A model data processing method, simulation device, storage medium, and testing system

Abstract

This invention belongs to the field of electronic control system technology, and particularly relates to a model data processing method, simulation device, storage medium and testing system; the state vector dimension and the quantity to be estimated are determined by prior knowledge of the physical system (900); the first state estimation step / unit obtains the first initial state value of the quantity to be estimated based on historical data; the second state enhancement step / unit performs a second preprocessing on the original data and uses it for solving explicit ordinary differential equations; the third system decomposition step / unit decomposes the physical system (900) into a set of subsystems and implements the aforementioned processing; the fourth state focusing step / unit limits the range of the loss function value (010) according to the fifth rate of change boundary value; the method and product can be used as a general modeling tool to improve the system simulation performance, and achieves more ideal generalization performance, accuracy and inference speed than the related neural network ordinary differential equation (NODEs) scheme.

Description

Technical Field

[0001] This invention belongs to the field of electronic control system technology, and in particular relates to a model data processing method, simulation device, storage medium and testing system. Background Technology

[0002] Model-Based Development (MBD) has been widely applied in product design, testing, and other stages. By using a sufficiently detailed system model in the early stages of design, it can replace the actual hardware to complete necessary improvements, simulations, and verifications. The modeling stage is the foundation for system design and verification using the above methods. However, traditional modeling processes face significant challenges when dealing with certain specific systems. In particular, it is often difficult to obtain satisfactory results in the modeling of nonlinear and complex systems.

[0003] Furthermore, related technologies also employ Neural Ordinary Differential Equations (NODEs) to combine deep learning with ordinary differential equations to describe complex system dynamics, exhibiting excellent generalization ability and accuracy. Particularly for modeling nonlinear and complex systems, NODEs can acquire the inherent laws of the system through an automatic learning process, eliminating the need for massive amounts of knowledge in specific domains and manual adjustments, greatly simplifying the modeling process and improving its efficiency and accuracy. However, the inference speed of NODEs is relatively slow, which has become a technical bottleneck for their deep application, necessitating improvements to the information and data processing processes. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention improves and optimizes the NODEs technology: First, by introducing an initial state estimator, i.e., through a first state estimation step / unit, the system's state at the start of simulation / simulation operations is estimated more accurately, providing a more accurate starting point for the model; this improvement is particularly important for systems where the initial state is unknown or difficult to measure directly, and can significantly improve the accuracy and stability of the model.

[0005] On the other hand, in order to capture the dynamic behavior of the system more comprehensively, the present invention can introduce augmented states (AS) into the NODEs model through a second state augmentation step / unit; these augmented states can represent unobserved or difficult-to-observe states in the system, thereby enabling the model to describe the dynamic process of the system more comprehensively.

[0006] Accordingly, the innovation of this invention enables the model to learn not only based on observable input and output data, but also to take into account factors that have an important impact on system behavior but are often ignored in traditional models, thereby further improving the generalization and accuracy of the system.

[0007] Furthermore, for scenarios where the system state changes discontinuously, this invention can decompose the system into multiple subsystems in an event-driven manner, and enable the states between the subsystems to inherit from each other, thus ensuring the accuracy of the overall system during mode switching.

[0008] In order to further improve the training efficiency and inference speed of the model, the present invention can also add penalty terms for state values ​​and state derivatives to the loss function, namely the first constraint on the state values ​​and the second constraint on the state derivatives. These penalty terms help guide the network to pay more attention to the key features of the system dynamics during the training process, thereby accelerating the convergence of the network and improving the performance of the model.

[0009] Based on the aforementioned key improvements, this invention demonstrates higher performance and faster inference speed than the original NODEs-like systems through its embodiments; it also makes this invention have significant application value in fields such as system modeling.

[0010] Specifically, this invention discloses a model data processing method, the core of which is the use of a first state estimation step and / or a second state enhancement step. First, based on the prior knowledge of the physical system, the physical system needs to be represented as an explicit ordinary differential equation. This explicit ordinary differential equation is independent of time and the dimension of the state vector is determined by a preset dimension parameter N.

[0011] Among them, physical quantities that cannot be directly measured can be identified as quantities to be estimated, and the first state estimation step obtains the first initial state value of the quantities to be estimated at the beginning of the simulation process based on historical data; while the second state enhancement step uses a preset data processing procedure to perform a second preprocessing on the original data, and uses the output of the second preprocessing for solving the explicit ordinary differential equation.

[0012] Furthermore, by collecting preset parameters of the physical system, a first raw data segment of length L can be obtained, such that the number of the first raw data segments is M; and its length L is measured by the time length; where parameters M and N are both positive integers; then, the F-th state set of future time F can be predicted according to the preset neuron differential equation, and the F-th state set can be mapped by the second function (060), and the second function (060) can be determined according to similar prior knowledge.

[0013] Furthermore, the model data processing method may also include a third system decomposition step; if the physical system has a discontinuous change process, the third system decomposition step can be used to decompose the physical system into a preset number of subsystem sets; similarly, each system in the subsystem set can be simulated and / or processed using the model data processing method described above.

[0014] Specifically, the second preprocessing can employ a data segmentation and enhancement process and / or a standardization process; the segmentation and enhancement process divides each first original data segment of length L into a second data segment of length S, and expands the length of each first original data segment to the sum of the difference between L and S (LS) divided by the time interval I and 1; the standardization process can convert the first original data segment into intermediate data under a preset probability distribution.

