Engine health assessment method and system based on one-dimensional shock tube flow

Abstract

The invention discloses an engine health assessment method and system based on one-dimensional shock tube flow, and relates to the technical field of engine fault intelligent diagnosis. The method comprises the following steps: acquiring the running state and corresponding running data of each device in the engine; working condition classification in the Riemann problem of the one-dimensional Euler equation is matched according to the operation state and the operation data of the device, and the working condition classification comprises strong shock waves, weak shock waves, strong expansion, weak expansion and contact interruption leading; constructing an engine state prediction model based on a depth operator network according to the working condition classification, and training the engine state prediction model by using a data set corresponding to the working condition classification; and performing health assessment on the engine device by using the trained engine state prediction model. According to the method, the flow problem of the equivalent one-dimensional shock tube is solved by using the depth operator network, so that the accurate evaluation of the health state of the engine is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent engine fault diagnosis technology, and in particular to an engine health assessment method and system based on one-dimensional shock tube flow. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] In engine development, health status assessment is a crucial step. Engines operate under complex multi-condition scenarios, increasing the difficulty of online health monitoring. For example, transient pressure waves in the intake manifold directly affect the engine's charging efficiency (volume efficiency) and scavenging effect, which are key factors determining engine power, torque, and emissions. Similarly, transient pressure waves in the exhaust system not only directly affect the engine's scavenging efficiency and pumping losses but also determine the availability of exhaust energy (e.g., turbocharging) and the thermal shock and mechanical stress of aftertreatment devices (e.g., three-way catalytic converters, particulate filters). However, traditional assessment methods rely on one-dimensional CFD software (such as GT-POWER and AVL BOOST) for simulation. While more efficient than three-dimensional CFD, multi-parameter, multi-condition optimization designs (e.g., manifold length, diameter, structural optimization) still require repeated solving of the Euler / Navier-Stokes equations, resulting in high computational costs. This makes real-time control or online health monitoring difficult, hindering rapid iterative design.

[0004] Computational fluid dynamics is a discipline that uses numerical methods and computers to solve fluid dynamics governing equations, thereby simulating and analyzing fluid flow phenomena. It divides the continuous physical space (and time) into finite, discrete grids or point sets, transforms the partial differential equations to be solved on these discrete elements or nodes, and converts the partial differential equations into a system of algebraic equations about the physical quantities on the grid elements or nodes. Then, it solves this large system of linear or nonlinear equations on a computer.

[0005] The one-dimensional shock tube problem is a classic benchmark problem in computational fluid dynamics. In a straight tube with a constant cross-sectional area filled with gas, at an initial moment (t=0), the tube is divided into two regions at a certain location (e.g., x=0) by an infinitely thin diaphragm. These two regions are filled with gases of different states. When the diaphragm is instantaneously removed, a strong unsteady flow is triggered due to the initial pressure discontinuity. This flow process includes a series of typical wave phenomena, such as shock waves, contact discontinuities, and expansion waves. The core difficulty of the shock tube problem stems from the fact that its solution is a hybrid solution governed by nonlinear partial differential equations and containing motion discontinuities. Numerical algorithms in computational fluid dynamics need to be able to accurately simulate the shock tube problem.

[0006] Equating the engine intake and exhaust processes and key transient events to a "shock tube" problem can effectively alleviate the computational cost pressure of existing methods in handling complex operating conditions. However, how to accurately characterize the engine operation process using the one-dimensional shock tube flow principle is a problem that urgently needs to be solved by existing technologies. In addition, in the process of predicting engine state based on the one-dimensional shock tube problem, traditional numerical methods, such as the finite difference method, finite volume method, and finite element method, require fine mesh generation for high computational accuracy, resulting in long computation time, and are prone to numerical oscillations when dealing with high gradient regions (such as shock waves). Traditional neural network methods, such as fully connected neural networks and convolutional neural networks, typically learn point-to-point mappings. When initial or boundary conditions change, the network needs to be retrained, and their ability to learn function-to-function mappings (i.e., operator learning) is limited. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the present invention aims to provide an engine health assessment method and system based on one-dimensional shock tube flow. By utilizing a deep operator network to solve the equivalent one-dimensional shock tube flow problem, an accurate assessment of the engine's health status is achieved.

[0008] To achieve the above objectives, the present invention is implemented through the following technical solution: The first aspect of this invention provides an engine health assessment method based on one-dimensional shock tube flow, comprising the following steps: Acquire the operating status and corresponding operating data of each component in the engine; Based on the device's operating status and operating data, the operating condition classification in the one-dimensional Euler equation Riemann problem is matched. The operating condition classification includes strong shock wave, weak shock wave, strong expansion, weak expansion, and contact discontinuity dominance. An engine condition prediction model is constructed based on deep operator networks according to the operating condition classification, and the engine condition prediction model is trained using the dataset corresponding to the operating condition classification. The trained engine condition prediction model is used to perform health assessments on engine components.

[0009] Furthermore, the specific steps for classifying operating conditions in the one-dimensional Euler equation Riemann problem based on the device's operating status and operating data are as follows: Establish the equivalent relationship between the operating states of various components in the engine and the one-dimensional shock tube flow process; Based on the equivalence relation, the engine health assessment problem is transformed into a one-dimensional Euler equation Riemann problem to be solved.

[0010] Furthermore, the specific conditions for classifying operating conditions are as follows: Strong shock wave conditions are characterized by high temperature and high pressure on the left and low temperature and low pressure on the right; The weak shock wave conditions are: moderate temperature and moderate pressure on the left side, and temperature on the right side is close to that on the left side, while pressure is slightly lower than that on the left side. Strong expansion conditions are characterized by high temperature and low pressure on the left and low temperature and high pressure on the right. The weak expansion condition is characterized by moderate temperature and low pressure on the left and low temperature and moderate pressure on the right. The contact interruption conditions are medium temperature and low pressure on the left and low temperature and medium pressure on the right.

