Method and device for predicting flow field data of pusher kiln based on digital twin

Abstract

The present invention discloses a method, device, electronic device and storage medium for predicting the flow field data of a pusher kiln based on digital twin. The method includes obtaining the CFD flow field data of the pusher kiln; performing dimensionality reduction processing on the CFD flow field data to obtain low-dimensional data; training an initial flow field prediction model based on the low-dimensional data to obtain a trained flow field prediction model; predicting the operation data of the pusher kiln collected in real time based on the trained flow field prediction model to obtain base state prediction data; and reconstructing based on the base state prediction data and the low-dimensional data to predict the flow field data of the pusher kiln. This method can provide rapid response prediction at the production site to obtain the flow field data of the pusher kiln.

Description

Technical Field

[0001] The present invention belongs to the field of data processing, and relates to a method, device, electronic device and storage medium for predicting the flow field data of a pusher kiln based on digital twin. Background Art

[0002] A pusher kiln is a continuous heating and sintering device. By arranging the required temperature according to the process requirements of the sintered product to meet the thermal demand of the product, its working process is a typical multi-disciplinary coupling process, involving relevant theoretical knowledge such as physics, mechanics and heat transfer. The states of the temperature field and pressure field inside the pusher kiln have important application values for the remote monitoring and real-time adjustment of production. Furnace kiln failures will cause huge losses in productivity and profits. However, current manufacturing industries usually only determine whether the pusher kiln is operating normally by measuring the temperature at limited points through sensors, which is very unintuitive, has a low degree of digitization, and cannot represent the flow field distribution inside the entire pusher kiln. The concept of digital twin provides an effective technical tool for the modeling, analysis and monitoring of pusher kilns.

[0003] Digital twin is essentially a simulation model of a physical object. Different from static simulation, it can evolve in real time by receiving data from the physical object, so as to remain consistent with the physical object throughout its life cycle. Digital twin - an exact virtual copy of a machine or system - is revolutionizing industries. In the prior art, Wang et al. constructed a digital twin five-dimensional model taking a die-cutting machine as an example, connecting isolated machines to an interconnected system to display the real-time state of the machines; Xie et al. proposed a digital twin-driven data flow framework for each state of the tool life cycle to ensure the process and potential for continuous improvement of the tool.

[0004] High-precision simulation is the basis for realizing digital twin. With the development of computer technology and numerical calculation technology, computational fluid dynamics (CFD) provides an alternative method for studying fluid flow in complex systems and is widely used in many fields, including aerospace engineering, environmental science, biomedicine and industrial design. Refaey et al. used the ANSYS FLUENT software package to simulate the convective heat and fluid flow in the cooling zone of a tunnel kiln. Compared with physical manufacturing systems, the cost and safety of CFD are significantly reduced, and it is an important tool for solving early design and engineering problems of products and production development. However, the numerical simulation of this type of solution has a large amount of calculation, and the high-fidelity modeling and simulation of important equipment, high-confidence simulation prediction and high-real-time data interaction are challenges in the process of establishing digital twin during the production process. The computational process of high-precision numerical simulation requires a large amount of computing resources, especially when the number of flow fields to be solved is huge. In engineering, the time consumption is usually unacceptable.

[0005] In recent years, the combination of CFD and machine learning has promoted the development of CFD, bringing new perspectives and research methods to this field, endowing the field with new vitality and good development prospects. Machine learning requires a large amount of data. The combination of CFD and high-performance computers for a large number of calculations to form a database and then perform modeling is a relatively large impetus to CFD technology. Gockel et al. established a CFD model of an aircraft using flight parameters recorded during flight to evaluate the aircraft's ability to meet mission requirements; Sun et al. used an airfoil database to train an artificial neural network to achieve data-driven modeling and rapid inverse design of airfoils. Roberto Molinaro et al. used polynomial regression, multivariate adaptive regression splines, random forests, and fully connected neural networks for data-driven modeling of simulated digital twins; Gianmarco Aversano et al. used the POD-Kriging method to construct a digital twin furnace operating under flameless combustion conditions to predict combustion data under new operating conditions. However, in the emerging digital twin, it is necessary to rely on means including simulation, actual measurement, and data analysis to perceive the state of physical entities and monitor equipment in real time. The existing combination schemes of CFD and machine learning cannot achieve the effect of perceiving the state of physical entities and monitoring equipment in real time. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method for predicting the flow field data of a pusher kiln based on digital twin, which can quickly and accurately predict the flow field data of the digital twin pusher kiln.

[0007] The technical solution of the invention is as follows:

[0008] In a first aspect, a method for predicting the flow field data of a pusher kiln based on digital twin is proposed, including:

[0009] Obtain the CFD flow field data of the pusher kiln;

[0010] Perform dimensionality reduction processing on the CFD flow field data to obtain low-dimensional data;

[0011] Train an initial flow field prediction model based on the low-dimensional data to obtain a trained flow field prediction model;

[0012] Predict the operating data of the pusher kiln collected in real time based on the trained flow field prediction model to obtain base state prediction data;

[0013] Reconstruct based on the base state prediction data and the low-dimensional data to predict the flow field data of the pusher kiln.

[0014] Further, the obtaining of the CFD flow field data of the pusher kiln includes:

[0015] Obtain the geometric data of the pusher kiln and obtain the historical data during the operation of the pusher kiln;

[0016] Simulate the physical process of the pusher kiln based on the geometric data and the historical data to generate the initial CFD flow field data of the pusher kiln;

[0017] Post-process the initial CFD flow field data to obtain the CFD flow field data.

[0018] Further, training the initial flow field prediction model based on the low-dimensional data to obtain a trained flow field prediction model includes:

[0019] Standardize the low-dimensional data and preprocess the standardized low-dimensional data to obtain preprocessed data;

[0020] Extract features from the preprocessed data based on multi-scale convolution to obtain multi-scale features;

[0021] Train the initial flow field prediction model based on the multi-scale features to obtain a trained flow field prediction model.

