Temperature time sequence prediction method and device for mechanism-data reconstructed particle flow system

Abstract

The invention provides a mechanism-data reconstructed particle flow system temperature time sequence prediction method and device, and the method comprises the steps: constructing a brand new physical information neural network (PINN) method through combining normalization time, dimensionless particle temperature and a local weight adjustment method; a theoretical comprehensive loss function is constructed by a control equation, an initial condition and limited observation data sample points, and corresponding comprehensive loss functions are determined one by one by taking a simple uniform cooling particle flow system as a case according to three inverse problems of missing energy continuous dissipation coefficients, missing initial particle temperatures and missing both of the energy continuous dissipation coefficients and the initial particle temperatures; and solving and iteratively updating by taking the minimum specific comprehensive loss function as a target, and sequentially and accurately determining respective missing values in the three types of inverse problems and a rule that the particle temperature changes along with time. According to the invention, the PINN combines experimental data with a simulation method, and is high in accuracy and robustness in predicting flow field information, a closed control equation and an initial boundary value condition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of reconstruction of flow field characteristics of particle flow system and mining of physical laws, and particularly relates to a mechanism-data reconstruction particle flow system temperature time sequence prediction method and device. BACKGROUND

[0002] Particle flow systems exist widely in nature and industrial production: rapid prediction of debris flow phenomena in nature can directly protect the safety of tens of thousands of lives and property, and full study of particle flow phenomena in debris flow can provide effective debris flow impact force prediction and effective ideas for protection structure design; in industrial production, such as chemical processes, the flow behavior of catalysts directly determines the efficiency of catalytic reactions and the yield of final products; in the field of metallurgical minerals, particle flow characteristics directly determine the efficiency of ore separation and metallurgical efficiency; in the pharmaceutical process, ensuring that the binder and the pharmaceutical preparation are fully mixed is crucial to the quality of the drug. Therefore, in-depth and systematic study of the flow characteristics of particle flow systems is of great significance for reducing the impact of natural disasters, improving industrial production efficiency and economic benefits.

[0003] Common methods for understanding the flow characteristics of particle flow systems mainly include experimental measurement and numerical simulation methods. Experimental measurement methods mainly include invasive and non-invasive measurement. Invasive measurement mainly uses probe measurement, such as Prior Art 1 (Publication No.: CN114367319B) and Prior Art 2 (CN103097015B). This invasive probe measurement method significantly affects the basic flow information of the particle flow field, because the inserted probe directly limits the range of particle motion, which is quite different from the particle motion characteristics in the real system. Non-invasive measurement, such as PIV technology in Prior Art 3 (Publication No.: CN113311186B), can only measure relatively simple two-dimensional flow field information, and cannot accurately capture three-dimensional flow characteristics. In summary, common experimental measurement methods can only obtain local and two-dimensional particle flow characteristics, and cannot accurately describe the overall flow field characteristics. With the rapid development of computer technology, numerical simulation methods have become one of the main technical means for understanding the flow characteristics of particle flow systems. The coarse-grained calculation method in Prior Art 4 (Publication No.: CN112131633B) can reduce the number of discrete units by packing a group of small particles with larger coarse particles, but the interaction mechanism between coarse particles is still difficult to determine by analytical means, which may reduce the accuracy of the coarse-grained method.

[0004] In summary, considering the importance of particle flow systems, but the existing measurement and numerical simulation methods have limitations, therefore, it is urgent to develop a new method to systematically study the flow field characteristics and internal physical mechanisms of particle flow systems. SUMMARY

[0005] One purpose of the present application is to provide a mechanism-data reconstructed granular flow system temperature time sequence prediction method, which can comprehensively reconstruct the granular flow full flow field motion information, combine numerical experimental data with incomplete control equations and initial conditions based on a physical information neural network method, accurately predict the granular temperature time sequence characteristics in the granular flow system, and assist in determining the complete control equation form and initial conditions, and has high robustness, accuracy, universality and scalability. Another purpose of the present application is to provide a mechanism-data reconstructed granular flow system temperature time sequence prediction device. Still another purpose of the present application is to provide a computer readable medium. Still another purpose of the present application is to provide a computer device.

[0006] In order to achieve the above purposes, the present application discloses a mechanism-data reconstructed granular flow system temperature time sequence prediction method, comprising:

[0007] The granular flow system is simulated by the discrete element method to generate experimental granular temperatures at different times, and the reconstructed granular flow system is a uniformly cooled smooth granular flow system;

[0008] A theoretical comprehensive loss function is constructed based on the experimental granular temperatures at different times and the physical control equation by the constructed physical information-based neural network;

[0009] According to the system missing condition, the target comprehensive loss function is determined according to the theoretical comprehensive loss function, and the network parameters of the physical information-based neural network are iteratively updated to minimize the target comprehensive loss function, and a granular temperature time sequence characteristic prediction model is constructed. The system missing condition includes missing energy dissipation coefficient, missing initial time granular temperature or missing energy dissipation coefficient and initial time granular temperature;

[0010] Based on the granular temperature time sequence characteristic prediction model, the granular temperature is predicted according to the input predicted time information to obtain the target granular temperature time sequence.

[0011] Preferably, the granular flow system is simulated by the discrete element method to generate experimental granular temperatures at different times, comprising:

[0012] According to the preset granular velocity distribution function, and based on Newton's law of motion and the linear interaction force model between particles, the granular motion information in the reconstructed granular flow system is simulated, and the experimental granular velocities at different times are counted;

[0013] The experimental granular fluctuation velocity is determined according to the experimental granular velocities at different times combined with the average granular velocity;

[0014] According to the experimental granular fluctuation velocity, the average value of the fluctuation energy of all particles in the system is counted to generate experimental granular temperatures at different times.

[0015] Preferably, a theoretical comprehensive loss function is constructed according to the experimental particle temperature at different time and the physical control equation through the constructed neural network based on physical information, including:

[0016] The experimental particle temperature at different time is dimensionless processed to obtain dimensionless experimental particle temperature;

[0017] A data sample loss function and an initial condition loss function are constructed according to the dimensionless experimental particle temperature through the neural network based on physical information;

[0018] The time information at different time is normalized to obtain normalized time information;

[0019] A mechanism-data dual-driven control equation loss function is constructed according to the preset local weight coefficient, the normalized time information, the dimensionless experimental particle temperature and the physical control equation through the neural network based on physical information, and the local weight coefficient is determined according to the reciprocal of the predicted dimensionless particle temperature at the corresponding time;

[0020] The data sample loss function, the initial condition loss function and the mechanism-data dual-driven control equation loss function are weighted according to the preset global weight coefficient to generate the theoretical comprehensive loss function.

[0021] Preferably, a data sample loss function and an initial condition loss function are constructed according to the dimensionless experimental particle temperature through the neural network based on physical information, including:

[0022] The dimensionless experimental particle temperature is randomly sampled to obtain dimensionless sparse particle temperature;

[0023] A data sample loss function is constructed according to the dimensionless sparse particle temperature and the corresponding time at different time through the neural network based on physical information;

[0024] An initial condition loss function is constructed according to the dimensionless experimental particle temperature at the initial time through the neural network based on physical information.

[0025] Preferably, the system missing condition is missing energy dissipation coefficient;

[0026] According to the theoretical comprehensive loss function, a target comprehensive loss function is determined according to the system missing condition, and the network parameters of the neural network based on physical information are iteratively updated to minimize the target comprehensive loss function, and a particle temperature time sequence feature prediction model is constructed, including:

[0027] The theoretical comprehensive loss function is determined as the target comprehensive loss function;

[0028] The neural network based on physical information is trained and optimized by a preset optimization algorithm, network parameters are iteratively updated, and a particle temperature time series feature prediction model and an optimized energy dissipation coefficient are generated, the network parameters including network weights and network biases.

[0029] Preferably, the system missing condition is missing the initial time particle temperature.

[0030] According to the system missing condition, the target comprehensive loss function is determined according to the theoretical comprehensive loss function, and the network parameters of the neural network based on physical information are iteratively updated to build a particle temperature time series feature prediction model, including:

[0031] The initial condition loss function is removed from the theoretical comprehensive loss function to generate the target comprehensive loss function.

[0032] The neural network based on physical information is trained and optimized by a preset optimization algorithm, network parameters are iteratively updated, and a particle temperature time series feature prediction model and an optimized energy dissipation coefficient are generated, the network parameters including network weights and network biases.

[0033] Preferably, the system missing condition is missing the energy dissipation coefficient and the initial time particle temperature.

[0034] According to the system missing condition, the target comprehensive loss function is determined according to the theoretical comprehensive loss function, and the network parameters of the neural network based on physical information are iteratively updated to build a particle temperature time series feature prediction model, including:

[0035] The initial condition loss function is removed from the theoretical comprehensive loss function to generate the target comprehensive loss function.

