Casting virtual auxiliary system
Through the casting virtual assistance system, combined with deep learning and reinforcement learning technologies, the shortcomings of traditional casting systems in simulation efficiency, defect prediction and process adjustment are solved, and efficient optimization and real-time defect tracing of the aluminum alloy casting process are achieved, thereby reducing the scrap rate.
Patent Information
- Application Number
- CN202510735260.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-04
- Publication Date
- 2025-09-16
AI Technical Summary
Traditional casting systems have shortcomings in simulation efficiency, defect prediction, and process adjustment, making it difficult to achieve real-time response and efficient optimization. This is especially true in the aluminum alloy casting process, where scrap rates are high and defect tracing is difficult.
By adopting a physical data dual-modal engine, a multi-source data spatiotemporal synchronization module, a dynamic process decision engine and a defect evolution prediction subsystem, combined with deep learning and reinforcement learning technologies, an intelligent control system for the entire casting process is constructed to achieve real-time interaction and closed-loop feedback between simulation and measured data.
It significantly improves the efficiency of process parameter optimization and defect prediction accuracy, shortens the process adjustment response time, reduces the scrap rate of aluminum alloy casting, and realizes high-precision fusion of multi-source data and real-time process adjustment.
Smart Images
Figure CN120654552A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of casting technology, and in particular to a casting virtual auxiliary system. Background Art
[0002] Casting is a fundamental manufacturing process, widely used in the automotive, aviation, and military industries. Traditional casting systems rely on empirical models and static parameter settings, and suffer from three core pain points: First, physical simulation efficiency is low. Traditional finite element and smoothed particle fluid dynamics methods struggle to balance computational accuracy and real-time performance, and complex flow field simulations can take hours. Second, defect prediction lags. Existing machine learning models are mostly trained based on offline data and cannot dynamically respond to temperature gradient changes and stress concentration during the casting process. Third, process adjustments are subject to lag. Traditional PID control struggles to adapt to nonlinear systems coupled with multiple parameters such as pouring speed and mold temperature. Studies have shown that for every 1% decrease in mold filling integrity during aluminum alloy casting, the scrap rate increases by 3%-5%. Defect tracing requires real-time monitoring of dozens of parameters, such as material viscosity and cooling rate. To this end, existing technologies have attempted to address these issues by introducing digital twins and industrial Internet of Things technologies, but they still face challenges such as strong heterogeneity in multi-source data, low algorithm collaboration efficiency, and high closed-loop feedback delays.
[0003] Therefore, a casting virtual assistance system is now developed in combination with the above-mentioned problems. Summary of the Invention
[0004] In order to overcome the shortcomings of the prior art, the present invention provides a casting virtual assistance system.
[0005] The technical solution of the present invention is:
[0006] A casting virtual auxiliary system, characterized by comprising:
[0007] A physical data bimodal engine that constructs a casting process model using a dynamic SPH-DL hybrid algorithm;
[0008] A multi-source data spatiotemporal synchronization module that integrates equipment parameters, material properties, and real-time monitoring data;
[0009] A dynamic process decision engine that generates real-time optimization parameters based on reinforcement learning;
[0010] The defect evolution prediction subsystem uses capsule networks and reverse process mapping algorithms to achieve defect tracing, and realizes closed-loop data flow of each module through digital threads.
[0011] As a preferred embodiment of the present invention, the steps of running the physical data bimodal engine include:
[0012] Particle field initialization: Set the initial particle spacing according to the simulation area size. Set the initial particle spacing d according to the casting volume. cell =0.01m, then the particle radius r = d cell / 2, V calculates the initial number of particles: N = V / v cell ,in r is the particle radius;
[0013] Dynamic mesh generation: Using the improved SPH algorithm and ASTM E8 standard, the shear stress-shear rate curve of the molten metal at the pouring temperature was measured using a rotational rheometer, and the viscosity μ was obtained by fitting. The kernel function is:
[0014] where h = k·μ 1 / 3 , k = 0.5, μ is the material viscosity;
[0015] Data fusion calculation: Calculate the weights of multi-source data through the attention mechanism:
[0016] Attention mechanism weight α i The query vector Q and key vector K are calculated through the multi-head attention mechanism and generated by projection of the fully connected layer. The feature dimension d = 256 is normalized after concatenating the sensor features and the simulation features.