[0022] Further, extracting multi-scale features from the preprocessed data based on multi-scale convolution includes:

[0023] Fuse the features output by the previous multi-scale convolution layer in the multi-scale convolution with the input features through skip connections to obtain fused features;

[0024] Use the fused features as the input features of the next multi-scale convolution layer in the multi-scale convolution to obtain the multi-scale features.

[0025] Further, the multi-scale convolution includes 3 multi-scale convolution layers, and each multi-scale convolution layer contains three kernel scales of 1*1, 3*1, and 5*1.

[0026] Further, reducing the dimension of the CFD flow field data to obtain low-dimensional data includes:

[0027] Perform standardization processing on the CFD flow field data to obtain standardized data;

[0028] Extract the eigenvalues and covariance matrix of the standardized data;

[0029] Generate a feature matrix based on the eigenvectors of the eigenvalues, and obtain the low-dimensional data based on the feature matrix and the covariance matrix.

[0030] Further, reconstructing based on the ground state prediction data and the low-dimensional data to predict the flow field data of the pusher kiln, including:

[0031] Obtain the pseudo-inverse matrix of the feature matrix;

[0032] Based on the pseudo-inverse matrix, the ground state prediction data, and the eigenvectors of the eigenvalues, perform reconstruction to predict the flow field data of the pusher kiln.

[0033] In a second aspect, a flow field data prediction model for a pusher kiln based on digital twin is proposed, including:

[0034] An operating data acquisition module configured to acquire the CFD flow field data of the pusher kiln;

[0035] A low-dimensional data acquisition module configured to perform dimensionality reduction processing on the CFD flow field data to obtain low-dimensional data;

[0036] A model training module configured to train an initial flow field prediction model based on the low-dimensional data to obtain a trained flow field prediction model;

[0037] An initial prediction module configured to predict the operating data of the pusher kiln collected in real time based on the trained flow field prediction model to obtain ground state prediction data;

[0038] A prediction data acquisition module configured to reconstruct based on the ground state prediction data and the low-dimensional data to predict the flow field data of the pusher kiln.

[0039] In a third aspect, an electronic device is proposed, including: one or more processors;

[0040] A storage device for storing one or more computer programs, which, when executed by the one or more processors, cause the electronic device to implement the above-mentioned method.

[0041] In a fourth aspect, a computer-readable storage medium is proposed, on which computer-readable instructions are stored, which, when executed by a processor of a computer, cause the computer to execute the above-mentioned method

[0042] In the technical solution provided by the embodiments of the present application, the flow field data of the pusher kiln can be predicted with fast response at the production site, the calculation speed is accelerated, the CFD calculation cost is saved, and at the same time, a high-precision predicted flow field of the pusher kiln can be obtained to realize the real-time evaluation of the operating state of the pusher kiln.

[0043] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present application. Description of the Drawings

[0044] The drawings herein are incorporated into and constitute a part of this specification, showing embodiments consistent with the present application, and are used together with the specification to explain the principles of the present application. Obviously, the drawings in the following description are only some embodiments of the present application, and those of ordinary skill in the art can obtain other drawings based on these drawings without creative efforts. In the drawings:

[0045] Figure 1 is a flowchart of a method for predicting the flow field data of a pusher kiln based on digital twin shown in this embodiment;

[0046] Figure 2 is an overall framework diagram of the digital twin shown in this embodiment;

[0047] Figure 3 is a workflow diagram during the operation of the virtual reality twin layer shown in this embodiment;

[0048] Figure 4 is a framework diagram of the neural network model shown in this embodiment;

[0049] Figure 5 is a structural diagram of the flow field prediction model shown in this embodiment;

[0050] Figure 6 is a structural diagram of the MCNN shown in this embodiment;

[0051] Figure 7(a) is a geometric structure diagram of the internal space of the pusher kiln shown in this embodiment;

[0052] Figure 7(b) is a front view of the pusher kiln shown in this embodiment;

[0053] Figure 7(c) is a side view of the pusher kiln shown in this embodiment;

[0054] Figure 8 is a relationship diagram between the basic mode number and energy shown in this embodiment;

[0055] Figure 9 is a relationship diagram between the basic mode number and the reconstruction error shown in this embodiment;

[0056] Figure 10 is a comparison diagram between the POD-MCNN prediction result and the CFD result of the temperature trend under the test set shown in this embodiment;

[0057] Figure 11 is a comparison diagram between the POD-MCNN prediction result and the CFD result of the pressure trend under the test set shown in this embodiment

[0058] Figure 12It is a comparison chart between the POD-MCNN prediction results and the CFD results of the temperature trend under new data conditions shown in this embodiment;

[0059] Figure 13 It is a comparison chart between the POD-MCNN prediction results and the CFD results of the pressure trend under new data conditions shown in this embodiment

[0060] Figure 14 It is a structural diagram of a push plate kiln flow field data prediction device based on digital twin shown in this embodiment. Detailed implementation manners

[0061] Here, the exemplary embodiments will be described in detail, and the examples are shown in the drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementation manners described in the following exemplary embodiments do not represent all the implementation manners consistent with the present application. On the contrary, they are only examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0062] The block diagrams shown in the drawings are only functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor devices and / or microcontroller devices.

[0063] The flowcharts shown in the drawings are only exemplary descriptions and do not necessarily include all the contents and operations / steps, nor do they necessarily need to be executed in the described order. For example, some operations / steps can be decomposed, and some operations / steps can be combined or partially combined. Therefore, the actual execution order may change according to the actual situation.

[0064] In the present application, "a plurality of" means two or more. "And / or" describes the association relationship of associated objects and indicates that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. The character " / " generally represents an "or" relationship between the associated objects before and after.

[0065] With the development of digitalization and simulation technologies, digital twin technology has become an emerging research hotspot in various industries. The prototype of the digital twin model, the "mirror space model", was first introduced by Michael Grieves in the Product Lifecycle Management (PLM) course at the University of Michigan in 2003. It constructs a digital object model through the actual data of physical objects, then predicts the physical objects through the data objects, and can optimize the physical objects through the predicted data. Subsequently, in cooperation with NASA and the US Air Force, this concept has been enriched. Elements such as model-based product performance prediction and optimization have been enhanced.

[0066] Subsequently, the academic and industrial communities have conducted extensive research and discussions on the digital twin concept. In 2011, NASA and the US Air Force Research Laboratory defined the concept of digital twin as an integrated multi-physics, multi-scale, and probabilistic simulation model that can be used to predict the health status and remaining service life of aircraft, etc. In 2012, Glaessgen et al. believed that digital twin is a complex system integrating multi-physics and multiple scales, as well as multi-probabilistic simulations based on physical models, historical data, and sensors to update data in real time. In 2017, Grieves et al. further explained digital twin as the virtual information structure of products. By constructing a digital twin model, all information about the actual inspected product can be obtained. In 2018, Tao et al. defined digital twin as a component of PLM, using physical, virtual, and interactive data in the product life cycle to draw a product map in real time. Haag et al. defined digital twin as a comprehensive digital description of products that can simulate the behavioral characteristics of real-world models.

[0067] In recent years, there has been an increasing amount of research on digital twin in the industrial field. Yi et al. proposed a three-layer intelligent assembly application framework based on DT. Zhang et al. proposed a DT-based production line design method and verified the effectiveness of the method with a glass production line as an example. Wang et al. proposed a new framework for data-driven intelligent customization enhanced by digital twin and applied it to common elevator customization problems. These studies have provided solutions for the creation of twins in the entire system, but no studies have been found that focus on the real-time simulation of digital twins for high-dimensional CFD data within key equipment.

[0068] When solving fluid flow and heat transfer problems, the widely used modeling method is the traditional CFD flow field calculation, which is based on the basic principle of numerically solving the differential equations that control fluid flow to derive the discrete distribution of the fluid flow field within a continuous region.

[0069] When the calculation conditions change, traditional CFD modeling methods need to perform complex and time-consuming CFD calculations again, which not only consumes a large amount of computing resources but also results in low efficiency. In fact, as a system, the flow field has its characteristics, but the practice of re-running CFD calculations each time ignores these characteristics.

[0070] With the advent of the era of artificial intelligence, data-driven modeling methods have made greater progress. Initially, these methods have been highly effective in the fields of computer vision, speech recognition, image recognition, and natural language processing. Against this background, machine learning, especially deep learning methods, have gradually been used in flow field reconstruction, flow field display, and flow field feature extraction, etc. Zhu et al. used an RBF neural network and three subsonic flow field sample data of the NACA0012 airfoil to achieve data-driven turbulence modeling. Gao et al. proposed a physically constrained CNN architecture for learning the solutions of partial differential equations with parameters on irregular regions. Thuerey et al. used the Unet architecture to predict the airfoil pressure and velocity distributions on the surface.

[0071] Compared with traditional modeling and simulation methods, neural network-based methods have low computational costs and fast prediction speeds. Hartmann et al. proposed that model simplification is a key technology for transferring highly detailed and complex simulation models to other fields and life cycle stages. Using a reduced-order model can effectively remove redundant features in the data, reduce the degrees of freedom, improve the model execution speed, while maintaining the required accuracy and predictability, providing solutions for the development of digital twins. Proper Orthogonal Decomposition (POD) analyzes the main features of the flow field by mapping high-dimensional flow field data to a low-dimensional orthogonal space to achieve the dimensionality reduction of the flow field.

[0072] Previous studies have mostly focused on the overall structure of manufacturing systems, and few studies have focused on the digital twin of the flow field in important equipment. Although there are many studies on machine learning in the field of computational fluid dynamics, how to apply it to establish digital twin models in industry has not been widely studied.

[0073] The following will further elaborate on the present invention in conjunction with the accompanying drawings and specific embodiments:

[0074] As Figure 1 shown, in this embodiment, a method for predicting the flow field data of a pusher kiln based on digital twin is proposed, and the specific implementation process is as follows:

[0075] S1: Obtain the CFD flow field data of the pusher kiln.

[0076] In this embodiment, the implementation of the method for predicting the flow field data of the pusher kiln based on digital twin can be as in Figure 2The overall framework of the digital twin shown is completed. The overall framework of this digital twin is divided into a Professional analysis layer, a Virtual-real twin layer, and a Basic support layer according to technical characteristics. The implementation of the push plate kiln flow field data prediction method based on digital twins is specifically completed in the Virtual-real twin layer.

[0077] Among them, the Basic support layer is the technology that supports the implementation of the push plate kiln flow field data prediction method based on digital twins, including Internet of Things Technology, CAE (Computer Aided Engineering in engineering design), AI (Artificial Intelligence), High performance parallel computing technology, Secure Connect Technology, etc.

[0078] The Physical object in the Virtual-real twin layer provides shape and size information, as well as Data Collection of operation data information. Among them, Physical Equipment includes push-plate kiln, sensor, etc. The operation data information collects History data and Real-time working condition data. The data of the real-time operation of the physical entity is transmitted through the MQTT protocol.

[0079] The Twinned object in the Virtual-real twin layer generates the CFD flow field solution in real-time prediction. The predicted data can be visualized for the prediction results through the Professional analysis layer, providing service decisions for the Smart decision making and control of the physical entity to perform Optimized control.

[0080] Specifically, the twin is an integration of a geometric model, a rule model, a physical model, and a neural network model. The purpose of the geometric model is to create a virtual geometry for a physical object and generate a mesh file for the physical model for subsequent process solving. The purpose of establishing the rule model is to set the rules for the simulation boundary conditions of the physical model. The physical model is used for simulation calculation and generating CFD big data. The generated data cannot be directly input into the neural network. Post-processing is required to extract the data we need. Finally, the neural network model performs POD dimensionality reduction of high-dimensional data, as well as MCNN training and prediction. In the neural network model, the working condition (which can be regarded as the operation data of the pusher kiln collected in real time) is subjected to multi-scale convolution by building a surrogate model for predicting the flow field to obtain the flow field.

[0081] Specifically, the workflow during the operation of the virtual reality twin layer can be referred to Figure 3 , and this process can be divided into two stages: online training and offline analysis. During online training, through the collaboration of the four models, non-linear fitting of the neural network is performed on the big data flow field. In offline analysis, as long as the on-site condition parameters are input into the trained model, the twin model can track the on-site data of the physical world for real-time update and dynamic evolution, so as to monitor and evaluate the equipment in real time.

[0082] In one embodiment, in the geometric model, the shape, outlet, size, and inlet of the physical entity are measured. ICEM creates points, lines, and surfaces to build 3D models, and then creates parts such as inlets and outlets, boundary conditions, etc. Create blocks and associate them with the parts, then mesh and refine the mesh, and confirm whether the mesh meets the conditions. If it meets the requirements, export the mesh file. If it does not meet the requirements, continue to refine the mesh. Specifically, the measured data is used to create and mesh the 3D model of the physical entity using ANSYS ICEM CFD preprocessing software to restore the physical entity at a 1:1 scale. In subsequent simulations, the geometric model realizes a complete mirror image of the actual device, and the geometric model finally outputs a mesh file.

[0083] In the regular model, since the production operation of the pusher kiln has a certain process, based on the historical data of the sensors during the operation of the actual physical entity, the process curve of the on-site operation of the pusher kiln can be obtained, and the normal values of each variable during the production process can be summarized to determine the reasonable value range of each parameter; specifically, it is to obtain the physical entity sensing data through the sensor, then remove the abnormal data from the sensing data, then summarize the reasonable range of the key process parameters, and output based on the range and the sensing data as the boundary conditions of the physical model.

[0084] In the physical model, based on the geometric model and the rule model, the ANSYS FLUENT solver is used to simulate the physical process. An appropriate physical solution model is selected, and the behavior rules of parameters are set according to the rule model. The behavior rules are reflected in various parameter settings in the software, such as boundary conditions. In this part, the FLUENT automation script is used to generate flow field data, and TECPLOT is used for post-processing the data. Specifically, it involves obtaining the mesh, forming the parameter space with process parameters, performing Parameter space LHS sampling, setting the solution model and boundary conditions in FLUENT (i.e., generating flow field data), then calculating the data generated by FLUENT, judging whether the calculation is complete. If it is complete, post-processing is carried out through TECPLOT and added to the Simulation Database; otherwise, the process is repeated through FLUENT.

[0085] The neural network model is the key module for realizing the offline dynamic analysis of CFD. In the neural network model (abbreviated as POD-MCNN), the large flow field data generated by the physical model is divided into a training set and a test set for the design and training of the neural network model. The trained model can quickly respond to input parameters and output the corresponding CFD results. Specifically, it includes:

[0086] The generation of the Training set and Test set, data dimensionality reduction (POD dimensionality reduction) through the test set and training set, training MCNN with the reduced-dimensional data (MCNN training), obtaining the trained Trained POD-MCNN model, and outputting the prediction results through this POD-MCNN model.

[0087] Based on Figure 2 And Figure 3, the operation data in S1 can be regarded as the training set and test set data. The CFD flow field data of the pusher kiln is obtained, including: obtaining the geometric data of the pusher kiln and the historical data during the operation of the pusher kiln; simulating the physical process of the pusher kiln based on the geometric data and historical data to generate the initial CFD flow field data of the pusher kiln; and post-processing the initial CFD flow field data to obtain the CFD flow field data.

[0088] Among them, the geometric data is obtained through a geometric model, and the historical data is obtained through a rule model. The physical process of the pusher kiln is simulated in a physical model to obtain the initial CFD flow field data, and the initial CFD flow field data is post-processed by TECPLOT in the physical model to obtain the CFD flow field data.

[0089] S2: Perform dimensionality reduction processing on the CFD flow field data to obtain low-dimensional data;

[0090] Steps S2 - S5 in this embodiment are all completed in the neural network model. The framework of this neural network model can refer to Figure 4 , combining the advantages of POD and CNN, enabling the digital twin to provide a wider range of operating conditions than simulation and using it for real-time monitoring of the internal flow field of the pusher kiln.

[0091] Specifically, the processing process in this neural network model can be summarized into three steps: POD dimensionality reduction, MCNN prediction, and POD reconstruction. First, for the input parameter space, extract the physical field characteristics of the simulation database through POD, map the high-dimensional flow field to a low-dimensional space (POD decomposition), obtain the basic modes and coefficients. Then, establish a surrogate model of the input parameters and the low-dimensional flow field characteristics (i.e., Coefficients) through a one-dimensional multi-scale convolutional neural network, that is, the MCNN model, to respond to new input parameters. Finally, reconstruct the data using the basic modes and the MCNN predicted data (Predicted coefficients) to obtain the final data.

[0092] In a specific embodiment, S2 is the POD dimensionality reduction process. By performing standardization processing on the CFD flow field data, standardized data is obtained; the eigenvalues and covariance matrix of the standardized data are extracted; a feature matrix is generated based on the eigenvectors of the eigenvalues, and low-dimensional data is obtained based on the feature matrix and the covariance matrix.

[0093] In a specific embodiment, the parameter space C (i)Obtained by performing Latin - hypercube sampling on the parameters, the CFD flow - field data (flow - field data) is \(X = [x_1,x_2,\cdots,x r \in R n×r , where \(n\) is the number of nodes in the computational domain and \(r\) is the number of samples in the parameter space. The flow - field data is high - dimensional. Therefore, directly mapping the input parameter \(C (i) to the flow - field data \(X (i) is not practical. Therefore, it is necessary to first reduce the dimension of the CFD flow - field data.

[0094] The average value of the temperature field is calculated as follows:

[0095]

[0096] The normalized data is:

[0097]

[0098] Calculate the covariance matrix and \(n\) eigenvalues of the normalized data and arrange them from largest to smallest. Take the eigenvectors \(q (1) ,q (2) ,\cdots,q (m) corresponding to the first \(m\) eigenvalues to form the eigen - matrix \(U\);

[0099] U=[q (1) ,q (2) ,\cdots,q (m) T

[0100] Each eigenvector represents the basic mode of the temperature field, and the reduced - order matrix \(Y\in R m×r can be obtained;

[0101]

[0102] When \(m\) takes an appropriate value, it can be ensured that:

[0103]

[0104] where \(U'\) represents the pseudo - inverse matrix of \(U\). The \(m\) - dimensional data \(Y (i) can be used to characterize the high - dimensional flow - field data \(X (i) with a low error rate. Through POD modeling, the basic modes and the corresponding coefficient matrix can be obtained, where the basic modes Modes are used for flow - field reconstruction, and the coefficient matrix Coefficients together with the parameter - space matrix are used for MCNN modeling and flow - field prediction.

[0105] S3: Train the initial flow - field prediction model based on the low - dimensional data to obtain the trained flow - field prediction model; ​

[0106] The flow field prediction model in this embodiment is the MCNN model, that is, the surrogate model. In a specific embodiment, the structure of the flow field prediction model can be referred to Figure 5 . The input feature map is passed through multi-scale convolution (Conv) to obtain the feature maps of each layer respectively, and then feature fusion (concat) is performed, and finally the prediction result (Output) is input.

[0107] In a specific embodiment, the multi-scale convolutional neural network model (MCNN) is designed to complete the mapping of the input parameter C (i) to the basic mode coefficient matrix Y (i) . The same feature map will have different expression forms in two different scale spaces. The single-scale convolution process uses the same-scale convolution kernel to extract features from the input. It is easy to ignore other precision features, thus obtaining complete feature expression information. In contrast, the multi-scale convolutional neural network can use multiple convolution kernels to convolve with the input at different scales to extract features at different scales. Then these features are fused to achieve the purpose of complementary information.

[0108] In this embodiment, the low-dimensional data is normalized, and the normalized low-dimensional data is preprocessed to obtain preprocessed data; based on multi-scale convolution, feature extraction is performed on the preprocessed data to obtain multi-scale features; based on the multi-scale features, the initial flow field prediction model is trained to obtain the trained flow field prediction model; among them, based on multi-scale convolution, feature extraction is performed on the preprocessed data to obtain multi-scale features, including: fusing the features output by the previous multi-scale convolution layer in the multi-scale convolution with the input features through skip connection to obtain fused features; using the fused features as the input features of the next multi-scale convolution layer in the multi-scale convolution to obtain multi-scale features.

[0109] Specifically, the model input C=(c1, c2,..., c n ) are parameters obtained from the physical entity sensor through the MQTT protocol, which can be regarded as low-dimensional data. In order to eliminate the influence of different dimensions between the parameters and accelerate the network gradient descent to find the optimal solution, C is normalized. Then, the input parameters are preprocessed through a fully connected neural network C'=f(C) to obtain preprocessed data. Of course, before performing multi-scale convolution on the preprocessed data, the dimension of the low-dimensional input signal, that is, the preprocessed data, can also be increased, which is beneficial to extracting the multi-scale features of the input signal through multi-scale convolution operations.

[0110] The features extracted by the convolutional kernels at different scales are different. Convolutions with larger kernel sizes have larger receptive fields but are not sensitive to detailed features. Convolutions with smaller kernel sizes can capture detailed features more effectively. Three multi-scale convolutional layers are designed in the model, and each layer contains three kernel scales of 1*1, 3*1, and 5*1. These convolutional kernels are used for multi-scale feature extraction of the upper-layer output. In addition to the kernel size F, the convolutional output is also affected by the kernel stride S, the number of channels n, and the padding size P. The padding operation adds zeros around the boundary of the input matrix. The size of the output matrix after the convolutional operation can be calculated as:

[0111] N l+1 =(N l -F + 2P) / S + 1

[0112] where N l is the size of the l-th layer features, and N l+1 is the size of the l+1-th layer features.

[0113] To ensure the high non-linearity of the proposed model, the neural network uses a multi-layer deep structure. In addition, the model also fuses the original features with the multi-scale features through skip connections as the input of the next layer. The features extracted by the multi-scale convolutional layers are finally output through the fully connected layer. Figure 6 Shows the structure of the MCNN, where "Multi-scale conv" is the multi-scale convolution, "Fully connected" is the fully connected layer, "flatten" is the flattening operation, and Table 1 lists the main parameters of the network.

[0114] The activation function enhances the non-linear expression ability of the network. LeakyRelu is selected as the activation function.

[0115]

[0116] When a is 0.01, LeakyReLU solves the possible neuron death problem in Relu. It has a small positive slope in the negative region, so it can perform backpropagation even for negative input values. It solves the problem of the vanishing gradient of the Relu activation function.

[0117] Generally, there is a pooling layer after convolution in a convolutional neural network. The pooling layer in a CNN can reduce the dimension of the features to a certain extent, but this is not a necessary operation for a CNN. In this experiment, there are not many features after convolution. There is no need to reduce the dimension. The pooling process will result in the loss of features extracted by the convolutional layer, which will affect the performance of the model. Therefore, the MCNN structure does not use a pooling layer. Experiments have shown that the structure without a pooling layer performs better in the flow field prediction model task.

[0118] Table 1 Main Parameters of the Multi-Scale Convolutional Neural Network

[0119] Network layer name Output shape Kernel size Number of kernels Stride Padding Fully connected 1 64 - - - - Multi-scale conv 1-1 3*64 1 3 1 0 Multi-scale conv 1-2 3*64 3 3 1 1 Multi-scale conv 1-3 2*64 5 2 1 2 Concat 9*64 - - - - Multi-scale conv 2-1 6*64 1 6 1 0 Multi-scale conv 2-2 5*64 3 5 1 1 Multi-scale conv 2-3 5*64 5 5 1 2 Concat 17*64 - - - - Multi-scale conv 3-1 11*64 1 11 1 0 Multi-scale conv 3-2 11*64 3 11 1 1 Multi-scale conv 3-3 10*64 5 10 1 2 Concat 33*64 - - - - Conv+flatten 256 3 2 1 - Fully connected 2 10 - - - -

[0120] For such a deep model, all operation layers can be regarded as a complex function G. The network output is:

[0121]

[0122] where θ is the parameter to be learned in the network. The training of this initial flow field prediction model is an iterative process driven by the model loss. The model parameters are updated by minimizing the difference between the network output and the ground truth and propagating the error back to each layer of the network. The loss function of the model consists of a data loss and a regularization term. The mean square error is used for the data loss function, and the l2 norm is used for the regularization term. The purpose of including the regularization term is to prevent the model from overfitting, which may lead to a decline in generalization performance.

[0123]

[0124] where is the network output, Y is the ground truth, λ is the regularization coefficient used to control the strength of regularization. θ i ∈θ is the parameter to be learned in the network. M is the number of parameters.

[0125] During the training process, the purpose of the network is to find a set of optimal parameters θ to minimize the loss function, expressed as:

[0126]

[0127] To effectively train the model, we adopt a mini-batch based learning strategy and set the batch size to 64. Mini-batch training has been proven to effectively improve the generalization performance and significantly reduce the memory cost. The Adam optimization algorithm is used as the optimization algorithm for the model.

[0128] S4: Based on the trained flow field prediction model, predict the operation data of the pusher kiln collected in real time to obtain the ground state prediction data;

[0129] When the flow field prediction model is trained, for the newly input parameters, that is, the operation data of the pusher kiln collected in real time, the MCNN model can predict the basic modal coefficient Y s , and then perform POD reconstruction on it to obtain the predicted flow field data.

[0130] S5: Based on the ground state prediction data and the low-dimensional data, perform reconstruction to predict the flow field data of the pusher kiln.

[0131] In this embodiment, the pseudo-inverse matrix of the feature matrix is obtained; based on the pseudo-inverse matrix, the ground state prediction data, and the eigenvectors of the eigenvalues, reconstruction is performed to predict the flow field data of the pusher kiln:

[0132]

[0133] The reconstruction result can be directly used as high-precision flow field data to achieve the real-time response of the simulation digital twin to the input parameters.

[0134] This paper combines CFD modeling and simulation, Internet of Things, and deep learning technologies to propose a method for predicting the flow field data of a pusher kiln based on digital twin. This method realizes the real-time simulation of physical and twin objects by integrating geometric, rule, physical, and neural network models, and innovatively embeds proper orthogonal decomposition (POD) and multi-scale convolutional neural network (MCNN). POD is used to map high-dimensional data to low-dimensional features, and MCNN is used to construct a model for predicting low-dimensional features to achieve fast flow field prediction. This method can quickly predict the high-dimensional flow field distribution based on sensing data to achieve the real-time evaluation of the operating state of the pusher kiln.

[0135] In a specific embodiment, the structure of a pusher kiln can be referred to Figure 7(a)-Figure 7(c) , which is electrically heated. The effective size of the internal space of the furnace body is XYZ = 4000mm × 1100mm × 455mm, where X is the kiln length, Y is the kiln width, and Z is the kiln height. Since the kiln is very long, the model is shown in an appropriate scale. Figure 7(a) shows the geometric structure of the inner space of the kiln (The geometry of the inner space of the push-plate). According to the temperature, the kiln can be divided into a preheating zone, a high temperature zone, and a cooling zone. Figures 7(b) and (c) show the front view and side view of the pusher kiln, which include an exhaust duct 1, a flue gas outlet 2, insulation materials 3, a fluid zone 4, a product zone 5, a kiln head 6, a kiln tail 7, and electric heating 8

[0136] The outer wall of the pusher kiln is made of heat-insulating materials. There are several electric heating rods above and below the kiln body for heating the pusher kiln. During the production process, a large amount of nitrogen will be blown into the kiln from the kiln tail through a pipeline and flow out from the outlet. When predicting the flow field data of the pusher kiln based on digital twin, after analyzing the production line data, the temperature values and inlet gas flow velocity values at six process critical points have a greater impact on the production process. Therefore, the 7D parameters in Table 2 are selected as the variable parameters for generating the simulation dataset.

[0137] Table 2 Key process parameter ranges

[0138] Parameter type Parameter range Temperature point 1 (T1) [540,720]℃ Temperature point 2 (T2) [720,810]℃ Temperature point 3 (T3) [900,1080]℃ Temperature point 4 (T4) [1035,1170]℃ Temperature point 5 (T5) [1215,1395]℃ Temperature point 6 (T6) [1300,1400]℃ Gas flow rate (Q) <![CDATA[[200,300]m 3 / h]]>

[0139] In this embodiment, it is aimed to quickly generate the flow field solution of region 4 using CFD digital twin. To generate the simulation database, the parameter space is sampled using Latin Hypercube Sampling (LHS). The CFD simulation is carried out using the Ansys Fluent 19.3 automation script. The mesh consists of approximately 147k cells. During the operation of the pusher kiln, the heat transfer process in the kiln is a complex heat transfer process with the comprehensive functions of heat conduction, convective heat transfer, and radiative heat transfer. A corresponding CFD mathematical model of the heat exchange and gas flow in the kiln is established. The standard k-ε model is used for the gas flow, and the discrete ordinate (DO) method is used for the radiation model.

[0140] First, perform POD analysis on the flow field data of the generated samples to obtain the base mode flow field and the corresponding eigenvalues. Arrange the eigenvalues in descending order, and define the ratio of the sum of the eigenvalues corresponding to the first q order base modes to the sum of all eigenvalues as the energy proportion ∑E of the first q order base modes i , and the energy proportion characterizes the amount of flow field information contained in the first q order base modes. Based on this, it can be determined how many orders of modes can contain most of the information in the original flow field. Figure 8 Shows the relationship between the basic mode number and energy. Figure 8 Is the change trend of the total energy share of the first n order (n < 15) POD base modes. The abscissa is the order of the base mode, and the ordinate is the total energy proportion.

[0141] Figure 8 In [], the energy of the first ten basic modes accounts for more than 99.95%, indicating that the flow field characteristics are mainly concentrated in the first ten POD basic modes. 90% of the dataset is used as the training sample for POD reduction, and the remaining 10% is used as the test sample for POD reconstruction. To quantitatively analyze the reconstruction effect, the reconstruction error of the test sample is defined as:

[0142]

[0143] Among them, Is the predicted data at the i-th grid point. is the actual data of the i-th high-precision grid point calculated by computational fluid dynamics (CFD), and N is the number of grid points. The total error of the T test sample is defined as follows.

[0144]

[0145] The relationship between the number of basic modes and the reconstruction error Error can be referred to Figure 9 , obviously, the flow field characteristics are mainly concentrated in the first ten orders. This is consistent with Figure 8 the information in. Therefore, after the error analysis of the POD reconstruction, the number of basic modes is selected as 10.

[0146] Then, the prediction performance of POD-MCNN can be evaluated based on the processed data above. Three cases are selected in the test set to test the model. The data of these three cases can be referred to Date in Table 3. The specific explanations of the 7D parameters in this case can be referred to Table 2. The mean absolute error calculations of POD-MCNN prediction and CFD calculation are performed for the three cases. The results are shown in Table 3. It can be seen from the data in Table 2 that the mean absolute temperature error (MAE t) in the three cases is only 1.4045, and the mean absolute pressure error (MAE p) is only 0.5140. The error can be ignored in engineering applications. Therefore, it can be known that this network model performs well in the flow field prediction task of test data.

[0147] Table 3 Parameters and MAE of three cases in test data

[0148] Case Data MAE t MAE p Case 1 [665,738,950,1104,1345,1318,268] 0.9144 0.4492 Case 2 [590,723,1068,1167,1365,1352,206] 1.5922 0.7042 Case 3 [594,787,901,1138,1331,1387,241] 1.7068 0.3887

[0149] For the working pusher kiln, the trend of temperature change is very important. Figure 11 and Figure 12 show the comparison between the POD-MCNN prediction results and the CFD results of the temperature and pressure trends in the three test sets formed by the data obtained in the three cases when Z = 350 mm and Y = 550 mm. It can be seen from the figure that the data trend output by POD-MCNN is consistent with the data trend of CFD simulation and almost overlaps.

[0150] After training, this model performs well in the temperature field and pressure field prediction tasks of test data and can accurately predict the flow field data of the test set.

[0151] To evaluate the interpolation performance of the model, three new cases were generated to test the model. The parameters of the new cases 1-3 can be referred to Date in Table 4. Among them, the parameter selections of cases 1 and 2 are still within the parameter space, but the parameter selection of case 3 is not within the parameter space. Under this new data, the comparison graphs between the POD-MCNN prediction results and the CFD results of the temperature trend and the comparison graphs between the POD-MCNN prediction results and the CFD results of the pressure trend can be referred to Figure 12 and Figure 13 , from the data in Table 4 and Figure 12 and Figure 13 it can be seen that the MAE t of case 1 is 1.3420 and 0.4836. Therefore, the prediction accuracy of the model for new data in the parameter space is not necessarily worse than that of the test set. The MAEs of case 3 are 3.9226 and 1.6723 respectively, and its error is relatively high because the parameter selection of case 3 is not within the parameter range.

[0152] Table 4 Parameters and MAEs of three cases in the new data

[0153] Case Data MAE t MAE p Case 1 [546,786,1046,1144,1226,1317,207] 1.3420 0.4836 Case 2 [714,809,1028,1126,1364,1354,290] 2.5835 0.8289 Case 3 [514,707,926,1029,1234,1247,190] 3.9226 1.6723

[0154] In this embodiment, the calculation speed of POD-MCNN was also tested, and the test results shown in Table 5 were obtained. For the prediction of unknown parameters, the trained model only takes 0.2 seconds to obtain the results, while the CFD simulation takes 1470 seconds, greatly saving the calculation cost and time consumption.

[0155] Table 5 Model acceleration ratio performance

[0156]

[0157] The MCNN model was compared with other models, which demonstrated the importance of each part of the push plate kiln flow field data prediction method based on digital twin. The three comparison models are the multi-scale pooling convolutional neural network (MPCNN), the single-scale convolutional neural network (CNN), and the fully connected neural network (FNN). These three models can be interpreted as degenerate versions of the MCNN model. MPCNN retains the pooling operation of the convolutional neural network, CNN uses a convolutional kernel of one size, and FNN does not use convolutional operations.

[0158] The training process adopted the same batch size, training time, initial learning rate, and weight decay as the MCNN training. The mean square error on the test set was used to measure the prediction accuracy of the model:

[0159]

[0160] where is the predicted value output by the network, and Y is the true label value. The experimental results are shown in Table 6. The proposed model has the smallest mean square error on the test set. Using convolution instead of fully connected reduces the number of network parameters. At the same time, the overfitting phenomenon is improved, and the error is reduced by 35%. Using multi-scale instead of single-scale convolution makes the features extracted by convolution richer, and the error is reduced by 15.5%. Abandoning the pooling operation will retain all the features extracted by multi-scale convolution, and the error is reduced by 27.2%.

[0161] Table 6 Performance comparison of different models

[0162] Model Number of parameters MSE FNN 8k 0.00343 CNN 6k 0.00264 MPCNN 6k 0.00306 MCNN 6k 0.00223

[0163] Among them, FNN, CNN, and MPCNN are different neural network models.

[0164] In summary, a multi-scale convolutional neural network is used for surrogate modeling. Using multi-scale convolutional kernels can extract features of different fineness and improve the feature learning ability of the model. The pooling operation is not required in the model, which allows the data features extracted by the convolutional layer to be retained, thereby improving the accuracy of the network. Based on high-fidelity simulation and artificial intelligence technologies, the digital twin can track the data in the physical world for real-time updates, providing a data basis for the monitoring, evaluation, optimization, and maintenance of on-site equipment.

[0165] As Figure 14 shown, in this embodiment, a pusher kiln flow field data prediction model based on digital twin is also proposed, including: an operation data acquisition module 141 configured to acquire CFD flow field data of the pusher kiln; a low-dimensional data acquisition module 142 configured to perform dimensionality reduction processing on the CFD flow field data to obtain low-dimensional data; a model training module 143 configured to train an initial flow field prediction model based on the low-dimensional data to obtain a trained flow field prediction model; an initial prediction module 144 configured to predict the operation data of the pusher kiln collected in real time based on the trained flow field prediction model to obtain base state prediction data; and a prediction data acquisition module 145 configured to reconstruct based on the base state prediction data and the low-dimensional data to predict the flow field data of the pusher kiln.

[0166] It should be noted that the pusher kiln flow field data prediction device based on digital twin provided in the above embodiment belongs to the same concept as the pusher kiln flow field data prediction method based on digital twin provided in the above embodiment. The specific manners in which each module and unit perform operations have been described in detail in the method embodiment, and will not be elaborated here.

[0167] The units involved in the embodiments described in this application can be implemented in software or in hardware, and the described units can also be provided in a processor. Among them, the names of these units do not, in some cases, constitute a limitation on the unit itself.

[0168] Another aspect of this application also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the aforementioned method for predicting the flow field data of a pusher kiln based on digital twins. The computer-readable storage medium can be included in the electronic device described in the above embodiments, or can exist alone without being assembled into the electronic device.

[0169] Another aspect of this application also provides a computer program product or a computer program. The computer program product or the computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the prediction of the flow field data of the pusher kiln based on digital twins provided in the above various embodiments.

[0170] The above content is only a preferred exemplary embodiment of this application and is not used to limit the implementation of this application. Those of ordinary skill in the art can easily make corresponding adaptations or modifications according to the main concept and spirit of this application. Therefore, the protection scope of this application should be subject to the protection scope required by the claims.

Claims

1. A method for predicting the flow field data of a pusher kiln based on digital twin, characterized in that, Including: Obtain the CFD flow field data of the pusher kiln; Perform dimensionality reduction processing on the CFD flow field data to obtain low-dimensional data; Train the initial flow field prediction model based on the low-dimensional data to obtain the trained flow field prediction model; Predict the operation data of the pusher kiln collected in real time based on the trained flow field prediction model to obtain the ground state prediction data; wherein, the ground state prediction data is the basic modal coefficient Y s ; Reconstruct based on the ground state prediction data and the low-dimensional data, and predict the flow field data of the pusher kiln through the following formula: Wherein, U′ represents the pseudo-inverse matrix of the feature matrix, and the feature matrix includes the covariance matrix of the standardized data arranged from large to small and the eigenvectors of the first m eigenvalues among n eigenvalues. The standardized data is obtained by standardizing the CFD flow field data. is the average value of the temperature field, and X is the flow field data of the pusher kiln obtained by prediction. s ​ 2. The method according to claim 1, characterized in that, The obtaining of the CFD flow field data of the pusher kiln includes: Obtain the geometric data of the pusher kiln and obtain the historical data during the operation of the pusher kiln; Simulate the physical process of the pusher kiln based on the geometric data and the historical data to generate the initial CFD flow field data of the pusher kiln; Perform post-processing on the initial CFD flow field data to obtain the CFD flow field data.

3. The method according to claim 1, wherein The training of the initial flow field prediction model based on the low-dimensional data to obtain the trained flow field prediction model includes: Standardize the low-dimensional data and preprocess the standardized low-dimensional data to obtain preprocessed data; Extract multi-scale features from the preprocessed data based on multi-scale convolution; Train the initial flow field prediction model based on the multi-scale features to obtain the trained flow field prediction model.

4. The method according to claim 3, wherein The extracting of multi-scale features from the preprocessed data based on multi-scale convolution includes: Fuse the features output by the previous multi-scale convolution layer in the multi-scale convolution with the input features through skip connection to obtain fused features; Use the fused features as the input features of the next multi-scale convolution layer in the multi-scale convolution to obtain the multi-scale features.

5. The method according to any one of claims 3-4, characterized in that, The multi-scale convolution includes 3 multi-scale convolution layers, and each multi-scale convolution layer contains three kernel scales of 1*1, 3*1, and 5*1.

6. The method according to claim 1, wherein The dimensionality reduction processing of the CFD flow field data to obtain low-dimensional data includes: Perform standardization processing on the CFD flow field data to obtain standardized data; Extract the eigenvalues and covariance matrix of the standardized data; Generate a feature matrix based on the eigenvectors of the eigenvalues, and obtain the low-dimensional data based on the feature matrix and the covariance matrix.

7. The method according to claim 6, characterized in that, The reconstruction based on the ground state prediction data and the low-dimensional data to predict the flow field data of the pusher kiln includes: Obtain the pseudo-inverse matrix of the feature matrix; Reconstruct based on the pseudo-inverse matrix, the ground state prediction data, and the eigenvectors of the eigenvalues to predict the flow field data of the pusher kiln.

8. A push plate kiln flow field data prediction device based on digital twin, characterized in that, Including: An operating data acquisition module configured to obtain the CFD flow field data of the pusher kiln; A low-dimensional data acquisition module configured to perform dimensionality reduction processing on the CFD flow field data to obtain low-dimensional data; A model training module configured to train the initial flow field prediction model based on the low-dimensional data to obtain the trained flow field prediction model; The initial prediction module is configured to predict the operation data of the pusher kiln collected in real time based on the trained flow field prediction model, and obtain the ground state prediction data; wherein, the ground state prediction data is the basic mode coefficient Y s ; A prediction data acquisition module configured to reconstruct based on the ground state prediction data and the low-dimensional data, and predict the flow field data of the pusher kiln through the following formula: Wherein, U′ represents the pseudo-inverse matrix of the feature matrix, and the feature matrix includes the covariance matrix of the standardized data arranged from large to small and the eigenvectors of the first m eigenvalues among n eigenvalues. The standardized data is obtained by standardizing the CFD flow field data. is the average value of the temperature field, and X is the flow field data of the pusher kiln obtained by prediction. s ​ 9. An electronic device, characterized in that, Including: One or more processors; A storage device for storing one or more computer programs which, when executed by the one or more processors, cause the electronic device to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Computer-readable instructions are stored thereon which, when executed by a processor of a computer, cause the computer to execute the method according to any one of claims 1 to 7.

Images