[0036] The neural network based on physical information is trained and optimized by a preset optimization algorithm, network parameters are iteratively updated, and a particle temperature time series feature prediction model, an optimized energy dissipation coefficient and an initial time optimized particle temperature are generated, the network parameters including network weights and network biases.

[0037] Preferably, based on the particle temperature time series feature prediction model, the particle temperature is predicted according to the input predicted time information to obtain a target particle temperature time series, including:

[0038] The predicted time information is normalized to generate normalized predicted time information.

[0039] The normalized to-be-predicted time information is calculated through a particle temperature time sequence prediction model to generate a dimensionless particle temperature time sequence, which is a corresponding relationship between the dimensionless particle temperature and the normalized to-be-predicted time information.

[0040] The dimensionless particle temperature is time de-normalized and the particle temperature is dimensioned according to the normalized to-be-predicted time information to generate a target particle temperature time sequence.

[0041] Preferably, the method further comprises:

[0042] The predicted target particle temperature time sequence and the obtained real particle temperature time sequence are compared to generate a particle temperature time sequence feature error result.

[0043] Preferably, the method further comprises:

[0044] The optimized energy dissipation coefficient and the pre-calculated theoretical energy dissipation coefficient are compared to generate an energy dissipation coefficient error result, and the optimized energy dissipation coefficient is obtained through iterative calculation based on the physical control equation in the particle temperature time sequence prediction model.

[0045] Preferably, the method further comprises:

[0046] The initial time optimized particle temperature and the obtained initial time real particle temperature are compared to generate an initial time particle temperature error result, and the initial time optimized particle temperature is obtained through iterative calculation based on the particle temperature time sequence prediction model.

[0047] The application also discloses a mechanism-data reconstructed particle flow system temperature time sequence prediction device, comprising:

[0048] A numerical simulation unit is configured to perform numerical simulation on the reconstructed particle flow system through a discrete element method to generate experimental particle temperatures at different times, and the reconstructed particle flow system is a uniformly cooled smooth particle flow system.

[0049] A theoretical comprehensive loss function construction unit is configured to construct a theoretical comprehensive loss function according to the experimental particle temperatures at different times and the physical control equation through a constructed neural network based on physical information.

[0050] A particle temperature time sequence feature prediction model training unit is configured to determine a target comprehensive loss function according to the theoretical comprehensive loss function under a system missing condition, and iteratively update network parameters of the neural network based on physical information to construct a particle temperature time sequence feature prediction model, with the minimum target comprehensive loss function as the target, and the system missing condition includes missing energy dissipation coefficient, missing initial time particle temperature, or missing energy dissipation coefficient and initial time particle temperature.

[0051] The particle temperature time sequence feature prediction unit is configured to predict particle temperature based on a particle temperature time sequence feature prediction model according to input time information to be predicted, and obtain a target particle temperature time sequence.

[0052] The application further discloses a computer readable medium, which stores a computer program, and the program is executed by a processor to realize the method.

[0053] The application further discloses a computer device, which comprises a memory and a processor, the memory is configured to store information comprising program instructions, and the processor is configured to control execution of the program instructions, and the processor realizes the method when executing the program.

[0054] The application further discloses a computer program product, which comprises computer program / instructions, and the computer program / instructions are executed by a processor to realize the method.

[0055] The application generates experimental particle temperatures at different times by numerically simulating a reconstructed granular flow system by using a discrete element method, and reconstructs the granular flow system into a uniform cooling smooth granular flow system; a neural network based on physical information is constructed by time normalization, particle temperature dimensionless and local weight design, a theoretical comprehensive loss function is constructed according to experimental particle temperatures at different times, physical control equations and initial particle temperatures; according to a system missing condition, the target comprehensive loss function is determined according to the theoretical comprehensive loss function, and the network parameters of the neural network based on physical information are iteratively updated to minimize the target comprehensive loss function, a particle temperature time sequence feature prediction model is constructed, and accurate prediction and supplement of the missing condition are realized; the system missing condition includes missing energy dissipation coefficient, missing initial time particle temperature or missing energy dissipation coefficient and initial time particle temperature; based on the particle temperature time sequence feature prediction model, particle temperature prediction is performed according to input time information to be predicted, and a target particle temperature time sequence is obtained, which can fully reconstruct the granular flow field motion information, and the physical information neural network method combines numerical experimental data with incomplete control equations and initial conditions, which can not only accurately predict the particle temperature time sequence feature in the granular flow system, but also assist in determining the complete control equation and constitutive relation, energy dissipation loss parameter and initial condition, and has high robustness, accuracy, universality and expansibility. BRIEF DESCRIPTION OF DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings described below are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0057] Figure 1 A flow chart of a temperature time series prediction method of a mechanism-data reconstructed granular flow system provided for an embodiment of the present application;

[0058] Figure 2 A flow chart of another temperature time series prediction method of a mechanism-data reconstructed granular flow system provided for an embodiment of the present application;

[0059] Figure 3 A structure schematic diagram of a PINN suitable for uniformly cooling a granular flow system provided for an embodiment of the present application;

[0060] Figure 4 A comparison diagram of a target granular temperature time series and a real granular temperature time series when the energy dissipation coefficient A is missing provided for an embodiment of the present application;

[0061] Figure 5 A time series diagram of the relative error between a target granular temperature time series and a real granular temperature time series when the energy dissipation coefficient A is missing provided for an embodiment of the present application;

[0062] Figure 6 A structure schematic diagram of a temperature time series prediction device of a mechanism-data reconstructed granular flow system provided for an embodiment of the present application;

[0063] Figure 7 A structure schematic diagram of a computer device provided for an embodiment of the present application. DETAILED DESCRIPTION

[0064] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0065] It should be noted that the mechanism-data reconstructed particle flow system temperature time sequence prediction method and device disclosed in the application can be used in the field of artificial intelligence technology, and can also be used in any field other than the field of artificial intelligence technology. The application field of the mechanism-data reconstructed particle flow system temperature time sequence prediction method and device disclosed in the application is not limited.

[0066] In order to facilitate understanding of the technical solutions provided in the present application, the related contents of the technical solutions of the present application will be described first. Accurate and comprehensive capture of the flow field characteristics of the particle flow system is of great significance to ensure the safety of daily life and improve the quality and efficiency of industrial production. With the vigorous development of machine learning technology, the Physics-Informed Neural Networks (PINN) method based on physical information is one of the typical data and physical law driven machine learning methods. The method combines local measurement data with part / all physical laws in the numerical simulation process, and through reasonable architecture of the neural network form and fusion of the flow field characteristics and other physical information composed of data and equations into the learning process, it is expected to quickly realize accurate reconstruction of the flow characteristics of the particle flow system, and further improve the physical laws satisfied by the system through the learning results.

[0067] The present application simulates the uniform cooling smooth particle flow system by the discrete element method, generates the size of the particle temperature at different times in the system, and uses it as a high-precision numerical experiment point. The normalized time sequence is used as the input layer, the dimensionless particle temperature is used as the output layer, the fully connected neural network is combined, the incomplete particle temperature control equation, the initial particle temperature and the particle temperature at part of the time are coupled into the loss function as physical information, and the local weight of the sample point is adjusted based on the physical characteristics. The one-to-one correspondence between the normalized time and the dimensionless particle temperature is determined by the optimization parameter algorithm, and the physical information neural network method is further used to realize the change relationship of the particle temperature with time in the uniform cooling smooth particle flow system, the missing physical law and the initial particle temperature. The present application combines numerical experimental data with part of the physical law such as the incomplete control equation and the missing initial condition. Not only can the change of the particle temperature in the particle flow system be accurately described, but also the complete control equation form and the initial condition can be determined, the accuracy of the prediction result is improved, and the generalization ability is good. The present application provides a solution for industrial complex particle flow systems.

[0068] It is worth noting that, Figure 1 Or Figure 2 The application scenario of the method shown is a uniform cooling smooth particle flow system.

[0069] The implementation process of the mechanism-data reconstructed particle flow system temperature time sequence prediction method provided by the embodiments of the present application will be described below by taking a mechanism-data reconstructed particle flow system temperature time sequence prediction device as an execution subject. It can be understood that the execution subject of the mechanism-data reconstructed particle flow system temperature time sequence prediction method provided by the embodiments of the present application includes but is not limited to the mechanism-data reconstructed particle flow system temperature time sequence prediction device.

[0070] Figure 1 The flowchart of the mechanism-data reconstructed particle flow system temperature time sequence prediction method provided by the embodiments of the present application is shown in Figure 1 The method comprises the following steps.

[0071] In step 101, the discrete element method (DEM) is used to perform numerical simulation on the reconstructed particle flow system to generate experimental particle temperatures at different times.