[0015] The aforementioned preset probability distribution can be a standard normal distribution Norm(0,1); the intermediate data can be obtained by dividing the difference between the original data Xraw and the arithmetic mean by the standard deviation.

[0016] In addition, its second preprocessing can also divide the intermediate data and / or the second data segment into a third training set and a fourth test set; its second state augmentation step can also calculate the upper and lower bounds of the preset first variable before model training, and calculate the fifth rate of change boundary values ​​of the first difference and the second difference of the first variable over time.

[0017] Furthermore, the model data processing method may also include a fourth state focusing step; the fourth state focusing step may limit the value of the loss function value Loss according to the aforementioned fifth rate of change boundary value; wherein, the loss function value may be obtained by linearly combining the average absolute error of the first measurement dataset and the model prediction value with a preset regularization term.

[0018] Specifically, its first measurement dataset includes M first raw data segments, and its regularization term includes a first constraint on the state value and a second constraint on the state derivative, which can be obtained by solving the third training set mentioned above.

[0019] The solution of the above explicit ordinary differential equations can be achieved using a preset integrator, and the input control can be achieved using a preset interpolation process.

[0020] Specifically, its explicit ordinary differential equation can be driven by a neural network structure, which can be a graph neural network (GNN) and / or a convolutional neural network (CNN); at the same time, the neural network structure can be used to construct the neural network ordinary differential equation NODE and / or to generate the initial code.

[0021] Similarly, embodiments of the present invention also disclose a simulation device, the core of which includes a first state estimation unit and / or a second state enhancement unit; in practice, the physical system can be represented as an explicit ordinary differential equation based on prior knowledge of the physical system, the explicit ordinary differential equation is independent of time and the dimension of the state vector can be determined by a preset dimension parameter N.

[0022] Among them, physical quantities that cannot be directly measured can be identified as quantities to be estimated, and the first state estimation unit obtains the first initial state value of the quantities to be estimated at the beginning of the simulation process based on historical data; while the second state enhancement unit performs a second preprocessing on the original data using a preset data processing procedure, and uses the output of the second preprocessing for solving the above explicit ordinary differential equation.

[0023] Specifically, preset parameters of the physical system can be collected to obtain a first raw data segment of length L, the number of which is M; its length L can be measured by the time length, and M and N are both positive integers; the set of the Fth state at future time F can also be predicted according to the preset neuron differential equation, the set of the Fth state can be obtained by mapping the second function, and the second function can be determined according to similar prior knowledge; the second function can be represented as g, or called the state-observation mapping function; and g is independent of prior knowledge and can be a general nonlinear function, such as a neural network.

[0024] Furthermore, the simulation device may also be equipped with a third system decomposition unit; if the physical system has a discontinuous change process, the third system decomposition unit can be used to decompose the physical system into a preset number of subsystem sets; wherein, each system in the subsystem set can be simulated and / or data processed using the simulation device described above.

[0025] The second preprocessing includes a data segmentation and enhancement process and / or a standardization process. The segmentation and enhancement process can divide each first original data segment of length L into a second data segment of length S, and make the length of each first original data segment expanded to the sum of the difference between L and S (LS) divided by the time interval I and 1. The standardization process can convert the first original data segment into intermediate data under a preset probability distribution.

[0026] Specifically, its preset probability distribution can adopt the standard normal distribution Norm(0,1); and its intermediate data can be obtained by dividing the difference between the original data Xraw and the arithmetic mean by the standard deviation; its second preprocessing can also divide the intermediate data and / or the second data segment into a third training set and a fourth test set.

[0027] The second state enhancement unit can also calculate the upper and lower bounds of the preset first variable before model training, and calculate the fifth rate of change boundary values ​​of the first difference and the second difference of the first variable over time.

[0028] Furthermore, the simulation device may also be equipped with a fourth state focusing unit; the fourth state focusing unit can limit the value of the loss function value Loss according to the fifth rate of change boundary value; wherein, the loss function value can be obtained by linear combination of the average absolute error of the first measurement dataset and the model prediction value with a preset regularization term.

[0029] Specifically, its first measurement dataset includes M first raw data segments, and its regularization term includes a first constraint on the state value and a second constraint on the state derivative, which can be obtained by solving the third training set mentioned above.

[0030] The solution of the explicit ordinary differential equation can be achieved using a preset integrator and can be input controlled using a preset interpolation process; while the explicit ordinary differential equation can be driven by a neural network structure, which can be a graph neural network (GNN) and / or a convolutional neural network (CNN); its neural network structure can be used to construct the neural network ordinary differential equation NODE and / or to generate the initial code.

[0031] Similarly, embodiments of the present invention also disclose a computer storage medium and a testing system to solve the same technical problem; wherein, the computer storage medium includes a storage medium body for storing a computer program; when the computer program is executed by a microprocessor, it can implement the model data processing method as described above; and the testing system includes any of the simulation devices and / or computer storage media as described above.