[0011] Furthermore, the engine state prediction model includes a conditional encoding branch and a spatiotemporal processing branch. The conditional encoding branch is used to encode the initial and boundary conditions for the working condition classification and extract features related to the global flow state. The spatiotemporal processing branch is used to encode any specified spatial and temporal coordinates and decouple them to generate independent spatiotemporal features corresponding to specific physical quantities.

[0012] Furthermore, the specific steps for training the engine condition prediction model using the dataset corresponding to the operating condition classification are as follows: Design a comprehensive loss function, which includes data fitting loss and physical constraint loss; The engine condition prediction model is trained in stages using the dataset corresponding to the operating condition classification and the comprehensive loss function.

[0013] Furthermore, the specific steps for conducting health assessments of engine components using the trained engine condition prediction model are as follows: The system acquires the real-time operating status and corresponding operating data of each component of the engine, determines the operating condition classification, and obtains the initial conditions and corresponding spatiotemporal coordinates. The initial conditions are encoded using a conditional coding branch to generate high-dimensional conditional features, and the spatiotemporal processing branch is used to generate spatiotemporal features of multiphysics. The high-dimensional conditional features and spatiotemporal features are then fused and analyzed to obtain the prediction results of multiphysics.

[0014] A second aspect of the present invention provides an engine health assessment system based on one-dimensional shock tube flow, comprising: The data acquisition module is configured to acquire the operating status and corresponding operating data of various components in the engine; The equivalent module is configured to match the operating condition classification in the one-dimensional Euler equation Riemann problem based on the device's operating status and operating data. The operating condition classification includes strong shock wave, weak shock wave, strong expansion, weak expansion and contact discontinuity dominance. The model training module is configured to build an engine state prediction model based on a deep operator network according to the working condition classification, and train the engine state prediction model using the dataset corresponding to the working condition classification. The condition assessment module is configured to perform health assessments on engine components using a trained engine condition prediction model.

[0015] A third aspect of the present invention provides a computer-readable storage medium storing a computer program adapted for loading by a processor and executing steps in the engine health assessment method based on one-dimensional shock tube flow as described in the first aspect of the present invention.

[0016] A fourth aspect of the present invention provides a computer device comprising: A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the engine health assessment method based on one-dimensional shock tube flow as described in the first aspect of the present invention.

[0017] A fifth aspect of the present invention provides a computer program product or computer program comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps in the engine health assessment method based on one-dimensional shock tube flow as described in the first aspect of the present invention.

[0018] The above one or more technical solutions have the following beneficial effects: This invention discloses an engine health assessment method and system based on one-dimensional shock tube flow. By fully exploring the relationship between the engine operation process and the one-dimensional shock tube flow process, the complex operating conditions of the engine are equivalent to a one-dimensional Euler equation Riemann problem. The analysis is carried out under different operating conditions, thereby realizing efficient online health assessment of the engine.

[0019] This invention presents a numerical computation method for one-dimensional unsteady flow equations based on a deep operator network framework. This method solves the one-dimensional unsteady flow equations using a deep operator network architecture, overcoming the shortcomings of traditional numerical methods, such as reliance on fine mesh partitioning, long computation time, and susceptibility to numerical oscillations when dealing with high gradient regions (e.g., shock waves). It also addresses the limited ability of traditional neural network methods to map functions to functions, enabling the effective training of corresponding mapping sets from limited data samples.

[0020] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the depth operator network structure in Embodiment 1 of the present invention; Figure 2 This is a graph showing the loss function values ​​of the depth operator network in Embodiment 1 of the present invention; Figure 3 This is a comparison chart of the training results of the deep operator network and the theoretical exact solution results in Embodiment 1 of the present invention. Detailed Implementation

[0023] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0024] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0025] Example 1: Embodiment 1 of the present invention provides an engine health assessment method based on one-dimensional shock tube flow, comprising the following steps: Step 1: Obtain the operating status and corresponding operating data of each component in the engine.

[0026] Step 2: Match the operating conditions in the one-dimensional Euler equation Riemann problem with the device's operating status and operating data.

[0027] Step 2.1: Establish the equivalent relationship between the operating state of each component in the engine and the flow process in the one-dimensional shock tube.

[0028] In one specific implementation, this embodiment assumes that the engine flow is one-dimensional and defines the flow direction from the engine inlet to the engine outlet as positive. The positive flow direction within the engine's pipes is consistent with the positive flow direction of the engine. Based on the pipe flow direction, the pipe inlet is defined as the left side. Relevant physical parameters are indicated using subscripts. The right side indicates the export route, and relevant physical parameters are indicated using subscripts. This indicates that, based on the actual physical parameters of the target shock tube experiment or engineering scenario, the initial conditions of the computational domain are determined, namely the temperature on the left side of the tube. Flow velocity on the left side Left-side pressure Temperature on the right side Flow velocity on the right side Right-side pressure Define the initial condition vector Typically, the initial velocity Based on the pressure and temperature ratios of the left and right states, the type of Riemann problem to which this problem belongs can be estimated.

[0029] The operating conditions are classified into strong shock waves, weak shock waves, strong expansion (sparse waves), weak expansion, and contact discontinuity dominance, and a set of Riemann problem datasets for each type of operating condition is provided for pre-training the network.

[0030] The specific conditions for classifying operating conditions are as follows: The strong shock wave condition is characterized by high temperature and high pressure on the left and low temperature and low pressure on the right. The operating condition judgment condition is... >> In the pre-training dataset, the temperature range on the left side of the shock tube is 600-800K, the temperature range on the right side is 300-400K, the pressure range on the left side is 8-10 bar, and the pressure range on the right side is 1-2 bar.

