a Gaussian process as the basis for uncertainty inference. The Gaussian process employs a radial basis function kernel, the expression of which is: ,in For signal variance, The length scale parameters can all be learned from the data through maximum likelihood estimation. The deep learning predictive surrogate model output from the alternative solution in step S3 is then used. As a priori evaluation environment, the initial process parameter vector Used as the starting anchor point for Bayesian search.

[0108] The acquisition function is constructed using the expectation boosting criterion. This criterion guides the optimizer in exploring and utilizing the parameter space. In the nth probe loop, the acquisition function is calculated as follows:

[0109] The operator Z is defined as follows:

[0110] In the formula, Represents six-dimensional process parameter coordinates Expected improvement value at the location; Represents a vector node for the process parameters to be explored; This represents the optimal objective function value in the current iteration; and Let represent the mean and standard deviation of the Gaussian process's posterior prediction of the objective function at the nth iteration, respectively; This indicates the exploration of equilibrium hyperparameters, which are fixed at 0.01 in this embodiment; and Let represent the cumulative distribution function and probability density function of the standard normal distribution, respectively; This represents the normalization evaluation operator. The maximum number of iterations is set to 200 steps. If the problem still does not converge after reaching the maximum, the current optimal solution is taken.

[0111] Iterative computer search increases the absolute value of expected utility. Maximize the new parameter node coordinates and input them into the deep learning predictive surrogate model. Obtain a new set of prediction targets Update the posterior probability distribution of the Gaussian process until the preset maximum exploration step count of 200 steps is reached. Extract and determine the coordinates of the parameter nodes that minimize the final objective cost function as the optimal process parameter vector. This optimal process parameter vector Equivalent replacement of the optimal process parameter vector in Example 1 The process is directly transferred to step S5 to execute the heavy industry physical conversion action, generating a physical casting in the forming process state. .

[0112] Alternative step S6: Use a visual temperature measurement array and acoustic emission probe for online closed-loop fine-tuning.

[0113] To overcome the mirror contamination interference of the laser interferometer element caused by the intense thermal radiation environment, this alternative step modifies the sensor mode. A dual-color temperature measurement vision system and a high-frequency piezoelectric emission sensor are arranged within the physical execution space.

[0114] A dual-color temperature-sensing vision system captures physical castings during the forming process. The radiance of two adjacent narrow bands on the surface, with wavelengths set as follows: , The emissivity drift error is eliminated by Planck's law dual-band colorimetric method, and an absolute two-dimensional temperature array is output, defined as a real-time temperature image. A high-frequency piezoelectric acoustic emission sensor is attached to the metal support surface of the spatially addressed microchannel array cooling mechanism to capture the physical casting during the forming process. The high-frequency elastic wave released instantaneously by the microscopic cracking of the internal lattice under contraction strain, with a center frequency of 150kHz, is extracted by fast Fourier transform and its energy integral area is defined as the real-time acoustic feature. .

[0115] Real-time temperature image With real-time acoustic characteristics Data fusion is performed using a serial vector concatenation method, with the concatenation dimension matching the model input port. The fused vector is then injected into the nonlinear model predictive controller. The internal feedforward rolling prediction function of the controller is replaced with a deep learning predictive surrogate model. The model predicts and controls a quadratic cost function by minimizing it, and then outputs new parameters for fine-tuning the increment. Fine-tune the parameter increments. Immediately superimposed onto the optimal process parameter vector This completes the seamless closed-loop recycling of underlying data.

[0116] After demolding, data write-back and exception handling are performed in the same way as in Example 1.

[0117] In specific cases, the target of the application should be defined as follows: The solid valve body of the main steam isolation valve in a third-generation nuclear power plant. This solid valve body is designed to withstand a pressure of 17.5 MPa and an operating temperature of 360℃. It is made of ZG10MnMoNi5-4 low-alloy high-strength heat-resistant steel. The dimensions of the solid body are as follows: length, width, and height. Due to fluid dynamics requirements, the internal structure features an extremely irregular S-shaped constriction nozzle topology. The thickest part of the valve inlet flange has a wall thickness of 185mm, while the thinnest part of the internal guide ribs has a wall thickness of only 12mm. This extreme wall thickness difference of more than 15 times completely exceeds the molding capability limit of existing open-loop static processes.

