A method for coordinated control of mold temperature field based on digital twin

Abstract

This invention discloses a method for collaborative control of mold temperature field based on digital twins, belonging to the field of mold temperature control technology. The method includes: constructing a digital twin model of the mold temperature field covering macroscopic, mesoscopic, and microscopic scales; deploying distributed fiber optic temperature sensors to achieve continuous temperature field sensing; establishing a physical-digital real-time synchronous mapping through ensemble Kalman filtering; implementing model predictive control based on the digital twin; using augmented reality technology to achieve three-dimensional visualization of the temperature field and multimodal human-computer interaction; and using reinforcement learning to enable the digital twin to have autonomous optimization capabilities. This invention, through the deep integration of physical and digital spaces, achieves accurate sensing, real-time mapping, and intelligent control of the mold temperature field, significantly improving temperature control accuracy and system adaptability.

Description

Technical Field

[0001] This invention belongs to the field of mold temperature control technology, and in particular relates to a method for coordinated control of mold temperature field based on digital twin. Background Technology

[0002] Mold temperature control is a critical step in manufacturing processes such as injection molding and die casting. Precise control of the temperature field directly affects product quality and production efficiency.

[0003] In existing technologies, mold temperature control mainly relies on traditional PID controllers combined with point temperature sensors to achieve feedback regulation. However, this approach has the following shortcomings: First, the control response is lagging, making it difficult to adapt to rapidly changing production conditions; second, point temperature measurement cannot obtain continuous temperature field distribution information, resulting in monitoring blind spots; third, it lacks the ability to predict the evolution trend of the temperature field, limiting control accuracy; and fourth, the physical state is not visible, making it difficult for operators to intuitively grasp the temperature field distribution.

[0004] Therefore, there is an urgent need for a new control method that can achieve real-time mapping, accurate prediction and coordinated control of the temperature field. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for coordinated control of mold temperature field based on digital twins. By constructing a high-fidelity digital twin of the mold, real-time synchronous mapping and intelligent control of the physical mold and the digital model can be achieved.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for coordinated control of mold temperature field based on digital twin, the method comprising the following steps:

[0008] S1. Construct a multi-scale digital twin basic model of the mold, and establish a digital model of the mold temperature field from three spatial scales: macroscopic, mesoscopic and microscopic. The macroscopic scale describes the overall temperature distribution based on the three-dimensional unsteady heat conduction equation. The mesoscopic scale introduces convective heat transfer boundary conditions in key areas for local refinement. The microscopic scale considers the temperature dependence of material thermal properties. The two-way information transfer between the three scales is realized through boundary mapping, volume averaging and homogenization methods. The mesh density is dynamically adjusted by an adaptive meshing technique based on the temperature gradient norm.

[0009] S2. Deploy a continuous temperature field sensing network to achieve continuous monitoring of the mold temperature field through distributed optical fiber temperature sensors. The sensing is based on the Raman scattering principle, and temperature information is obtained by measuring the intensity ratio of Stokes light and anti-Stokes light. The spatial location of the measurement point is determined by optical time-domain reflectometry. The original signal is filtered by a Butterworth low-pass filter, and outliers are identified and corrected by local statistical tests. A dual-path redundancy mechanism is configured to ensure the continuity of temperature measurement.

[0010] S3. Establish a physical-digital real-time synchronous mapping mechanism, use the ensemble Kalman filter algorithm to fuse the sensing data into the digital twin model, use N ensemble members to represent the state uncertainty, use model operators to predict the state, use observation operators to map the model space to the observation space, and correct each ensemble member according to the Kalman gain; introduce an anomaly detection mechanism based on Mahalanobis distance, and trigger online correction of model parameters when an anomaly is detected.

[0011] S4. Implement predictive control based on digital twins. Model predictive control is performed based on the synchronized digital twin. The control sequence is obtained by solving a constrained optimization problem. The optimization objectives include temperature tracking accuracy, control increment, and weighted sum of slack variables. A hierarchical control architecture is adopted, with a global coordination layer for overall temperature field optimization and a local execution layer for tracking the setpoint through PID control. The two layers are coupled through an adaptive coordination factor. The optimization problem is solved using a sequential quadratic programming algorithm, and only the first control action is executed according to the rolling time domain principle.

[0012] S5. It achieves augmented reality visualization of the temperature field, employing a ray casting algorithm to perform volumetric rendering of the temperature field and generating a color cloud map through a temperature-color mapping function. The opacity is adaptively adjusted according to the temperature gradient. The mold pose is determined through marker recognition, and the virtual temperature field is superimposed onto the physical mold. It supports three interaction methods: gesture recognition, voice commands, and touch operation. Gestures are captured by a depth camera and recognized by a convolutional neural network; voice commands trigger control actions through keyword matching; and touch operation allows for precise parameter settings through a graphical interface. It also provides a multi-user collaboration mechanism and a timeline playback function.

[0013] S6. Perform autonomous learning optimization using a digital twin, modeling temperature control as a Markov decision process. The state space includes historical temperature, historical control data, and environmental disturbances, while the action space represents the adjustment parameters for each temperature control zone. The reward function integrates temperature deviation, energy consumption, temperature uniformity, and production cycle. A deep neural network parameterized policy network outputs the action distribution, and a value network estimates the expected return. The policy is updated using a proximal policy optimization algorithm, and the advantage function is calculated using generalized advantage estimation. Training is conducted within the digital twin, and the new policy can only be deployed after meeting the conditions for performance improvement and variance reduction.

