Vacuum isothermal forging mechanical hand digital twin monitoring method and device based on physical constraint neural network

Abstract

A vacuum isothermal forging mechanical hand digital twin monitoring method and device based on physical constraint neural network, the main steps include: combining the collectable data with the equipment mechanism model, using neural network and mechanism model to obtain the data of difficult monitoring position, optimizing and correcting the simulation data of the whole equipment, obtaining the temperature and stress change of the whole structure of the mechanical hand in different working states, and mapping it to the virtual twin body of the mechanical hand; taking the minimum contact distance error and the maximum clamping stiffness prediction accuracy as the objective function, setting the reasonable range of the mechanical hand parameters based on the historical data or expert experience, and obtaining the recommended parameters through the Gaussian process modeling and the expected improvement method. The method can establish a more accurate vacuum isothermal forging mechanical hand digital twin model, and give the work parameter recommendation of carrying new products, which can effectively shorten the generation cycle and adapt to small batch and multi variety production type.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent operation and maintenance of high-end equipment, and particularly relates to a vacuum isothermal forging manipulator digital twin monitoring method based on a physically constrained neural network. BACKGROUND

[0002] The processing of aerospace equipment components requires precise control of material organization and performance. However, many key components have materials with narrow processing parameter range, large deformation resistance, and organization and performance that are very sensitive to hot processing. It is difficult to obtain ideal forming results and organization and performance using traditional forging processes. Therefore, the metal blank and the mold need to be heated to the same temperature (1000-1200℃) in a fully enclosed vacuum environment to reduce the deformation resistance of the material, avoid the generation of oxidation skin, and improve the plasticity of the material, so that the material can be processed into a product with uniform organization and excellent performance.

[0003] The material transfer manipulator, which transfers forging products between different chambers (forging chamber, heating chamber, vacuum channel, and in-out chamber) of the vacuum isothermal forging equipment, is the core functional component of the vacuum isothermal forging equipment. Its particularity lies in the need to complete high-precision (±2mm) material transfer in a vacuum, high-temperature, periodic thermal shock, and heavy-load working environment. Its working state directly affects the production efficiency and working process. The repeated positioning accuracy of the manipulator decreases and its stability decreases due to periodic thermal shock and heat accumulation caused by continuous operation. In addition, due to the harsh working environment, sensors cannot be installed in the ideal position, and the state sensing of the manipulator is under test.

[0004] Digital twin technology is an advanced technology that realizes virtual-real space interaction, breaks through the limitations of traditional simulation and test conditions, and can master the actual running state of equipment at the fastest speed and optimal cost. However, for hot processing equipment, the digital twin model not only needs to reflect its inherent shape and working state, but also needs to reflect its mechanical properties and temperature field data, so as to become a powerful tool for studying the reliability of hot processing equipment. However, the existing equipment digital twin reflects the mechanical properties and temperature field, and the simulation data is often displayed, making it difficult to intuitively feel the structural mechanical property information of the equipment during operation. Therefore, it is necessary to monitor the state at multiple key positions, preprocess the measured data, and use the processed results to optimize and correct the simulation results, so as to realize effective monitoring of the equipment state. SUMMARY

[0005] The present application provides a vacuum isothermal forging manipulator digital twin monitoring method based on a physically constrained neural network, which uses collected data and a mechanism model of the manipulator to obtain inference data, and then optimizes and calibrates simulation data, and maps it to a digital twin body to realize visual monitoring of the running state of the manipulator.

[0006] At the same time, due to the vacuum isothermal forging equipment often produces multiple batches of small batch products, which makes the manipulator prone to positioning accuracy attenuation after carrying several products, and needs to reset the working parameters; if the new product is carried from the forging chamber to the feeding chamber, in order to ensure the accuracy of the carrying, it is often necessary to wait for the product to cool down before carrying, which greatly increases the production time. In order to solve this problem, the method provided by the present disclosure will combine the historical data of the manipulator carrying the product (including characteristic parameters, working parameters and sensor data), minimize the contact distance error and maximize the clamping stiffness prediction accuracy as the objective function, set a reasonable range of manipulator position, speed and other parameters based on historical data or expert experience, and obtain recommended parameters through Gaussian process modeling and expected improvement method to adapt to the production type of small batch and multiple varieties.