[0118] The 3D optical scanning equipment was activated, and the raster projection period was set to 8 frames per second. After 45 spatial pose transformations, the binocular camera captured a total of... Each discrete spatial coordinate point. After running the ICP registration and denoising algorithm for 120 seconds, clean 3D point cloud data is output. 3D point cloud data It uses 685MB of memory.

[0119] Enter the octree voxelization partitioning space. Set the voxel edge length constant. 3D point cloud data Mapping generation Individual voxel units, of which those classified as boundary voxel sets Total peripheral voxels indivual.

[0120] The core computing engine extracts wall thickness parameters. : Center coordinates of the thick area of ​​the flange A Euclidean distance minimization search was performed, and the nearest boundary voxel distance was 92.5 mm. Multiplying this by two, the local wall thickness at this point was output as 185.0 mm. The coordinates of the center point of the guide rib were then determined. Performing isomorphic calculations, the output local wall thickness at this point is 12.3 mm. All local wall thickness values ​​form a tensor with dimension [missing value]. Wall thickness parameter matrix .

[0121] The calculation engine extracts flow channel parameters. The skeleton extraction algorithm identified a total of 1540 central topological points in the S-shaped flow channel. The friction derivative of the cross-sectional area decreased drastically from +0.85 to -1.24. This array of 1540 points constitutes the flow channel parameters. .

[0122] The calculation engine extracts thin-walled parameters. : Preset wall thickness safety threshold For the wall thickness parameter matrix Perform binarization mask filtering to capture common data. The coordinates of each spatial point have a thickness less than 15.0 mm, and this portion of the coordinate space occupies 4.2% of the total surface area. This set of mask points, together with the percentage scalar, constitutes the thin-wall parameter. .

[0123] Finally, the wall thickness parameter matrix is... Flow channel parameters Thin-wall parameters Encapsulated in memory, defined as a unique set of structural characteristic parameters. Set of structural feature parameters All generated data is written to the SSD temporary storage area and packaged there. .

[0124] Loading structural feature parameter set Connecting to a historical technology knowledge base It includes 14,520 valid historical records.

[0125] The weighted Euclidean distance calculator is set with the following weighting coefficients: A line-by-line comparison operation occurs. For record number 8972, since its historical wall thickness tensor mean and spatial volume covariance matrix are similar to the current set of structural characteristic parameters... It exhibits a high degree of geometric homomorphism, weighted Euclidean distance The calculated score is the absolute minimum of 0.0415. Lock record number 8972 and extract the six process parameters from the latter half of its value set: Historical casting temperature Historical casting speed Historical cooling gas flow rate Historical cooldown time Historical vibration frequency Historical amplitude .

[0126] The system assigns the above six numerical values ​​without modification and instantiates and outputs the initial process parameter vector in the memory stack. The entire data link fully inherits the computation results from the previous node.

[0127] Read in the initial process parameter vector The central composite design logic matrix generator expands by 15% positively and negatively across six dimensions to construct a design matrix comprising 77 experimental nodes. .

[0128] For the experimental matrix Node number 1 in the diagram has the following settings: The thermohydrodynamic solver was invoked to perform finite element analysis for 2 hours, outputting the spatial and temporal physical fields. The predicted volume of the thick, loosened area is: The deformation deflection at the thin-walled section is 1.8 mm. After three-dimensional Boolean subtraction, the output objective function response set is extracted. Including dimensional deviations tissue uniformity The remaining 76 nodes execute in parallel, resulting in a complete set of 77 pairs of input-output scatter points.

[0129] The least squares method is used to perform second-order polynomial space surface fitting operations to calculate the specific regression model. This addresses dimensional deviations. Dimensional response surface prediction model For the specific expression, please refer to the core fitting equation mentioned above. Here is a numerical example of the truncation term:

[0130] In the formula, This indicates that the pouring temperature parameter should be substituted. This indicates that the pouring speed parameter is substituted in; This indicates that the cooling gas flow rate parameter is substituted. This analytical expression model is passed to the optimization engine in step S4 as a continuous prediction interface, and step S3 strictly completes the data dimensionality upgrade and transformation tasks.

[0131] Step S4: Data Evolution – Sequential Quadratic Programming Algorithm Initiated. The initial starting coordinate anchor point is located at… .