[0014] Furthermore, in step S1, the macroscopic model is discretized using the finite element method, and the matrix equation is obtained through the Galerkin weighted residual method. The time discretization adopts the implicit Euler scheme. The mesoscopic convective heat transfer coefficient is determined by the Nusselt number, and the flow state is judged according to the Reynolds number. The calculation is performed using the Graetz solution, the Dittus-Boelter correlation, or linear interpolation, respectively. The thermal conductivity, specific heat capacity, and density at the microscopic scale are all expressed as functions of temperature.

[0015] Furthermore, in step S2, the optical fiber is deployed using a spiral winding method, and the spatial resolution is determined by the pulse width; outliers are corrected by weighted interpolation of adjacent points, and the weighting coefficient is determined according to the spatial distance; dual-path switching is based on statistical verification of the temperature of the primary and backup paths, and the switching process uses a smooth transition function.

[0016] Furthermore, in step S3, the model perturbation and observation perturbation are generated based on the error statistical characteristics, and the prediction error covariance is estimated through ensemble statistics; the cross covariance and observation prediction covariance are calculated through ensemble members; the model parameter correction adopts a gradient-based optimization method, and the loss function is the weighted observation error within the time window.

[0017] Furthermore, in step S4, the prediction function is calculated recursively using a digital twin; the global layer optimization objective includes a weighted sum of temperature tracking, temperature uniformity, and energy consumption; the coordination factor is adaptively adjusted according to the relative magnitude of global and local control deviations; and the Hessian matrix is ​​updated using a quasi-Newton method.

[0018] Furthermore, in step S5, the camera intrinsic parameter matrix is ​​obtained through calibration; the gesture control parameter mapping includes the product of gesture gain, movement amplitude, and direction vector; multi-user collaboration is achieved through the conditional selection of master controller, slave controller, and current value; and timeline playback is reconstructed through linear interpolation of historical data.

[0019] Furthermore, in step S6, the policy network outputs the mean and standard deviation of a normal distribution; the temporal difference error is used to calculate the dominance function; the importance sampling ratio is used for policy gradient estimation; the pruning parameter prevents the policy update from being too large; and the deployment criteria comprehensively consider the performance improvement and stability. Beneficial effects

[0020] 1. Enhanced overall sensing capability: Continuous temperature field monitoring is achieved through distributed fiber optic temperature measurement, eliminating the blind spots of traditional point-based temperature measurement.

[0021] 2. Significantly improved control accuracy: Model predictive control based on digital twins effectively overcomes the lag of traditional feedback control.

[0022] 3. Improved human-computer interaction efficiency: Augmented reality visualization and multimodal interaction methods enhance the intuitiveness and convenience of operation.

[0023] 4. Strong system self-adaptability: Reinforcement learning enables the system to have autonomous optimization capabilities, and the control performance continues to improve over time.

[0024] 5. High accuracy of virtual-real synchronization: Data assimilation technology ensures that the digital twin accurately reflects the real-time status of the physical model.

[0025] 6. Optimized computational efficiency: Adaptive mesh and hierarchical control architecture improve computational efficiency while ensuring accuracy. Attached Figure Description

[0026] Figure 1 A flowchart illustrating the steps of the method described in this invention is shown. Detailed Implementation

[0027] Exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0028] Combination Figure 1 This invention provides a method for collaborative control of mold temperature field based on digital twins. By constructing a high-fidelity digital mapping of the physical mold, real-time monitoring, accurate prediction, and intelligent control of the temperature field are achieved. The core of the method lies in establishing a digital twin that can accurately reflect the evolution law of the physical mold temperature field, and based on this, achieving deep integration and collaborative optimization of physical and digital spaces.

[0029] Step S1: Construct a multi-scale digital twin basic model of the mold

[0030] The multi-scale digital twin basic model comprehensively describes the temperature field of the mold from three spatial scales: macroscopic, mesoscopic, and microscopic. It achieves effective transmission of information at different scales through a scale coupling mechanism.

[0031] The macroscopic model is based on the three-dimensional unsteady heat conduction theory, and its governing equations are:

[0032]

[0033] in: Density of the mold material; Specific heat capacity of the material; For temperature; For time; This is the partial derivative of temperature with respect to time; Thermal conductivity; For Hamiltonian operators; For divergence operators; For temperature gradient; This represents the density of the internal heat source.

[0034] The control equations are spatially discretized using the finite element method, and the mold domain is divided into... The temperature field is divided into a finite number of elements, and within each element, a shape function is used to approximate the temperature field:

[0035]

[0036] in: Spatial coordinates and time Temperature at that location; For the first Shape functions of nodes; For the first Each node in time Temperature value; This represents the number of unit nodes. This is the summation symbol.

[0037] Applying the Galerkin weighted residual method, the discretized matrix equation is obtained:

[0038] in: This is the heat capacity matrix; The node temperature vector; The derivative of the nodal temperature vector with respect to time; This is the heat conduction matrix; This is the thermal load vector.