[0072] In the embodiments of the present application, the reconstructed particle flow system is a uniformly cooled smooth particle flow system. Specifically, the uniformly cooled smooth particle flow system is simulated by DEM to generate experimental particle temperatures at different times, which are used as high-precision numerical experimental points.

[0073] In step 102, a physical information based neural network (PINN) is constructed, and a theoretical comprehensive loss function is constructed according to the experimental particle temperatures at different times and the physical control equation.

[0074] In the embodiments of the present application, the data sample loss function of the dimensionless particle temperature is constructed by a non-replacement random sampling method Based on the initialized PINN, the initial condition loss function is constructed according to the initial time particle temperature predicted by the output layer of the network The input time data set of the PINN is obtained by a uniform distribution strategy, that is, the normalization is performed before the hidden layer at different times, and the mechanism-data dual-drive control equation loss function is determined according to the physical control equation of the uniformly cooled particle flow system The data sample loss function, the initial condition loss function and the mechanism-data dual-drive control equation loss function are weighted and summed to obtain the theoretical comprehensive loss function.

[0075] In step 103, the target comprehensive loss function is determined according to the theoretical comprehensive loss function under the system missing condition, and the network parameters of the PINN are iteratively updated to construct a particle temperature time sequence feature prediction model.

[0076] In the embodiments of the present application, the system missing condition includes missing energy dissipation coefficient, missing initial time particle temperature or missing energy dissipation coefficient and initial time particle temperature.

[0077] It is worth noting that if the system missing condition is missing the initial time particle temperature or missing the energy dissipation coefficient and the initial time particle temperature, the initial condition loss function does not exist, the initial condition loss function is removed from the theoretical comprehensive loss function to obtain the target comprehensive loss function; the network weight and the network bias of the PINN are iteratively updated to minimize the target comprehensive loss function, and a particle temperature time sequence feature prediction model is constructed.

[0078] In step 104, based on the particle temperature time sequence feature prediction model, the particle temperature is predicted according to the input time information to be predicted to obtain the target particle temperature time sequence.

[0079] In the embodiment of the application, the particle temperature time sequence feature prediction model can identify a complete uniform cold particle flow system and has good generalization performance, which can successfully predict the system missing condition, such as the energy dissipation coefficient and / or the initial time particle temperature; on the other hand, the particle temperature time sequence feature prediction model is applied to reconstruct the particle flow system flow field characteristic prediction, and the target particle temperature time sequence can be obtained according to the input time information to be predicted, which is the law of the evolution of the particle temperature in the system with time.

[0080] In the technical scheme provided by the embodiment of the application, the discrete element method is used to numerically simulate the reconstructed particle flow system to generate experimental particle temperatures at different times, and the reconstructed particle flow system is a uniform cooling smooth particle flow system; the theoretical comprehensive loss function is constructed based on the physical information neural network according to the experimental particle temperatures at different times and the physical control equation; the target comprehensive loss function is determined according to the theoretical comprehensive loss function according to the system missing condition, and the network parameters of the physical information neural network are iteratively updated to minimize the target comprehensive loss function, and the particle temperature time sequence feature prediction model is constructed, the system missing condition includes missing the energy dissipation coefficient, missing the initial time particle temperature or missing the energy dissipation coefficient and the initial time particle temperature; based on the particle temperature time sequence feature prediction model, the particle temperature is predicted according to the input time information to be predicted to obtain the target particle temperature time sequence, which can fully reconstruct the particle flow full flow field motion information, and the physical information neural network method combines the numerical experimental data with the incomplete control equation and the initial condition, which can not only accurately predict the particle temperature time sequence feature in the particle flow system, but also can assist in determining the complete control equation form and the initial condition, and has high robustness, accuracy, universality and scalability.

[0081] Figure 2 The flowchart of another mechanism-data reconstructed particle flow system temperature time sequence prediction method provided by the embodiment of the application is shown as Figure 2 The method comprises the following steps.

[0082] Step 201, according to the preset particle velocity distribution function, and based on Newton's law of motion and the linear interaction force model between particles, simulate the particle motion information in the reconstructed granular flow system, and count the experimental particle velocity at different times.

[0083] In the embodiment of the application, each step is executed by a mechanism-data reconstructed granular flow system temperature time sequence prediction device.

[0084] In the embodiment of the application, the reconstructed granular flow system is a uniformly cooled smooth granular flow system, and the system characteristic parameters include but are not limited to simulation domain length, particle number, particle diameter, particle density, regional solid content, recovery coefficient, total simulation time, Young's modulus, Poisson's ratio, DEM time step and simulation initial temperature.

[0085] For example: the simulation domain length L is 0.064 m, the particle number N is 5w, the particle diameter size is 0.001 m, the particle density is 1500 kg / m 3 , the regional solid content is 0.1, the recovery coefficient is 0.95, the total simulation time is 5 s, the Young's modulus is 1×10 8 , the Poisson's ratio is 0.3, the DEM time step is s, and the simulation initial temperature is 0.01 m 2 / s 2 .

[0086] Specifically, under the condition of three-dimensional periodic boundary conditions, the smooth particles are uniformly distributed, and the particle velocity distribution function is:

[0087]

[0088] wherein, is the particle velocity distribution function, which gives the number density of particles at a specific velocity c; n is the number density; c is the particle velocity, which is a velocity vector containing the size and direction of particle motion; is the experimental particle temperature.

[0089] The initial particle temperature in the system is determined to ensure that the average particle velocity is 0.

[0090] It is worth noting that DEM is a method based on Newton's law of motion combined with a linear interaction force model between particles, which is used to describe the particle motion information in the system.

[0091] In this process, the particles in the system are uniformly distributed over a period of time. In the reconstructed granular flow system, the experimental particle velocity at different times is counted.

[0092] Step 202, generating experimental particle temperatures at different times according to experimental particle velocities at different times.

[0093] Step 202 specifically includes:

[0094] Step 2021, determining experimental particle fluctuation velocities according to experimental particle velocities at different times and average particle velocities.

[0095] In the embodiment of the present application, by , the numerical experimental particle velocities and the average particle velocities at different times are calculated to obtain the particle fluctuation velocities. Wherein, N is the number of particles; is the experimental particle fluctuation velocity of the i-th particle; is the experimental particle velocity of the i-th particle; is the average particle velocity.

[0096] Step 2022, generating experimental particle temperatures at different times according to experimental particle fluctuation velocities.

[0097] In the embodiment of the present application, with the passage of time, due to the free inelastic collision between particles, the particle fluctuation energy in the system, i.e. the particle temperature, will gradually decrease over time. The particle temperature is defined as the average kinetic energy level of the particle velocity fluctuation in the system, and this energy loss reflects the energy attenuation process caused by inelastic collision inside the particle system. This phenomenon can be quantitatively described by the change of particle temperature, and the particle temperature is defined as:

[0098]

[0099] Wherein, is the experimental particle temperature; is the fluctuation velocity; N is the number of particles; is the fluctuation velocity of the i-th particle; is the average value of the fluctuation energy of all particles.

[0100] In the embodiment of the present application, the experimental particle temperatures at different times simulated by the discrete element method above are taken as high-precision numerical experimental points.

[0101] Step 203, dimensionless processing of experimental particle temperatures at different times to obtain dimensionless experimental particle temperatures.

[0102] In the embodiment of the present application, by , the experimental particle temperatures at different times are dimensionless processed to obtain dimensionless experimental particle temperatures. Wherein, is the dimensionless particle temperature, is the experimental particle temperature, is the particle temperature at the initial time.

[0103] Step 204, constructing a data sample loss function and an initial condition loss function according to the dimensionless experimental particle temperature through the PINN.

[0104] In the embodiment of the application, the PINN is pre-constructed, and the PINN is composed of a full connection neural network, including an input layer, a plurality of hidden layers and an output layer; the input layer receives normalized time information, the number of neurons of each layer is matched with the feature dimension, and the PINN model can fully capture the law of evolution of the average particle temperature of the system with time; the hidden layer performs nonlinear transformation through an activation function, and extracts and converts high-level abstract representations of input features layer by layer; the output layer adopts a Softplus activation function to smoothly output the data from the hidden layer, and ensures that the predicted particle temperature is positive and conforms to the physical reality. The constructed PINN can be initialized for subsequent training and application.