[0007] The vacuum isothermal forging manipulator digital twin monitoring method based on the physical constraint neural network provided by the present disclosure mainly includes the following steps:

[0008] S1, according to the structure and motion trajectory of the manipulator, determine the key placement position of the stress sensor and the temperature sensor, and arrange the sensors to collect the stress and temperature data of the manipulator in different working states and different positions;

[0009] S2, set the simulation parameters according to the working state and process parameters, and obtain the stress simulation data of the manipulator in different working states and the temperature simulation data in different positions;

[0010] S3, for the parts that cannot directly place sensors, establish a heat transfer model and a thermal-mechanical coupling mechanism model of the manipulator, and use PINN (Physics-Informed Neural Networks, Physics-Informed Neural Networks) network to infer the data of the difficult-to-monitor position through the sensor data of the known position;

[0011] S4, map the sensor point coordinates to the simulation grid nodes, and optimize and correct the simulation results of the equipment in different working states through the data in different working states, including the directly collected data and the data inferred by the PINN network;

[0012] S5, use the RBF (Radial Basis Function, Radial Basis Function) proxy model to map the optimized simulation results to the virtual twin according to the running state of the manipulator in real time.

[0013] Further, the step S3 specifically includes:

[0014] A heat conduction mechanism model is established by Fourier's law, taking the material thermal conductivity and the heat dissipation coefficient as the core parameters:

[0015]

[0016] Where T(x,t) is the temperature field, ρ is the material density, and c p λ is the specific heat capacity, k(T) is the nonlinear thermal conductivity, P is the contact heat power, A is the contact area, λ is the attenuation coefficient; x is the spatial coordinate; t is the time; d is the distance. Represents the spatial rate of change of the temperature field;

[0017] Simultaneously, by combining the thermal expansion effect and mechanical load to calculate stress and strain, a thermo-mechanical coupling mechanism model is established;

[0018] σ=c:(ε-α(T)ΔTI)

[0019]

[0020] Where σ is the stress tensor; ε is the strain tensor; u is the displacement field; C is the elastic stiffness matrix; and α(T) is the coefficient of thermal expansion.

[0021] By utilizing PINN networks and mechanistic models, data from hard-to-monitor locations can be acquired through data collected by sensors.

[0022] Furthermore, it also includes the following steps:

[0023] S6 records the robot arm parameters, product characteristics (geometric and physical features), and temperature and vibration data for each product handling operation into a database to establish a parameter database. With the objective function of minimizing contact distance error and maximizing the prediction accuracy of clamping stiffness, reasonable ranges for robot arm position and speed parameters are set based on historical data or expert experience. Recommended parameters are obtained through Gaussian process modeling and expected improvement methods, and the operation is pre-performed in the robot arm digital twin system to reduce physical trial and error costs.

[0024] Furthermore, step S6 specifically includes:

[0025] (1) Collect data from temperature and vibration sensors when the robot is handling products; use the temperature data to infer the contact distance between the robot and the workpiece through a PINN network.

[0026] The PINN architecture is designed as follows: Input layer: time, clamping end temperature, and sensor temperature; Hidden layer: CompleX-PINN architecture using complex variable activation functions, reconstructing the output space through the Cauchy integral formula; Output layer: Contact distance; Non-contact area temperature conforms to the natural convection cooling model, serving as a boundary constraint; Loss function includes residual terms from the transient heat conduction equation.

[0027]

[0028] Where ρ is the material density, c p Where is the specific heat capacity, and k(T) is the nonlinear thermal conductivity; Represents the spatial rate of change of the temperature field;

[0029] (2) Obtain the clamping stiffness using vibration sensor data, combine the changes in vibration sensor data during the process of the robot handling the product, capture the time nodes of data mutation, and establish a mapping model between clamping stiffness and signal root mean square value.