[0039] The elements of the heat capacity matrix are calculated using unit integration:

[0040] in: For the heat capacity matrix, the first Line 1 Column elements; Number the units; The total number of units; In the first Integral over a unit domain; For the first A shape function; For the first A shape function; It is a volumetric infinitesimal element.

[0041] The elements of the heat conduction matrix are:

[0042] in: For the heat conduction matrix, the first Line 1 Column elements; For the first The gradient of a shape function; For the first The gradient of a shape function; This is the dot product operation.

[0043] The time discretization employs an implicit Euler scheme to ensure numerical stability.

[0044]

[0045] in: For the first Temperature vector at any given time; For the first Temperature vector at any given time; For time step; For the first The thermal load vector at time t.

[0046] The mesoscale model refines the model of key local regions based on the macroscopic model. These key regions include the vicinity of the gate, thin-walled structures, hot spots, and the area around cooling pipes. A convective heat transfer boundary condition is introduced at the interface between the mold and the cooling medium.

[0047]

[0048] in: Heat flux density; This is the derivative of temperature along the outward normal of the wall. This indicates that the value is taken at the wall surface; The convective heat transfer coefficient; The wall temperature; This refers to the temperature of the cooling fluid.

[0049] The convective heat transfer coefficient Through Nusel Sure:

[0050] in: For Nusselt numbers; The fluid's thermal conductivity; It is the hydraulic diameter.

[0051] The calculation of the Nusselt number needs to be determined based on the flow conditions. The Reynolds number is defined as:

[0052]

[0053] in: It is the Reynolds number; For fluid density; For flow rate; This refers to dynamic viscosity.

[0054] when At that time, the flow is in a laminar state, and the Nusselt number is obtained using the Graetz solution:

[0055]

[0056] in: The diameter of the pipe; This refers to the length of the pipe. It is a Prandtl number; The critical Reynolds number for the transition from laminar to transitional flow is determined based on pipe roughness and flow conditions, and is generally taken as 2300; the exponent It represents the power of three squared.

[0057] when At this point, the flow is in a fully developed turbulent state, and the Dittus-Boelter correlation is used:

[0058] in: The critical Reynolds number for the transition from transitional flow to turbulent flow is typically taken as 10,000; the exponent 0.8 represents the power of 0.8. The temperature correction index is 0.4 when the fluid is heated and 0.3 when it is cooled.

[0059] Transition area Linear interpolation is used:

[0060]

[0061] in: For laminar flow, the Nusselt number is used. denoted as the Nusselt number for turbulent conditions.

[0062] The microscale model considers the influence of the material's microstructure on its thermal properties. The temperature dependence of the thermal conductivity is expressed as:

[0063]

[0064] in: For temperature Thermal conductivity at that time; Reference temperature Thermal conductivity at the specified value; The first temperature coefficient of thermal conductivity; It is the second temperature coefficient of thermal conductivity; The coefficient represents the square of the temperature difference; the coefficient is determined by material thermophysical property experiments or obtained from a material database.

[0065] The relationship between specific heat capacity and temperature is as follows:

[0066]

[0067] in: For temperature Specific heat capacity at time; Specific heat capacity at reference temperature; This is the primary temperature coefficient of specific heat capacity; It is the second temperature coefficient of specific heat capacity.

[0068] The temperature dependence of density is described by the coefficient of thermal expansion:

[0069] in: For temperature Density at time; Density at reference temperature; is the coefficient of volumetric thermal expansion.

[0070] Two-way information transfer between the three scales is achieved through a scale bridging algorithm. Downlink transfer from macroscopic to mesoscopic scales employs boundary condition mapping.

[0071] in: Temperature for the mesoscopic model; Indicates the boundary of the mesoscopic model Take the value above; For the first One interpolation basis function; For the macroscopic model Temperature of each node; The number of nodes used for interpolation.

[0072] Downward transfer from mesoscopic to microscopic scales is established through volume averaging:

[0073]

[0074] in: The volume average temperature; The volume of a representative volume unit; This is the integral over a representative volume element; Temperature distribution at the microscopic scale; It is a volumetric infinitesimal element.

[0075] The upward transfer from the microscopic to the mesoscopic is achieved through homogenization:

[0076]

[0077] in: It is the equivalent thermal conductivity; For the microscale in coordinate and temperature The local thermal conductivity is below.

[0078] The upward transmission from mesoscopic to macroscopic levels is achieved through the superposition of contributions from sub-models:

[0079]

[0080] in: For the load vector of the macroscopic model; This is the initial load vector; Number the mesoscopic sub-models; The number of mesoscopic sub-models; For the first The projection matrix of each sub-model; This is the transpose of the projection matrix; For the first The load contribution of each mesoscopic sub-model.

[0081] To improve computational efficiency, an adaptive meshing technique is employed. Mesh adjustment is based on the temperature gradient norm.

[0082]

[0083] in: Let be the Euclidean norm of the temperature gradient; For temperature Partial derivatives of coordinates; For temperature Partial derivatives of coordinates; For temperature Partial derivatives of coordinates; This is for square root operations.

[0084] The grid size adjustment strategy is as follows:

[0085]

[0086] in: The adjusted grid size; This is the current grid size; The refinement / coarsening scaling factor typically has a value of 2; To refine the threshold; The threshold is used for coarsening; the threshold is determined through numerical experiments.