[0105] As an optional solution, the input layer of the PINN is in the form of layer=[1, 128, 128, 128, 128, 1], indicating that the input layer is composed of only one variable of time, the time step is 0.01s in the training process, the hidden layer is 4 layers (L=4), and the output layer is a particle temperature related variable layer; the hidden layer performs nonlinear transformation through an activation function, and extracts and converts high-level abstract representations of input features layer by layer; the output layer adopts a Softplus activation function to smoothly output the data from the hidden layer and keep it positive, which conforms to the physical reality, and the form of the entire full connection neural network is as follows:

[0106]

[0107]

[0108]

[0109] wherein, is the input normalized time information, is the output of the kth layer, is the network weight of the kth layer, is the output of the input layer, is the output of the (k-1)th layer, is the network bias of the kth layer, is the activation function, and L is the total number of layers, is a nonlinear transformation function.

[0110] In the embodiment of the application, step 204 specifically includes:

[0111] ​​Step 2041, random sampling is performed on the dimensionless experimental particle temperature to obtain the dimensionless sparse particle temperature.

[0112] In the embodiment of the present application, the dimensionless experimental particle temperature is collected by the non-replacement random sampling method, and the dimensionless experimental particle temperature corresponding to a specified number of different time points is obtained; the dimensionless experimental particle temperature corresponding to the different time points is determined as the dimensionless sparse particle temperature.

[0113] It is worth noting that the specified number can be determined according to actual needs, and the embodiment of the present application does not limit this. As an optional solution, the specified number is 21, that is, 21 dimensionless experimental particle temperatures corresponding to different time points are randomly sampled as the dimensionless sparse real particle temperature.

[0114] Step 2042, a data sample loss function is constructed by PINN according to the dimensionless sparse particle temperature and the corresponding different time points.

[0115] In the embodiment of the present application, the different time points corresponding to the dimensionless sparse particle temperature are normalized to obtain normalized sparse time information, and the normalized sparse time information is used as the input of PINN. According to the dimensionless particle temperature prediction value output by PINN and the dimensionless sparse real particle temperature, the data sample loss function constructed is as follows:

[0116]

[0117] Wherein, is the data sample loss function, N Data is the number of dimensionless sparse particle temperatures, is the dimensionless particle temperature prediction value of the i-th time point predicted, is the real value of the i-th time point of the dimensionless particle temperature randomly sampled from the dimensionless experimental particle temperature.

[0118] In the embodiment of the present application, the data residual set of the data sample loss function is composed of the data loss of the dimensionless sparse particle temperature corresponding to only a small number of random time points.

[0119] Step 2043, an initial condition loss function is constructed by PINN according to the dimensionless initial time point experimental particle temperature.

[0120] In the embodiment of the present application, according to the initial time point particle temperature prediction value output by PINN and the dimensionless sparse particle temperature, the initial condition loss function constructed is as follows:

[0121]

[0122] wherein, is the initial condition loss function, is the dimensionless particle temperature prediction value at the initial moment, 1 is the dimensionless particle temperature true value at the initial moment, that is, the dimensionless experimental particle temperature at the initial moment.

[0123] Step 205, normalizing different moments to obtain normalized time information.

[0124] In the embodiment of the application, by , the current moment, the maximum moment and the minimum moment are calculated to generate normalized time information. Wherein, is the normalized time information, t is the current moment, t min is the minimum moment in different moments, t max is the maximum moment in different moments.

[0125] In the embodiment of the application, by normalizing the time information, the optimal input range of the PINN hidden layer activation function can be met, so as to improve the model convergence speed and prediction accuracy.

[0126] Step 206, constructing a mechanism-data double-drive control equation loss function by PINN according to the preset local weight coefficient, the normalized time information, the dimensionless experimental particle temperature and the physical control equation.

[0127] In the embodiment of the application, the physical control equation, that is, the particle temperature equation. According to the normalized time information and the dimensionless experimental particle temperature, the corresponding uniform cooling state particle temperature control equation is generated as follows:

[0128]

[0129] wherein, t min is the minimum moment in different moments, t max is the maximum moment in different moments, is the differential function of the dimensionless experimental particle temperature, A is the energy dissipation coefficient, is the particle temperature at the initial moment, is the dimensionless experimental particle temperature.

[0130] It is worth noting that the energy dissipation coefficient A can be taken as an unknown value participating in PINN iteration optimization when missing. If it needs to be closed in advance, it can be obtained by solving the following kinetic theory of particles:

[0131]

[0132]

[0133] wherein, is the energy dissipation coefficient obtained by solving the kinetic theory of granular flow, e is the restitution coefficient of the granules, is the average solid content in the system, is the radial distribution function, is the diameter of the granules.

[0134] In the embodiment of the present application, the loss function related to the mechanism-data dual-driven control equation is constructed according to the normalized time information, the dimensionless experimental granule temperature, the physical control equation and the local weight coefficient preset according to the physical characteristics of the system as follows:

[0135]

[0136] wherein, is the loss function of the mechanism-data dual-driven control equation, A is the energy dissipation coefficient, is the initial granule temperature, is the i-th dimensionless predicted experimental granule temperature, is the number of granules, is the differential function of the i-th dimensionless predicted experimental granule temperature, t min is the minimum time in different time, t max is the maximum time in different time, is the local weight coefficient.

[0137] It is worth mentioning that the local weight coefficient is determined according to the reciprocal of the dimensionless granule temperature predicted at the corresponding time, and the local weight coefficient preset according to the physical characteristics of the system is specifically that the physical model of the local weight is selected as the reciprocal of the dimensionless granule temperature predicted at the corresponding time.

[0138] The present application embeds physical information by incorporating physical laws as constraint conditions into the loss function, part of which includes the mechanism-data dual-driven control equation loss function corresponding to the control equation, and the physical law in the uniform cooling granule flow system is the granule temperature equation. According to the partial differential equation term given by the physical law, the state variable differential is calculated in turn by using automatic differentiation technology, and then the mechanism-data dual-driven control equation loss function combined with data and physics is formed. The local weight coefficient of the physically driven loss part is adjusted to the reciprocal of the dimensionless granule temperature predicted at the corresponding time, so as to enhance the contribution of the corresponding control equation when the granule temperature is low, thereby being able to well capture the granule temperature variation law when the granule temperature is low, and further improve the network prediction accuracy.

[0139] Step 207, according to the preset global weight coefficient, the data sample loss function, the initial condition loss function and the mechanism-data double drive control equation loss function are weighted to generate a theoretical comprehensive loss function.

[0140] In the embodiment of the application, the global weight coefficient includes a data sample weight coefficient, an initial condition weight coefficient and a mechanism-data double drive control equation weight coefficient.

[0141] It should be noted that the global weight coefficient can be set according to actual needs, and the embodiment of the application does not limit this.

[0142] Specifically, the data sample loss function, the initial condition loss function and the mechanism-data double drive control equation loss function are weighted and summed to obtain the theoretical comprehensive loss function as follows:

[0143]

[0144] Wherein, is the mechanism-data double drive control equation weight coefficient, is the initial condition weight coefficient, is the data sample weight coefficient, A is the energy dissipation coefficient, is the initial particle temperature, is the dimensionless particle temperature true value of the i-th particle, is the number of input layer at different times, is the differential function prediction value of the i-th dimensionless experimental particle temperature, t min is the minimum time in different times, t max is the maximum time in different times, is the local weight coefficient, N Data is the number of dimensionless sparse particle temperature, is the dimensionless particle temperature prediction value at the i-th time, is the dimensionless sparse particle temperature true value at the i-th time, is the dimensionless particle temperature prediction value at the initial time, 1 is the dimensionless particle temperature true value at the initial time.

[0145] Step 208, according to the system missing condition, the target comprehensive loss function is determined according to the theoretical comprehensive loss function, and the network parameters of the neural network based on physical information are iteratively updated to construct a particle temperature time series feature prediction model.

[0146] In the embodiment of the present application, the specific process of determining the target comprehensive loss function and model training is described in turn for three types of system missing conditions. The system missing conditions include: missing energy dissipation source term parameters, resulting in incomplete control equations as physical laws, missing initial particle temperature, and missing both energy dissipation source term parameters and initial particle temperature.

[0147] If the system missing condition is missing energy dissipation coefficient, step 208 specifically includes:

[0148] Step 2081, determine the theoretical comprehensive loss function as the target comprehensive loss function.

[0149] In the embodiment of the present application, if the system missing condition is missing energy dissipation coefficient, there is no influence on the theoretical comprehensive loss function, and the theoretical comprehensive loss function with missing energy dissipation coefficient is determined as the target comprehensive loss function.

[0150] Step 2082, through a preset optimization algorithm, the neural network based on physical information is trained and optimized to minimize the target comprehensive loss function, the network parameters are iteratively updated, and the particle temperature time series feature prediction model and the optimized energy dissipation coefficient are generated.

[0151] In the embodiment of the present application, the optimization algorithm iteratively updates and adjusts the network parameters to minimize the target comprehensive loss function. The optimization algorithm strategy is coupled with ADAM and L-BFGS-B, specifically, the ADAM optimizer is used to make the parameters globally converge quickly, and then the L-BFGS-B algorithm is used to realize fine optimization and parameter adjustment, finally the model converges, and the particle temperature time series feature prediction model and the optimized energy dissipation coefficient are generated.