[0032] In summary, this invention determines the state vector dimension and the quantity to be estimated based on prior knowledge of the physical system. A first state estimation step / unit obtains the first initial state value of the quantity to be estimated from historical data. A second state enhancement step / unit performs a second preprocessing on the original data and uses it for solving explicit ordinary differential equations. A third system decomposition step / unit decomposes the physical system into a set of subsystems and performs the aforementioned processing. A fourth state focusing step / unit limits the range of the loss function value based on a fifth rate of change boundary value. The method and product can serve as a general modeling tool to improve system simulation performance, achieving more ideal generalization performance, accuracy, and inference speed compared to related neural network ordinary differential equation (NODE) ​​schemes.

[0033] It should be noted that the terms "first," "second," and similar terms used in this article are merely for describing the constituent elements of the technical solution and do not constitute a limitation on the technical solution, nor should they be interpreted as an indication or implication of the importance of the corresponding elements; elements with terms such as "first," "second," or similar terms indicate that at least one of the elements is included in the corresponding technical solution. Attached Figure Description

[0034] To more clearly illustrate the technical solution of the present invention and facilitate a further understanding of its technical effects, features, and objectives, the present invention will be described in detail below with reference to the accompanying drawings. The drawings constitute an essential part of the specification and are used together with Embodiment 1 of the present invention to illustrate the technical solution of the present invention, but do not constitute a limitation on the present invention.

[0035] The same reference numerals in the attached figures represent the same components, as detailed in the figures below.

[0036] Figure 1 This is a schematic diagram of the first embodiment of the present invention, namely the catalyst system.

[0037] Figure 2 This is a schematic diagram of the modeling of the second embodiment of the present invention, namely the vehicle thermal management system.

[0038] Figure 3 This is a simplified schematic diagram of the first embodiment of the present invention.

[0039] Figure 4 The power system model 071 is the first embodiment of the present invention.

[0040] Figure 5 This is a schematic diagram of the system topology of the second embodiment of the present invention.

[0041] Figure 6 The power system model 072 is the second embodiment of the present invention.

[0042] Figure 7 This invention provides a time-independent neural network-driven explicit ordinary differential equation system.

[0043] Figure 8 The data provided are a predetermined number of velocity transfer matrices and their first-order Markov chain model data for this invention.

[0044] Figure 9 This is a schematic diagram of the general form of the neuron differential equation model according to an embodiment of the present invention.

[0045] Figure 10 This is a schematic diagram of the input temperature jump of the heater core in an embodiment of the present invention.

[0046] Figure 11This is a model structure with event-driven (pattern) in an embodiment of the present invention.

[0047] Figure 12 This is the loss function structure in an embodiment of the present invention.

[0048] Figure 13 This is a comparison chart of the modeling prediction and actual measurement results of the catalyst in the first embodiment of the present invention.

[0049] Figure 14 This is a comparison chart of the modeling prediction and actual measurement results of the vehicle thermal system according to the second embodiment of the present invention.

[0050] Figure 15 This is a circuit diagram of the vehicle thermal system test box according to the second embodiment of the present invention.

[0051] Figure 16 This is a schematic diagram of the process of an embodiment of the method of the present invention.

[0052] Figure 17 This is a schematic diagram of the composition structure of an embodiment of the simulation device of the present invention.

[0053] Figure 18 This is a schematic diagram of the layout structure of an embodiment of the product of the present invention. Figure 1 .

[0054] Figure 19 This is a schematic diagram of the layout structure of an embodiment of the product of the present invention. Figure 2 .

[0055] Figure 20 This is a schematic diagram of the layout structure of an embodiment of the product of the present invention. Figure 3 .

[0056] The meanings of the labels / symbols in the attached figures are as follows.

[0057] 001 - Example of a three-way catalytic converter system;

[0058] 002 - Example of a vehicle thermal management model;

[0059] 010 - Loss function value;

[0060] 011 - Mean absolute error (between measured data and model predictions) (used to represent the predictive power of the model);

[0061] 012 - Loss Weight;

[0062] 013 - First constraint, i.e., state value constraint term;

[0063] 015 - Second constraint, namely the state derivative constraint term;

[0064] 017 - Mean Absolute Error, or MSE (Mean Squared Error).

[0065] 021 - First Local Model;

[0066] 022 - Second Local Model;

[0067] 02N - The Nth local model;

[0068] 030 - Model input quantity;

[0069] 039 - Model output;

[0070] 040 - (Thermal Management System) Topology;

[0071] 050 - First function;

[0072] 051 - First parameter;

[0073] 052 - First function, first parameter;

[0074] 053 - First function, second parameter;

[0075] 054 - The third parameter of the first function, for example, the model input is 030;

[0076] 055 - The fourth parameter of the first function, such as the time parameter;

[0077] 060 - Second function;

[0078] 061 - Second parameter;

[0079] 062 - First parameter of the second function;

[0080] 063 - Second function, second parameter;

[0081] 064 - The third parameter of the second function is equivalent to the third parameter of the first function;

[0082] 071-First Model;

[0083] 072-Second Model;

[0084] 073 - Third Model, with Time Parameters;

[0085] 081 - Velocity transfer matrix;

[0086] 083 - First-order Markov chain model data;

[0087] 091 - Initial State Estimator;

[0088] 095 - Indication of Input Temperature Jump in Heater Unit Figure 1 ;

[0089] 097 - Indication of Input Temperature Jump in Heater Unit Figure 2 ;