[0087] Mesh quality is monitored using the distortion index:

[0088] in: For grid quality indicators; This represents the actual unit volume; The volume is the ideal unit volume.

[0089] when Local mesh reconstruction is triggered at time, where This is the critical distortion degree, which is generally taken as 0.1.

[0090] Step S2: Deploy a continuous temperature field sensing network

[0091] Based on the digital model framework constructed in step S1, continuous monitoring of the temperature field of the physical mold is achieved by deploying distributed fiber optic temperature sensors. The distributed fiber optic sensing is based on the Raman scattering principle; when a laser pulse propagates in the fiber, the scattered light generated by its interaction with fiber molecules carries temperature information.

[0092] The relationship between temperature and scattered light intensity is as follows:

[0093] in: Absolute temperature; It is Planck's constant; For Raman frequency shift; Boltzmann's constant; It is the natural logarithm; These are system calibration constants; Stokes light intensity; This represents the intensity of the anti-Stokes light.

[0094] The spatial location of the measurement point was determined using optical time-domain reflectometry.

[0095] in: The measurement point is located at the distance from the starting end of the optical fiber; It is the speed of light in a vacuum; The round-trip time of the laser pulse; is the effective refractive index of the optical fiber; factor 2 represents the round-trip path of the optical pulse.

[0096] Spatial resolution is determined by pulse width:

[0097] in: Spatial resolution; This represents the laser pulse width.

[0098] The raw temperature signal needs to be filtered to improve the signal-to-noise ratio. A Butterworth low-pass filter is used.

[0099]

[0100] in: The filter transfer function; For Laplace variables; The cutoff angular frequency; The filter order; express of Power of 1.

[0101] The filtered signal is used for outlier detection, and the judgment criteria are as follows:

[0102]

[0103] in: For the first Temperature at each measurement point; This represents the local average temperature. This is for absolute value operations; The detection coefficient is set according to the noise characteristics, and is generally taken as 3; This represents the local standard deviation.

[0104] When outliers are detected, corrections are made using interpolation between neighboring points:

[0105]

[0106] in: For the first The corrected temperature; For the first Temperature at point; For the first Temperature at point; For the first The weighting coefficient of the point; For the first The weight coefficients of the points satisfy the following conditions: .

[0107] The weighting coefficients are determined based on spatial distance:

[0108]

[0109] in: For the first Point and the Spatial distance between points; For the first Point and the Spatial distance between points.

[0110] In a dual-path redundancy configuration, the decision to switch between primary and backup paths is based on statistical tests:

[0111]

[0112] in: To switch criteria; The average temperature of the main road; The average temperature of the backup road; The standard deviation of the main road temperature; The standard deviation of the backup road temperature; This is for square root operations.

[0113] when The switching is triggered on time, where The preset switching threshold is set according to system reliability requirements, and generally ranges from 2 to 3. A smooth transition function is used during the switching process to avoid sudden data changes.

[0114]

[0115] in: for Temperature during the switching process; for Main road temperature at all times; for Always be prepared for road temperature; It is a transition function that smoothly transitions from 0 to 1 within the switching time window.

[0116] The transition function takes the form of sigmoid:

[0117]

[0118] in: It is an exponential function; For the transition rate parameter; This is the midpoint of the switching time window.

[0119] The optical fiber is laid using a helical winding method, and the parametric equation of the helix is:

[0120]

[0121] in: , , Angle on the spiral Cartesian coordinates at the location; The radius of the helix; It is a cosine function; It is a sine function; Pitch; Pi is the mathematical constant of a circle.

[0122] Step S3: Establish a physical-digital real-time synchronization mapping mechanism

[0123] Based on the digital model constructed in step S1 and the sensing data obtained in step S2, real-time synchronization between physical space and digital space is achieved through a data assimilation algorithm. The data assimilation employs an ensemble Kalman filter method, which can effectively handle nonlinear systems and quantify uncertainties.

[0124] The state-space model is represented as:

[0125]

[0126] in: For the first The state vector at any given time contains the temperature values ​​of all grid nodes; For the first The state vector at any given time; This is the model operator, namely the multi-scale heat conduction model in step S1; For the first Model error at time step; For the first The observation vector at time point is derived from the fiber optic temperature measurement data in step S2; For the observation operator, the state space is mapped to the observation space; For the first The observation error at a given time.

[0127] pass Each set member represents the uncertainty of the state:

[0128]

[0129] in: For the first The set members are based on the first Time information for the first The predicted value at any given time; For the first The set member in the th ... Analysis values ​​at any given time; For the first The member in the The model perturbation at any given time; The number of members in the set is determined based on computing resources and accuracy requirements, and is generally between 50 and 100.

[0130] The model perturbation is generated based on the statistical characteristics of model error:

[0131]

[0132] in: This indicates that the mean is 0 and the covariance matrix is... The multivariate normal distribution; For the first The model error covariance matrix at time t.

[0133] The prediction error covariance is estimated using ensemble statistics:

[0134]

[0135] in: For the first The first time information The covariance matrix of the prediction error at any given time; The average of the ensemble predictions; superscript This indicates the matrix transpose.