[0030] (3) Taking the minimization of contact distance error and the maximization of clamping stiffness prediction accuracy as the objective function, a unified objective function is constructed using the weighted sum method. At the same time, based on historical data or expert experience, a reasonable range of robot position and speed parameters is set.

[0031] (4) Use Gaussian process to model the relationship between parameters and objective function, select RBF kernel + white noise kernel to capture nonlinear relationship, and select the next set of parameter combination through expectation improvement;

[0032] (5) Use the recommended parameters of the robot arm to perform a pre-operation of building a twin in Unity3D software to reduce the cost of physical trial and error and adapt to small-batch, multi-variety production types.

[0033] A digital twin monitoring device for a vacuum isothermal forging robot based on a physical constraint neural network, applying the above method, mainly includes:

[0034] Data acquisition module: includes stress sensor and temperature sensor, used to collect stress and temperature data of the robot arm under different working conditions and at different positions;

[0035] Data inference module: For parts where sensors cannot be directly placed, based on the robot's heat transfer model and thermo-coupling mechanism model, it uses PINN network to infer data for difficult-to-monitor locations using sensor data at known locations;

[0036] Simulation data calibration module: used to set simulation parameters according to working conditions and process parameters, and obtain stress simulation data and temperature simulation data of the robot under different working conditions; map sensor point coordinates to simulation mesh nodes, and optimize and calibrate the simulation results of the equipment under different working conditions by using data from different working conditions, including directly acquired data and data obtained by inference using PINN network;

[0037] State mapping module: Used to map the optimized simulation results to the virtual twin in real time based on the operating state of the robot arm using the RBF proxy model;

[0038] Robotic Arm Parameter Recommendation Module: This module records the robotic arm parameters, product characteristics, and temperature and vibration data for each product handling operation into a database, establishes a parameter map, and sets the reasonable range for the robotic arm position and speed parameters based on historical data or expert experience. Recommended parameters are obtained through Gaussian process modeling and desired improvement methods.

[0039] Robotic arm motion pre-simulation module: A twin is built in Unity3D software, recommended process parameters are obtained using the robotic arm parameter recommendation module, and the operation is pre-simulated in the robotic arm digital twin system to reduce physical trial and error costs and adapt to small-batch, multi-variety production types.

[0040] Compared with existing technologies, the beneficial effects of this disclosure are: ① By collecting sensor data from multiple positions under different working conditions of the equipment, and by using sensor data from known positions and mechanistic models to infer data from difficult-to-monitor positions, the simulation data of the equipment is optimized and corrected using actual data and inferred data, and mapped to a digital twin, which allows equipment maintenance personnel to have a more intuitive understanding of the robot's operating status; ② With minimizing contact distance error and maximizing clamping stiffness prediction accuracy as objective functions, and based on historical data or expert experience, reasonable ranges for parameters such as robot position and speed are set, and recommended parameters are obtained through Gaussian process modeling and expected improvement methods, which can effectively shorten the production cycle; ③ It is suitable for small-batch, multi-variety production types. Attached Figure Description

[0041] The above and other objects, features and advantages of this disclosure will become more apparent from the more detailed description of exemplary embodiments of this disclosure taken in conjunction with the accompanying drawings, in which the same reference numerals generally represent the same components.

[0042] Figure 1 This is a flowchart illustrating an exemplary embodiment of the present disclosure. Detailed Implementation

[0043] Preferred embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While preferred embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure will be thorough and complete, and will fully convey the scope of the present disclosure to those skilled in the art.

[0044] This disclosure provides a digital twin monitoring method for a vacuum isothermal forging robot based on a physically constrained neural network. This method is a multimodal state monitoring and adaptive process parameter recommendation method for material handling robots in vacuum isothermal forging equipment. By deploying different types of sensors at multiple locations on the robot, monitoring data of the robot under different working states is collected. Deep learning constrained by a mechanistic model is used to infer monitoring data for locations where sensors are difficult to deploy, optimizing the simulation results of the equipment under different working states, and mapping them to a digital twin to achieve visualized monitoring of the robot's operating state. Then, the recommended working parameters for the robot to handle different products are obtained using the data.