[0090] 100 - First state estimation step;

[0091] 200 - Second state enhancement step;

[0092] 300 - Third System Disassembly Steps;

[0093] 400 - Fourth State Focusing Step;

[0094] 600 - Simulation device;

[0095] 610 - First State Estimation Unit;

[0096] 620 - Second State Enhancement Unit;

[0097] 630 - Third System Disassembly Unit;

[0098] 640 - Fourth State Focusing Unit;

[0099] 810 - First group of prediction comparison chart (optimal);

[0100] 820 - Second group of prediction comparison chart (worst);

[0101] 861 - Sixth measurement value;

[0102] 862 - Sixth predicted value;

[0103] 871 - Seventh measurement value;

[0104] 872 - Seventh predicted value;

[0105] 881 - Eighth measurement value;

[0106] 882 - Eighth predicted value;

[0107] 890 - Test box plot;

[0108] 900 - Vehicle, physical model, or system under test;

[0109] 901 - Test System;

[0110] 903 - Computer storage media. Detailed Implementation

[0111] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Of course, the specific embodiments described below are merely illustrative of the technical solutions of the present invention, and not intended to limit the invention. Furthermore, the parts described in the embodiments or drawings are merely illustrative examples of relevant parts of the present invention, and not the entirety of the invention.

[0112] like Figure 16 The model data processing method shown includes a first state estimation step 100 and / or a second state enhancement step 200; firstly, it can be based on, as follows: Figure 18 The prior knowledge of the physical system 900 shown is used to represent the physical system 900 as an explicit ordinary differential equation. This explicit ordinary differential equation is independent of time and the dimension of the state vector is determined by a preset dimension parameter N.

[0113] Among them, physical quantities that cannot be directly measured can be identified as quantities to be estimated, and the first state estimation step 100 obtains the first initial state value of the quantity to be estimated at the beginning of the simulation process based on historical data; while the second state enhancement step 200 uses a preset data processing procedure to perform a second preprocessing on the original data, and uses the output of the second preprocessing for the solution of the above explicit ordinary differential equation.

[0114] Specifically, the parameters preset by the physical system 900 can be collected to obtain a first raw data segment of length L, and the number of the first raw data segments is M; its length L can be measured by the time length, and M and N are both positive integers; in addition, the set of the Fth state at future time F can be predicted according to the preset neuron differential equation, and the set of the Fth state can be obtained by mapping the second function 060, which can be determined according to prior knowledge.

[0115] Furthermore, such as Figure 16 As shown, the model data processing method also includes a third system decomposition step 300; if the physical system 900 has a discontinuous change process, the third system decomposition step 300 can be used to decompose the physical system 900 into a preset number of subsystem sets.

[0116] Each of the above subsystems can be simulated and / or processed using the model data processing methods described above.

[0117] Furthermore, the second preprocessing can employ a data segmentation and enhancement process and / or a standardization process; the segmentation and enhancement process can divide each first original data segment of length L into a second data segment of length S, and make the length of each first original data segment expanded to the sum of the difference between L and S (LS) divided by the time interval I and 1; the standardization process can convert the first original data segment into intermediate data under a preset probability distribution.

[0118] The preset probability distribution includes the standard normal distribution Norm(0,1); the intermediate data can be obtained by dividing the difference between the original data Xraw and the arithmetic mean by the standard deviation.

[0119] Specifically, the second preprocessing can further divide the intermediate data and / or the second data segment into a third training set and a fourth test set; the second state enhancement step 200 can also calculate the upper and lower bounds of the preset first variable before model training, and can calculate the fifth rate of change boundary values ​​of the first difference and the second difference of the first variable over time.

[0120] Furthermore, to further improve the inference speed, the model data processing method may also include a fourth state focusing step 400; this fourth state focusing step 400 is then limited according to its fifth rate of change boundary value, such as... Figure 12 The loss function values ​​shown are 0 and 10.

[0121] The loss function value 010 can be obtained by linearly combining the mean absolute error 011 of the first measurement dataset and the model prediction values ​​with a preset regularization term; the first measurement dataset includes M first original data segments, and the regularization term includes a first constraint 013 on the state value and a second constraint 015 on the state derivative, which can be obtained by solving the third training set mentioned above.

[0122] Specifically, the solution of its explicit ordinary differential equations can be achieved using a preset integrator, and the input control can be achieved using a preset interpolation process.

[0123] Furthermore, its explicit ordinary differential equations can be driven by a neural network structure, which can be a graph neural network (GNN) and / or a convolutional neural network (CNN); this neural network structure can also be used to construct neural network ordinary differential equations (NODEs) and / or to generate initial codes.

[0124] Similarly, such as Figure 17 The simulation device 600 shown may be equipped with a first state estimation unit 610 and / or a second state enhancement unit 620. Similarly, based on prior knowledge of the physical system 900, the physical system 900 may be represented as an explicit ordinary differential equation, which is independent of time and whose state vector dimension can be determined by a preset dimension parameter N. Physical quantities that cannot be directly measured may also be identified as quantities to be estimated, and the first state estimation unit 610 may obtain the first initial state value of the quantities to be estimated at the start of the simulation process based on historical data. The second state enhancement unit 620 may perform a second preprocessing on the original data using a preset data processing procedure, and may use the output of the second preprocessing for solving the explicit ordinary differential equation.

