A method and system for modeling a gas exchange process of a marine engine based on a thermodynamic information neural operator

Abstract

The application provides a marine engine gas exchange process modeling method and system based on a thermodynamic information neural operator, and belongs to the technical field of digital twinning of marine power. To solve the contradiction between real-time, accuracy, physical interpretability and cross-type generality of the existing model. The method comprises constructing a thermodynamic information neural operator model, which comprises a branch network receiving quasi-static operating condition parameters and a trunk network receiving instantaneous state parameters. The model is trained using a composite loss function, which comprises a data loss term between the model predicted value and the real data, and a physical loss term derived based on the zero-dimensional charging and discharging method. The physical loss term is specifically defined as the mass conservation equation residual and the energy conservation equation residual. The application embeds the physical law as a strong constraint into the operator network, realizes high-fidelity, real-time, physically interpretable and cross-type generalizable engine gas exchange process simulation.

Description

Technical Field

[0001] This invention belongs to the field of digital and intelligent technology of marine power systems, specifically relating to a modeling method and system for gas exchange processes in marine engines based on thermodynamic information neural operators. Background Technology

[0002] The overall performance of marine diesel engines, including their fuel economy, power output density, and emission control, largely depends on the efficiency of their gas exchange process. This process (typically referring to the period from exhaust valve opening to intake valve closing) is a highly complex nonlinear and unsteady hydrodynamic process. It determines the scavenging efficiency of the high-temperature residual exhaust gas in the cylinder, the fresh air charge coefficient, and the initial thermodynamic state (pressure, temperature, and composition) of the working fluid in the cylinder at the start of the compression stroke. Therefore, establishing a real-time, accurate gas exchange process model that can be generalized to different engine models within the engine's digital twin system is a key prerequisite and core technological bottleneck for achieving high-fidelity simulation.

[0003] However, existing technical solutions in this field for simulating this gas exchange process all have insurmountable and mutually restrictive limitations, as follows: High-precision physical models: These models, such as advanced models based on one-dimensional computational fluid dynamics or the method of characteristics, can accurately simulate pressure wave dynamics, gas inertial effects, and complex flow phenomena within intake and exhaust manifolds with extremely high fidelity. However, their solution process involves complex systems of nonlinear partial differential equations, consuming enormous computational resources. This high computational cost makes them completely unsuitable for the real-time requirements of digital twin systems, especially for in-loop simulations or shipboard deployments requiring rapid response.

[0004] Simplified physical models: The most typical type of this type is the lumped parameter model based on the zero-dimensional charge / discharge method. These models significantly improve computational speed by simplifying the complex intake and exhaust systems (manifold, valves) into a series of control volumes and orifices, allowing them to approach or meet real-time requirements. However, the accuracy of these models heavily relies on a large number of empirical or semi-empirical sub-models and parameters, especially key parameters such as valve flow coefficient and cylinder wall heat transfer coefficient. These parameters are not only difficult to obtain accurately under all operating conditions, but they also drift nonlinearly with changes in engine wear, aging, carbon buildup, and even fuel quality. Obtaining these parameters typically requires expensive and time-consuming engine bench dynamic testing or pre-calibration using high-precision one-dimensional CFD simulations. This results in extremely poor versatility for different engine models, long deployment cycles, high calibration costs, and a lack of predictive capability for uncalibrated operating conditions or new engine configurations.

[0005] Purely data-driven models: These models include traditional machine learning models (such as support vector machines and random forests) or standard deep neural networks. Once trained, these models have extremely fast forward inference speeds, fully meeting real-time requirements. However, their inherent "black box" nature results in almost zero physical interpretability. The model only learns the statistical correlation between input and output, rather than the physical causal relationship, making it highly susceptible to overfitting noise or spurious correlations in the training data. More seriously, these models are essentially high-dimensional function fitters with extremely weak generalization ability. When engine operating conditions (such as speed and load) exceed the manifold range covered by the training data, or when directly applied to new engine models with untrained geometry, their prediction accuracy drops sharply, even producing catastrophic errors that severely deviate from physical reality.