[0152] It is worth noting that the initial learning rate and the number of training steps of the ADAM optimizer are determined by specific examples, and the embodiment of the present application does not limit this.

[0153] In the embodiment of the present application, the network parameters include network weights and network biases.

[0154] Further, based on the comparison between the optimized energy dissipation coefficient and the energy dissipation coefficient determined in advance by the kinetic theory of granular particles, the energy dissipation coefficient error result is generated. The optimized energy dissipation coefficient is obtained by iterative calculation based on the physical control equation in the particle temperature time series feature prediction model, and the theoretical energy dissipation coefficient is obtained by solving the kinetic theory of granular particles.

[0155] In one specific embodiment, the calculated optimized energy dissipation coefficient is -57.3452, and the theoretical energy dissipation coefficient obtained by solving the kinetic theory of granular particles is -57.3452. The relative error is 0.0044%, which proves that the present application has the ability to accurately predict the core constitutive relationship parameters.

[0156] If the system missing condition is the missing particle temperature at the initial time, step 208 specifically comprises:

[0157] Step 3081, removing the initial condition loss function from the theoretical comprehensive loss function to generate a target comprehensive loss function.

[0158] In the embodiment of the application, if the system missing condition is the missing particle temperature at the initial time, there is no initial condition loss function, and the target comprehensive loss function obtained by removing the initial condition loss function from the theoretical comprehensive loss function is as follows:

[0159]

[0160] wherein, is a mechanism-data double-drive control equation weight coefficient, is a data sample weight coefficient, A is an energy dissipation coefficient, is the predicted value of the experimental particle temperature at the i-th time, N t is the number of different time points in the input layer, is the predicted value of the differential function of the experimental particle temperature at the i-th time, t min is the minimum time in different time points, t max is the maximum time in different time points, is a local weight coefficient, N Data is the number of sparsely sampled particle temperatures, is the predicted value of the particle temperature at the i-th time, is the true value of the particle temperature at the i-th time after sparse sampling.

[0161] Step 3082, by a preset optimization algorithm, taking minimizing the target comprehensive loss function as the goal, model training and optimization are performed on the neural network based on physical information, network parameters are iteratively updated, a particle temperature time series feature prediction model and an optimized particle temperature at the initial time are generated.

[0162] In the embodiment of the application, the network parameters are iteratively updated and adjusted by using the optimization algorithm, so that the target comprehensive loss function is minimized, wherein only the energy dissipation coefficient is an unknown parameter for network iterative optimization. The optimization algorithm strategy is coupled by ADAM and L-BFGS-B, specifically, the ADAM optimizer is used to make the parameters globally converge quickly, and then the L-BFGS-B algorithm is used to realize fine optimization and parameter adjustment, so that the final model converges, and a particle temperature time series feature prediction model and an optimized particle temperature at the initial time are generated.

[0163] It is worth noting that the initial learning rate and the number of training steps of the ADAM optimizer are determined by specific examples, and the embodiment of the application does not limit this.

[0164] The network parameters include network weights and network biases in the embodiment of the application.

[0165] Further, the initial time particle temperature error result is generated by comparing the optimized initial time particle temperature and the obtained initial time real particle temperature, the optimized initial time particle temperature is obtained based on iterative calculation of the particle temperature time sequence feature prediction model, and the initial time real particle temperature is actually obtained.

[0166] If the system missing condition is missing energy dissipation coefficient and initial time particle temperature, step 208 specifically includes:

[0167] Step 4081, removing the initial condition loss function from the theoretical comprehensive loss function to generate a target comprehensive loss function.

[0168] In the embodiment of the application, if the energy dissipation coefficient and the initial time particle temperature are missing, the initial condition loss function does not exist, the initial condition loss function is removed from the theoretical comprehensive loss function of the missing energy dissipation coefficient to obtain a target comprehensive loss function, that is, the data sample loss function and the mechanism-data dual-drive control equation loss function containing the missing energy dissipation coefficient are integrated to generate a target comprehensive loss function. The generated target comprehensive loss function is as follows:

[0169]

[0170] Wherein, is a mechanism-data dual-drive control equation weight coefficient, is a data sample weight coefficient, A is an energy dissipation coefficient, is an experimental particle temperature prediction value at the i th time, N t is the number of different times in the input layer, is a differential function prediction value of the experimental particle temperature at the i th time, t min is the minimum time in different times, t max is the maximum time in different times, is a local weight coefficient, N Data is the number of sparsely sampled particle temperatures, is a particle temperature prediction value at the i th time, is a real value of the particle temperature at the i th time after sparse sampling.

[0171] Step 4082, by a preset optimization algorithm, taking minimizing the target comprehensive loss function as the goal, the PINN is model trained and optimized, the network parameters are iteratively updated, the particle temperature time sequence feature prediction model, the optimized energy dissipation coefficient and the optimized initial time particle temperature are generated.

[0172] Figure 3A structure diagram of a PINN suitable for a uniform cooling particle flow system is provided for an embodiment of the present application, as shown in Figure 3 The time information t enters the PINN, and first time normalization processing is performed on the normalized time information The normalized time information is input into a neural network (Neural Networks) for calculation, and a dimensionless particle temperature prediction value is output; differential calculation (AD) is performed on the normalized time information to obtain a differential calculation result , and a comprehensive loss function Loss is constructed; model training and optimization are performed on the PINN with the goal of minimizing the comprehensive loss function, and if the loss function result is less than a preset loss threshold , the particle temperature time series feature prediction model is constructed; if the loss function result is greater than or equal to the preset loss threshold , the network parameter bias result is calculated and the network parameters W k and b k are iteratively updated until the loss function result is less than the preset loss threshold . Further, the dimensionless particle temperature prediction value at any time i and the corresponding real particle temperature prediction value have a linear mapping relationship , and the trained particle temperature time series feature prediction model can output the particle temperature prediction value .

[0173] In an embodiment of the present application, the network parameters are iteratively updated and adjusted using an optimization algorithm to minimize the target comprehensive loss function. The optimization algorithm strategy is coupled with ADAM and L-BFGS-B, specifically, the ADAM optimizer is used to make the parameters globally converge quickly, and then the L-BFGS-B algorithm is used to realize fine optimization and parameter adjustment, and finally the model converges to generate the particle temperature time series feature prediction model, the optimized energy dissipation coefficient, and the optimized particle temperature at the initial time.

[0174] It should be noted that the initial learning rate and the number of training steps of the ADAM optimizer are determined by specific examples, and the present application does not limit this.

[0175] In an embodiment of the present application, the network parameters include network weights and network biases.

[0176] Further, based on the comparison between the optimized energy dissipation coefficient and the theoretically calculated energy dissipation coefficient, an energy dissipation coefficient error result is generated, the optimized energy dissipation coefficient is obtained by iterative calculation based on the physical control equation in the particle temperature time series feature prediction model, and the theoretically calculated energy dissipation coefficient is obtained by solving the kinetic theory of particles.

[0177] Further, based on the initial time point optimized particle temperature and the obtained initial time point real particle temperature, an initial time point particle temperature error result is generated, the initial time point optimized particle temperature is obtained based on the particle temperature time sequence feature prediction model iterative calculation, and the initial time point real particle temperature is actually obtained.

[0178] Step 209, normalizing the to-be-predicted time information to generate normalized to-be-predicted time information.

[0179] In the embodiment of the application, by , the current time, the maximum time and the minimum time in the to-be-predicted time information are calculated to generate the normalized to-be-predicted time information. Wherein, is the normalized to-be-predicted time information, t is the current time in the to-be-predicted time information, t min is the minimum time in the to-be-predicted time information, t max is the maximum time in the to-be-predicted time information.

[0180] Step 210, calculating the normalized to-be-predicted time information by the particle temperature time sequence feature prediction model to generate the dimensionless particle temperature time sequence.

[0181] In the embodiment of the application, the dimensionless particle temperature time sequence specifically refers to the dimensionless particle temperature size at each normalized time, and the dimensionless particle temperature time sequence is obtained by arranging the dimensionless particle temperature in sequence according to the dimensionless time. The dimensionless particle temperature time sequence is the corresponding relationship between the dimensionless particle temperature and the normalized to-be-predicted time information.

[0182] Specifically, the normalized to-be-predicted time information is input into the particle temperature time sequence feature prediction model for prediction, and the dimensionless particle temperature time sequence is output.

[0183] Step 211, performing reverse normalization processing on the dimensionless particle temperature time sequence to generate a target particle temperature time sequence.