[0017] As a preferred embodiment of the present invention, the operation steps of the multi-source data spatiotemporal synchronization module include:
[0018] The time alignment of multi-source data spatiotemporal synchronization uses cubic spline interpolation function The weight function The sensor timestamp is synchronized by the GPS timing module, and the simulation time step Δt sim =0.1ms and the sensor data timestamp error is ≤1ms. The spatial registration uses the ICP algorithm to align the X-ray imaging point cloud with the simulation point cloud, and removes abnormal matching pairs with a spacing of >2mm. The transformation matrix R and t are solved by SVD decomposition. The maximum number of iterations is k max =50, linear regression is used to calibrate the sensor clock drift for error compensation, and the spatial coordinate deviation Δx≤0.05mm.
[0019] As a preferred embodiment of the present invention, the reinforcement learning optimization step of the dynamic process decision engine includes:
[0020] Reinforcement learning reward function of dynamic process decision engine Reinforcement learning reward function of dynamic process decision engine R(s,a)=0.6·η fill +0.3·η cool -0.1·η defectIn the middle, filling completeness η fill By volume filling rate V filled / V total Calculation, cooling uniformity η cool Take the temperature gradient The reciprocal of the defect risk index η defect The defect probability p is predicted by the capsule network defect Mapping to 1-p defect , the state space includes the pouring speed v and the mold temperature T m , where v is calculated by PLC reading the hydraulic cylinder pressure sensor WIKA CPH7000 to calculate the flow rate v = Q / A, Q is the volume flow rate, A is the gate cross-sectional area, the strategy network is updated using the PPO algorithm, and the loss function Advantage function The calculation was performed by generalized advantage estimation with a discount factor of 0.99, GAE hyperparameter λ = 0.95, KL divergence coefficient β initial value 3.0 and β new =β old ·e -0.01t attenuation.
[0021] As a preferred embodiment of the present invention, the capsule network operation steps of the defect evolution prediction subsystem include:
[0022] Iterative update protocol vector of capsule network dynamic routing protocol for defect evolution prediction subsystem The termination condition is max(b ij )>0.9 or the number of iterations reaches k max =5, low-level capsule output c i Normalization by Squash function The temperature coefficient is initialized to 3.0 and linearly decays to 0.5. The feature pyramid network extracts features from the bottom up through ResNet-50, downsampling with a step size of 2 and an output channel number of 256. The top-down path is upsampled by bilinear interpolation and spliced with horizontal connections to generate multi-scale features. The input X-ray image is histogram equalized with a contrast stretching factor of 0.5 and is input to the network after CLAHE enhancement. The real-time acoustic emission signal is decomposed by 3 layers of DB4 wavelet and then input to the capsule network.
[0023] As a preferred embodiment of the present invention, the operation steps of the digital twin self-evolution module include:
[0024] Knowledge graph construction: The TransE algorithm is used to construct the material-process relationship triple (h, r, t), and the loss function is:
[0025] Meta-learning parameter initialization: Update the task loss function through the MAML algorithm:
[0026]
[0027] As a preferred embodiment of the present invention, the steps of operating the AR-assisted interactive interface include:
[0028] Virtual-real registration: The virtual casting trajectory is registered with the real scene through the ARKit framework, with a registration error of ≤2mm;
[0029] Gesture recognition: Use the MediaPipe framework to detect hand key points and call the update view matrix when performing rotation / scaling operations.
[0030] As a preferred embodiment of the present invention, the cross-factory data collaborative federated learning step includes:
[0031] Differential privacy noise: Add Gaussian noise to the gradient parameters:
[0032] Model aggregation: FedAvg algorithm is used to calculate global model parameters:
[0033] The present invention has the following advantages: 1. The present invention constructs an intelligent control system covering the entire casting process by deeply integrating physical mechanism models and data-driven algorithms. The system adopts a dynamic multimodal engine to realize real-time interaction between simulation and measured data, and combines reinforcement learning and deep learning technologies to effectively improve the process parameter optimization efficiency and defect prediction accuracy.