[0031] The weak shock wave condition is characterized by moderate temperature and pressure on the left side, and a temperature on the right side close to that on the left side, while the pressure is slightly lower. The operating condition judgment criteria are as follows: > In the pre-training dataset, the temperature range on the left side of the shock tube is 400-500K, the temperature range on the right side is 350-450K, the pressure range on the left side is 2-3 bar, and the pressure range on the right side is 1-2 bar.

[0032] The strong expansion condition is characterized by high temperature and low pressure on the left and low temperature and high pressure on the right. In the pre-training dataset, the temperature range on the left side of the shock tube is 800-1200K, the temperature range on the right side is 300-400K, the pressure range on the left side is 1-2 bar, and the pressure range on the right side is 8-10 bar.

[0033] The weak expansion condition is characterized by moderate temperature and low pressure on the left and low temperature and moderate pressure on the right. In the pre-training dataset, the temperature range on the left side of the shock tube is 400-600K, the temperature range on the right side is 300-450K, the pressure range on the left side is 1-2 bar, and the pressure range on the right side is 2-4 bar.

[0034] The contact discontinuity conditions are medium temperature and low pressure on the left and low temperature and medium pressure on the right. In the pre-training dataset, the temperature range on the left side of the shock tube is 400-600K, the temperature range on the right side is 300-450K, the pressure range on the left side is 1-2 bar, and the pressure range on the right side is 2-4 bar.

[0035] Step 2.2: Based on the equivalence relation, the engine health assessment problem is transformed into a one-dimensional Euler equation Riemann problem to be solved.

[0036] Step 3: Construct an engine state prediction model based on deep operator networks according to the operating condition classification, and train the engine state prediction model using the dataset corresponding to the operating condition classification.

[0037] Step 3.1: Construct an engine state prediction model based on deep operator networks according to the operating condition classification.

[0038] In one specific implementation, such as Figure 1 As shown, this embodiment provides a method for solving the one-dimensional shock tube problem based on a deep operator network. It deeply couples and explicitly decouples the network output from the Euler equations of the shock tube problem and the characteristics of its solutions. A multi-output engine state prediction model is constructed based on the DeepONet architecture. The engine state prediction model includes a conditional encoding branch and a spatiotemporal processing branch. The conditional encoding branch encodes the initial and boundary conditions for operating condition classification, extracting features related to the global flow state, i.e., conditional features. The spatiotemporal processing branch encodes any specified spatial and temporal coordinates and decouples them to generate independent spatiotemporal features corresponding to specific physical quantities. The final prediction of the engine state prediction model is synthesized through the tensor inner product operation of the conditional features and the spatiotemporal features. In this embodiment, the specific physical quantities refer to temperature, flow velocity, pressure, and density.

[0039] In this embodiment, the structure of the depth operator network is determined based on the classification of working conditions. Specifically, it is determined based on the number of discrete points in the initial conditions. The minimum number of physical quantities that characterize a physical field Determine the input dimension of the conditional coding branch. The input dimension of the spatiotemporal processing branch is determined based on the dimension of the spatiotemporal field. Based on the problem complexity, determine the number of hidden layers, the number of neurons in the hidden layers, and the dimension of the final output features.

[0040] More specifically, the conditional coding branch includes: The conditional coding branch consists of multiple fully connected layers, serving as a conditional encoder. Its input is a tensor. ,in The corresponding dimension of the initial condition parameter vector, This represents the number of sample batches contained in the input tensor, i.e. It is the field of real numbers Includes There are n samples, and the dimension of the initial condition parameter vector for each sample is n. The tensor is represented in the same way for subsequent tensors. This input is passed through a hidden layer containing multiple fully connected layers, and ReLU or Tanh activation functions are used to extract a shared conditional feature tensor. ,in This corresponds to the output dimension of the shared conditional feature. This feature is then passed through a linear output layer and mapped to a high-dimensional conditional feature tensor. ,in The output dimension corresponds to the high-dimensional conditional features. The technical effect of this design is that it encodes the initial conditions into a high-dimensional space and explicitly structures them into four parallel feature subspaces, laying the foundation for accurate matching with four independent physical fields and effectively improving the model's ability to represent complex initial conditions.

[0041] .

[0042] The spatiotemporal processing branch specifically includes: The spatiotemporal processing branch consists of a shared feature extraction module and a physics field decoupling output module. Its input is a tensor. ,in The number of spatiotemporal coordinate points. Corresponding to one-dimensional unsteady normalized spatial coordinates x and time coordinates The input first passes through a shared feature extraction module consisting of multiple fully connected layers to generate a shared spatiotemporal feature tensor. ,in The output dimension corresponding to the shared spatiotemporal features.

[0043] .

[0044] The physics field decoupling output module is the key innovation of this embodiment. It consists of four independent linear layers (TempLayer, Flow Layer, Pressure Layer, Density Layer) operating in parallel, sharing spatiotemporal features. They are simultaneously fed into these four linear layers and independently mapped to feature tensors corresponding to temperature, flow rate, pressure, and density, respectively. : ,in The output dimensions corresponding to the characteristics of each physical quantity.

[0045] .

[0046] Among them, subscript , , , These represent temperature, flow rate, pressure, and density, respectively.

[0047] The technical advantage of this module lies in achieving explicit decoupling guided by physics. The shared layer learns the common dynamics of the flow, while the four independent output layers allow the network to customize feature representations based on the unique evolution laws and sensitivities of each physical quantity. This overcomes the challenge of a single output head accurately fitting multiple physical quantities with vastly different dimensions and magnitudes simultaneously, significantly improving the accuracy in capturing complex structures such as shock waves and contact discontinuities, and enhancing the numerical stability of the model.

[0048] Step 3.2: Preprocess the dataset corresponding to the working condition classification.

[0049] In one specific implementation, the dataset that best matches the estimated type of the one-dimensional Euler equation Riemann problem is selected from a pre-built classification pre-training dataset library. If the actual conditions fall between two categories, one dataset can be loaded for pre-training and then fine-tuned, or two adjacent datasets can be used for mixed training to cover the characteristics of the target working condition.