[0125] Among them, the parameters preset by the physical system 900 can be collected to obtain a first raw data segment of length L, and the number of the first raw data segments is M; its length L can be measured by the time length, and M and N are both positive integers; it can also predict the F-th state set of future time F according to the preset neuron differential equation, and the F-th state set can be obtained by mapping the second function 060, which can be determined according to prior knowledge.

[0126] Similarly, the simulation device 600 may also be provided with a third system decomposition unit 630; if the physical system 900 has a discontinuous change process, the third system decomposition unit 630 can also be used to decompose the physical system 900 into a preset number of subsystem sets; wherein, each system in the subsystem set can also be simulated and / or processed by the simulation device 600 as described above.

[0127] Specifically, the second preprocessing includes a data segmentation and enhancement process and / or a standardization process; the segmentation and enhancement process divides each first original data segment of length L into a second data segment of length S, such that the length of each first original data segment is expanded to the sum of the difference between L and S (LS) divided by the time interval I and 1; the standardization process can convert the first original data segment into intermediate data under a preset probability distribution.

[0128] The preset probability distribution can adopt the standard normal distribution Norm(0,1); the intermediate data can be obtained by dividing the difference between the original data Xraw and the arithmetic mean by the standard deviation; the second preprocessing can also divide the intermediate data and / or the second data segment into a third training set and a fourth test set; and the second state enhancement unit 620 can also calculate the upper and lower bounds of the preset first variable before model training, and can calculate the fifth rate of change boundary values ​​of the first difference and the second difference of the first variable over time.

[0129] Furthermore, the simulation device 600 may also be provided with a fourth state focusing unit 640; the fourth state focusing unit 640 may limit the value of the loss function 010 according to the fifth rate of change boundary value; wherein, the loss function value 010 may be obtained by linear combination of the average absolute error 011 of the first measurement dataset and the model prediction value with a preset regularization term.

[0130] The first measurement dataset includes M first original data segments, whose regularization terms include a first constraint 013 on the state value and a second constraint 015 on the state derivative. The first constraint 013 and the second constraint 015 are obtained by solving the third training set.

[0131] Specifically, the solution of its explicit ordinary differential equation can be achieved using a preset integrator, and the input control can be performed using a preset interpolation process; the explicit ordinary differential equation can be driven by a neural network structure, which includes graph neural networks (GNN) and / or convolutional neural networks (CNN); its neural network structure can be used to construct neural network ordinary differential equations (NODE) ​​and / or to generate initial codes.

[0132] In addition, such as Figure 19 , Figure 20 The computer storage medium 903 shown includes a storage medium body for storing computer programs; when the computer program is executed by a microprocessor, it can implement any of the model data processing methods described above.

[0133] Similarly, such as Figures 18 to 20 The test system 901 shown includes any of the simulation devices 600 and / or computer storage media 903 as described above, which can be used to solve the same technical problem.

[0134] In practical applications, the basic dataset for the model must first be created through steps such as system analysis and variable definition; for different model instances, the selection of variables, i.e., their definition, is as follows: Figure 4 , Figure 6 , Figure 7 The process of handling various relevant parameters; specifically, for example... Figure 1 The catalyst system modeling shown and as Figure 2 The modeling of the vehicle thermal management system shown is briefly described below; the system can be abstracted based on the prior knowledge of the above system.

[0135] The mechanism model of the catalyst can be determined by physicochemical equations. Typically, chemical reactions can be modeled as a set of ordinary differential equations, and then a set of differential-algebraic equations (DAEs) can be formed based on the law of conservation of mass.

[0136] like Figure 3 As shown, its submodule blk (block) represents a small local model of the catalyst cross section; the input of each model (i.e., model input 030) is the output of the previous module (i.e., model output 039); specifically, it is used to represent the gas composition.

[0137] Each blk contains the system state (i.e., the first parameter 052 of the first function), including the molar ratio of the catalyst, temperature, etc. As the state of the system, the molar ratio of the catalyst cannot be directly measured during the prediction process and is an estimated value (quantity).

[0138] Furthermore, this catalyst model can be viewed as a dynamic system formed by the dynamic changes in catalyst composition, and specifically represented as follows: Figure 4 The first model 071 represents the form; specifically, it includes the first function 050 and the second function 060.

[0139] For the aforementioned vehicle thermal management system, the following can be adopted: Figure 5 The topology 040 shown; the dynamic system model of this structure can be similarly represented as follows: Figure 6 The second model 072 represents the form; and specifically includes similar first functions 050 and second functions 060.

[0140] Specifically, the aforementioned dynamic system can be determined by the first parameter 051 and the second parameter 061. After determining the form of its functional relationship f and g, the first parameter 051 and the second parameter 061 can be determined by solving the inverse problem, i.e., the parameter identification process. The first parameter 051 and the second parameter 061 that minimize the error between the predicted value and the measured signal (e.g., the mean absolute error MSE) are obtained.

[0141] The above embodiments have the advantage of a small number of parameters, but the form of the functional relationship f and g needs to be determined according to its physical principles. If an overly simplified functional relationship is used, the prediction performance may deteriorate. However, an overly complex functional relationship will increase the difficulty of parameter identification. Secondly, different forms of mathematical equations will face different processing methods when solving inverse problems, which is not conducive to the calculation of functions. This will lead to difficulties in predicting the development cycle.

[0142] Furthermore, the system can be modeled as follows: Figure 7 The above is a time-independent neural network-driven explicit ordinary differential equation system. The first function's third parameter, i.e., the model input 30, includes the concentration values ​​(ppm) of CO, CO2, NOx, HC, and O2, the gas mass flow rate, and the average temperature of the catalyst. The first parameter of the first function, 052, represents the system state as an N-dimensional vector. The second parameter of the second function is the output, which includes the concentration values ​​(ppm) of the four waste gases: CO, CO2, NOx, and HC.

[0143] Similarly, for such Figure 8The system shown has the following parameters: the third parameter of the first function, i.e., the model input 30, includes engine heat input, low-temperature radiator inlet gas velocity, engine side throttle valve opening, heater core inlet gas velocity, heater core inlet gas temperature, positive temperature coefficient (PTC) heater power, actual vehicle speed, battery heating power, engine-heater core three-way valve opening, and heater core-battery three-way valve opening; the first parameter of the first function, 052, i.e., the system state, is also an N-dimensional vector; the second parameter of the second function, i.e., the output, includes engine outlet temperature, low-temperature radiator outlet temperature, passenger compartment temperature, PTC inlet temperature, cooler outlet temperature, heater core outlet temperature, battery outlet temperature, and battery inlet temperature.

[0144] It should be noted that in the two embodiments above, the dimension N of the system state vector is a hyperparameter that can be selected by human experience.

[0145] Furthermore, data is collected and processed; for example... Figure 1 In the illustrated embodiment, catalytic converter data was collected from a real vehicle, yielding 17 time-series data points with durations ranging from 340 seconds to 677 seconds; for example... Figure 2 In the embodiment shown, the thermal management system obtained a total of 68 time-series data points from the simulation software Amesim.

[0146] Specifically, in order to improve the representativeness of the data, in such cases... Figure 2 In the illustrated embodiment, data acquisition was performed step-by-step using Amesim.

[0147] On the one hand, such as Figure 8 As shown, the velocity transfer matrices of 55 common road spectra were statistically analyzed and saved as a first-order Markov chain model. Simultaneously, ambient temperature, passenger cabin requirements, and powertrain controller parameters were randomly set, and simulations were performed using Amesim. To facilitate convergence and improve the generalization ability of the neural network model, preprocessing of the original data was required, specifically through the following two steps: First, data segmentation enhancement was performed, dividing each original time-series data (length L) into short trajectories of length S, such that the time interval between the starting points of each short trajectory is I. Thus, each time-series data was divided into the difference between L and S, divided by I, rounded down, and then incremented by 1. Second, data standardization was performed, using the z-score normalization method to convert the original data into a standard 0,1 distribution.

[0148] Specifically, the arithmetic mean and standard deviation need to be calculated separately for each variable, and then the standardized data are obtained by using the standardization formula (i.e., the difference between the original data and the mean divided by the standard deviation).

[0149] This requires dividing the dataset into training and testing sets; for example, Figure 1 In the illustrated embodiment, the 13 data points from the above-described 17 datasets, after the aforementioned processing, must be divided into a training set and a test set of 4 data points; for example, Figure 2 In the example shown, the 34 data points after the above processing of the 68 datasets are divided into a training set and a test set.

[0150] Furthermore, the boundaries of the training set variables and the boundary of the rate of change can be calculated. Before training, the upper and lower bounds of the variables, as well as the approximate boundary of the rate of change over time after first and second differencing, can be calculated for subsequent training. Figure 12 The loss function shown is 010.

[0151] Furthermore, system models and neural networks can be trained; such as... Figure 9 This is a general form of the neuron differential equation model; it includes an initial state estimator 091, which can estimate the initial state of the system and the backbone NODE model of the system model based on historical data, and use it to predict the state at future times; such as Figure 9 As shown, its output is obtained by mapping the state through a function g.

[0152] The vehicle's thermal system is controlled by two three-way valves (one of which is normally open); therefore, four modes need to be simulated: such as Figure 10 As shown, the system state undergoes discontinuous jumps at the moment the three-way valve opens and closes (from 0 to non-0 or from non-0 to 0); as can be seen from the figure, the input temperature of the heater core changes at the moment the three-way valve of the engine-heater core opens after 500 seconds of operation.

[0153] To address this issue, embodiments of the present invention can divide the system into six subsystems based on the opening and closing status of the three-way valve. Specifically, variables are input into the corresponding subsystems according to the opening and closing status of the three-way valve. These subsystems include: engine operating independently, WPTC-Heatcore operating independently, battery pack operating independently, engine-WPTC-Heatcore interconnected, WPTC-Heatcore-battery pack interconnected, and engine-WPTC-Heatcore-battery pack interconnected.

[0154] Specifically, the decomposition of subsystems in the embodiments of the present invention is not limited to, for example... Figure 2 The thermal management system shown can be implemented using this method or product for any system with discontinuous processes.

[0155] The loss function can be adopted as follows: Figure 12The structure shown includes the mean absolute error 011 of the measured data and model predictions (representing the model's predictive power) and a regularization term; the regularization term includes constraints on the state values ​​(state value constraint term 013) and constraints on the state derivatives (state derivative constraint term 015), the range of which can be calculated in advance using the training set in the previous step.

[0156] Specifically, its integrator can be implemented using forward Euler or Dopri5 or other numerical integrators; if an adaptive step-size integrator (such as the Runge-Kutta method) is used, the control input is interpolated using the pchip function (other interpolation methods can also be used); neural network parameter optimization can be achieved using the Adam solver (other neural network optimization algorithms can also be used); its gradient calculation can be performed using backpropagation or the adjoint sensitivity algorithm.

[0157] In summary, the NODE network is a fully connected multilayer perceptron, whose core structure includes an initial state estimator, augmented states, event-driven modeling, and loss function design.

[0158] Accordingly, the above embodiments can be verified and evaluated from the perspectives of prediction accuracy and inference speed. For prediction accuracy, the evaluation index 1-NRMSE (standardized root mean square error)*100% is used, which is an index that does not change with linear standardization operations. The value is between negative infinity and 100%, and the closer it is to 100%, the stronger the prediction ability.

[0159] like Figure 13 As shown, the catalyst modeling results are presented; for example, Figure 1 The illustrated embodiment uses the predicted and measured results of four key output signals—HC, CO2, CO, and NOx—from catalyst modeling to compare with, for example... Figure 9 The first set of prediction comparison chart 810 shows the best prediction performance in the test set; the two columns on the right, namely the second set of prediction comparison chart 820, show the worst prediction performance.

[0160] Among them, the model demonstrated relatively accurate predictive ability in 67% of the average metrics across all test sets, while the average inference time on both the training and test sets was 0.13 seconds, making it suitable for real-time applications.

[0161] Similarly, for such Figure 2 The vehicle thermal system model shown is in Figure 14The paper also provides relevant comparative data; among them, it presents a comparison between the prediction of eight temperature output signals for thermal system modeling and the Amesim results; on the entire test set, the model's average index of 92% shows extremely accurate prediction ability, and the average trajectory inference time of more than 4000 seconds of real time on both the training and test sets is less than 2 seconds, which can also be used to realize real-time functions.

[0162] In addition, in the modeling of the vehicle's thermal system, besides the aforementioned open-loop testing (referring to testing on the collected data), software-in-the-loop (SIL) testing was also performed. For this embodiment of the invention, the controller of the actual vehicle interacts with the agent model (referring to the improved NODE in this embodiment) to obtain actual fuel consumption, state of charge (SOC), and temperature signals. The results are compared with those of the Amesim model as follows... Figure 14 Give; from such Figure 15 The test box diagram of the whole vehicle thermal system shown also reveals that: under the test condition of continuously predicting the future temperature for 960 seconds, the median absolute value of fuel consumption error is 0.00223L / 100km, the median absolute value of SOC error is 0.00098 (where SOC takes values ​​within [0,1]), and the median of the average absolute error of the temperature of the 8 nodes in the thermal system is also within 3 degrees Celsius.

[0163] Furthermore, the embodiments of this invention also have the following variations: Firstly, different encoding structures can be adopted, and the extended schemes include using different neural network structures; for example, alternative schemes such as Graph Neural Networks (GNN) and Convolutional Neural Networks (CNN) can be used for initial value encoding or as the network structure for the NODE. Secondly, different training data generation and standardization methods can also be used; different data preprocessing and enhancement techniques can be used; for example, Generative Adversarial Networks (GAN) or Variational Auto-Encoders (VAE) can be used to generate training data, or other different data standardization methods, such as the MinMax method, can be used. Thirdly, different loss function designs and training methods can also be used to change the model's training strategy; for example, different optimization algorithms, loss functions, or regularization techniques can be used. In fact, the methods and products of this invention can be applied across fields; according to their general modeling process, they can be extended to other fields or scenarios.

[0164] It should be noted that the above embodiments are only for more clearly illustrating the technical solution of the present invention. Those skilled in the art will understand that the implementation of the present invention is not limited to the above content. Any obvious changes, substitutions or replacements made based on the above content do not exceed the scope of the technical solution of the present invention. Other implementations will also fall within the scope of the present invention without departing from the concept of the present invention.

Claims

1. A method for processing model data for a catalytic converter system, characterized in that... It includes a first state estimation step (100), a second state enhancement step (200), a third system disassembly step (300), and a fourth state focusing step (400). Based on the prior knowledge of the physical system (900), the dimension of the state vector is determined by the preset dimension parameter N; among which, physical quantities that cannot be directly measured are determined as quantities to be estimated. The first state estimation step (100) obtains the first initial state value of the quantity to be estimated at the start of the simulation process based on historical data. The second state enhancement step (200) uses a preset data processing procedure to perform a second preprocessing on the original data, and uses the output of the second preprocessing for solving the explicit ordinary differential equation; the second state enhancement step (200) also calculates the upper and lower bounds of the preset first variable before model training, and calculates the fifth rate of change boundary values ​​of the first difference and the second difference of the first variable over time. In the third system disassembly step (300), if the physical system (900) has a discontinuous change process, the physical system (900) is decomposed into a preset number of subsystem sets. The fourth state focusing step (400) limits the value of the loss function value (010) Loss according to the fifth rate of change boundary value; wherein: the loss function value (010) is obtained by linear combination of the mean absolute error (011) of the first measurement dataset and the model prediction value with a preset regularization term; the first measurement dataset includes M first original data segments, and the regularization term includes a first constraint (013) on the state value and a second constraint (015) on the state derivative, and the first constraint (013) and the second constraint (015) are calculated by the third training set.