[0132] Comprehensive objective cost function The weighting coefficient is set to .

[0133] First iteration, Calculating the eigenvalues ​​and gradient vector of the Hessian matrix at this coordinate point reveals that the gradient of the cooling parameters decreases most rapidly. A quadratic programming subproblem generates the step direction, updating the parameter coordinates to... .

[0134] Second to forty-second iterations: The algorithm continuously uses the response surface prediction model Verify the target penalty value of the feedforward coordinates.

[0135] Iterations 43 to 45: Synthetic objective cost function The values ​​remained at 0.15881, 0.15875, and 0.15872, representing three consecutive absolute differences. , All are lower than Determine the threshold.

[0136] Extract the parameter vector node corresponding to the termination of the iteration and determine it as the optimal process parameter vector. This indicates that for this nuclear power valve model, the theoretical physical extreme conditions require cooler cast metal and more sustained and intense forced gas heat dissipation to suppress the thermal throttling effect caused by abrupt changes in wall thickness.

[0137] Step S5: Data Evolution – Optimal Process Parameter Vector The digital controller is sent to the heavy-duty electrical drive cabinet.

[0138] First action instruction: The programmable logic controller sends the target constant temperature threshold to the medium frequency induction melting furnace. The power device heats the ZG10MnMoNi5-4 steel liquid through pulse width modulation and maintains it at a constant molten state of 1548℃.

[0139] Second action command: The servo hydraulic pump valve locks the vacuum casting flow rate at 7.8 kg / s according to the received voltage analog signal, and the liquid metal smoothly passes over the thin-walled area of ​​the guide rib.

[0140] The third action command: Once the molten metal level reaches the apex of the riser in the mold shell, all the electromagnetic proportional servo valves of the 320 spatially addressed microchannel array cooling mechanisms distributed around the mold shell open. The nitrogen compressor outputs a high-speed airflow at a given rate of 425 L / min, forcibly impacting the thick outer ceramic shell of the flange. The forced heat exchange time is precisely cut off by a timer after reaching 615 seconds.

[0141] Fourth action command: When the shell temperature crosses 200°C below the solidus line of the material, the bottom piezoelectric ceramic servo stage excites a stable 68Hz sinusoidal mechanical wave, and the displacement extreme value is locked at 2.1mm by the laser limiter. The mechanical stress wave oscillates back and forth inside the shell, completely crushing the ceramic material attached to the surface.

[0142] The aforementioned physical parameters work together to successfully transform a set of numerical variables into an objectively existing physical component with a real three-dimensional heat transfer field, which is defined as a physical casting in the process of forming. .

[0143] The molding process has entered its most severe period of environmental disturbance. The current clock cycle is set to... It is currently in the third cooling stage of solidification phase transition. Due to the sudden opening of the south-facing double doors of the workshop, a gust of cold ambient air intruded, disrupting the original heat dissipation boundary.

[0144] An infrared thermal imaging sensor array is aligned with a physical casting in the forming process. The scan revealed that the temperature of the two-dimensional surface nodes on the windward side was 32°C lower than that on the leeward side, forming an unexpected hotspot gradient. The solution output was defined as the real-time temperature slice. Simultaneously, the laser interferometric displacement sensor detected a slight warping at the free end of the flange cantilever beam, with a linear displacement offset of 0.18 mm. The calculated output was defined as real-time displacement data. .

[0145] The nonlinear model predictive controller is activated. Real-time temperature data is then transferred. With real-time displacement data Forced injection of the response surface prediction model after fitting in step S3 Perform feedforward rolling prediction.

[0146] The prediction results show that without intervention, in At that point, the hot crack is about to break through the critical point of intergranular bonding force.

[0147] The model predictive control quadratic cost function solver initiates a matrix minimum search. For the current state, it calculates the compensation value for the six-dimensional process vector, and the output is defined as the parameter fine-tuning increment. This indicates that the best corrective decision made by the controller is to keep other parameters unchanged, but immediately reduce the forced cooling gas flow rate to reduce the excessively rapid cooling rate of the windward side and smooth out the temperature gradient.

[0148] The underlying actuator immediately sends the optimal process parameter vector. With parameter fine-tuning increment Superimpose and synthesize the final physical correction command quantity The electromagnetic proportional servo valve instantly closed from 425 L / min to 340 L / min.

[0149] After closed-loop cancellation, At that moment, the warping displacement of the flange cantilever beam returned to within the 0.02mm tolerance zone. Thanks to its rigorous data transmission logic, the system not only eliminated empiricism but also successfully captured and suppressed external random thermal disturbances, eliminating the physical possibility of hot cracking defects and verifying the industrial practicality of this technical architecture's absolute closed-loop design and high-precision molding.

[0150] Example 3: This example aims to solve the physical failure problem caused by the aforementioned single scalar cooling when dealing with castings with extreme thickness differences. This example extends the time-dimensional closed loop in Example 1 into a spatiotemporal dual matrix closed loop.

[0151] Clustering partitioning of the spatially addressed cooling matrix (corresponding to step S5 for further refinement): The central programmable logic controller no longer directly issues a single cooling flow rate scalar, but instead calls the K-means clustering algorithm to process the wall thickness parameter matrix extracted in step S1. Perform spatial discretization calculations. The number of clusters N is determined based on the ratio of the maximum to the minimum wall thickness of the casting. Adaptive determination: when When, N is 4; when In the first case, N is 3; otherwise, N is 2. In this embodiment... Therefore, the number of clusters N is fixed at 4. The K-means clustering iteration terminates when the change in cluster centers is less than 0.1 mm.

[0152] The clustering objective function is defined as:

[0153] In the formula, This represents the clustering objective function value, in units of... ; This represents the wall thickness value at the i-th spatial coordinate point, in mm; This represents the center wall thickness of the j-th cluster, which is the arithmetic mean of the wall thicknesses of all points in that cluster, in mm. The number of coordinate points included in the j-th cluster is dimensionless; 4 represents the number of clusters.

[0154] After the algorithm converges, the surface of the ceramic shell is strictly divided into four physically isolated independent cooling zones: Zone A (Severe Hot Spot Zone): Wall thickness > 100mm, encased in imported flange.

[0155] Zone B (Main Transition Zone): 50mm < wall thickness ≤ 100mm.

[0156] Zone C (Standard Pressure Zone): 15mm < wall thickness ≤ 50mm.

[0157] Zone D (ultra-thin wall zone): wall thickness ≤ 15mm, encased in internal S-shaped flow guide ribs.

[0158] The absolute value of the optimal cooling gas flow rate calculated in step S4, for example, 425 L / min, is orthogonally decomposed based on the mass heat capacity ratio of each zone to generate a four-dimensional initial cooling vector. The orthogonal decomposition formula is: the traffic of the j-th partition. ,in The thermal capacity of the metal encased in the ceramic shell of the j-th partition is given. The four scalar values ​​correspond to the electromagnetic proportional servo valves driving the four sets of spatially addressed microchannel array cooling mechanisms encased in partitions A to D.

[0159] Extraction of the real-time spatial thermal gradient tensor (corresponding to step S6 in detail): Upon entering the cooling stage, at any sampling time k, after the system captures the discrete matrix of the two-dimensional temperature field on the surface, it immediately solves for the spatial partial derivatives of the temperature field in a three-dimensional spatial coordinate system that matches the voxel mesh using the central difference method, thereby synthesizing the real-time spatial thermal gradient tensor. The expression for calculating partial derivatives using the central difference method is:

[0160] In the formula, The spatial gradient tensor of the surface temperature field is expressed in °C / cm. , , Let X, Y, and Z represent the partial derivatives of temperature along the X, Y, and Z coordinate axes, respectively, and let Z represent the spatial difference step size and voxel edge length. Consistent.

[0161] exist At that time, the calculation engine detected the magnitude of the tangential temperature gradient at the boundary coordinate domain between partition A (flange) and partition D (internal rib). The temperature reached 85℃ / cm, far exceeding the preset critical phase transformation thermal stress gradient threshold of 50℃ / cm. This threshold was calculated using a phase transformation thermal stress model based on the material's high-temperature fracture stress and thermal expansion coefficient. For carbon steel, the threshold is typically 45-55℃ / cm, and for stainless steel, it is typically 50-60℃ / cm. This coordinate domain was immediately mathematically isolated, and a local chilling coordinate set was extracted and generated. .

[0162] Minimization of MPC control under tensor constraints (corresponding to dynamic feedback calculation): The controller receives a local chill coordinate set. At this point, the cooling parameter in the parameter fine-tuning increment is no longer a single value, but requires the output of a four-dimensional space correction control vector. .

[0163] Six-dimensional optimal process parameter vector The third part is the cooling gas flow rate component; replace this component with... This forms the corrected optimal process parameter vector.

[0164] Solving the quadratic cost function of model predictive control At this time, the system triggers a spatial penalty mechanism. For partition D (the extremely thin-walled region) that encloses the chilled coordinate set, the system applies a weighting matrix to suppress the amplitude of the control action. The corresponding diagonal elements are injected with dynamic decay penalty weights. The adaptive calculation formula for the penalty weight is as follows:

[0165] in, This represents the factor by which the local thermal gradient exceeds the limit. ,and ,corresponding The range of values ​​is In this embodiment, because the gradient exceedance factor reaches 1.7 times, the calculated value is... Rounded down to .

[0166] The extremely large mathematical penalty value forces the quadratic programming solver to instantly reduce the computational load during iteration. The output is calculated to be -10 L / min, and at the same time... The output is +35L / min.

[0167] The underlying actuator receives a four-dimensional spatially corrected control vector and completes the spatially differentiated reconfiguration of multiple valves in just 0.2 seconds. In the next cycle of verification, due to the accelerated cooling of the thick wall and the cessation of forced heat dissipation of the thin wall, the temperature gradient modulus at the interface rapidly drops to the safe range of 38℃ / cm.

[0168] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0169] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for precision forming of valve body by investment casting using a parameter iteration algorithm, characterized in that, Includes the following steps: Step S1: Use a 3D scanning device to acquire 3D point cloud data of the physical valve body, perform geometric analysis on the 3D point cloud data to extract a set of structural feature parameters, and store the set of structural feature parameters in the valve body structural feature parameter library. Step S2: Based on the set of structural feature parameters in the valve body structure feature parameter library, and combined with the historical process knowledge base, perform similarity matching calculation to generate an initial process parameter vector; Step S3: Using the initial process parameter vector as the input node, perform finite element simulation based on the central composite design method, extract the objective function response set, and fit to generate a response surface prediction model; wherein, the boundary conditions of the finite element simulation are set based on the boundary voxel set obtained in step S1. Step S4: Invoke the response surface prediction model and perform sequential quadratic programming iteration guided by the objective function response set to calculate and output the optimal process parameter vector; Step S5: Drive the melting equipment, casting equipment, cooling system and vibration demolding system according to the optimal process parameter vector to generate a physical casting in the forming process state; Step S6: A sensor array is used to collect real-time temperature and displacement data of the physical casting in the forming process. The real-time temperature and displacement data are input into the response surface prediction model to calculate the parameter fine-tuning increment. The parameter fine-tuning increment is superimposed on the optimal process parameter vector for dynamic feedback control until the vibration acceleration signal collected by the sensor is lower than the preset vibration acceleration safety threshold, at which point demolding is determined to be complete. After demolding, the set of structural feature parameters of qualified valve bodies in this batch and the optimal process parameter vector are written back to the valve body structural feature parameter library and the historical process knowledge base.

2. The method for precision forming of valve body by investment casting using a parameter iteration algorithm according to claim 1, characterized in that, The step S1, which involves geometric analysis of the 3D point cloud data to extract a set of structural feature parameters, further includes: The 3D point cloud data is converted into a 3D voxel mesh using an octree spatial partitioning algorithm; Calculate the minimum Euclidean distance from the center point of any voxel in the three-dimensional voxel grid to the physical boundary of the three-dimensional point cloud data, and multiply the minimum Euclidean distance by two to generate the wall thickness parameter in the corresponding spatial coordinates.

3. The method for precision forming of valve body by investment casting using a parameter iteration algorithm according to claim 2, characterized in that, The step of extracting the set of structural feature parameters in step S1 further includes: Based on the region growing algorithm, the hollow flow channel connected domains within the three-dimensional voxel mesh are identified, and the topological skeleton of the hollow flow channel connected domains is extracted as flow channel parameters. A binarization filtering extraction operation below a preset wall thickness safety threshold is performed on the wall thickness parameter, and the area ratio and spatial coordinate domain set of the region below the preset wall thickness safety threshold are extracted as the thin-wall parameter; The wall thickness parameter, the flow channel parameter, and the thin wall parameter are assembled together to form the set of structural feature parameters.

4. The method for precision forming of valve body by investment casting using a parameter iteration algorithm according to claim 1, characterized in that, The step of generating the initial process parameter vector in step S2 further includes: Extract the historical feature dimension vector stored in the historical process knowledge base; The distance matching value between the set of structural feature parameters and each of the historical feature dimension vectors is calculated using a weighted Euclidean distance metric. Extract the six values ​​bound to the target historical process record with the smallest distance matching value: pouring temperature, pouring speed, cooling gas flow rate, cooling time, vibration frequency, and amplitude. Map and assign these six values ​​to generate the initial process parameter vector.

5. The method for precision forming of valve body by investment casting using a parameter iteration algorithm according to claim 1, characterized in that, The step of extracting the objective function response set in step S3 further includes: Under each experimental matrix node generated by the central composite design, the thermohydrodynamic solver is invoked to calculate the spatiotemporal physical field of the molten metal filling and solidification process; The coordinates of the boundary points of the thermal stress deformation field are extracted from the spatiotemporal physical field. The coordinates of the boundary points of the thermal stress deformation field are then subjected to a three-dimensional Boolean difference operation with the three-dimensional point cloud data. The output size deviation is used as the first element of the objective function response set.

6. The method for precision forming of valve body by investment casting using a parameter iteration algorithm according to claim 5, characterized in that, The step of extracting the objective function response set in step S3 further includes: The solidification latent heat release rate span is extracted from the spatiotemporal physical field, and the grain size standard deviation distribution rate is calculated based on the non-uniform nucleation model. The uniformity of the microstructure is then output as the second element of the objective function response set. The dimensional deviation and tissue uniformity are assembled into a tensor to form the objective function response set.

7. The method for precision forming of valve body by investment casting using a parameter iteration algorithm according to claim 6, characterized in that, The regression equation for fitting and generating the response surface prediction model in step S3 is as follows: In the formula, The objective function response set represents the predicted output, including two output dimensions: size deviation and fabric uniformity. , This represents the 6-dimensional process input variables, which correspond to pouring temperature, pouring speed, cooling gas flow rate, cooling time, vibration frequency, and amplitude, respectively. This represents the intercept constant term in the regression model; This represents the partial regression coefficients of the first-order linear main effects; This represents the partial regression coefficient for a second-order pure quadratic effect; This represents the partial regression coefficient of the two-parameter cross-coupling effect; i and j represent the process parameter dimension indices, and satisfy the following conditions: .

8. The method for precision forming of valve body by investment casting using a parameter iteration algorithm according to claim 1, characterized in that, The step S4, which involves performing sequential quadratic programming iteration guided by the objective function response set, includes: The elements in the response set of the objective function are weighted and integrated to construct a comprehensive objective cost function. In each internal iteration loop, the comprehensive objective cost function is expanded by Taylor second order to construct an approximate Hessian matrix of the Lagrange function to solve the quadratic programming subproblem, and an optimization search step size direction vector is generated to update the parameter vector nodes.

9. The method for precision forming of valve body by investment casting using a parameter iteration algorithm according to claim 8, characterized in that, The step of calculating and outputting the optimal process parameter vector in step S4 includes: when the absolute value of the difference change of the comprehensive objective cost function is strictly lower than the preset convergence threshold, stopping the iterative calculation, and forcibly outputting the parameter vector node corresponding to the stopping time as the optimal process parameter vector.

10. The method for precision forming of valve body by investment casting using a parameter iteration algorithm according to claim 1, characterized in that, The step S5, which involves driving the various devices based on the optimal process parameter vector, includes: The values ​​of pouring temperature and pouring speed in the optimal process parameter vector are analyzed and converted into electrical drive signals to control the smelting equipment and the pouring equipment; The cooling gas flow rate and cooling time values ​​in the optimal process parameter vector are analyzed, and the cooling system is driven to perform forced heat exchange. The vibration frequency and amplitude values ​​in the optimal process parameter vector are analyzed and the vibration demolding system is driven to perform mechanical peeling.

Images