[0050] Specifically, given a pre-training dataset, its samples It can be represented as ,in This is the initial condition vector. For the corresponding high-precision reference solution flow field, Indicates the first One sample, The corresponding physical parameters include temperature. Flow rate ,pressure ,density The dataset is dimensionless and normalized, and reference parameters, mean, and variance are obtained. The specific steps are as follows: Select appropriate dimensionless reference values ​​based on the units and dimensions of common initial conditions. , And calculate based on the ideal gas law and the speed of sound equation. , subscript Indicates a reference value.

[0051] , .

[0052] Dimensionless calculations were performed to obtain the dimensionless temperature. Flow rate ,pressure ,density .

[0053] .

[0054] Based on dimensionless normalization, calculate the mean values ​​of each physical quantity used for normalization. ( , , , ) and standard deviation ( , , , ).

[0055] , , , .

[0056] Perform normalization calculations to obtain the normalized temperature. Flow rate ,pressure ,density .

[0057] .

[0058] Record and solidify the reference values ​​used for dimensionless transformation, as well as the mean and standard deviation of each physical quantity used for normalization. These values ​​must be used consistently in the subsequent model inference stage to process the input conditions and the denormalization output results.

[0059] Step 3.3: Train the engine condition prediction model using the dataset corresponding to the operating condition classification.

[0060] Step 3.3.1: Design the comprehensive loss function, which includes data fitting loss. and physical constraint loss .

[0061] In one specific implementation, the form of the loss function and the values ​​of the hyperparameters are first determined. Then, a data fitting loss is selected. Loss weights of the four physical quantities , , , Select and The weight hyperparameter.

[0062] During training, a comprehensive loss function consisting of multiple parts is used for optimization: Data fitting loss The mean squared error is used to calculate the difference between the four physical fields predicted by the network and the reference solution, which is the basic part of the loss function.

[0063] .

[0064] in, This indicates the number of sample batches contained in the input tensor. The number of spatiotemporal coordinate points. The index represents the loss weight of the corresponding physical quantity. Indicates the network's predicted output value, subscript Represents the true value of a physical quantity, that is Indicates the first The sample at the th spatiotemporal coordinates Solution flow field at the location The predicted value. Data fitting loss. That is The mean square error between the predicted and actual values ​​of all nodes in a batch of samples is multiplied by the corresponding physical quantity loss weight.

[0065] To enhance the physical consistency of the solution, a batch of "residual points" are randomly sampled in the training domain. The flow field predicted by the network is substituted into the one-dimensional Euler equations, and the mean square error of the residuals is calculated as a regularization term.

[0066] .

[0067] in This represents the residual operator for the Euler equation. This loss term guides the network to satisfy the underlying physical laws, significantly improving its generalization ability and the reasonableness of the solution in sparse regions of the training data.

[0068] Ultimately, the total loss function of the network model is a weighted sum of the losses described above: .

[0069] in and These are hyperparameters. Training employs a phased strategy, initially using... The primary goal is rapid convergence, with subsequent steps gradually introducing [other technologies]. Fine-tune the model until it converges on the validation set.

[0070] Step 3.3.2: Use the dataset corresponding to the working condition classification and the comprehensive loss function to train the engine condition prediction model in stages.

[0071] In one specific implementation, the engine state prediction model is pre-trained using a preprocessed dataset. A deep operator network is trained to solve the numerical solution of the flow equation. New loss function values ​​are continuously obtained during training. Training ends when the loss function values ​​converge to a threshold, thereby achieving the solution of the one-dimensional unsteady flow equation.

[0072] Specifically, initial training is performed using a large learning rate and high data loss weights. Once the loss has initially decreased, the learning rate is reduced, and the physical loss weights are gradually increased to introduce physical constraints. The total loss value on the validation set is then monitored. Training ends when the loss function value decreases by less than a preset threshold over multiple consecutive epochs, and the error between the predicted key features of the flow field (such as shock wave velocity and location) and the verification reference solution stabilizes within an acceptable range (e.g., <2%). The result is as follows: Figure 2 As shown in the figure, the trained DeepONet model is deployed to the application scenario. Input data is acquired through sensors or other means, and then the DeepONet model can be used for fast prediction. The comparison figure between the training results of the deep operator network and the theoretical exact solution results in this embodiment is shown in the figure. Figure 3 As shown.

[0073] Step 4: Use the trained engine condition prediction model to perform a health assessment of engine components.

[0074] Step 4.1: Obtain the real-time operating status and corresponding operating data of each component of the engine, determine the operating condition classification, and obtain the initial conditions and corresponding spatiotemporal coordinates.

[0075] Step 4.2: Encode the initial conditions using the conditional coding branch and generate high-dimensional conditional features. Generate the spatiotemporal features of the multiphysics field using the spatiotemporal processing branch. Combine the high-dimensional conditional features and the spatiotemporal features for analysis to obtain the prediction results of the multiphysics field.

[0076] In one specific implementation, the conditionally coded branch is output. It is split into four conditional feature vectors in the second dimension. ,in Subsequently, these four conditional eigenvectors are compared with the feature tensors of the four corresponding physical fields output by the spatiotemporal processing branch. Perform the inner product point by point, where This operation efficiently integrates specific initial condition information with arbitrary spatiotemporal coordinate information, ultimately generating four independent tensors. This refers to the full-field distribution of temperature, flow rate, pressure, and density directly predicted by the network. The formula is as follows: For each physical field The calculation for the overall prediction is as follows: .

[0077] The technical advantages of this synthesis method are as follows: First, its mathematical form guarantees that the output solution is a continuous function of spatiotemporal coordinates, providing flow fields of arbitrary resolution and breaking through the resolution limitations of traditional discrete numerical methods. Second, it achieves "one-time training, instantaneous solution across the entire domain." After the model is trained, for any new problem, a single forward propagation is sufficient to obtain the full-field solution. Compared to traditional computational fluid dynamics methods that require iterative solutions, this significantly improves efficiency and is suitable for scenarios such as rapid parameter scanning and real-time prediction.

[0078] To better illustrate the superiority of the method in this embodiment, the flow in the engine intake manifold and the exhaust manifold will be used as examples.

[0079] (1) Solution for the flow in the engine intake manifold, the steps are as follows: S1: In this embodiment, pressure and temperature sensors are installed on the intake manifold. Initially, it is assumed that a sudden change in throttle opening or the intervention of the exhaust gas turbocharger causes an increase in pressure at the front section (left side) of the intake manifold. Therefore, the initial condition vector consists of the states on both the left and right sides: the pressure on the left side (after the throttle / turbocharger, on the high-pressure side). and temperature The pressure on the right side (before the intake manifold, low-pressure side) is directly measured by a sensor. and temperature It is also measured by the sensor at this location. The initial condition vector of the computational domain is determined. ,initial and The estimation is set to 0. Based on the pressure ratio and temperature ratio of the left and right states, the problem is predicted to be classified as a "weak shock wave" or "contact discontinuity-dominated" condition, that is, high-pressure stationary gas drives low-pressure stationary gas, which will generate a rightward shock wave, contact discontinuity and leftward expansion wave.

[0080] S2: Based on the Riemann problem type predicted in S1, select a weak shock wave dataset from the pre-built classification pre-training dataset library.

[0081] S3: Perform dimensionless and normalized transformation on the dataset in S2, and obtain the reference parameters, mean, and variance of the dataset.

[0082] S7: Select appropriate dimensionless reference values ​​based on the units and dimensions of common initial conditions. , And calculate based on the ideal gas law and the speed of sound equation. , .

[0083] , .

[0084] S8: Perform dimensionless calculations.

[0085] .

[0086] S9: Based on dimensionless normalization, calculate the mean of each physical quantity used for normalization. , , , ) and standard deviation ( , , , ).

[0087] , , , .

[0088] S10: Perform normalization calculations.

[0089] .

[0090] S11: Record and solidify the reference values ​​used for dimensionless transformation and the mean and standard deviation of each physical quantity used for normalization. These values ​​must be used consistently in the subsequent model inference stage to process the input conditions and the denormalization output results.

[0091] S4: Determine the input dimension of the network based on the input conditions, and select appropriate parameters such as the number of layers and neurons in the depth operator network.

[0092] S12: Determine the input dimension of the conditional coding branch based on the fact that there are 2 sensors and 3 minimum physical quantities representing the physical field. The value is 6. The input dimension of the spatiotemporal processing branch is determined based on the dimension of the one-dimensional unsteady spatiotemporal field. The value is 2.

[0093] S13: Determine the network structure based on the complexity of the "weak shock wave" problem. The conditional coding branch consists of 4 hidden layers, each with 128 neurons, and the final output layer is mapped to a shape of... The characteristic tensor, where The value is set to 64. The shared feature extraction module of the spatiotemporal processing branch consists of three hidden layers, each with 64 neurons; its physical field decoupling output module contains four independent linear layers, corresponding to temperature, flow rate, pressure, and density, respectively, mapping the shared features to their respective values. 3D feature tensor.

[0094] S5: Determine the form of the loss function and the values ​​of the hyperparameters.

[0095] S14: Select data fitting loss The loss weights for the four physical quantities are determined. Considering the crucial roles of pressure p and velocity u in shock wave capture and flow rate calculation, they are given higher weights. The weight vector is set as follows:

[0096] S15: Select and The weight hyperparameters are set. A phased training strategy is adopted, with the initial training phase primarily data-driven. and Physical information loss The residual terms of the one-dimensional Euler equation are used to calculate the residuals at randomly sampled points within the training domain.

[0097] S6: Pre-train the network in S4 using the dataset from S3. The deep operator network is trained to solve the numerical solution of the flow equation. New loss function values ​​are continuously obtained during training. When the loss function value converges to the threshold, the training ends, thereby realizing the solution of the one-dimensional unsteady flow equation.

[0098] S16: First use a larger learning rate (10e-3) and higher data loss weights ( , This allows for initial training and a rapid reduction in data fitting loss.

[0099] S17: After the loss initially decreases, reduce the learning rate to 10e-4 and gradually adjust the loss weights. , By introducing stronger physical constraints, the network is guided to make its predicted solutions more consistent with the physical laws described by the one-dimensional Euler equation while satisfying the data, thereby improving its generalization ability in untrained areas.

[0100] S18: Total loss value on the monitoring validation set Training ends when the loss function value decreases by less than 10e-6 over 50 consecutive epochs, and the error between the predicted key features of the flow field (such as shock wave velocity and location) and the verification reference solution stabilizes within an acceptable range (e.g., <2%).

[0101] S19: Deploy the converged DeepONet model to an embedded system of the engine electronic control unit (ECU) or a high-performance real-time simulation platform. In actual engine operation, the model will perform the following key tasks: real-time flow prediction and intake valve timing optimization under engine transient conditions.

[0102] The model receives real-time measurements from the intake manifold pressure and temperature sensors, which form the initial condition vector. Initial velocity and It is usually set to 0 or estimated based on the throttle opening. For the current engine crankshaft angle... corresponding time The model is based on the aforementioned processing. and the spatial location of concern (For example, the intake valve inlet) is used as input, and the pressure at that location in the next moment is inferred within milliseconds. and speed By combining the ideal gas law to calculate the instantaneous mass flow rate of the gas about to enter the cylinder, and integrating the instantaneous flow rate during the intake valve opening period, the in-cylinder charge of this cycle can be predicted.

[0103] Based on the pressure fluctuation waveform at the end of the intake manifold (near the intake valve) predicted by the model, the ECU can predict the optimal timing for the pressure peak (high-pressure pulse) to reach the intake valve. The ECU then dynamically adjusts the intake valve opening timing accordingly, ensuring that the intake valve opens near the pressure peak. By utilizing the dynamic effect of pressure fluctuation (inertial boost), the ECU maximizes the amount of air entering the cylinder, thereby increasing the engine's low-speed torque.

[0104] The inflation volume predicted by the model can be cross-validated with the inflation volume estimated based on the speed-density method or a thermodynamic model. If a significant deviation exists, the ECU can trigger online fine-tuning of the model parameters (such as updating initial conditions). The estimation strategy can be used to determine the air-fuel ratio and combustion stability, or it can be used as a feedback signal to correct related control parameters such as fuel injection quantity and ignition advance angle.

[0105] S20: The intake system may produce abnormal noises under transient conditions due to strong pressure fluctuations or airflow separation. This application solution can assist in the digital analysis and diagnosis of these abnormal noises.

[0106] The high-frequency pressure fluctuation data predicted by the model are analyzed in the frequency domain. The pressure pulsation amplitude at a specific frequency is extracted using Fast Fourier Transform, and this amplitude is strongly correlated with the sound pressure level of airflow noise. For specific intake system structures (such as resonant cavities and air filter housings), a "normal" threshold library of pressure pulsation amplitudes at key frequencies (such as Helmholtz resonant frequencies and pipe resonant frequencies) under different operating conditions is established. When the amplitude of a frequency component predicted by the model exceeds the threshold multiple times consecutively, the system can issue a warning of potential abnormal airflow noise risks caused by structural loosening, component aging, or design defects.

[0107] When designing a new intake system, this model can be used to quickly simulate the pressure fluctuation spectrum under different geometric schemes (such as changing the intake pipe length and diameter, adding a Helmholtz resonator, etc.). By comparing the spectrum, the design can be optimized to actively suppress pressure pulsations at specific frequencies that may cause unpleasant noise, thereby reducing intake noise in the digital prototype stage.

[0108] S21: For turbocharged engines, this model can be applied to the coordinated control of the turbocharging system and the early warning of surge boundary.

[0109] When the throttle opens rapidly, the model can predict the dynamic process of the intake manifold pressure rise. Based on this, the ECU can precisely control the actuation rate of the wastegate valve or the variable geometry turbocharger, so that the boost pressure build-up process is better matched with the intake demand (flow prediction), reducing turbo lag and avoiding the impact of intake pressure overshoot on the intercooler and piping.

[0110] By comparing the pressure ratio and flow rate relationship predicted by the model at the compressor outlet (represented as the "left" high-pressure end) with that at the throttle valve (represented as the "right" low-pressure end), the ECU can take early intervention measures (such as slightly opening the wastegate valve or adjusting valve timing to increase flow) to prevent the compressor from entering the surge zone, protect the turbocharger, and maintain smooth engine operation when the model-predicted operating point approaches the surge boundary.

[0111] (2) Solution for the flow in the engine exhaust pipe, the steps are as follows: S1: In this embodiment, pressure and temperature sensors are installed on the exhaust pipe, and the estimated in-cylinder flow rate at the moment the exhaust valve closes can be obtained through the engine electronic control unit (ECU). The initial conditions simulate the pressure fluctuation caused by the sudden closure of the exhaust valve at the end of the exhaust stroke. Pressure in the left-side state (closer to the engine cylinder side). and temperature The initial velocity is measured instantaneously by a sensor at the cylinder head outlet just before the valve closes. The pressure can be estimated by the ECU based on engine speed and displacement; the pressure in the right-side condition (downstream, such as before the three-way catalytic converter). and temperature The initial velocity was measured by the sensor at that location. The density can be approximated as 0 in the initial stage of the shutdown event. The density is calculated using the equation of state. Based on the pressure and temperature ratios of the left and right states, the problem is predicted to approximate a "strong shock wave" condition.

[0112] S2: Select a strong shock wave dataset from the pre-built classification pre-training dataset library based on the Riemann problem type predicted in S1.

[0113] S3: Perform dimensionless and normalized transformation on the dataset in S2, and obtain the reference parameters, mean, and variance of the dataset.

[0114] S7: Select appropriate dimensionless reference values ​​based on the units and dimensions of common initial conditions. , And calculate based on the ideal gas law and the speed of sound equation. , .

[0115] , .

[0116] S8: Perform dimensionless calculations.

[0117] .

[0118] S9: Based on dimensionless normalization, calculate the mean of each physical quantity used for normalization. , , , ) and standard deviation ( , , , ).

[0119] , , , , S10: Perform normalization calculations.

[0120] .

[0121] S11: Record and solidify the reference values ​​used for dimensionless transformation and the mean and standard deviation of each physical quantity used for normalization. These values ​​must be used consistently in the subsequent model inference stage to process the input conditions and the denormalization output results.

[0122] S4: Determine the input dimension of the network based on the input conditions, and select appropriate parameters such as the number of layers and neurons in the depth operator network.

[0123] S12: Determine the input dimension of the conditional coding branch based on the fact that there are 2 sensors and 3 minimum physical quantities representing the physical field. The value is 6. The input dimension of the spatiotemporal processing branch is determined based on the dimension of the one-dimensional unsteady spatiotemporal field. The value is 2.

[0124] S13: Given the increased complexity of the "strong shock wave" problem (high shock wave intensity, steep gradient), a deeper network structure was determined to improve expressive power. The conditional coding branch was designed with 5 hidden layers, each with 256 neurons. The shared feature extraction module of the spatiotemporal processing branch is set to 128. It consists of four hidden layers, each with 128 neurons. The shared feature extraction module of the spatiotemporal processing branch comprises three hidden layers, each with 64 neurons; its physical field decoupling output module contains four independent linear layers, corresponding to temperature, flow rate, pressure, and density respectively, mapping the shared features to their respective... 3D feature tensor.

[0125] S5: Determine the form of the loss function and the values ​​of the hyperparameters.

[0126] S14: Select data fitting loss The loss weights for the four physical quantities are determined. Considering the crucial roles of pressure p and velocity u in shock wave capture and flow rate calculation, and the significant impact of exhaust pipe pressure fluctuations on engine performance, they are assigned higher weights. The weight vector is set as follows: .

[0127] S15: Select and The weighting hyperparameters are defined. Considering that the physical equations in the strong shock wave region may be more difficult to implicitly satisfy due to numerical format issues, certain weights are assigned to the physical constraints in the initial stage. The initial weights are set to... and Physical information loss The residual terms of the one-dimensional Euler equation are used to calculate the residuals at randomly sampled points within the training domain.

[0128] S6: Pre-train the network in S4 using the dataset from S3. The deep operator network is trained to solve the numerical solution of the flow equation. New loss function values ​​are continuously obtained during training. When the loss function value converges to the threshold, the training ends, thereby realizing the solution of the one-dimensional unsteady flow equation.

[0129] S16: First use a larger learning rate (10e-3) and higher data loss weights ( , Initial training was conducted to enable the network to stably learn the basic structure of strong shock waves.

[0130] S17: After the loss initially decreases, reduce the learning rate to 10e-4 and gradually adjust the loss weights. , Further strengthen the constraints of physical laws to ensure that physical relationships such as the Rankine-Hugoniot condition are satisfied before and after the shock wave, thereby improving the physical rationality and smoothness of the solution.

[0131] S18: Total loss value on the monitoring validation set Training ends when the loss function value decreases by less than 10e-6 over 50 consecutive epochs, and the error between the predicted key features of the flow field (such as shock wave velocity and location) and the verification reference solution stabilizes within an acceptable range (e.g., <2%).

[0132] S19: Deploy the converged DeepONet model to the engine ECU or turbocharger control unit. During actual engine operation, the model performs the following key tasks: dynamic prediction of exhaust back pressure and real-time management of pulse turbine energy.

[0133] The model receives real-time measurements from the exhaust manifold pressure sensor (back pressure sensor) and temperature sensor, and combines these with the instantaneous exhaust valve closing velocity estimated by the ECU to form the initial condition vector. .

[0134] Regarding the location of the exhaust system, which is of concern (such as turbine inlet, exhaust manifold end) and future moments The model is based on the aforementioned processing. and Using this as input, the pressure history at that location can be inferred within milliseconds. By using the difference between the exhaust pressure waveform and the intake pressure predicted by integration, the pumping work under transient conditions can be evaluated in real time, providing a basis for optimizing the engine operating point. The ECU can then fine-tune the valve overlap angle or exhaust gas recirculation (EGR) rate to actively optimize the exhaust pressure waveform while meeting emission requirements, thereby reducing pumping losses and improving thermal efficiency.

[0135] The model predicts the pressure pulse waveform, velocity, and temperature at the turbine inlet. Based on this, the instantaneous energy (enthalpy drop) delivered to the turbine rotor by each exhaust pulse can be estimated more accurately. The ECU uses the predicted pulse energy timing to dynamically optimize the variable geometry turbine blade opening or the exhaust bypass valve opening. For example, the variable geometry turbine nozzles are slightly narrowed during the peak of the pulse energy to improve turbine speed and boost response; the exhaust bypass valve is opened earlier during the trough of the pulse to avoid turbine blockage and excessive back pressure. This feedforward control strategy maximizes the utilization of exhaust pulse energy, significantly reduces turbo lag, and improves low-speed torque.

[0136] S20: The converged DeepONet model was used for thermal shock early warning of ternary catalytic converters and particulate traps.

[0137] Combined with model-predicted exhaust velocity and temperature This allows for the calculation of transient convective heat flux density reaching the front end of the catalyst or trap. It also establishes heat flux density thresholds that post-treatment devices with different materials and structures can withstand. The system issues a warning when the peak or rate of change of the heat flux density predicted by the model exceeds the safety threshold.

[0138] Upon receiving a warning, the ECU can initiate protective intervention measures. For example, by delaying the ignition timing, enriching the air-fuel ratio (within permissible limits), or briefly adjusting valve timing, it can reduce the peak exhaust temperature or slow its rate of rise, thereby protecting the aftertreatment system from thermal cracking or sintering damage. For active regeneration of the particulate filter, the model can predict additional pressure fluctuations and temperature field changes caused by the regeneration process (such as fuel injection and combustion enhancement) to optimize the regeneration fuel injection strategy and airflow control, ensuring efficient and safe regeneration and avoiding carrier melting due to localized overheating.

[0139] S21: The converged DeepONet model is used for exhaust system noise, vibration and acoustic roughness analysis and abnormal noise source localization.

[0140] The pressure fluctuation data at key locations in the exhaust system (such as muffler inlets and pipe bends) predicted by the model are subjected to spectral analysis. The amplitude of pressure pulsations at specific frequencies (such as half-order and pipe standing wave frequencies) is correlated with the housing vibration signals measured by accelerometers. When the pressure pulsations at a certain frequency predicted by the model increase abnormally, and the vibration sensor also detects an anomaly at that frequency, the source of the abnormal noise caused by airflow excitation can be accurately located.

[0141] When designing exhaust systems, this model can be used to quickly simulate the inlet and outlet pressure fluctuation spectrum and transmission loss under different muffler structures (such as changing the expansion chamber length, insertion pipe position, and porous material layout). By comparing the spectrum and transmission loss curves, the muffler design can be optimized in the digital prototype stage, and noise radiation at target frequencies (such as specific engine order noise) can be suppressed in a targeted manner, significantly shortening the trial and error cycle based on physical prototypes.

[0142] Example 2: Embodiment 2 of the present invention provides an engine health assessment system based on one-dimensional shock tube flow, comprising: The data acquisition module is configured to acquire the operating status and corresponding operating data of various components in the engine; The equivalent module is configured to match the operating condition classification in the one-dimensional Euler equation Riemann problem based on the device's operating status and operating data. The operating condition classification includes strong shock wave, weak shock wave, strong expansion, weak expansion and contact discontinuity dominance. The model training module is configured to build an engine state prediction model based on a deep operator network according to the working condition classification, and train the engine state prediction model using the dataset corresponding to the working condition classification. The condition assessment module is configured to perform health assessments on engine components using a trained engine condition prediction model.

[0143] Example 3: Embodiment 3 of the present invention provides a computer-readable storage medium storing a computer program adapted for loading by a processor and executing the steps in the engine health assessment method based on one-dimensional shock tube flow as described in Embodiment 1 of the present invention.

[0144] Example 4: Embodiment 4 of the present invention provides a computer device, the device comprising: A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the steps in the engine health assessment method based on one-dimensional shock tube flow as described in Embodiment 1 of the present invention.

[0145] Example 5: Embodiment 5 of the present invention provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps in the engine health assessment method based on one-dimensional shock tube flow as described in Embodiment 1 of the present invention.

[0146] The steps and methods involved in Examples 2, 3, 4 and 5 above correspond to those in Example 1. For specific implementation methods, please refer to the relevant description section of Example 1.

[0147] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application. In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. The computer-readable storage medium can be any available medium that a computer can access or a data processing device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium, an optical medium, or a semiconductor medium, etc. The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An engine health assessment method based on one-dimensional shock tube flow, characterized in that, Includes the following steps: Acquire the operating status and corresponding operating data of each component in the engine; Based on the device's operating status and operating data, the operating condition classification in the one-dimensional Euler equation Riemann problem is matched. The operating condition classification includes strong shock wave, weak shock wave, strong expansion, weak expansion, and contact discontinuity dominance. An engine condition prediction model is constructed based on deep operator networks according to the operating condition classification, and the engine condition prediction model is trained using the dataset corresponding to the operating condition classification. The trained engine condition prediction model is used to perform health assessments on engine components.

2. The engine health assessment method based on one-dimensional shock tube flow as described in claim 1, characterized in that, The specific steps for classifying operating conditions in the Riemann problem based on the device's operating status and data are as follows: Establish the equivalent relationship between the operating states of various components in the engine and the one-dimensional shock tube flow process; Based on the equivalence relation, the engine health assessment problem is transformed into a one-dimensional Euler equation Riemann problem to be solved.

3. The engine health assessment method based on one-dimensional shock tube flow as described in claim 1, characterized in that, The specific conditions for classifying operating conditions are as follows: Strong shock wave conditions are characterized by high temperature and high pressure on the left and low temperature and low pressure on the right; The weak shock wave conditions are: moderate temperature and moderate pressure on the left side, and temperature on the right side is close to that on the left side, while pressure is slightly lower than that on the left side. Strong expansion conditions are characterized by high temperature and low pressure on the left and low temperature and high pressure on the right. The weak expansion condition is characterized by moderate temperature and low pressure on the left and low temperature and moderate pressure on the right. The contact interruption conditions are medium temperature and low pressure on the left and low temperature and medium pressure on the right.

4. The engine health assessment method based on one-dimensional shock tube flow as described in claim 1, characterized in that, The engine state prediction model includes a conditional encoding branch and a spatiotemporal processing branch. The conditional encoding branch is used to encode the initial and boundary conditions for the working condition classification and extract features related to the global flow state. The spatiotemporal processing branch is used to encode any specified spatial and temporal coordinates and decouple them to generate independent spatiotemporal features corresponding to specific physical quantities.

5. The engine health assessment method based on one-dimensional shock tube flow as described in claim 1, characterized in that, The specific steps for training the engine condition prediction model using the dataset corresponding to the operating condition classification are as follows: Design a comprehensive loss function, which includes data fitting loss and physical constraint loss; The engine condition prediction model is trained in stages using the dataset corresponding to the operating condition classification and the comprehensive loss function.

6. The engine health assessment method based on one-dimensional shock tube flow as described in claim 4, characterized in that, The specific steps for performing health assessments of engine components using the trained engine condition prediction model are as follows: The system acquires the real-time operating status and corresponding operating data of each component of the engine, determines the operating condition classification, and obtains the initial conditions and corresponding spatiotemporal coordinates. The initial conditions are encoded using a conditional coding branch to generate high-dimensional conditional features, and the spatiotemporal processing branch is used to generate spatiotemporal features of multiphysics. The high-dimensional conditional features and spatiotemporal features are then fused and analyzed to obtain the prediction results of multiphysics.

7. An engine health assessment system based on one-dimensional shock tube flow, characterized in that, include: The data acquisition module is configured to acquire the operating status and corresponding operating data of various components in the engine; The equivalent module is configured to match the operating condition classification in the one-dimensional Euler equation Riemann problem based on the device's operating status and operating data. The operating condition classification includes strong shock wave, weak shock wave, strong expansion, weak expansion and contact discontinuity dominance. The model training module is configured to build an engine state prediction model based on a deep operator network according to the working condition classification, and train the engine state prediction model using the dataset corresponding to the working condition classification. The condition assessment module is configured to perform health assessments on engine components using a trained engine condition prediction model.

8. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the engine health assessment method based on one-dimensional shock tube flow as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded by a processor and executed as described in any one of claims 1-6, for the engine health assessment method based on one-dimensional shock tube flow.

10. A computer device, characterized in that, include: A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the engine health assessment method based on one-dimensional shock tube flow as described in any one of claims 1-6.

Images