[0136] The set average is calculated as follows:

[0137]

[0138] in: The arithmetic mean of the predicted values ​​for all set members.

[0139] The Kalman gain matrix is ​​calculated as follows:

[0140]

[0141] in: For the first Kalman gain matrix at time step; This is the cross-covariance matrix between the state and the observation; To predict the covariance matrix for observation; The observation error covariance matrix; This represents the matrix inversion operation.

[0142] The cross-covariance is calculated as follows:

[0143]

[0144] in: For the first The predicted state of each set member; The ensemble average of state predictions; For the first The observation predictions obtained by mapping the observation operators for each member; It is the ensemble average of the observed predictions.

[0145] The observed-predicted covariance is calculated as follows:

[0146]

[0147] The update steps correct each set member using Kalman gain:

[0148]

[0149] in: For the first The updated analysis values ​​for each member; For the first The observed perturbations of each member satisfy .

[0150] The anomaly detection mechanism is implemented using Mahalanobis distance:

[0151]

[0152] in: The Mahalanobis distance; The new information covariance matrix; It is the inverse of the new information covariance matrix.

[0153] when Online correction of model parameters is triggered at specific times, where For degrees of freedom Confidence level is The critical value of the chi-square distribution; Let be the dimension of the observation vector; It is usually taken as 0.05 or 0.01.

[0154] Online calibration of model parameters uses the gradient descent method.

[0155]

[0156] in: For the first The model parameter vector at time step; For the first Parameters updated in real time; The learning rate is set according to the required convergence speed. For loss function For parameters The gradient.

[0157] The loss function is defined as follows:

[0158]

[0159] in: The length of the time window; For the time index within the time window; For the first The observation vector at time; For parameters Next State estimation at time; It is the inverse of the observation error covariance matrix.

[0160] Step S4: Implement predictive control based on digital twins

[0161] Using the real-time synchronous digital twin established in step S3, model predictive control is implemented to achieve precise regulation of the temperature field. The predictive control simulates the future evolution of the temperature field based on the digital twin, and obtains the optimal control strategy by solving an optimization problem.

[0162] The optimization problem of model predictive control is formulated as follows:

[0163]

[0164] in: The objective function for model predictive control; For control sequences; To predict the time index within the time domain; To predict the length of the time domain; To control the length of the time domain; For the first The first time information prediction Temperature at any given time; For the first Reference temperature at any given time; Indicates It is the squared weighted L2 norm of the weight matrix; For the first Control increment at any given moment; Indicates It is the squared weighted L2 norm of the weight matrix; It is a relaxation factor; These are slack variables.

[0165] The constraints include temperature constraints:

[0166]

[0167] in: The lower limit of temperature; This is the upper limit of the temperature range; It is a vector consisting entirely of 1s; Represents element-wise comparisons in vector inequalities.

[0168] Control input constraints:

[0169]

[0170] in: To control the lower limit of input; To control the upper limit of input; For the first Time-based control input.

[0171] Control Incremental Constraints:

[0172]

[0173] in: To control the maximum allowable value of the increment.

[0174] The control increment is defined as:

[0175]

[0176] in: For the first Time-based control input.

[0177] The prediction model is based on the digital twin from step S3:

[0178]

[0179] in: For multi-step prediction functions of digital twins; For the first Temperature field estimation at time.

[0180] The prediction function is implemented through recursive calculation using a digital twin:

[0181]

[0182] in: For the first Time information for the first Temperature prediction at any given time; For the first Time information for the first Temperature prediction at any given time; To predict the time step; This is the temperature field evolution function; For the first Time-based control input.

[0183] In a layered control architecture, the optimization objective of the global coordination layer is:

[0184]

[0185] in: The objective function is the global layer function. Index for temperature-controlled areas; This represents the total number of temperature-controlled zones. For the first The average temperature of the region; The target temperature; Represents the square of the Euclidean norm; This is the weighting coefficient for temperature uniformity; The average temperature across all regions; For the first The square of the deviation between the regional temperature and the average temperature; Energy consumption weighting coefficient; Total energy consumption.

[0186] The average temperature is calculated as follows:

[0187]

[0188] The local execution layer uses PID control to track the global layer setpoint:

[0189]

[0190] in: For the first The area is The amount of control at any given moment; For the first The proportional gain of the region; For the first The area is Tracking error at any given moment; For the first The integral gain of the region; To track error from 0 to Integral at time step; For integration variables; For the first Differential gain of the region; To track the derivative of the error with respect to time.

[0191] Tracking error is defined as:

[0192]

[0193] in: For the global layer, the first The area is Temperature setpoint at any time; For the first The area is The actual temperature at any given moment.

[0194] The global layer and the local layer are coupled through a coordination factor:

[0195]

[0196] in: For the final control output; As a coordinating factor; For global layer control variables; This is a local layer control variable.

[0197] The coordination factor is adaptively adjusted based on the control deviation:

[0198]

[0199] in: It is an exponential function; To adjust the parameters and control the slope of the sigmoid function; This is for global control deviation; For local control deviation; This represents the Euclidean norm of a vector.

[0200] The optimization problem is solved using a sequential quadratic programming algorithm, which linearizes the nonlinear optimization problem at the current point:

[0201]

[0202] in: To control the increment vector; The Hessian matrix; To control the transpose of the increment vector; The gradient vector; This is the transpose of the gradient vector.

[0203] The Hessian matrix is ​​updated using the BFGS quasi-Newton method:

[0204]

[0205] in: For the first The Hessian matrix of the next iteration; For the first The Hessian matrix of the next iteration; Let be the step size vector of the decision variables; The gradient difference vector; For the first The gradient of the objective function in the next iteration; For the first The gradient of the objective function in the next iteration.

[0206] According to the rolling time domain principle, only the first control action obtained from the optimization is executed, and the optimization problem is solved again in the next time step.

[0207] Step S5: Achieve augmented reality visualization of the temperature field

[0208] Based on the temperature field data of the digital twin in step S4, three-dimensional visualization and human-computer interaction are achieved through augmented reality technology.

[0209] The volume rendering of the temperature field uses a ray casting algorithm:

[0210]

[0211] in: screen coordinates Pixel intensity at that location; The length of the light ray; For light parameters; For position The opacity function at the location; For position Temperature The corresponding color value; It is an exponential function; From 0 to Opacity accumulation; For integration variables; It is a light element.

[0212] The opacity function is adaptively adjusted according to the temperature gradient:

[0213]

[0214] in: Based on opacity; For visualization of adjustment coefficients; For position Temperature gradient norm at that location; This represents the maximum value of the temperature gradient.

[0215] The mapping from temperature to color uses a piecewise linear function:

[0216]

[0217] in: For temperature The corresponding RGB color value; The RGB value is for blue. The RGB value for cyan; The RGB value is green. The RGB value is yellow; The RGB value is for red. The lowest temperature; This is the highest temperature; , , For temperature breakpoints, satisfying .

[0218] Augmented reality overlay requires precise spatial registration, and the pose of the mold in the camera coordinate system is determined through marker recognition:

[0219]

[0220] in: It is a 4×4 homogeneous transformation matrix, representing the transformation from the mold coordinate system to the camera coordinate system; It is a 3×3 rotation matrix; It is a 3×1 translation vector; It is the transpose of the 1×3 zero vector.

[0221] The projection of the virtual temperature field onto the screen is achieved through camera intrinsics:

[0222]

[0223] in: Scale factor; These are screen pixel coordinates; These are the coordinates of a three-dimensional point in the mold coordinate system. This is a 3×3 camera intrinsic parameter matrix.

[0224] The camera intrinsic parameter matrix is ​​defined as follows:

[0225]

[0226] in: The focal length is in the x-direction; The focal length is in the y-direction; The coordinates of the main point.

[0227] Gesture recognition uses a depth camera to acquire 3D information about the hand and then employs a convolutional neural network for classification.

[0228]

[0229] in: The predicted gesture type; Indicates the index that maximizes the probability. ; For a given depth image Gesture type The conditional probability; For gesture type index.

[0230] The mapping relationship of gesture control parameters is as follows:

[0231]

[0232] in: The amount of parameter adjustment caused by the gesture; Gain is controlled by gestures; This refers to the range of motion of the gesture. This is the unit vector for the gesture direction.

[0233] Voice commands are implemented through keyword recognition. The speech signal is converted into Mel-frequency cepstral coefficient features and then matched against a predefined command template. The matching degree is calculated as follows:

[0234]

[0235] in: A score for the degree of matching; The extracted MFCC feature vector; The template feature vector; This is the transpose of the MFCC eigenvectors; This represents the Euclidean norm of a vector.

[0236] when The corresponding control action is triggered at the time, among which This is the matching threshold.

[0237] Touch interaction enables precise parameter settings through a graphical user interface:

[0238]

[0239] in: Set the new parameter value; Set a value for the slider; The value is the input field. The original parameter value; Adjust the step size for the button.

[0240] Multi-user collaboration is achieved through permission management and conflict resolution mechanisms:

[0241]

[0242] in: This is the final control value; Set values ​​for the main control user; Set values ​​for secondary control users; This is the current value.

[0243] The timeline playback function reconstructs the temperature field at any given moment by interpolating historical data:

[0244]

[0245] in: for Temperature field that is constantly replayed; For the first The temperature field of a historical moment; For the first The temperature field of a historical moment; For the first A timestamp of a historical moment; For the first A timestamp of a historical moment; For the moment to be played back, to satisfy .

[0246] Step S6: Perform autonomous learning optimization of the digital twin

[0247] Based on the operational data accumulated in the aforementioned steps, reinforcement learning algorithms are used to continuously optimize the control strategy, giving the digital twin the ability to evolve autonomously.

[0248] The temperature control problem is modeled as a Markov decision process, with the state space defined as follows:

[0249]

[0250] in: for The state vector at any given time; for Temperature field at any given moment; For the front The historical temperature field at a specific moment; For the front Historical control actions at any given moment; for Constant environmental disturbances; The length of the temperature history; To control the length of the history.

[0251] The range of motion refers to the adjustment amount of each temperature control zone:

[0252]

[0253] in: for Action vector at any given moment; for Time of the first Adjustment range for each temperature control zone; This represents the total number of temperature-controlled zones.

[0254] The reward function integrates multiple optimization objectives:

[0255]

[0256] in: for Instant rewards for each moment; for Time of the first The temperature of the area; For the first The area's reference temperature; The square of the temperature deviation; Temperature deviation weighting; for Energy consumption at any given moment; Energy consumption weighting; for Temperature variance over time; The weights are uniform. This refers to the actual cycle time. The target cycle time; Penalties will be imposed for exceeding the time limit during the cycle. The periodic time weight is used.

[0257] The policy network uses deep neural network parameterization to output the probability distribution of actions:

[0258]

[0259] in: For parameters The strategy in the state Select action The probability density; It follows a multivariate normal distribution; This is the action mean vector output by the mean network; This is the covariance matrix output by the covariance network.

[0260] Expected return for estimating the state in a value network:

[0261]

[0262] in: For parameters The value function; Indicating in strategy The following expectations; Indexing future moments; This is a discount factor, with a value ranging from 0.95 to 0.99; for Momentary rewards; for The state at any given moment.

[0263] The policy network is updated using the proximal policy optimization algorithm, and its objective function is:

[0264]

[0265] in: The objective function for pruning; Indicates time Expectations; This is an operation to find the minimum value; Importance sampling ratio; This is the old strategy; For the estimation of the advantage function; For the clipping function, Limited to interval; This is the trimming parameter, with a value range of 0.1 to 0.2.

[0266] The advantage function is calculated using generalized advantage estimation:

[0267]

[0268] in: for Generalized advantage estimation at any given moment; Index for future steps; The moment when the trajectory terminates; This is a GAE parameter used to balance bias and variance, with a value of 0.95. for The timing difference error at each moment; express of Power of 1.

[0269] Timing difference error is defined as:

[0270]

[0271] in: for The timing difference error at each moment; for Value estimation of state at any given time; for Value estimation of the state at any given moment.

[0272] The loss function of the value network is:

[0273]

[0274] in: For the value function loss; The square of the difference between the predicted value and the target value; The target value is estimated using Monte Carlo returns.

[0275] The policy network parameters are updated as follows:

[0276]

[0277] in: For the first The policy network parameters for the next iteration; The learning rate of the policy network; For the trimming loss parameters The gradient.

[0278] The value network parameters are updated as follows:

[0279]

[0280] in: For the first Value network parameters for the next iteration; The learning rate of the value network; For the parameter of value loss The gradient.

[0281] The new strategy can only be deployed to the physical control system if both of the following conditions are met.

[0282] First, conditions for performance improvement:

[0283]

[0284] in: For strategy Performance metrics; For the new strategy; The original strategy; This represents the absolute value of the original strategy's performance metrics; The performance improvement threshold ranges from 0.05 to 0.1.

[0285] Second, conditions for improving stability:

[0286]

[0287] in: For strategy The performance variance reflects the control stability.

[0288] In summary, the mold temperature field collaborative control method based on digital twin provided by this invention forms a complete closed-loop system of "sensing-modeling-synchronization-control-interaction-optimization" through the organic coordination of six key steps.

[0289] Compared to existing technologies, this invention improves the accuracy of temperature field description through multi-scale modeling, eliminates monitoring blind spots through distributed sensing, ensures precise synchronization between virtual and real data through data assimilation, overcomes the lag of traditional feedback control through predictive control, enhances human-computer interaction efficiency through augmented reality, and endows the system with adaptive evolutionary capabilities through reinforcement learning. The method does not require large-scale modifications to existing equipment and can be widely applied to manufacturing fields requiring precise temperature control, such as injection molding, die casting, and extrusion, demonstrating good engineering practicality and application value.

[0290] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A mold temperature field collaborative regulation method based on digital twinning, characterized in that, The method comprises the following steps: S1, constructing a mold multi-scale digital twin basic model: a temperature field digital model is established from three spatial scales of macro, meso and micro. The macro scale describes the overall temperature distribution based on a three-dimensional unsteady heat conduction equation. The meso scale introduces a convective heat transfer boundary condition in the key area. The micro scale establishes a temperature-dependent relationship of material thermal physical properties. The scale bridging algorithm is used to realize information transmission among the three scales. The scale bridging algorithm comprises: the macro-to-meso scale adopts boundary condition mapping, and the macro temperature is mapped to the meso boundary through an interpolation base function; the meso-to-micro scale is connected through volume averaging, and the temperature is integrated and averaged in a representative volume element; the micro-to-meso scale is realized through a homogenization method, and the micro thermal physical properties are converted into equivalent parameters; the meso-to-macro scale is realized through submodel contribution superposition, and the meso load is transmitted to the macro model through a projection matrix; S2, deploying a continuous temperature field perception network: a distributed optical fiber temperature sensor is deployed to form a continuous perception network. The temperature is measured based on the Raman scattering principle. The measurement point position is determined through optical time domain reflection technology. The original signal is filtered and outlier corrected; S3, establishing a physical-digital real-time synchronization mapping mechanism: a set Kalman filtering algorithm is used to establish a physical-digital real-time synchronization mapping. The perception data is fused into the digital twin model through multiple set members representing uncertainty. The abnormal detection based on Mahalanobis distance and the online correction of model parameters are introduced; S4, implementing digital twin-based predictive control: model predictive control is implemented based on the synchronized digital twin. The control sequence is obtained by solving a multi-objective optimization problem. A hierarchical architecture is used to realize the coupling of global coordination and local execution; S5, realizing temperature field augmented reality visualization: three-dimensional visualization of the temperature field is realized through a ray casting algorithm. Augmented reality technology is used for superimposed display. Multi-modal human-computer interaction such as gestures, voice and touch is supported; S6, executing digital twin autonomous learning optimization: a reinforcement learning algorithm is used to optimize the control strategy. The temperature control is modeled as a Markov decision process. Autonomous learning is realized through a policy network and a value network.