2. The model data processing method as described in claim 1, wherein: The physical system (900) collects preset parameters to obtain a first raw data segment of length L, and the number of the first raw data segments is M; the length L is measured by the time length, and M and N are both positive integers; the set of the Fth state at future time F is predicted according to the preset neuron differential equation, and the set of the Fth state is mapped by the second function (060), which is determined according to the prior knowledge.

3. The model data processing method as described in claim 2, wherein: The second preprocessing includes a data segmentation and enhancement process and / or a normalization process; The segmentation enhancement process divides each of the first original data segments of length L into second data segments of length S, such that the length of each first original data segment is expanded to the sum of the difference between L and S (LS) divided by the time interval I and 1; the standardization process converts the first original data segments into intermediate data under a preset probability distribution.

4. The model data processing method as described in claim 3, wherein: The preset probability distribution includes the standard normal distribution Norm(0,1); the intermediate data is obtained by dividing the difference between the original data Xraw and the arithmetic mean by the standard deviation.

5. The model data processing method as described in claim 3 or 4, wherein: The second preprocessing also divides the intermediate data and / or the second data segment into a third training set and a fourth test set.

6. The model data processing method as described in claim 1, wherein: The solution of the explicit ordinary differential equation is achieved using a preset integrator and input control is performed using a preset interpolation process.

7. The model data processing method as described in claim 6, wherein: The explicit ordinary differential equation is driven by a neural network structure, which includes a graph neural network (GNN) and / or a convolutional neural network (CNN); the neural network structure is used to construct the neural network ordinary differential equation NODE and / or to generate initial codes.

8. A simulation device (600) for a catalyst system, comprising a first state estimation unit (610), a second state enhancement unit (620), a third system disassembly unit (630) and a fourth state focusing unit (640). Based on prior knowledge of the physical system (900), the physical system (900) is represented as an explicit ordinary differential equation, which is time-independent and whose state vector dimension is determined by a preset dimension parameter N; wherein, Physical quantities that cannot be directly measured are identified as quantities to be estimated; The first state estimation unit (610) obtains the first initial state value of the quantity to be estimated at the start of the simulation process based on historical data. The second state enhancement unit (620) performs a second preprocessing on the original data using a preset data processing procedure, and uses the output of the second preprocessing for solving the explicit ordinary differential equation; The second state enhancement unit (620) also calculates the upper and lower bounds of the preset first variable before model training, and calculates the fifth rate of change boundary values ​​of the first difference and the second difference of the first variable over time; If the physical system (900) has a discontinuous change process, the third system disassembly unit (630) will decompose the physical system (900) into a preset number of subsystem sets. The fourth state focusing unit (640) limits the value of the loss function value (010) Loss according to the fifth rate of change boundary value; wherein: the loss function value (010) is obtained by linear combination of the mean absolute error (011) of the first measurement dataset and the model prediction value with a preset regularization term; the first measurement dataset includes M first original data segments, and the regularization term includes a first constraint (013) on the state value and a second constraint (015) on the state derivative, and the first constraint (013) and the second constraint (015) are calculated by the third training set.

9. The simulation apparatus (600) as described in claim 8, wherein: Collect the preset parameters of the physical system (900) to obtain a first raw data segment of length L, and the number of the first raw data segments is M; The length L is measured by the time length, and M and N are both positive integers; the set of the Fth state at future time F is predicted according to the preset neuron differential equation, and the set of the Fth state is obtained by mapping the second function (060), which is determined according to the prior knowledge.

10. The simulation apparatus (600) as described in claim 8, wherein: The second preprocessing includes a data segmentation and enhancement process and / or a normalization process; The segmentation enhancement process divides each of the first original data segments of length L into second data segments of length S, such that the length of each first original data segment is expanded to the sum of the difference between L and S (LS) divided by the time interval I and 1; the standardization process converts the first original data segments into intermediate data under a preset probability distribution.

11. The simulation apparatus (600) as described in claim 10, wherein: The preset probability distribution includes the standard normal distribution Norm(0,1); the intermediate data is obtained by dividing the difference between the original data Xraw and the arithmetic mean by the standard deviation; the second preprocessing also divides the intermediate data and / or the second data segment into a third training set and a fourth test set.

12. The simulation device (600) as described in claim 1, wherein: The solution of the explicit ordinary differential equation is implemented using a preset integrator and input control is performed using a preset interpolation process; the explicit ordinary differential equation is driven by a neural network structure, which includes a graph neural network (GNN) and / or a convolutional neural network (CNN); the neural network structure is used to construct the neural network ordinary differential equation NODE and / or to generate initial codes.

13. A computer storage medium (903) comprising a storage medium body for storing a computer program; wherein the computer program, when executed by a microprocessor, implements the model data processing method as described in any one of claims 1 to 7.

14. A test system (901) comprising a simulation device (600) as claimed in any one of claims 8 to 12 and / or a computer storage medium (903) as claimed in any one of claims 13.

Images