[0034] 2. The present invention breaks the heterogeneous barriers of multi-source data through a spatiotemporal synchronization mechanism, establishes a high-precision digital twin model, uses capsule networks and inverse mapping algorithms to achieve defect traceability and process closed-loop feedback, significantly shortens process adjustment response time, and relies on federated learning and knowledge graph technology to complete cross-regional production data privacy protection and knowledge reuse. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 It is a structural schematic diagram of the present invention. DETAILED DESCRIPTION
[0036] Reference herein to an embodiment means that a particular feature, structure, or characteristic described in connection with the embodiment may be included in at least one embodiment of the present invention. The appearance of such a phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.
[0037] A virtual casting assistance system, integrating physical simulation with real-time data to create a closed-loop control chain, consists of four core modules: a physical data dual-modal engine, a multi-source data spatiotemporal synchronization module, a dynamic process decision engine, and a defect evolution prediction subsystem. System operation begins in the pre-pouring stage. The physical-data dual-modal engine dynamically generates an initial particle field based on material parameters and pouring volume, using an improved SPH kernel function to construct a non-uniform grid. The initial simulation results are calibrated using the multi-source data spatiotemporal synchronization module. The timestamps of sensor data (piezoelectric vibration signal sampling rate 50kHz, X-ray imaging resolution 2048×2048) and simulation data are aligned using a GPS timing module. Spatial deviations are registered using the ICP algorithm to align the X-ray point cloud with the simulated particle field. Anomalous matching pairs with a spacing greater than 2mm are eliminated, ultimately achieving a coordinate error of ≤0.05mm.
[0038] The data after time and space synchronization (such as pouring speed v = 0.8m / s, mold temperature T_m = 320℃) are input into the dynamic process decision engine, and the 18-dimensional parameters (including pouring speed, temperature gradient, etc.) are reduced by PCA. ) is compressed into 10-dimensional principal components to drive the update of the reinforcement learning policy network. The PPO algorithm updates the policy parameters based on the generalized advantage estimate. The KL divergence coefficient β is adjusted from 3.0 according to exponential decay, and finally generates adjustment instructions (such as increasing the pouring speed to 0.9m / s and increasing the cooling water flow by 15%).
[0039] While the adjusted process parameters are transmitted to the casting actuator, real-time X-ray images and acoustic emission data enter the defect evolution prediction subsystem. The capsule network iteratively updates the protocol vector via a dynamic routing protocol, locating the defect position (error ≤ 1.5mm) and inversely correcting the simulation model parameters (e.g., viscosity μ adjusted by ±5%). The system feeds the optimized model parameters back to the physical data dual-modal engine via a digital thread, forming a closed-loop chain of "simulation prediction → data calibration → decision optimization → defect feedback," supporting the dynamic regulation of the casting process and the continuous evolution of process knowledge.
[0040] Specifically, such as Figure 1 As shown, the system specifically includes:
[0041] Physical data dual-modal engine, which builds a casting process model through a dynamic SPH-DL hybrid algorithm;
[0042] Multi-source data spatiotemporal synchronization module: The multi-source data spatiotemporal synchronization module realizes the integration of equipment parameters, material characteristics and real-time monitoring data;
[0043] Dynamic process decision engine, which generates real-time optimization parameters based on reinforcement learning;
[0044] Defect evolution prediction subsystem: The defect evolution prediction subsystem uses capsule network and reverse process mapping algorithm to realize defect tracing, and realizes closed-loop flow of data of each module through digital thread.
[0045] The physical data bimodal engine operates in the following steps:
[0046] Particle field initialization: Set the initial particle spacing according to the simulation area size. Set the initial particle spacing d according to the casting volume. cell =0.01m, then the particle radius r = d cell / 2, V calculates the initial number of particles: N = V / v cell ,in r is the particle radius;
[0047] Dynamic mesh generation: Using the improved SPH algorithm and ASTM E8 standard, the shear stress-shear rate curve of the molten metal at the pouring temperature was measured using a rotational rheometer, and the viscosity μ was obtained by fitting. The kernel function is:
[0048] where h = k·μ 1 / 3 , k = 0.5, μ is the material viscosity;
[0049] Data fusion calculation: Calculate the weights of multi-source data through the attention mechanism:
[0050] Attention mechanism weight α i The query vector Q and key vector K are calculated through the multi-head attention mechanism and generated by projection of the fully connected layer. The feature dimension d = 256 is normalized after concatenating the sensor features and the simulation features.