[0184] In the embodiment of the application, the dimensionless particle temperature is time-reverse normalized with the normalized to-be-predicted time information and the dimensionless particle temperature is dimensioned to generate the target particle temperature time sequence.

[0185] In the embodiment of the application, the reverse normalization strategy is: . Wherein, is the target particle temperature time sequence, is the mapping function of the PINN network, is the normalized to-be-predicted time information, is the initial time point particle temperature, The dimensionless particle temperature time series.

[0186] It is worth noting that the initial particle temperature can be the initial optimized particle temperature or the initial real particle temperature.

[0187] Further, the predicted target particle temperature time series and the obtained real particle temperature time series are compared to generate a particle temperature time series feature error result. To verify the ability of the PINN to solve the inverse problem of the uniform cooling particle flow system and the generalization performance of the verifier, the real particle temperature time series of the uniform cooling particle flow system is obtained according to the law of change with time:

[0188]

[0189] wherein, is the real particle temperature time series, is an intermediate parameter, is the initial particle temperature, t is the time, and A is the energy dissipation coefficient.

[0190] The target particle temperature time series predicted by the particle temperature time series feature prediction model trained based on the PINN at each time is compared with the real particle temperature time series, and the relative error between the two is defined as:

[0191]

[0192] wherein, R(t i ) is the relative error between the target particle temperature time series and the real particle temperature time series at time t, is the target particle temperature time series at time t, is the real particle temperature time series at time t. The average error of the two is:

[0193]

[0194]

[0195] wherein, is the average error between the target particle temperature time series and the real particle temperature time series at time t, N t is the number of particles, is the target particle temperature time series at time t, is the real particle temperature time series at time t.

[0196] ​​​​​​In a specific embodiment, the input layer time interval involved in the PINN training here is 0.01s. But the time interval involved in the generalization process is further encrypted to 0.005s.

[0197] In a specific embodiment, Figure 4 A comparison chart between the target particle temperature time series and the real particle temperature time series in the uniform cooling particle flow system when the energy dissipation coefficient A is missing is provided for the embodiment of the present application, as shown in Figure 4 The particle recovery coefficient e is 0.95, the average solid content in the system is 0.1, the number of particles N is 50000, and the particle temperature at the initial time is 0.01 m 2 / s 2 . The horizontal axis is time, unit: s, range: [10 -4 ,10 0 ]; the vertical axis is the temperature time series, unit: m 2 / s 2 , range: [10 -4 ,10 -2 ]; the red solid line represents the target particle temperature time series predicted by the particle temperature time series feature prediction model based on PINN training, and the black dotted line represents the real particle temperature time series.

[0198] In a specific embodiment, Figure 5 A time series chart of the relative error between the target particle temperature time series and the real particle temperature time series in the uniform cooling particle flow system when the energy dissipation coefficient A is missing is provided for the embodiment of the present application, as shown in Figure 5 The horizontal axis is time, unit: s, range: [0,5]; the vertical axis is the relative error, unit: %, range: [0.0,0.4].

[0199] As can be seen from Figure 4 and Figure 5 , the prediction accuracy of the particle temperature of the present application is high. The average relative error of the particle temperature determined in the training process for relatively discrete data sample points is 0.0433%, and the average relative error of the particle temperature obtained by generalization for relatively dense data sample points is only 0.0437%. The generalization error is comparable to the training error, and the difference from the overall average relative error 0.0435% is not large, which shows that the training and generalization ability of the particle temperature time series feature prediction model based on PINN training is good.

[0200] Further, the inverse problem of the uniform cooling particle flow system with missing initial particle temperature and energy dissipation coefficient and initial conditions is investigated in turn in the embodiment of the present application. The overall loss function form involved in both is:

[0201]

[0202] wherein, is the mechanism-data dual-driven control equation weight coefficient, is the data sample weight coefficient, A is the energy dissipation coefficient, is the predicted value of the experimental particle temperature at the i th moment, N t is the number of different moments in the input layer, is the predicted value of the differential function of the experimental particle temperature at the i th moment, t min is the minimum moment in different moments, t max is the maximum moment in different moments, is the local weight coefficient, N Data is the number of sparsely sampled particle temperatures, is the predicted value of the particle temperature at the i th moment, is the true value of the particle temperature at the i th moment after sparse sampling.

[0203] Further, for the missing initial particle temperature, the global weight is set to , the ADAM initial learning rate is set to 0.0001, and the model iteration number (Niter) is 5x10 4 times.

[0204] Further, for the missing energy dissipation coefficient and the initial particle temperature at the initial moment, the global weight is set to , the ADAM initial learning rate is set to 0.001, and the model iteration number (Niter) is 2x10 5 times.

[0205] In summary, the results of the above two inverse problem solutions for the uniform cooling particle flow system are as follows:

[0206]

[0207] The application adopts the strategies of input information normalization, particle temperature dimensionless, and increasing local physical constraint weight at low temperature moment, and the solving accuracy can be improved by 2-3 orders of magnitude compared with the standard PINN, and the generalization performance is good; compared with the traditional method which can only solve the “complete control equation and initial condition” direct problem, PINN can effectively reconstruct the flow field and identify the incomplete control equation and the full flow field information based on the physical prior knowledge and sparse data of the flow field; for the incomplete control equation group or missing initial condition in this embodiment, the full flow field information can still be obtained through network iteration optimization, and the complete control equation is identified; for the trained PINN network, the learning process has strong interpretability and excellent generalization ability, so the full flow field information can be accurately obtained by directly inputting the time information.

[0208] It should be noted that the acquisition, storage, use, processing, etc. of data in the technical solutions in the present application comply with relevant provisions of laws and regulations. The user information in the embodiments of the present application is obtained through legal and compliant channels, and the acquisition, storage, use, processing, etc. of the user information is authorized and agreed by the client.

[0209] It should be noted that the information collected in the present application is information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application, etc. of the relevant data comply with relevant laws, regulations and standards of countries and regions, necessary security measures are taken, do not violate public order and good customs, and provide corresponding operation portals for users to choose to authorize or refuse.

[0210] It should be noted that the technical solutions provided by the present application provide corresponding operation portals for users to choose to agree or refuse the automatic decision result; if the user chooses to refuse, the expert decision process is entered.

[0211] Figure 6 A mechanism-data reconstructed granular flow system temperature time sequence prediction device structure diagram is provided for the embodiments of the present application, which is used to execute the mechanism-data reconstructed granular flow system temperature time sequence prediction method, as shown in Figure 6 The device comprises:

[0212] The numerical simulation unit 11 is used for numerically simulating the reconstructed granular flow system through the discrete element method to generate experimental particle temperatures at different times, and the reconstructed granular flow system is a uniformly cooled smooth granular flow system.

[0213] The theoretical comprehensive loss function construction unit 12 is used for constructing a theoretical comprehensive loss function according to the experimental particle temperatures at different times and the physical control equation through the constructed neural network based on physical information.

[0214] The particle temperature time sequence feature prediction model training unit 13 is used for determining a target comprehensive loss function according to the theoretical comprehensive loss function, and iteratively updating the network parameters of the neural network based on physical information to construct a particle temperature time sequence feature prediction model, with the goal of minimizing the target comprehensive loss function, and the system missing conditions include missing energy dissipation coefficient, missing initial time particle temperature or missing energy dissipation coefficient and initial time particle temperature.

[0215] The particle temperature time sequence feature prediction unit 14 is used for performing particle temperature prediction according to the input predicted time information based on the particle temperature time sequence feature prediction model to obtain a target particle temperature time sequence.

[0216] In the embodiment of the present application, the numerical simulation unit 11 is specifically configured to simulate the particle motion information in the reconstructed granular flow system according to the preset particle velocity distribution function, and based on Newton's law of motion and a linear inter-particle force model, and to count the experimental particle velocities at different times; according to the experimental particle velocities at different times, the average value of the fluctuation energy of all particles in the system is counted, and the experimental particle fluctuation velocity is determined in combination with the average particle velocity; and according to the experimental particle fluctuation velocity, the experimental particle temperatures at different times are generated.

[0217] In the embodiment of the present application, the theoretical comprehensive loss function construction unit 12 is specifically configured to perform dimensionless processing on the experimental particle temperatures at different times to obtain dimensionless experimental particle temperatures; through the neural network based on physical information, a data sample loss function and an initial condition loss function are constructed according to the dimensionless experimental particle temperatures; the different times are normalized to obtain normalized time information; through the neural network based on physical information, a mechanism-data dual-driven control equation loss function is constructed according to the preset local weight coefficient, the normalized time information, the dimensionless experimental particle temperature and the physical control equation, and the local weight coefficient is determined according to the reciprocal of the dimensionless particle temperature predicted at the corresponding time; the data sample loss function, the initial condition loss function and the mechanism-data dual-driven control equation loss function are weighted according to the preset global weight coefficient to generate the theoretical comprehensive loss function.