[0045] The flowcharts of exemplary embodiments according to this disclosure are attached. Figure 1 As shown, the main steps include:

[0046] (1) Based on the structure and motion trajectory of the robot, determine the key placement positions of the stress sensor and temperature sensor, and collect stress and temperature data of the robot under different working conditions and at different positions;

[0047] Simulation parameters are set according to working conditions and process parameters to obtain stress simulation data and temperature simulation data of the robot under different working conditions and at different positions.

[0048] (2) For locations where sensors cannot be placed directly:

[0049] A heat conduction mechanism model with material thermal conductivity and heat dissipation coefficient as core parameters is established using Fourier's law;

[0050]

[0051] Where T(x,t) is the temperature field, ρ is the material density, and c p λ is the specific heat capacity, k(T) is the nonlinear thermal conductivity, P is the contact heat power, A is the contact area, λ is the attenuation coefficient; x is the spatial coordinate; t is the time; d is the distance. Represents the spatial rate of change of the temperature field;

[0052] Simultaneously, a thermo-mechanical coupling mechanism model is established by combining the thermal expansion effect and mechanical load to calculate stress and strain.

[0053] σ=C:(ε-α(T)ΔTI)

[0054] Where σ is the stress tensor; ε is the strain tensor; u is the displacement field; C is the elastic stiffness matrix; α(T) is the coefficient of thermal expansion; I is the second-order unit tensor, representing the isotropic nature of thermal expansion, ensuring that the strain caused by temperature changes is equal in magnitude in all directions;

[0055] By utilizing PINN networks and mechanistic models, data from hard-to-monitor locations can be acquired through data collected by sensors.

[0056] (3) Map the sensor point coordinates to the simulation grid nodes, and optimize the simulation results of the calibration device under different working conditions by using data from different working conditions (including: directly acquired data and data obtained by inference using PINN network);

[0057] (4) Using the RBF proxy model, the optimized simulation results are mapped in real time to the simulation model under different states as a virtual twin of the robot, so as to realize the visualization monitoring of the robot's operating status.

[0058] (5) Collect temperature and vibration sensor data when the robot is handling the product. Use the temperature data to infer the contact distance between the robot and the product through the PINN network. At the same time, capture the time nodes of sudden changes in vibration sensor data, establish a mapping model between clamping stiffness and signal root mean square value, and perform pre-operation in a virtual twin environment to reduce physical trial and error costs.

[0059] (6) Recommended process parameters for small-batch, multi-variety production.

[0060] Application Examples

[0061] Using robotic arms to handle high-temperature forgings:

[0062] 1. Install strain gauges and temperature sensors at the transmission and easily deformable positions of the robotic arm;

[0063] 2. Using sensor data at known locations as nodes, and combining the heat transfer mechanism model with the thermo-mechanical coupling equation, the stress and temperature inference results for difficult-to-monitor parts are obtained by using PINN output.

[0064] 3. Simulation was performed using the weight of the forging, temperature, movement speed of the robot, and stiffness as initial parameters. 100 sets of simulation data and 20 sets of measured data were collected. The simulation results were optimized and corrected using a residual learning network based on physical constraints.

[0065] 4. Construct a twin in Unity3D software, use the RBF proxy model to map the simulation results onto the twin, and synchronize the position, stress, and temperature data of the robot arm in real time;

[0066] 5. Record the robot arm parameters, product characteristics (geometric and physical features), temperature, and vibration data for each product handling operation into the database and establish a parameter map. Using minimizing contact distance error and maximizing the accuracy of clamping stiffness prediction as objective functions, and based on historical data or expert experience, set reasonable ranges for robot arm position, speed, and other parameters. Obtain recommended parameters through Gaussian process modeling and desired improvement methods. Simulate 10 handling processes in Unity3D to determine the reliability of the recommended parameters. Once the success rate reaches 100%, the parameters are sent to the physical robot arm.

[0067] In this embodiment, based on the structure and working nature of the equipment, key positions are determined. Multiple different types of sensors are used to collect data on the manipulator at multiple key positions under different working states. The collected data and a neural network with physical constraints are used to optimize and correct the simulation results, obtaining the temperature and stress changes of the entire manipulator structure under different working states. A surrogate model is used to map these changes onto a virtual twin of the manipulator. With minimizing contact distance error and maximizing clamping stiffness prediction accuracy as objective functions, reasonable ranges for parameters such as manipulator position and speed are set based on historical data or expert experience. Recommended parameters are obtained through Gaussian process modeling and desired improvement methods.

[0068] The above technical solutions are merely exemplary embodiments of the present invention. For those skilled in the art, based on the application methods and principles disclosed in the present invention, it is easy to make various types of improvements or modifications, and not limited to the methods described in the specific embodiments of the present invention. Therefore, the methods described above are merely preferred and not restrictive.

Claims

1. A digital twin monitoring method for a vacuum isothermal forging robot based on a physical constraint neural network, characterized in that, Includes the following steps: S1. Based on the structure and motion trajectory of the robot, determine the key placement positions of the stress sensor and temperature sensor, and arrange the sensors to collect stress and temperature data of the robot under different working conditions and at different positions. S2, set simulation parameters according to working conditions and process parameters to obtain stress simulation data of the robot under different working conditions and temperature simulation data at different positions; S3. For parts where sensors cannot be directly placed, establish a heat transfer model and a thermo-coupling mechanism model for the robotic arm. Using sensor data from known locations, infer data from difficult-to-monitor locations using a PINN network. S4 maps the sensor point coordinates to the simulation grid nodes. By using data from different operating states, including directly acquired data and data obtained through PINN network inference, the simulation results of the calibration device under different operating states are optimized and corrected. S5 uses the RBF proxy model to map the optimized simulation results into the virtual twin in real time according to the operating status of the robot arm.

2. The method according to claim 1, characterized in that, It also includes the following steps: S6 records the robot arm parameters, product characteristics, and temperature and vibration data for each product handling operation into the database to establish a parameter database. With the objective function of minimizing contact distance error and maximizing the prediction accuracy of clamping stiffness, reasonable ranges for robot arm position and speed parameters are set based on historical data or expert experience. Recommended parameters are obtained through Gaussian process modeling and expected improvement methods. The operation is then rehearsed in the robot arm digital twin system to reduce physical trial and error costs.

3. A digital twin monitoring device for a vacuum isothermal forging robot based on a physical constraint neural network, using the method described in claim 1 or 2, characterized in that, include: Data acquisition module: It includes stress sensors and temperature sensors, used to collect stress and temperature data of the robot arm under different working conditions and at different positions; Data inference module: For parts where sensors cannot be directly placed, based on the robot's heat transfer model and thermo-coupling mechanism model, it uses PINN network to infer data for difficult-to-monitor locations using sensor data at known locations; Simulation data calibration module: used to set simulation parameters according to working conditions and process parameters, and obtain stress simulation data and temperature simulation data of the robot under different working conditions; map sensor point coordinates to simulation mesh nodes, and optimize and calibrate the simulation results of the equipment under different working conditions by using data from different working conditions, including directly acquired data and data obtained by inference using PINN network; State mapping module: Used to map the optimized simulation results to the virtual twin in real time based on the operating state of the robot arm using the RBF proxy model; Robotic Arm Parameter Recommendation Module: This module records the robotic arm parameters, product characteristics, and temperature and vibration data for each product handling operation into a database to create a parameter map. It uses minimizing contact distance error and maximizing the accuracy of clamping stiffness prediction as objective functions, while setting reasonable ranges for robotic arm position and speed parameters based on historical data or expert experience. Recommended parameters are obtained through Gaussian process modeling and desired improvement methods. Robotic arm motion pre-simulation module: A twin is built in Unity3D software, recommended process parameters are obtained using the robotic arm parameter recommendation module, and the operation is pre-simulated in the robotic arm digital twin system to reduce physical trial and error costs and adapt to small-batch, multi-variety production types.

Images