[0051] The operation steps of the multi-source data spatiotemporal synchronization module include:
[0052] The time alignment of multi-source data spatiotemporal synchronization uses cubic spline interpolation function The weight function The sensor timestamp is synchronized by the GPS timing module, and the simulation time step Δt sim =0.1ms and the sensor data timestamp error is ≤1ms. The spatial registration uses the ICP algorithm to align the X-ray imaging point cloud with the simulation point cloud, and removes abnormal matching pairs with a spacing of >2mm. The transformation matrix R and t are solved by SVD decomposition. The maximum number of iterations is k max =50, linear regression is used to calibrate the sensor clock drift for error compensation, and the spatial coordinate deviation Δx≤0.05mm.
[0053] The reinforcement learning optimization steps of the dynamic process decision engine include:
[0054] Reinforcement learning reward function of dynamic process decision engine Reinforcement learning reward function of dynamic process decision engine R(s,a)=0.6·η fill +0.3·η cool -0.1·η defect In the middle, filling completeness η fill By volume filling rate V filled / V total Calculation, cooling uniformity η cool Take the temperature gradient The reciprocal of the defect risk index η defect The defect probability p is predicted by the capsule network defect Mapping to 1-p defect , the state space includes the pouring speed v and the mold temperature T m , where v is calculated by PLC reading the hydraulic cylinder pressure sensor WIKA CPH7000 to calculate the flow rate v = Q / A, Q is the volume flow rate, A is the gate cross-sectional area, the strategy network is updated using the PPO algorithm, and the loss function Advantage function The calculation was performed by generalized advantage estimation with a discount factor of 0.99, GAE hyperparameter λ = 0.95, KL divergence coefficient β initial value 3.0 and β new =β old ·e -0.01t attenuation.
[0055] The capsule network operation steps of the defect evolution prediction subsystem include:
[0056] Iterative update protocol vector of capsule network dynamic routing protocol for defect evolution prediction subsystem The termination condition is max(b ij )>0.9 or the number of iterations reaches k max =5, low-level capsule output c i Normalization by Squash function The temperature coefficient is initialized to 3.0 and linearly decays to 0.5. The feature pyramid network extracts features from the bottom up through ResNet-50, downsampling with a step size of 2 and an output channel number of 256. The top-down path is upsampled by bilinear interpolation and spliced with horizontal connections to generate multi-scale features. The input X-ray image is histogram equalized with a contrast stretching factor of 0.5 and is input to the network after CLAHE enhancement. The real-time acoustic emission signal is decomposed by 3 layers of DB4 wavelet and then input to the capsule network.
[0057] The operation steps of the digital twin self-evolution module include:
[0058] Knowledge graph construction: The TransE algorithm is used to construct the material-process relationship triple (h, r, t), and the loss function is:
[0059] Meta-learning parameter initialization: Update the task loss function through the MAML algorithm:
[0060]
[0061] The operation steps of the AR-assisted interactive interface include:
[0062] Virtual-real registration: The virtual casting trajectory is registered with the real scene through the ARKit framework, with a registration error of ≤2mm;
[0063] Gesture recognition: Use the MediaPipe framework to detect hand key points and call the update view matrix when performing rotation / scaling operations.
[0064] As a preferred embodiment of the present invention, the steps of cross-factory data collaborative federated learning include:
[0065] Differential privacy noise: Add Gaussian noise to the gradient parameters:
[0066] Model aggregation: FedAvg algorithm is used to calculate global model parameters:
[0067] Specifically, during the specific experimental process:
[0068] Step 1: Aluminum alloy 6061-T6 was subjected to shear stress-shear rate testing using ASTM E8 standard. The rotational rheometer was used to measure the shear stress at 680°C. The viscosity μ was fitted to 0.003 Pa\cdotps. According to the formula h = 0.5 μ 1 / 3 Calculate the smoothing kernel radius h = 0.5 (0.003) 1 / 3 =0.144mm, the simulation area size is 1m×0.5m×0.3m, and the initial particle spacing d cell =0.01m, particle radius r = d cell / 2 = 0.005m, initial number of particles N = 15000;
[0069] Step 2: During the pouring process, dynamically adjust the particle spacing. When the particle spacing d ij >2h when the particles are merged. ij When <h / 2, the particles are split and the kernel function weight W(r,h) is calculated according to the formula, where r is the distance between the two particles and h = 0.144 mm. For example, when r = 0.1 mm, W = 0.023.