[0006] Standard Physical Information Neural Network (PINN): In recent years, PINN technology has been proposed, which enhances the physical interpretability of data-driven models to some extent by introducing residual terms of physical equations into the loss function of the neural network. However, standard PINN is essentially still a solver for specific problems (i.e., specific engines and specific operating ranges). It learns a function solution under specific boundary conditions and specific geometries. It lacks the ability to learn "operators," that is, it lacks the ability to take boundary conditions, geometric parameters, or operating parameters themselves as input. Therefore, when faced with various engine models with different engine geometric parameters (such as different bores, strokes, and valve positions) or different operating boundaries (such as different intake and exhaust pressures), standard PINN models must be costly retrained for each new situation, and their cross-model versatility, generalization ability, and rapid deployment capability still face significant challenges.

[0007] In summary, there is an urgent need in this field for a novel modeling method for engine gas exchange processes. This method must be able to fundamentally solve the aforementioned technical pain points simultaneously, namely: it must possess the interpretability and physical consistency of a high-precision physical model, the real-time computational capability of a data model, and the powerful generalization capability across operating conditions and models. Summary of the Invention

[0008] This invention addresses the shortcomings of existing technologies in simulating the gas exchange process of marine engines, particularly in terms of real-time performance, accuracy, and physical properties. To address the contradiction between interpretability and cross-model versatility, a modeling method and system for marine engine gas exchange processes based on thermodynamic information neural operators is proposed.

[0009] The technical solution adopted by the present invention to solve the above problems is as follows: This invention proposes a modeling method for gas exchange processes in marine engines based on thermodynamic information neural operators, comprising the following steps: A thermodynamic information neural operator model is constructed, which includes a branch network and a backbone network. The thermodynamic information neural operator model is trained using a composite loss function; the composite loss function includes data loss and physical loss terms; the physical loss terms include the residuals of the mass conservation equation and the energy conservation equation derived based on the thermodynamic conservation law of the zero-dimensional charge-discharge method.

[0010] Furthermore, the input to the branch network includes a condition parameter vector characterizing the quasi-static operating condition of the engine, and the condition parameter vector includes at least one of the following parameters: engine speed, engine load, or an encoding characterizing the valve operating state.

[0011] Furthermore, the inputs to the backbone network include a characterization of the engine at the current computation time. The instantaneous state vector includes at least one of the following parameters: instantaneous intake valve flow area, instantaneous exhaust valve flow area, intake manifold pressure, intake manifold temperature, exhaust manifold pressure, exhaust manifold temperature, instantaneous cylinder volume, instantaneous rate of change of cylinder volume, or the previous calculation time. The cylinder pressure and cylinder temperature.

[0012] Furthermore, the output of the thermodynamic information neural operator model is the next calculation time step. The predicted values ​​of cylinder pressure and cylinder temperature are obtained by summing the dot product of the output of the branch network and the output of the trunk network.

[0013] Furthermore, the residual of the mass conservation equation is calculated as the difference between a first rate of mass change and a second rate of mass change; the first rate of mass change is the next time step predicted by the model. The cylinder pressure and temperature are used, and the in-cylinder working fluid mass change rate is calculated using the ideal gas law; the second mass change rate is based on the current calculation time input from the backbone network. The instantaneous state vector is used to calculate the net inflow mass flow rate through the intake and exhaust valves, which is obtained by combining the compressible gas flow equation.

[0014] Furthermore, the residual of the energy conservation equation is calculated as the difference between a first rate of energy change and a second rate of energy change; the first rate of energy change is the next time step predicted by the model. The rate of change of the total thermodynamic energy of the working fluid in the cylinder is calculated based on the in-cylinder temperature and the mass of the working fluid; the second energy change rate is based on the current calculation time input from the backbone network. The instantaneous state vector is calculated as the algebraic sum of the power terms, which include at least the inflow enthalpy power, the outflow enthalpy power, and the piston work power.

[0015] Furthermore, the inflow enthalpy power and outflow enthalpy power in the second energy change rate calculation are calculated based on the second mass change rate and the temperature of the intake and exhaust manifolds; the piston power is calculated based on the cylinder pressure and the instantaneous change rate of cylinder volume in the instantaneous state vector.