[0218] In the embodiment of the present application, the theoretical comprehensive loss function construction unit 12 is specifically configured to randomly sample the dimensionless experimental particle temperatures to obtain dimensionless sparse particle temperatures; through the neural network based on physical information, a data sample loss function is constructed according to the dimensionless sparse particle temperatures and the corresponding different times; through the neural network based on physical information, an initial condition loss function is constructed according to the dimensionless experimental particle temperature at the initial time.

[0219] In the embodiment of the present application, the system missing condition is the missing energy dissipation coefficient; the particle temperature time series feature prediction model training unit 13 is specifically configured to determine the theoretical comprehensive loss function as a target comprehensive loss function; through a preset optimization algorithm, the neural network based on physical information is model trained and optimized with the minimum target comprehensive loss function as the target, the network parameters are iteratively updated, a particle temperature time series feature prediction model and an optimized energy dissipation coefficient are generated, and the network parameters include network weights and network biases.

[0220] In the embodiment of the present application, the system missing condition is missing the particle temperature at the initial moment; the particle temperature time sequence feature prediction model training unit 13 is specifically configured to remove the initial condition loss function from the theoretical comprehensive loss function, generate a target comprehensive loss function; through a preset optimization algorithm, the neural network based on physical information is trained and optimized to minimize the target comprehensive loss function, iteratively update the network parameters, generate the particle temperature time sequence feature prediction model and the optimized particle temperature at the initial moment, and the network parameters include network weights and network biases.

[0221] In the embodiment of the present application, the system missing condition is missing the particle temperature at the initial moment; the particle temperature time sequence feature prediction model training unit 13 is specifically configured to remove the initial condition loss function from the theoretical comprehensive loss function, generate a target comprehensive loss function; through a preset optimization algorithm, the neural network based on physical information is trained and optimized to minimize the target comprehensive loss function, iteratively update the network parameters, generate the particle temperature time sequence feature prediction model and the optimized particle temperature at the initial moment, and the network parameters include network weights and network biases.

[0222] In the embodiment of the present application, the particle temperature time sequence feature prediction unit 14 is specifically configured to normalize the to-be-predicted time information to generate normalized to-be-predicted time information; calculate the normalized to-be-predicted time information through the particle temperature time sequence feature prediction model to generate a dimensionless particle temperature time sequence, which is a corresponding relationship between the dimensionless particle temperature and the normalized to-be-predicted time information; and perform time de-normalization on the dimensionless particle temperature with the normalized to-be-predicted time information and dimensionality processing on the particle temperature to generate a target particle temperature time sequence.

[0223] In the embodiment of the present application, the device further comprises a first comparison unit 15.

[0224] The first comparison unit 15 is configured to compare the predicted target particle temperature time sequence with the obtained real particle temperature time sequence to generate a particle temperature time sequence feature error result.

[0225] In the embodiment of the present application, the device further comprises a second comparison unit 16.

[0226] The second comparison unit 16 is configured to compare the optimized energy dissipation coefficient with the pre-calculated theoretical energy dissipation coefficient to generate an energy dissipation coefficient error result, and the optimized energy dissipation coefficient is obtained by iteratively calculating the physical control equation in the particle temperature time sequence feature prediction model.

[0227] In the embodiment of the present application, the device further comprises a third comparison unit 17.

[0228] The third comparison unit 17 is configured to compare the optimized particle temperature at the initial time and the real particle temperature at the initial time obtained to generate a particle temperature error result at the initial time, and the optimized particle temperature at the initial time is obtained based on iterative calculation of the particle temperature time sequence feature prediction model.

[0229] In the scheme of the embodiment of the present application, the discrete element method is used to simulate the reconstructed granular flow system to generate experimental particle temperatures at different times, and the reconstructed granular flow system is a uniform cooling smooth granular flow system; the neural network based on physical information is constructed, the theoretical comprehensive loss function is constructed according to the experimental particle temperatures at different times and the physical control equation; the target comprehensive loss function is determined according to the theoretical comprehensive loss function under the system missing condition, and the network parameters of the neural network based on physical information are iteratively updated to minimize the target comprehensive loss function, and the particle temperature time sequence prediction model is constructed, and the system missing condition includes missing energy dissipation coefficient, missing particle temperature at the initial time or missing energy dissipation coefficient and particle temperature at the initial time; the particle temperature is predicted based on the particle temperature time sequence prediction model according to the input time information to be predicted to obtain the target particle temperature time sequence, the motion information of the whole flow field of the granular flow can be fully reconstructed, the numerical experimental data, the incomplete control equation and the initial condition are combined based on the physical information neural network method, not only the particle temperature time sequence feature in the granular flow system can be accurately predicted, but also the complete control equation form and the initial condition can be determined, and the robustness, accuracy, universality and scalability are high.

[0230] The system, device, module or unit illustrated in the above embodiments can be specifically implemented by a computer chip or entity, or by a product with certain functions. A typical implementation device is a computer device, and specifically, the computer device can be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an electronic mail device, a game console, a tablet computer, a wearable device or a combination of any of these devices.

[0231] The embodiment of the present application provides a computer device, which comprises a memory and a processor, the memory is used to store information including program instructions, and the processor is used to control the execution of the program instructions, when the program instructions are loaded and executed by the processor, each step of the above-mentioned mechanism-data reconstructed granular flow system temperature time sequence prediction method embodiment is implemented, and specific description can be referred to the above-mentioned mechanism-data reconstructed granular flow system temperature time sequence prediction method embodiment.

[0232] Reference is made below to Figure 7 which shows a structural schematic diagram of a computer device 600 suitable for being used to implement the embodiments of the present application.

[0233] like Figure 7 As shown, the computer device 600 includes a central processing unit (CPU) 601, which can perform various appropriate tasks and processes based on programs stored in read-only memory (ROM) 602 or programs loaded from storage section 608 into random access memory (RAM) 603. The RAM 603 also stores various programs and data required for the operation of the computer device 600. The CPU 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0234] The following components are connected to I / O interface 605: an input section 606 including a keyboard, mouse, etc.; an output section 607 including a cathode ray tube (CRT), liquid crystal feedback (LCD), etc., and speakers, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card such as a LAN card, modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to I / O interface 605 as needed. A removable medium 611, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 610 as needed so that computer programs read from it can be installed in storage section 608 as needed.

[0235] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program tangibly embodied on a machine-readable medium, the computer program including program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 609, and / or installed from removable medium 611.

[0236] Computer-readable media includes permanent and non-permanent, movable and non-movable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic tape disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible to a computing device. According to the definition herein, computer-readable media does not include transitory media such as modulated data signals and carriers.

[0237] For the convenience of description, the above device is described as various units described respectively in function. Of course, in the implementation of the present application, the functions of each unit can be implemented in the same or more software and / or hardware.

[0238] The present application is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device that implements the flowcharts and / or block diagrams. Figure 1 The flow or multiple flows and / or blocks Figure 1 The device that implements the functions specified in one or more blocks.

[0239] These computer program instructions can also be stored in a computer-readable memory that can guide the computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer-readable memory produce a product including instruction devices that implement the flowcharts and / or block diagrams. Figure 1 The flow or multiple flows and / or blocks Figure 1 The device that implements the functions specified in one or more blocks.

[0240] These computer program instructions can also be loaded into a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 Figure 1

[0241] It should also be noted that the term "comprising" or "including" or any other variation thereof is intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element preceded by "comprises a" does not, without more constraints, foreclose the existence of additional identical elements in the process, method, article, or apparatus that comprises the recited element.

[0242] In the technical solutions of the present application, the acquisition, storage, use, processing, etc. of data comply with relevant provisions of national laws and regulations.

[0243] It should be noted that in the embodiments of the present application, some software, components, models, etc. of the prior art can be mentioned, which should be considered as exemplary, and the purpose is only to illustrate the feasibility of the implementation of the technical solutions of the present application, but it does not mean that the applicant has or will necessarily use the scheme.

[0244] Those skilled in the art will appreciate that embodiments of the present application can be provided as a method, system or computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) containing computer-usable program code.

[0245] The present application can be described in the general context of computer-executable instructions, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. The present application can also be practiced in distributed computing environments where tasks are performed by remote processing devices that are connected through a communication network. In a distributed computing environment, program modules can be located in both local and remote computer storage media including storage devices.

[0246] ​​The various embodiments in the specification are described in progressive manner, and the same or similar parts among the various embodiments can be mutually referred to, and each embodiment focuses on the difference from other embodiments. In particular, for the system embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the part of the method embodiments.

[0247] The above only describes the embodiments of the present application and is not intended to limit the present application. The present application can have various modifications and changes for those skilled in the art. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the scope of claims of the present application.

Claims

1. A method for temperature time series prediction of a mechanism-data reconstructed granular flow system, characterized in that, The method comprises: numerical simulation of a reconstructed granular flow system by a discrete element method to generate experimental particle temperatures at different times, the reconstructed granular flow system being a uniformly cooled smooth granular flow system; construction of a theoretical comprehensive loss function based on the experimental particle temperatures at different times and a physical control equation by a constructed neural network based on physical information; determination of a target comprehensive loss function according to the theoretical comprehensive loss function under a system missing condition, and iterative updating of network parameters of the neural network based on physical information to minimize the target comprehensive loss function, to construct a particle temperature time series feature prediction model, the system missing condition including missing energy dissipation coefficient, missing initial time particle temperature, or missing energy dissipation coefficient and initial time particle temperature; particle temperature prediction based on the particle temperature time series feature prediction model according to inputted to-be-predicted time information to obtain a target particle temperature time series.

2. The method of claim 1, wherein, The numerical simulation of the reconstructed granular flow system by the discrete element method to generate the experimental particle temperatures at different times comprises: statistical acquisition of experimental particle velocities at different times according to a preset particle velocity distribution function and based on Newton's law of motion and a linear inter-particle force model to simulate particle motion information in the reconstructed granular flow system; determination of experimental particle fluctuation velocities according to the experimental particle velocities at different times and an average particle velocity; statistical acquisition of an average value of fluctuation energies of all particles in the system according to the experimental particle fluctuation velocities to generate experimental particle temperatures at different times.

3. The method of claim 1, wherein, The construction of the theoretical comprehensive loss function based on the experimental particle temperatures at different times and the physical control equation by the constructed neural network based on physical information comprises: dimensionless processing of the experimental particle temperatures at different times to obtain dimensionless experimental particle temperatures; construction of a data sample loss function and an initial condition loss function by the neural network based on physical information according to the dimensionless experimental particle temperatures; normalization of the different times to obtain normalized time information; construction of a mechanism-data dual-driven control equation loss function by the neural network based on physical information according to a preset local weight coefficient, the normalized time information, the dimensionless experimental particle temperatures, and the physical control equation, the local weight coefficient being determined according to an inverse of a dimensionless particle temperature predicted at a corresponding time; weighting of the data sample loss function, the initial condition loss function, and the mechanism-data dual-driven control equation loss function according to a preset global weight coefficient to generate the theoretical comprehensive loss function.

4. The method for predicting the temperature time series of granular flow systems based on mechanism-data reconstruction according to claim 3, characterized in that, The construction of the data sample loss function and the initial condition loss function by the neural network based on physical information according to the dimensionless experimental particle temperatures comprises: random sampling of the dimensionless experimental particle temperatures to obtain dimensionless sparse particle temperatures; construction of the data sample loss function by the neural network based on physical information according to the dimensionless sparse particle temperatures and corresponding different times. The initial condition loss function is constructed according to the dimensionless experimental particle temperature at the initial moment by the neural network based on physical information.

5. The method of claim 1, wherein, The system missing condition is missing the energy dissipation coefficient; The target comprehensive loss function is determined according to the theoretical comprehensive loss function under the system missing condition, and the network parameters of the neural network based on physical information are iteratively updated to minimize the target comprehensive loss function, so as to construct a particle temperature time sequence feature prediction model, including: The theoretical comprehensive loss function is determined as the target comprehensive loss function; The neural network based on physical information is trained and optimized by a preset optimization algorithm to minimize the target comprehensive loss function, iteratively update the network parameters, generate a particle temperature time sequence feature prediction model and an optimized energy dissipation coefficient, and the network parameters include network weights and network biases.

6. The method of claim 3, wherein the temperature time series prediction of the mechanism-data reconstructed granular flow system is based on a temperature time series prediction of a mechanism-data reconstructed granular flow system, wherein the mechanism-data reconstructed granular flow system is based on a mechanism reconstructed granular flow system and a data reconstructed granular flow system. The system missing condition is missing the particle temperature at the initial moment; The target comprehensive loss function is determined according to the theoretical comprehensive loss function under the system missing condition, and the network parameters of the neural network based on physical information are iteratively updated to minimize the target comprehensive loss function, so as to construct a particle temperature time sequence feature prediction model, including: The initial condition loss function is removed from the theoretical comprehensive loss function to generate the target comprehensive loss function; The neural network based on physical information is trained and optimized by a preset optimization algorithm to minimize the target comprehensive loss function, iteratively update the network parameters, generate a particle temperature time sequence feature prediction model and an optimized particle temperature at the initial moment, and the network parameters include network weights and network biases.

7. The method of claim 3, wherein the temperature time series prediction of the mechanism-data reconstructed granular flow system is based on a machine learning model. The system missing condition is missing the energy dissipation coefficient and the particle temperature at the initial moment; The target comprehensive loss function is determined according to the theoretical comprehensive loss function under the system missing condition, and the network parameters of the neural network based on physical information are iteratively updated to minimize the target comprehensive loss function, so as to construct a particle temperature time sequence feature prediction model, including: The initial condition loss function is removed from the theoretical comprehensive loss function to generate the target comprehensive loss function; The neural network based on physical information is trained and optimized by a preset optimization algorithm to minimize the target comprehensive loss function, iteratively update the network parameters, generate a particle temperature time sequence feature prediction model, an optimized energy dissipation coefficient and an optimized particle temperature at the initial moment, and the network parameters include network weights and network biases.

8. The method of claim 1, wherein, The particle temperature is predicted according to the inputted to-be-predicted time information based on the particle temperature time sequence feature prediction model, to obtain a target particle temperature time sequence, including: The to-be-predicted time information is normalized to generate normalized to-be-predicted time information; The normalized to-be-predicted time information is calculated by the particle temperature time sequence feature prediction model to generate a dimensionless particle temperature time sequence, which is a corresponding relationship between a dimensionless particle temperature and the normalized to-be-predicted time information. The dimensionless particle temperature is time de-normalized and particle temperature dimensioned with the normalized to-be-predicted time information to generate the target particle temperature time sequence.

9. The method of claim 1, wherein, The method further comprises: The predicted target particle temperature time sequence and the acquired real particle temperature time sequence are compared to generate a particle temperature time sequence feature error result.

10. The method of claim 5 or 7, wherein the temperature time series prediction of the mechanism-data reconstructed granular flow system is characterized by, The method further comprises: The optimized energy dissipation coefficient is compared with the pre-calculated theoretical energy dissipation coefficient to generate an energy dissipation coefficient error result, the optimized energy dissipation coefficient being obtained by iterative calculation based on the physical control equation in the particle temperature time sequence feature prediction model.

11. The method of claim 6 or 7, wherein the temperature time series prediction of the mechanism-data reconstructed granular flow system is characterized by, The method further comprises: The initial time optimized particle temperature is compared with the acquired initial time real particle temperature to generate an initial time particle temperature error result, the initial time optimized particle temperature being obtained by iterative calculation based on the particle temperature time sequence feature prediction model.

12. A mechanism-data reconstructed granular flow system temperature time series prediction device, characterized by, The device comprises: A numerical simulation unit configured to perform numerical simulation on a reconstructed granular flow system by a discrete element method to generate experimental particle temperatures at different times, the reconstructed granular flow system being a uniformly cooled smooth granular flow system; A theoretical comprehensive loss function construction unit configured to construct a theoretical comprehensive loss function based on the experimental particle temperatures at different times and a physical control equation by a constructed neural network based on physical information; A particle temperature time sequence feature prediction model training unit configured to determine a target comprehensive loss function according to the theoretical comprehensive loss function under a system missing condition, and to iteratively update network parameters of the neural network based on physical information to construct a particle temperature time sequence feature prediction model, the system missing condition including missing energy dissipation coefficient, missing initial time particle temperature, or missing energy dissipation coefficient and initial time particle temperature; A particle temperature time sequence feature prediction unit configured to perform particle temperature prediction based on the particle temperature time sequence feature prediction model according to input to-be-predicted time information to obtain a target particle temperature time sequence.

13. A computer readable medium having stored thereon a computer program, characterized in that The program is executed by the processor to implement the mechanism-data reconstructed granular flow system temperature time sequence prediction method of any one of claims 1 to 11.

14. A computer device comprising a memory for storing information including program instructions, and a processor for controlling execution of the program instructions, characterized in that, The program instructions are loaded and executed by the processor to implement the mechanism-data reconstructed granular flow system temperature time sequence prediction method of any one of claims 1 to 11.

15. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instructions are executed by the processor to implement the mechanism-data reconstructed granular flow system temperature time sequence prediction method of any one of claims 1 to 11.

Images