[0070] Step 3: Based on Courant stability conditions Where ρ = 2700 kg / m3 (density of aluminum alloy), g = 9.81 m / s2, calculation Δt = 0.0002 s, total simulation time T = 10 s, and 50,000 iteration steps;
[0071] Step 4: The vibration signal is collected by a piezoelectric sensor with a sampling rate of fs = 50 kHz and stored as a time series file. The casting is scanned by an X-ray imaging device with a resolution of 2048 × 2048 and a voxel size of 0.1 mm to generate a DICOM format image.
[0072] Step 5: The GPS timing module provides a unified time base for the sensor and simulation (accuracy ±1μs). The sensor data timestamps [1620000000.123, 1620000000.124, ...], simulation times = [1620000000.120, 1620000000.121, ...], are aligned using a cubic spline interpolation function. The weight function, at t = 1620000000.122, has interpolation weights w1 = 0.7, w2 = 0.3, and outputs y(t) = 0.7 y1 + 0.3 y2.
[0073] Step 6: The X-ray point cloud (500,000 points) and the simulated point cloud (100,000 points) are registered using the ICP algorithm. Abnormal matching pairs with a distance > 2 mm are eliminated. After 50 iterations, the coordinate deviations Δx = 0.03 mm, Δy = 0.02 mm, and Δz = 0.01 mm. The transformation matrix R (rotation) and the translation vector t are solved by SVD decomposition.
[0074] Step 7: The original state parameters include 18-dimensional data such as pouring speed (0.1-1.5 m / s) and mold temperature (200-600°C), which are reduced to 10-dimensional principal components through PCA;
[0075] Step 8: X-ray images (5000 images, 2048 × 2048 pixels) were histogram equalized (contrast stretch factor γ = 0.5) and enhanced with CLAHE (window 8 × 8, clip limit = 0.03). Real-time acoustic emission signals (NIPXIe-4497 acquisition card) were decomposed using the db4 wavelet 3-layer method to extract frequency domain energy features.
[0076] Step 9: The capsule network input is a 64-channel feature map, the output vector of the lower-level capsule i is normalized by the Squash function, and the dynamic routing is iterated 5 times;
[0077] Step 10: The feature pyramid network generates multi-scale features through upsampling and horizontal connection splicing, and finally locates the error.
[0078] The above embodiments are provided to persons familiar with the art for implementing or using the present invention. Personnel familiar with the art may make various modifications or changes to the above embodiments without departing from the spirit of the present invention. Therefore, the scope of protection of the present invention is not limited to the above embodiments, but should be the maximum scope of the innovative features mentioned in the claims.
Claims
1. A casting virtual assistance system, characterized in that: Includes: A physical data bimodal engine that constructs a casting process model using a dynamic SPH-DL hybrid algorithm; A multi-source data spatiotemporal synchronization module that integrates equipment parameters, material properties, and real-time monitoring data; A dynamic process decision engine that generates real-time optimization parameters based on reinforcement learning; The defect evolution prediction subsystem uses capsule networks and reverse process mapping algorithms to achieve defect tracing, and realizes closed-loop data flow of each module through digital threads.