2. The mold temperature field collaborative regulation method based on digital twinning according to claim 1, characterized in that, In the step S1, the macro scale model is discretized by the finite element method. The matrix equation is obtained by the Galerkin weighted residual method. The implicit Euler format is used for time discretization. The convective heat transfer coefficient of the meso scale is determined by the Nusselt number. The flow state is determined according to the Reynolds number. The Graetz solution, the Dittus-Boelter correlation or linear interpolation is used for calculation respectively. The thermal conductivity, specific heat capacity and density of the micro scale are represented as functions of temperature.

3. The mold temperature field collaborative regulation method based on digital twinning according to claim 1, characterized in that, The set Kalman filtering comprises: predicting the state of each set member through a model operator and adding a model disturbance to represent uncertainty; estimating the prediction error covariance and cross covariance through set statistics; calculating the Kalman gain matrix, which is equal to the product of the inverse matrix of the cross covariance and the sum of the observation covariance and the observation error covariance; updating each set member according to the Kalman gain, and the correction amount is the difference between the observation value and the predicted observation value multiplied by the Kalman gain.

4. The mold temperature field collaborative regulation method based on digital twinning according to claim 1, characterized in that, The optimization problem of the model predictive control is: the objective function contains the weighted quadratic term of the deviation of the predicted temperature from the reference temperature, the weighted quadratic term of the control increment, and the penalty term of the relaxation variable; the constraint conditions include the upper and lower temperature constraints, the control input constraints and the control increment constraints; the nonlinear problem is linearized at the current point by the sequential quadratic programming algorithm, and the Hessian matrix is updated by the quasi-Newton method.

5. The mold temperature field collaborative regulation method based on digital twinning according to claim 1 or 4, characterized in that, In the hierarchical architecture, the optimization objective of the global coordination layer includes the sum of squares of the deviations of the regional temperatures from the target temperature, the temperature uniformity index, and the total energy consumption; the local execution layer uses PID control to track the set value given by the global layer; the coordination factor is adaptively adjusted by the sigmoid function according to the relative size of the global control deviation and the local control deviation.

6. The mold temperature field co-regulation method based on digital twinning according to claim 1, characterized in that, In the three-dimensional visualization of the temperature field, the opacity function is adaptively adjusted according to the ratio of the temperature gradient norm to the maximum gradient; the temperature-to-color mapping uses a piecewise linear function to generate a gradient color by linear interpolation in different temperature intervals; the pose transformation matrix of the mold in the camera coordinate system is obtained by marker recognition, and the virtual temperature field is projected to the screen coordinates.

7. The mold temperature field co-regulation method based on digital twinning according to claim 1, characterized in that, The reinforcement learning includes: the state space contains the current and historical temperature fields, historical control actions, and environmental disturbances; the action space is the control adjustment amount of each temperature control region; the reward function is the negative weighted sum of the temperature deviation, energy consumption, temperature variance, and production cycle deviation; the proximal policy optimization algorithm is used to limit the policy update amplitude by clipping the importance sampling ratio; the advantage function is calculated by generalized advantage estimation, combined with the timing difference error and the discount factor.

8. The mold temperature field co-regulation method based on digital twinning according to claim 1, characterized in that, It also includes adaptive grid adjustment: the grid density requirement is determined according to the temperature gradient norm, the grid size is reduced to one-tenth of the original refinement ratio when the gradient exceeds the refinement threshold, and the grid size is expanded to the original coarsening ratio times when the gradient is below the coarsening threshold; the grid quality is monitored by the grid distortion index, and local grid reconstruction is triggered when the ratio of the actual cell volume to the ideal volume is less than a set value.

9. The mold temperature field co-regulation method based on digital twinning according to claim 1, characterized in that, The distributed fiber-optic temperature sensor uses a dual-redundancy configuration: the main path performs regular temperature measurement, and the backup path maintains a preheating state; the switching criterion is calculated by the ratio of the difference between the average temperatures of the main and backup paths to the square root of the sum of the standard deviations; the switching process uses a smooth transition function to smoothly transition from the main path to the backup path within a switching time window.

10. The mold temperature field co-regulation method based on digital twinning according to claim 1, characterized in that, The multi-modal human-computer interaction includes: gesture recognition acquires a sequence of hand depth images through a depth camera, identifies the gesture type using a convolutional neural network, and adjusts the control parameter adjustment amount as the product of the gesture amplitude, direction, and gain; voice instructions are implemented by extracting mel-frequency cepstral coefficient features and matching them with predefined templates; touch interaction provides precise parameter settings through sliders, input boxes, and buttons; multi-user collaboration is achieved through permission management, and the final control value is selected according to the operation state of the main control user and the secondary control user.

Images