[0016] This invention also proposes a modeling system for the gas exchange process of a marine engine based on a thermodynamic information neural operator, comprising: a storage module for storing a thermodynamic information neural operator model, the model including a branch network and a backbone network; and a processing module for training the thermodynamic information neural operator model using a composite loss function; wherein the composite loss function includes a data loss term and a physical loss term; the physical loss term includes the residuals of the mass conservation equation and the energy conservation equation derived based on the thermodynamic conservation law of the zero-dimensional charge-discharge method.

[0017] Furthermore, the branch network is configured to receive a vector of operating parameters characterizing the engine's quasi-static operating condition; the backbone network is configured to receive a vector of operating parameters characterizing the engine at the current calculation time. The instantaneous state vector.

[0018] Furthermore, when calculating the physical loss term, the processing module is further configured to: calculate the residual of the mass conservation equation, which is obtained by comparing a first mass change rate inferred from the model output and a second mass change rate calculated from the instantaneous state vector; and calculate the residual of the energy conservation equation, which is obtained by comparing a first energy change rate inferred from the model output and a second energy change rate calculated from the instantaneous state vector.

[0019] The beneficial effects of this invention are: A Balance Between High Precision and High Interpretability: This invention innovatively combines the powerful nonlinear mapping capabilities of deep operator networks with the conservation law constraints of physical information neural networks. The introduction of a physical loss term is akin to adding physical "prior knowledge" derived from first principles to the neural network, embedding thermodynamic laws (conservation of mass and energy) as strong constraints into the network weights. This enables the model to make physically sound predictions even in regions with sparse training data, completely avoiding the "black box" problem of traditional neural networks and giving the model a high degree of physical interpretability and credibility in simulating the gas exchange process within a cylinder.

[0020] Guaranteed real-time computing performance: Compared to one-dimensional CFD models that solve complex partial differential equation systems (whose computation time is measured in minutes or even hours), the thermodynamic information neural operator model described in this invention, once trained, involves only a series of highly optimized matrix operations in its forward inference process. It can complete the simulation of a single-cycle gas exchange process in milliseconds, far exceeding the real-time requirements of physics, fully meeting the stringent timeliness requirements of digital twin systems, in-loop simulation, and shipboard real-time control.

[0021] Exceptional cross-model versatility and adaptability to various operating conditions: this is the core advantage of this invention. The operator architecture (branch network / backbone network) of this invention allows its learning objective to shift from traditional function approximation (i.e., learning...) ) Upgrading to operator approximation (i.e., learning) The vector of operating parameters input to the branch network. This enables the model to learn the meta-learning relationship between "operating condition parameters" and "dynamic response mapping". Therefore, the model of this invention can not only accurately predict unseen operating conditions outside the training set, but also learn the meta-learning relationship between the operating condition parameter vectors in the branch network. By inputting different parameters (such as different cylinder bores and different valve timings), it is possible to quickly generalize to different engine models, greatly reducing the massive amount of work required to repeatedly model and calibrate for each new engine model.

[0022] Significantly Reduced Calibration Dependence: Compared to zero-dimensional charge / discharge models that heavily rely on empirical coefficients (such as valve flow coefficients), this invention's model, through a deep network structure, directly learns complex nonlinear fluid dynamics processes (including effects such as airflow throttling and inertia) end-to-end from data and physical constraints. This bypasses the cumbersome pre-calibration process for empirical parameters that are difficult to obtain, prone to drift, and costly to calibrate, significantly improving the model's deployment efficiency, robustness, and applicability. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the main framework principle of the present invention; Figure 2 This is a comparison chart of errors in the ablation experiment of this invention at different iteration numbers; Figure 3 This is a comparison chart of TINO's prediction results with other machine learning models. Detailed Implementation

[0024] Specific Implementation Method 1: This implementation method proposes a modeling method for the gas exchange process of marine engines based on thermodynamic information neural operators. This method is specifically designed to simulate the non-closed-loop stage of the engine from exhaust valve opening (EVO) to intake valve closing (IVC). The core of this method lies in constructing and using a thermodynamic information neural operator model.

[0025] The thermodynamic information neural operator model described above is a customized hybrid neural network architecture that deeply integrates physical information constraints and deep operator network theory. This architecture abandons the traditional neural network paradigm of fitting specific data points and instead focuses on learning the "operator" or "mapping relationship" between engine operating parameters (i.e., "input function") and its dynamic response process.

[0026] The thermodynamic information neural operator model is structurally decoupled into a branch network and a backbone network: Branch network: Input: A vector of operating parameters characterizing the engine's quasi-static or fixed properties. This vector parameterizes the operator itself, representing the selection of a specific dynamic system instance. Specifically, it includes: a normalized engine speed, a normalized engine load, and a one-hot encoded vector characterizing the current valve operating state (e.g., intake only, exhaust only, intake-exhaust overlap, fully closed, etc.). In other alternative embodiments, this vector may also include key parameters characterizing engine geometry or control characteristics, such as valve timing parameters, engine bore, and stroke.

[0027] Output: A set of basis function coefficients learned by a deep neural network.

[0028] Backbone network: Input: Characterizing the engine system at the current calculation time. The instantaneous state vector. This instantaneous state vector represents the query for the value of the operator at a specific "coordinate point" at a specific time. Specifically, it includes: the current time... The instantaneous flow area of ​​the intake valve, the instantaneous flow area of ​​the exhaust valve, the average pressure of the intake manifold, the average temperature of the intake manifold, the average pressure of the exhaust manifold, the average temperature of the exhaust manifold, the instantaneous volume of the cylinder at the current moment, the instantaneous rate of change of the cylinder volume, and the previous calculation time. The cylinder pressure and cylinder temperature.

[0029] Output: A set of basis functions learned by a deep neural network.

[0030] Model output: The final output of the model is defined as the sum of the dot products of the basis function coefficients output by the branch network and the basis functions output by the backbone network, with the option to add a trainable bias term.

[0031] The output result is the result of the calculation for the next time step. The predicted value of the engine cylinder state is specifically a two-dimensional vector containing cylinder pressure and cylinder temperature.

[0032] As the core innovation of this invention, to ensure that the thermodynamic information neural operator model possesses high-precision physical interpretability and strong generalization ability, its composite loss function during training is specifically designed as a weighted sum of two parts: a data loss term and a physical loss term. The mathematical expression of this composite loss function can be summarized as follows: in, This is a hyperparameter weight used to balance the relative contributions of the two parts of the loss.

[0033] In the above composite loss function, the (Data loss term) is defined as the model prediction (i.e., the next time step). The mean square error (MSE) or other statistical distance metric between the cylinder pressure and cylinder temperature data and the actual baseline data (which can come from high-fidelity one-dimensional CFD model simulation or engine bench test measurements). This is used to ensure the accuracy of the model predictions and the fidelity to observed facts.

[0034] In the above composite loss function, the The physical loss term is defined as the residual of the physical equations derived from the physical conservation laws (i.e., the first law of thermodynamics and the law of conservation of mass) based on the zero-dimensional charge-discharge method. This residual term forces the model's prediction process to strictly adhere to the laws of thermodynamic conservation, effectively applying a strong inductive bias derived from the first principles of physics to the neural network. This term ensures the physical interpretability of the model and its generalization ability in sparse regions of the training data. The physical loss term specifically includes: The residual of the mass conservation equation is constructed as the squared difference between the "first rate of change of mass" and the "second rate of change of mass." The "first rate of change of mass" refers to the next time step predicted by the thermodynamic information neural operator model. The cylinder pressure and cylinder temperature, combined with the ideal gas law (its expression is...) ), inferred from the model output arrive The "second rate of change" refers to the rate of change of the mass of the working fluid in the cylinder at any given time. The "second rate of change" refers to the rate of change of the mass of the working fluid at the current time, as input from the main network. The instantaneous state vectors (especially pipeline pressure, pipeline temperature, cylinder pressure, valve flow area, etc.) are directly calculated from the physical boundary conditions using the standard compressible gas flow equations. The net inflow mass flow rate through the intake and exhaust valves at all times.

[0035] The residual of the energy conservation equation is constructed as the squared difference between the "first rate of energy change" and the "second rate of energy change." The "first rate of energy change" refers to the next time step predicted by the thermodynamic information neural operator model. The cylinder temperature and the result calculated in conjunction with the aforementioned "first mass change rate" arrive The total thermodynamic energy of the working fluid in the cylinder at any given time (its expression is: The rate of change of energy. The "second rate of change of energy" refers to the rate of change of energy based on the current moment. The algebraic sum of the energy (power) terms calculated from the instantaneous state vector (i.e., physical boundary conditions) includes: the inflow enthalpy flow and outflow enthalpy flow calculated based on the "second mass change rate" and pipeline temperature; the piston power calculated based on the current cylinder pressure and cylinder volume change rate; and the cylinder wall heat transfer power calculated based on the temperature difference between the working fluid and the cylinder wall.

[0036] By simultaneously minimizing data loss and physical loss in the optimization objective, this invention forces deep operator networks to learn and strictly abide by the fundamental physical laws of mass conservation and energy conservation while learning the data distribution, thereby fundamentally solving the "black box" problem and physical unreliability of traditional data-driven models.

[0037] Specific Implementation Method 2: This implementation method proposes a marine engine gas exchange process modeling system based on thermodynamic information neural operators. The system includes: a storage module for storing thermodynamic information neural operator models, which include branch networks and backbone networks. The branch network is configured to receive a vector of operating parameters characterizing the engine's quasi-static operating condition; the backbone network is configured to receive a vector of operating parameters characterizing the engine at the current calculation time. The instantaneous state vector.

[0038] The processing module is used to train the thermodynamic information neural operator model using a composite loss function; wherein the composite loss function includes a data loss term and a physical loss term; the physical loss term includes the residuals of the mass conservation equation and the energy conservation equation derived based on the thermodynamic conservation law of the zero-dimensional charge-discharge method.

[0039] When calculating the physical loss term, the processing module is further configured to: calculate the residual of the mass conservation equation, which is obtained by comparing a first mass change rate inferred from the model output and a second mass change rate calculated from the instantaneous state vector; and calculate the residual of the energy conservation equation, which is obtained by comparing a first energy change rate inferred from the model output and a second energy change rate calculated from the instantaneous state vector.

[0040] Example: The technical solution of the present invention will be described in more detail and in a more complete manner below, so that those skilled in the art can understand and implement it.

[0041] This embodiment proposes a general modeling method for the gas exchange process of marine diesel engines. The core technology is to construct, train, and apply a thermodynamic information neural operator model. This model is specifically responsible for simulating the complete gas exchange phase of the engine from the exhaust valve opening moment (EVO) to the intake valve closing moment (IVC).

[0042] Module: Construction and Training of Thermodynamic Information Neural Operator Model The core objective of this module is to learn a nonlinear operator that can characterize the gas exchange process in an engine. The operator Receive a function / vector representing the operating condition. (Input from the branch network), and output a mapping (function) that represents the instantaneous state. (Input from the backbone network) is mapped to the instantaneous state of the system at the next moment. .

[0043] 1. Network Architecture In this specific embodiment, the thermodynamic information neural operator model adopts a deep operator network architecture, which consists of two parallel sub-networks (branch network and backbone network) typically composed of multilayer perceptrons (MLPs).

[0044] Branch network: its input It is a 5-dimensional working condition parameter vector, specifically including: [Standardized engine speed, standardized engine load, valve condition one-hot encoding (3D)] This one-hot encoding is used to explicitly inform the network which gas exchange substage it is currently in (e.g., exhaust only, overlapping intake and exhaust, intake only, etc.). The branch network consists of multiple fully connected neural networks, and its final output is a set of basis function coefficients, such as a... dimensional vector .

[0045] Backbone network: its input It is an 11-dimensional instantaneous state vector, at time... The collection specifically includes: [Instantaneous intake valve flow area, instantaneous exhaust valve flow area, average intake manifold pressure, average intake manifold temperature, average exhaust manifold pressure, average exhaust manifold temperature, instantaneous cylinder volume, instantaneous rate of change of cylinder volume, cylinder pressure at the previous time t, cylinder temperature at the previous time t] This vector comprehensively describes the instantaneous physical boundaries and thermodynamic initial states of gas exchange. The backbone network is also composed of multi-layer fully connected neural networks, and its final output is a set of basis functions, such as a... dimensional vector .

[0046] Output: Model's prediction for the next time step The predicted state values, i.e., [cylinder pressure (t+1), cylinder temperature (t+1)], are obtained by summing the dot product of the coefficient vector output by the branch network and the basis function vector output by the backbone network (its mathematical form is...). This is obtained by adding a learnable bias term.

[0047] 2. Introduction of physical constraints (core implementation example) The key to the success of this invention, and its core difference from the standard DeepONet, lies in the composite loss function used to train the model. This function can be expressed as a weighted sum as follows:

[0048] in, The term represents the mean squared error between the model's predicted values ​​and the training data (e.g., from a high-fidelity one-dimensional simulation model such as GT-suite or engine bench tests). This term ensures that the model fits the real data.

[0049] and This term is a physical residual term that ensures the model obeys the first law of thermodynamics (energy conservation) and mass conservation, constructed based on the conservation equations of the zero-dimensional charge-discharge method. This term guarantees the physical consistency of the model during prediction.

[0050] Detailed construction of the mass conservation residual: First mass change rate (network prediction side): Model outputs predicted pressure for the next time step. and predicted temperature According to the ideal gas law, i.e. The predicted in-cylinder mass can be calculated using the following formula: The internal volume of the cylinder is... and gas constant All are known. Therefore, the first rate of mass change (with respect to crankshaft angle) inferred from the network prediction is... The derivative of ( ) is defined as: ,in It is the known mass at the previous moment.

[0051] Second rate of mass change (physical equation side): based on the input from the backbone network. Time-state vector (Include , , , , , , The mass flow rate through the intake valve can be directly calculated using standard compressible gas flow equations (such as the Saint-Venant-Wantzer equation). and the mass flow rate through the exhaust valve Therefore, the second mass change rate (net flow rate) calculated from the physical boundary conditions is defined as: .

[0052] Residual: The residual due to mass conservation is the L2 norm (squared) of the difference between the two, and its expression is: .

[0053] Detailed construction of the energy conservation residual: First energy change rate (network prediction side): The total in-cylinder thermodynamic energy can be derived from... Obtained. The first rate of energy change inferred from network prediction. Defined as: .

[0054] Second rate of energy change (from the perspective of physical equations): According to the first law of thermodynamics in the form of an open system, the rate of energy change is equal to the sum of the powers of each component, and its expression is: Among them, intake enthalpy and exhaust enthalpy can be and (or cylinder temperature) The power of the piston is calculated (depending on the flow direction); can be and (All in) The cylinder wall heat transfer power was calculated using a (Chinese) method. It can be derived from the standard heat transfer formulas of Woschni et al. and Calculated. All these terms can be obtained from... instantaneous state vector at time t (i.e., physical boundary conditions) are calculated.

[0055] Residual: The energy conservation residual is the L2 norm (squared) of the difference between the two, and its expression is: .

[0056] Final physical loss It is defined as the sum of the mass conservation residual and the energy conservation residual, that is: .

[0057] By minimizing simultaneously during the optimization process and The thermodynamic information neural operator model described in this invention is forced to strictly adhere to the fundamental physical laws of mass conservation and energy conservation while fitting the observed data.

[0058] The verification of the TINO module is as follows: Figure 3 As shown, the model of this invention, under multiple (e.g., four) verification conditions that were not trained, showed that its prediction results for the dynamic process of in-cylinder pressure and in-cylinder temperature were in high agreement with the baseline values, demonstrating its high accuracy and generalization ability to unseen conditions.

[0059] Compared to traditional machine learning methods such as XGBoost, Support Vector Machine (SVM), and Random Forest (RF), the prediction accuracy of the model in this invention has a significant, order-of-magnitude advantage, and traditional methods struggle to capture such complex temporal dynamics.

[0060] Ablation experiments are a key technology verification method for this invention. For example... Figure 2 As shown, this experiment compared three models: (A) a standalone physical information neural network (PINN, i.e., without an operator architecture); (B) a standalone deep operator network (DeepONet, i.e., ... (A) Without physical constraints); (C) The thermodynamic information neural operator model of the present invention (TINO = PINN + DeepONet). Experimental results clearly show that the TINO model (C) of the present invention has the fastest convergence speed, the lowest final loss, and the highest prediction accuracy. In contrast, independent PINN (A) and independent DeepONet (B) both exhibit significant prediction bias. This strongly demonstrates the synergistic advantages and advancements of the proposed hybrid architecture: the operator architecture provides the necessary generalization, while physical constraints provide the necessary accuracy and physical consistency; the combination of the two produces a technical effect greater than 1+1>2.

[0061] In summary, this invention successfully provides a high-fidelity, high-efficiency, highly versatile, and physically interpretable digital twin modeling solution by applying a thermodynamic information neural operator model to the engine gas exchange process.

[0062] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. A method for modeling the gas exchange process of a marine engine based on thermodynamic information neural operators, characterized in that, This method is used to simulate the non-closed-loop phase of an engine from the opening of the exhaust valve to the closing of the intake valve, and the method includes the following steps: A thermodynamic information neural operator model is constructed, which includes a branch network and a backbone network. The input to the branch network includes a vector of operating parameters characterizing the quasi-static operating condition of the engine, and the operating parameter vector includes at least one of the following parameters: engine speed, engine load, or an encoding characterizing the valve operating state. The input to the backbone network includes a characterization of the engine at the current calculation time. The instantaneous state vector includes at least one of the following parameters: instantaneous intake valve flow area, instantaneous exhaust valve flow area, intake manifold pressure, intake manifold temperature, exhaust manifold pressure, exhaust manifold temperature, instantaneous cylinder volume, instantaneous rate of change of cylinder volume, or the previous calculation time. In-cylinder pressure and in-cylinder temperature; The thermodynamic information neural operator model is trained using a composite loss function; the composite loss function includes data loss and physical loss terms; the physical loss terms include the residuals of the mass conservation equation and the energy conservation equation derived based on the thermodynamic conservation law of the zero-dimensional charge-discharge method. The residual of the mass conservation equation is calculated as the difference between the first rate of change of mass and the second rate of change of mass. The first rate of change of mass is the next time step predicted by the model. The cylinder pressure and cylinder temperature are used to calculate the rate of change of the working fluid mass in the cylinder, which is derived from the ideal gas law. The second rate of change in quality is based on the current calculation time input from the backbone network. The instantaneous state vector is obtained, and the net inflow mass flow rate through the intake and exhaust valves is calculated by combining the compressible gas flow equation. The residual of the energy conservation equation is calculated as the difference between the first energy change rate and the second energy change rate; The first rate of energy change is the next time step predicted by the model. The rate of change of the total thermodynamic energy of the working fluid in the cylinder is calculated based on the in-cylinder temperature and the mass of the working fluid in the cylinder. The second rate of energy change is based on the current calculation time input from the backbone network. The instantaneous state vector is calculated as the algebraic sum of the power terms, which include at least the inflow enthalpy power, the outflow enthalpy power, and the piston work power.

2. The method for modeling the gas exchange process of a marine engine based on thermodynamic information neural operators according to claim 1, characterized in that, The output of the thermodynamic information neural operator model is the next calculation time. The predicted values ​​of cylinder pressure and cylinder temperature are obtained by summing the dot product of the output of the branch network and the output of the trunk network.

3. The method for modeling the gas exchange process of a marine engine based on thermodynamic information neural operators according to claim 2, characterized in that, The inflow enthalpy power and outflow enthalpy power in the second energy change rate calculation are obtained based on the second mass change rate and the temperature of the intake and exhaust manifolds; the piston power is calculated based on the cylinder pressure and the instantaneous change rate of cylinder volume in the instantaneous state vector.

4. A modeling system for a method of modeling gas exchange processes in marine engines based on thermodynamic information neural operators according to any one of claims 1 to 3, characterized in that, The system includes: The storage module is used to store a thermodynamic information neural operator model, which includes a branch network and a backbone network; The branch network is configured to receive a vector of operating parameters characterizing the engine's quasi-static operating condition; the backbone network is configured to receive a vector of operating parameters characterizing the engine at the current calculation time. The instantaneous state vector; The processing module is used to train the thermodynamic information neural operator model using a composite loss function; wherein the composite loss function includes a data loss term and a physical loss term; the physical loss term includes the residuals of the mass conservation equation and the energy conservation equation derived based on the thermodynamic conservation law of the zero-dimensional charge-discharge method; The processing module is further configured to: The residuals of the mass conservation equation are calculated by comparing the first rate of mass change inferred from the model output with that inferred from the model output. The second rate of change of mass is obtained from the instantaneous state vector calculation; The residual of the energy conservation equation is calculated by comparing the first rate of energy change inferred from the model output with that inferred from the model output. The second rate of energy change is derived from the instantaneous state vector calculation.

Images