2. A casting virtual assistance system according to claim 1, characterized in that: The operation steps of the physical-data dual-modal engine include: Particle field initialization: Set the initial particle spacing according to the simulation area size. Set the initial particle spacing d according to the casting volume. cell =0.01m, then the particle radius r = d cell / 2, V calculates the initial number of particles: N = V / v cell ,in r is the particle radius; Dynamic mesh generation: Using the improved SPH algorithm and ASTM E8 standard, the shear stress-shear rate curve of the molten metal at the pouring temperature was measured using a rotational rheometer, and the viscosity μ was obtained by fitting. The kernel function is: where h = k·μ 1 / 3 , k = 0.5, μ is the material viscosity; Data fusion calculation: Calculate the weights of multi-source data through the attention mechanism: Attention mechanism weight α i The query vector Q and key vector K are calculated through the multi-head attention mechanism and generated by projection of the fully connected layer. The feature dimension d = 256 is normalized after concatenating the sensor features and the simulation features.
3. A casting virtual assistance system according to claim 1, characterized in that: The operation steps of the multi-source data spatiotemporal synchronization module include: The time alignment of multi-source data spatiotemporal synchronization uses cubic spline interpolation function The weight function The sensor timestamp is synchronized by the GPS timing module, and the simulation time step Δt sim =0.1ms and the sensor data timestamp error is ≤1ms. The spatial registration uses the ICP algorithm to align the X-ray imaging point cloud with the simulation point cloud, and removes abnormal matching pairs with a spacing of >2mm. The transformation matrix R and t are solved by SVD decomposition. The maximum number of iterations is k max =50, linear regression is used to calibrate the sensor clock drift for error compensation, and the spatial coordinate deviation Δx≤0.05mm.
4. A casting virtual assistance system according to claim 1, characterized in that: The reinforcement learning optimization steps of the dynamic process decision engine include: Reinforcement learning reward function of dynamic process decision engine Reinforcement learning reward function of dynamic process decision engine R(s,a)=0.6·η fill +0.3·η cool -0.1·η defect In the middle, filling completeness η fill By volume filling rate V filled / V total Calculation, cooling uniformity η cool Take the temperature gradient The reciprocal of the defect risk index η defect The defect probability p is predicted by the capsule network defect Mapping to 1-p defect , the state space includes the pouring speed v and the mold temperature T m , where v is calculated by PLC reading the hydraulic cylinder pressure sensor WIKACPH7000 to calculate the flow rate v = Q / A, Q is the volume flow rate, A is the gate cross-sectional area, the strategy network is updated using the PPO algorithm, and the loss function Advantage function The calculation was performed by generalized advantage estimation with a discount factor of 0.99, GAE hyperparameter λ = 0.95, KL divergence coefficient β initial value 3.0 and β new =β old ·e -0.01t attenuation.
5. A casting virtual assistance system according to claim 1, characterized in that: The capsule network operation steps of the defect evolution prediction subsystem include: Iterative update protocol vector of capsule network dynamic routing protocol for defect evolution prediction subsystem The termination condition is max(b ij )>0.9 or the number of iterations reaches k max =5, low-level capsule output c i Normalization by Squash function The temperature coefficient is initialized to 3.0 and linearly decays to 0.
5. The feature pyramid network extracts features from the bottom up through ResNet-50, downsampling with a step size of 2 and an output channel number of 256. The top-down path is upsampled by bilinear interpolation and spliced with horizontal connections to generate multi-scale features. The input X-ray image is histogram equalized with a contrast stretching factor of 0.5 and is input to the network after CLAHE enhancement. The real-time acoustic emission signal is decomposed by 3 layers of DB4 wavelet and then input to the capsule network.
6. A casting virtual assistance system according to claim 1, characterized in that: The operation steps of the digital twin self-evolution module include: Knowledge graph construction: The TransE algorithm is used to construct the material-process relationship triple (h, r, t), and the loss function is: Meta-learning parameter initialization: Update the task loss function through the MAML algorithm:
7. A casting virtual assistance system according to claim 1, characterized in that: The operation steps of the AR auxiliary interaction interface include: Virtual-real registration: The virtual casting trajectory is registered with the real scene through the ARKit framework, with a registration error of ≤2mm; Gesture recognition: Use the MediaPipe framework to detect hand key points and call the update view matrix when performing rotation / scaling operations.
8. A casting virtual assistance system according to claim 1, characterized in that: The steps of cross-factory data collaborative federated learning include: Differential privacy noise: Add Gaussian noise to the gradient parameters: Model aggregation: FedAvg algorithm is used to calculate global model parameters: