An aviation engine modeling method, system, and storage medium

Abstract

The present invention belongs to the technical field of aero-engine modeling, and discloses a method, a system and a storage medium for aero-engine modeling. The thermodynamic process of the engine is deeply integrated with flight data, a neural network structure is used to construct an aero-engine model, and the constructed aero-engine model is trained and verified. Then, the aero-engine model based on the deep neural network is retrained and verified based on test run data or flight data. Experimental results and analysis show that, compared with traditional thermodynamic models, the models and algorithms of the present invention have higher accuracy and efficiency, can accelerate the iterative process of the aero-engine design process, improve the response time of the aero-engine control system and the fault diagnosis ability of the monitoring process.

Description

Technical Field

[0001] The present invention belongs to the technical field of aero-engine modeling, and in particular relates to an aero-engine modeling method, a modeling system, and a computer-readable storage medium based on a thermodynamic process and a deep neural network. Background Art

[0002] At present, a high-precision and high-efficiency aero-engine performance simulation model is the basis for engine design, control, and fault diagnosis. Most of the aero-engine models widely used in the industry at present adopt the idea of the component method: starting from the engine structure and function as the basic starting point, the engine as a whole is divided into several relatively independent but interrelated sub-components, and then corresponding mathematical models are established for the working performance of each sub-component. The input-output relationships of each component are solved by iteratively solving the balance equations such as gas flow, component cross-section pressure, and input-output power satisfied during the operation of the engine. However, the accuracy of the above modeling method highly depends on the component characteristic map of the engine, and the individual differences and component degradation of the engine reduce the accuracy of the component characteristic map, resulting in low accuracy of the mechanism model.

[0003] The correction method improves the modeling accuracy by correcting the component characteristic curve and introducing factors such as individual differences and performance degradation into the model. The main correction methods are the least squares method and the Kalman filtering method.

[0004] The least squares method optimizes the component characteristic curve to minimize the error between the model output and the corresponding parameters of the engine. The Kalman filtering method first converts the engine nonlinear model into a state space model, then introduces a health factor in the state equation to measure the degradation amount of each component characteristic, and finally uses the Kalman filtering method to perform minimum variance estimation on the observed quantity.

[0005] Problems and defects existing in the prior art are as follows:

[0006] (1) The modeling accuracy highly depends on the accuracy of the component characteristics. Although the correction method can reduce the dependence on the component characteristics to a certain extent through the fitting of test run data, the optimization accuracy still depends on the given value of the initial characteristic map. Due to factors such as component characteristic test errors, manufacturing tolerances, engine assembly errors, and component degradation, there are large differences between the component characteristic map and the actual working characteristics of the components. This leads to insufficient modeling accuracy of the method based on the component characteristic map.

[0007] (2) The above methods all use the method of iteratively solving the balance equation to calculate the engine operating point. The iterative process involves the solution of multiple nonlinear equations, and the calculation time is too long to meet the requirements of engine control and monitoring. Moreover, the calculation accuracy and efficiency of the existing technology engine model on flight data are relatively low.

[0008] The difficulty in solving the above problems and defects is as follows:

[0009] (1) Traditional thermodynamic modeling methods for aeroengines are difficult to obtain accurate individual engine component characteristic maps, resulting in the modeling accuracy being difficult to meet the requirements. At the same time, the thermodynamic model is difficult to adjust in real time for the degradation process of the engine. As the characteristics of each engine component continue to degrade, the modeling accuracy of the thermodynamic model also continuously decreases.

[0010] (2) The solution of the thermodynamic model is achieved by continuously iterating to obtain the Jacobian matrix. The iterative process is cumbersome and the calculation time is too long, making it difficult to achieve real-time response to engine performance and control.

[0011] The significance of solving the above problems and defects is as follows:

[0012] (1) By improving the modeling accuracy of aeroengines, a mapping from aeroengine control parameters to cross-section monitoring parameters can be effectively established. At a certain modeling accuracy, by comparing the estimated values of the cross-section monitoring parameters by the model with the measured values of the cross-section parameters by the sensors, the working state and health level of the aeroengine can be effectively evaluated, providing strong support for its use and maintenance work.

[0013] (2) With a certain modeling accuracy and calculation efficiency ensured, the aeroengine model can be used as a real-time airborne model to provide underlying model support for the aeroengine control system and monitoring system, improving the control and monitoring level of the aeroengine and maximizing the performance of the engine. Summary of the Invention

[0014] To overcome the problems existing in the related art, the disclosed embodiments of the present invention provide an aeroengine modeling method, modeling system, and computer-readable storage medium based on thermodynamic processes and deep neural networks. The technical solutions are as follows:

[0015] An aeroengine modeling method based on thermodynamic processes and deep neural networks includes:

[0016] Deeply integrating the engine thermodynamic process with flight data, constructing an aeroengine model using a neural network structure, training and validating the constructed aeroengine model, and then retraining and validating the aeroengine model based on deep neural networks using test run data or flight data.

[0017] Specifically, it includes: simulation of the input quantities of the thermodynamic model;

[0018] Construction process of the simulation data set under the thermodynamic model;

[0019] Construction process of the aeroengine model structure based on deep neural networks;

[0020] Training and verification of an aero - engine model based on a deep neural network;

[0021] The retraining and verification process of an aero - engine model based on a deep neural network using test run data or flight data.

[0022] In one embodiment, the simulation of the input quantities of the thermodynamic model includes: Generation of aero - engine operating points: within the flight envelope and the engine operating state range, a series of aero - engine operating data points are randomly generated according to a certain distribution rule;

[0023] Conversion of operating points into inputs of the thermodynamic model: According to the requirements of the input parameters of the thermodynamic model, the aero - engine operating data points are converted into the input quantities of the aero - engine thermodynamic model.

[0024] In one embodiment, the construction of the simulation data set under the thermodynamic model includes:

[0025] Calculation of aerodynamic - thermodynamic parameters: The generated input quantities are input into the selected thermodynamic model to calculate the output quantities of the aerodynamic - thermodynamic parameters at each section;

[0026] Generation of the simulation data set: For each input quantity, a mapping from it to the output of the thermodynamic model section is established to form a training sample, and all training samples are combined to form a simulation data set;

[0027] Partition of the simulation data set: The constructed simulation data set is divided into two non - overlapping parts - a simulation training set and a simulation verification set, which are used for the training and verification of the aero - engine neural network model respectively.

[0028] In one embodiment, the construction of the aero - engine model structure based on a deep neural network includes:

[0029] Determination of the component neural network structure: Taking the inputs of each component of the aero - engine thermodynamic model as inputs and the outputs as outputs, determine the number of neural network layers and the number of hidden nodes in each layer, and construct the neural network structure of each component of the aero - engine;

[0030] Construction of the aero - engine neural network model: According to the gas path flow sequence of the aero - engine, connect the constructed neural network structures of each component head - to - tail to form the neural network model of the aero - engine;

[0031] Determination of the inputs and outputs of the aero - engine neural network model: Determine that the input of the aero - engine neural network model is the input of the aerodynamic - thermodynamic model, and the output is the aerodynamic - thermodynamic parameters at each section.

[0032] In one embodiment, the training and verification of the aero - engine model based on a deep neural network includes:

[0033] Determine the configuration parameters of the training process: Determine the training step size, batch training capacity, and weight initialization method for the training process of the aero-engine model;

[0034] Construction of the thermodynamic loss function: Use the cross-section parameters output by the aero-engine neural network model to construct the balance equation satisfied by the co-operation of each component of the aero-engine, and take the difference between both sides of the balance equation as the thermodynamic loss function for neural network training;

[0035] Construction of the similarity loss function: Adopt a certain error metric to define the similarity loss function between the model output and the simulation dataset output during the training process;

[0036] Training of the model: Combine the thermodynamic loss and the similarity loss with a certain weight as the final loss function for the neural network training process, and use the simulation training set to train the aero-engine neural network model with the configuration parameters of the training process;

[0037] Verification of the model: Verify the accuracy and efficiency of the aero-engine neural network model on the simulation verification set.

[0038] In one embodiment, the retraining and verification of the aero-engine model based on the deep neural network for the test run data or flight data include:

[0039] Processing and segmentation of the test run data or flight data set: If the test run data or flight data of the aero-engine can be obtained, perform data preprocessing on the test run data and flight data in an appropriate processing manner to form an aero-engine test run data set or flight data set, and divide the above data set into two non-overlapping parts, the training set and the verification set;

[0040] Construction of the thermodynamic loss function: Define the thermodynamic loss function for training in the manner described above;

[0041] Construction of the similarity loss function: For the specific measurement cross-section data in the data set, select the corresponding aerodynamic and thermodynamic parameters from the aero-engine neural network model, and adopt a certain error metric to define the similarity loss function between the specific measurement cross-section data and the corresponding model output during the training process;

[0042] Training of the model: Combine the thermodynamic loss and the similarity loss with a certain weight as the final loss function for the neural network training process, and use the test run or flight data training set to train the aero-engine neural network model with the configuration parameters of the training process;

[0043] Verification of the model: Verify the accuracy and efficiency of the aero-engine neural network model on the test run or flight data verification set.

[0044] Another object of the present invention is to provide a thermodynamic process and deep neural network-based aeroengine modeling system for implementing the above-mentioned aeroengine modeling method based on thermodynamic processes and deep neural networks, including:

[0045] A model input quantity module for simulating the input quantities of the thermodynamic model;

[0046] A dataset construction module for constructing a simulation dataset under the thermodynamic model;

[0047] An engine model structure construction module for constructing an aeroengine model structure based on a deep neural network;

[0048] An engine model training and verification module for training and verifying an aeroengine model based on a deep neural network;

[0049] An engine model retraining and verification module for the retraining and verification process of an aeroengine model based on a deep neural network using test run data or flight data.

[0050] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to execute the steps of the aeroengine modeling method based on thermodynamic processes and deep neural networks.

[0051] Combining all the above technical solutions, the advantages and positive effects of the present invention are as follows:

[0052] (1) Within the flight envelope, a simulation dataset that conforms to the input and output of the engine thermodynamic model is established using a numerical simulation method. By simulating the characteristics of the engine thermodynamic model through the simulation data, it is beneficial for the data-driven method to quickly learn the engine characteristics.

[0053] (2) A deep neural network model based on the thermodynamic process of an aeroengine is constructed. This model fully considers the thermodynamic process of the aeroengine working. By building the network structure, it gets rid of the dependence on the component characteristic diagrams and improves the scalability and generalization of the engine model.

[0054] (3) A neural network pre-training method based on the simulation dataset. Using the constructed simulation dataset to pre-train the constructed neural network model can enable the neural network model to quickly complete the learning of the basic characteristics of the engine.

[0055] (4) The neural network model is trained using commissioning data and flight data. Through the training process, the neural network model can adapt to the individual characteristics and degradation levels of a single engine with minor adjustments, improving the modeling accuracy of the neural network model for it. During the training process, while ensuring the engine's thermodynamic process and thermal balance, the modeling accuracy of cross-sectional parameters is improved by adjusting the neural network weights.

[0056] (5) The neural network model converts the iterative solution process of the thermodynamic model into a feedforward calculation process of the neural network, avoiding the cumbersome iterative process of the thermodynamic model and greatly improving the calculation efficiency.

[0057] (6) The present invention includes the simulation of the input quantities of the thermodynamic model; the construction process of the simulation data set under the thermodynamic model; the construction process of the aero-engine model structure based on a deep neural network; the training and verification of the aero-engine model based on a deep neural network; the retraining and verification process of the aero-engine model based on a deep neural network using commissioning data or flight data.

[0058] (7) To solve the problems of insufficient modeling accuracy and low efficiency of the current aero-engine model based on component characteristic diagrams, the present invention proposes an aero-engine fusion modeling method based on thermodynamic processes and deep neural networks. In this model, a traditional thermodynamic modeling method is used as the framework to ensure that all cross-sectional parameters of the engine can be calculated, realizing the modeling of the engine's operating state. Neural networks are used to model components such as the low-pressure compressor, high-pressure compressor, combustion chamber, high- and low-pressure turbines, bypass duct, mixer, and nozzle respectively, and an accurate component characteristic mapping network is established through a large amount of flight data. Thus, the purpose of accurately modeling the aero-engine is achieved. At the same time, the entire modeling process adopts the form of a feedforward neural network, converting the iterative solution process of the thermodynamic equilibrium equation into an offline training process of the neural network; removing the iterative form in the online calculation process and improving the calculation efficiency.

[0059] (8) Experimental results and analysis show that compared with the traditional thermodynamic model, the model and algorithm of the present invention have higher accuracy and efficiency, can accelerate the iterative process of the aero-engine design process, improve the response time of the aero-engine control system and the fault diagnosis ability of the monitoring process. Moreover, the present invention can be used as the underlying model of the aero-engine in its design, control, and monitoring processes.

[0060] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the disclosure of the present invention. Brief Description of the Drawings

[0061] The accompanying drawings here are incorporated into the specification and form a part of this specification, showing embodiments in line with the disclosure of the present invention, and are used together with the specification to explain the principles disclosed by the present invention.

[0062] Figure 1 It is a flowchart of the aero-engine modeling method based on thermodynamic processes and deep neural networks provided by an embodiment of the present invention.

[0063] Figure 2 It is the cross-section numbering of the aero-engine provided by an embodiment of the present invention.

[0064] Figure 3 It is the component neural network structure diagram of step S31 provided by an embodiment of the present invention.

[0065] Figure 4 It is the overall schematic diagram of the aero-engine neural network of step S32 provided by an embodiment of the present invention.

[0066] Figure 5 It is the change curve of the maximum error of the training process T5 of step S44 provided by an embodiment of the present invention.

[0067] Figure 6 It is the verification set error diagram of the aero-engine neural network model for each cross-section parameter of step S45 provided by an embodiment of the present invention. Detailed implementation manners

[0068] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will describe the detailed implementation manners of the present invention in conjunction with the accompanying drawings. Many specific details are set forth in the following description in order to fully understand the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific implementations disclosed below.

[0069] It should be noted that when an element is referred to as being "fixed to" another element, it can be directly on the other element or there can also be an intermediate element. When an element is considered to be "connected" to another element, it can be directly connected to the other element or there may be an intermediate element at the same time. The terms "vertical", "horizontal", "left", "right" and similar expressions used in the present invention are only for the purpose of illustration and do not represent the only implementation manners.

[0070] Unless otherwise defined, all technical and scientific terms used in this invention have the same meanings as those commonly understood by those skilled in the technical field to which this invention belongs. The terms used in the description of this invention are only for the purpose of describing specific embodiments and are not intended to limit this invention. The term "and / or" used in this invention includes any and all combinations of one or more of the related listed items.

[0071] As Figure 1 shown, the method for modeling an aero-engine based on a thermodynamic process and a deep neural network provided by an embodiment of this invention includes:

[0072] S1: Combining the actual flight envelope of the aero-engine, using the Monte Carlo method to generate a series of input quantities for the thermodynamic model;

[0073] S2: Inputting the input quantities in S1 into the existing aero-engine aerodynamic thermodynamic model, calculating the aerodynamic thermodynamic parameters of each component section, and forming a series of section output quantities corresponding to the input quantities;

[0074] S3: Building an aero-engine model based on a deep neural network;

[0075] S4: Using the input-output quantities generated in S2 to train the aero-engine model based on a deep neural network built in S3;

[0076] S5: If there is aero-engine test run data or flight data available, further using the test run data or flight data to train the aero-engine model based on a deep neural network built in S3.

[0077] This invention also provides an aero-engine modeling system based on a thermodynamic process and a deep neural network for implementing the method for modeling an aero-engine based on a thermodynamic process and a deep neural network, including:

[0078] A model input quantity module for simulating the input quantities of the thermodynamic model;

[0079] A dataset construction module for constructing a simulation dataset under the thermodynamic model;

[0080] An engine model structure construction module for constructing the structure of an aero-engine model based on a deep neural network;

[0081] An engine model training and verification module for training and verifying the aero-engine model based on a deep neural network;

[0082] An engine model retraining and verification module for the retraining and verification process of the aero-engine model based on a deep neural network using the test run data or flight data.

[0083] The technical solution of the present invention will be further described below in combination with the detailed content of each step.

[0084] In an embodiment disclosed by the present invention, the above step S1 includes the following specific steps:

[0085] S11: Determine the flight envelope of the aeroengine and the engine operating state range. Taking the flight envelope and engine operating state range of a certain type of engine as an example, as shown in Table 1;

[0086] S12: Randomly generate a series of aeroengine flight data points within the flight envelope and the engine operating state range. According to at least one embodiment disclosed by the present invention, within the flight envelope, 16384 data points are randomly generated according to a uniform distribution.

[0087] S13: Convert the data points generated in S12 into the input quantities of the aeroengine thermodynamic model. According to at least one embodiment disclosed by the present invention, the input quantities of the aeroengine thermodynamic model are the total temperature at the engine inlet, the total pressure at the inlet, the ambient atmospheric pressure, and the high-pressure rotor speed. Using a general calculation method for the total pressure recovery coefficient of the inlet, the flight altitude and Mach number of the aeroengine in S12 are converted into the total temperature and total pressure at the inlet, and the simulation input data set Ω is obtained. in :

[0088] Ω in ={ω i =(T 2i ,P 2i ,Pamb 2i ,N 2i ,)|i = 1, 2, …, 16384}

[0089] In an embodiment disclosed by the present invention, the above step S2 includes the following specific steps:

[0090] S21: Select a set of mature aeroengine thermodynamic models. According to at least one embodiment disclosed by the present invention, select the dual-rotor mixed-exhaust aeroengine thermodynamic model of the Gasturb aeroengine simulation software;

[0091] S22: Input the input quantities generated in S1 into the thermodynamic model selected in S21, and calculate the output quantities of the aerodynamic and thermodynamic parameters of each section. Input the 15360 aeroengine operating points given in S13 into the thermodynamic model selected in S21, and calculate 15360 aerodynamic and thermodynamic parameters of each section of the aeroengine. mainly including: the total temperature of the section, the total pressure of the section, the static temperature of the section, the static pressure of the section, the Mach number of the section, and the mass flow of the section. The section division is as Figure 2 shown.

[0092] S23: For each input quantity in S1, establish its mapping to the output quantity in S22 to form a training sample. Combine all the training samples to form a simulation data set. Correlate the cross-sectional parameters obtained in S22 with their inputs to form the simulation data set Ω:

[0093] Ω = {ω i = (T 2i , P 2i , Pamb 2i , N 2i , sttn i ) | i = 1, 2, …, 15360}

[0094] wherein, sttn i represents the cross-sectional parameters of the thermodynamic model corresponding to the i-th simulation input.

[0095] S24: Divide the simulation data set constructed in S23 into two non-overlapping parts - a simulation training set and a simulation validation set, which are used for training and validating the aero-engine neural network model respectively.

[0096] According to at least one embodiment disclosed in the present invention, randomly divide the simulation data set Ω into a simulation training set Ω tr and a simulation validation set Ω val , wherein, there are 12288 data in the simulation training set Ω tr and 3072 data in the simulation validation set Ω val .

[0097] In one embodiment disclosed in the present invention, the above step S3 includes the following specific steps:

[0098] S31: Taking the inputs of each component of the aero-engine thermodynamic model as inputs and the outputs as outputs, determine the number of neural network layers and the number of hidden nodes in each layer, and construct the neural network structure of each component of the aero-engine. According to at least one embodiment disclosed in the present invention, the determined neural network structures of each component are as Figure 3 shown. Among them, the network inputs are the total temperature, total pressure, Mach number, air flow rate at the component inlet and other component-related input parameters (such as the low-pressure rotational speed needs to be input for the low-pressure compressor, the high-pressure rotational speed needs to be input for the high-pressure compressor, the fuel flow rate needs to be input for the combustor, the ambient pressure needs to be input for the nozzle, etc.); the number of neural network layers is 4, and the number of hidden nodes in each layer is xx.

[0099] S32: Connect the neural network structures of each component constructed in S31 head to tail in the gas path flow order of the aero-engine to form the neural network model of the aero-engine. The finally formed neural network model of the aero-engine is as Figure 4 shown.

[0100] S33: Determine that the input of the neural network model of the aero-engine is the input of the aerodynamic thermodynamics model, and the output is the aerodynamic thermodynamics parameters of each section. As can be seen from Figure 4 , the input of the aero-engine neural network model is: total temperature, total pressure, flow rate, ambient pressure, high-pressure speed, low-pressure speed, and fuel flow rate at the engine inlet; the output is: total temperature, total pressure, Mach number, and air / gas flow rate at each section.

[0101] In an embodiment disclosed by the present invention, the above step S4 includes the following specific steps:

[0102] S41: Determine the training step size, batch training capacity, and weight initialization method for the training process of the aero-engine model. According to at least one embodiment disclosed by the present invention, the training step size is 1e-3, and the training step size decays to 10% of the original every 100 generations; the number of training generations is 500 generations; the batch training capacity is 1024 training samples; the weight initialization uses a Gaussian distribution initialization method with a mean of 0 and a variance of 0.01.

[0103] S42: Use the section parameters output by the aero-engine neural network model constructed in S3 to construct the balance equation satisfied by the co-working of each component of the aero-engine, and use the difference between both sides of the balance equation as the thermodynamic loss function for neural network training. According to at least one embodiment disclosed by the present invention, for the co-working balance equation of a dual-rotor mixed-flow turbofan engine, the thermodynamic loss function for neural network training obtained is:

[0104] loss d = loss w + loss p

[0105] loss w = ∑(W in - W ex - W cl )

[0106] loss p = ∑(P t - P c - P ex )

[0107] In the formula, loss w is the flow balance loss: W in is the inlet flow of the network component unit, W ex is the outlet flow of the network component unit, W cl is the cooling and other air extraction flow rate. When the engine operates stably, the inlet and outlet flows of the component should be balanced, that is, W in - W ex - W cl = 0. Therefore, lossw Measure the approximation degree of the network output to the flow balance; loss p Is the power balance loss, P t Is the turbine output power calculated by the network, P c Is the power required by the compressor calculated by the network, P ex Is the power extraction. When the engine operates stably, the power balance on the same rotor, that is, P t -P c -P ex = 0. Therefore, loss can be used p Measure the approximation degree of the network to the power balance. The flow rate of each section and the power of the high and low pressure rotors in the above formula can be represented by the aerodynamic and thermodynamic parameters of each section output by the aeroengine neural network model. Therefore, the above formula can be transformed into a thermodynamic loss function regarding the output of the aeroengine neural network model.

[0108] S43: Adopt a certain error metric to define the similarity loss function between the model output and the output of the simulation data set generated by S2 during the training process. According to at least one embodiment disclosed in the present invention, the mean square error is adopted as the similarity loss function:

[0109]

[0110] Among them, loss m Is the similarity loss, y i Is the target value vector of the parameters of each section in the training sample, Is the output estimate value vector of the parameters of each section of the aeroengine neural network model.

[0111] S44: Combine the thermodynamic loss defined in S42 and the similarity loss defined in S43 with a certain weight as the loss function for the final neural network training process. Use the simulation training set constructed by S24 and the training process parameters given by S41 to train the aeroengine neural network model constructed by S32. According to at least one embodiment disclosed in the present invention, the final loss function is:

[0112] loss = loss m + loss d

[0113] On the simulation training set constructed by S24, use the Adam optimization method to train the aeroengine neural network model. During the training process, the variation of the maximum error of T5 with the number of training generations is as Figure 5 Shown, where the maximum error of T5 in each generation is defined as:

[0114]

[0115] S45: Verify the accuracy and efficiency of the aero-engine neural network model on the simulation verification set constructed in S24. The parameter errors of each cross-section obtained are as Figure 6 shown. It can be seen that the aero-engine neural network model has a good approximation result for the parameters of each cross-section; the comparison of the calculation efficiency between the neural network model and the thermodynamic model is shown in Table 1, and the calculation time is obtained on an i5-11300H (3.10 GHz 3.11 GHz) CPU. It can be seen that the calculation efficiency of the neural network model is much higher than that of the thermodynamic model widely used at present.

[0116] Table 1 Comparison of calculation efficiency of various engine models

[0117]

[0118] In an embodiment disclosed by the present invention, the above step S5 includes the following specific steps:

[0119] S51: If the test run data or flight data of the aero-engine can be obtained, perform data preprocessing on the test run data and flight data in an appropriate processing manner to form an aero-engine test run data set or flight data set, and divide the above data set into two non-overlapping parts, a training set and a verification set. According to at least one embodiment disclosed by the present invention, the actual flight data of this type of aero-engine for one year is obtained. 26,970 steady-state points are selected by ensuring that the change amplitude of the high-pressure rotational speed does not exceed 1% within 5 seconds, including flight altitude, Mach number, inlet total temperature, total pressure, ambient pressure, and turbine outlet temperature. The above steady-state points are randomly divided into two parts, one part is a flight data training set containing 2,000 steady-state points, and the other part is a flight data verification set containing 6,970 steady-state points.

[0120] S52: Define the thermodynamic loss function for training in the same way as in S42;

[0121] S53: For the specific measurement cross-section data in the data set, select the corresponding aerodynamic and thermodynamic parameters from the aero-engine neural network model, and define the similarity loss function using the method in S43. Since there is only one cross-section parameter, the turbine outlet temperature, in this flight data, the similarity loss function is defined as:

[0122]

[0123] In the formula, loss m is the similarity loss, T5 i is the target value of the turbine outlet temperature in the training sample, is the estimated value of the turbine outlet temperature output by the aero-engine neural network model.

[0124] S54: Train the aero-engine neural network model in the same way as S44. During the training process, the training step size becomes 1e-5 and decays to 80% of the original step size every 100 generations.

[0125] S55: Verify the model accuracy and efficiency in the same way as S45. The comparison of the model accuracy and efficiency with the calculation results of the current mainstream aerodynamic thermodynamics model on the flight data validation set given in S51 is shown in Table 2. Among them, the model accuracy is measured by the mean absolute error and the maximum absolute error, and the calculation time is obtained on an i5-11300H (3.10GHz 3.11GHz) CPU:

[0126]

[0127]

[0128] It can be seen that the aero-engine neural network model has a greater improvement than the current mainstream aerodynamic thermodynamics model in terms of both calculation accuracy and efficiency, verifying the effectiveness of the model.

[0129] Table 2 Comparison of calculation accuracy and efficiency of various engine models on flight data

[0130]

[0131] After considering the specification and practice disclosed herein, those skilled in the art will readily conceive of other embodiments of the present disclosure. This application is intended to cover any variations, uses, or adaptations of the present disclosure, which follow the general principles of the present disclosure and include the common general knowledge or conventional technical means in the technical field not disclosed in the present disclosure. The specification and examples are only illustrative, and the true scope and spirit of the present disclosure are pointed out by the appended claims.

[0132] It should be understood that the present disclosure is not limited to the exact structures described above and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of the present disclosure should be limited by the appended claims.

Claims

1. A method for modeling an aeroengine based on thermodynamic processes and deep neural networks, characterized in that, the method for modeling an aeroengine based on thermodynamic processes and deep neural networks includes: Deeply integrating the engine thermodynamic process with flight data, constructing an aeroengine model using a neural network structure, training and validating the constructed aeroengine model, and then retraining and validating the aeroengine model based on deep neural networks using test run data or flight data; The method for modeling an aeroengine based on thermodynamic processes and deep neural networks specifically includes: S1: Combining the actual flight envelope of the aeroengine, using the Monte Carlo method to generate a series of input quantities for the thermodynamic model; S2: Inputting the input quantities in S1 into the existing aeroengine aerothermodynamic model, calculating the aerothermodynamic parameters of each component section, and forming multiple section output quantities corresponding to the input quantities; S3: Building an aeroengine model based on a deep neural network; S4: Using the input-output quantities generated in S2 to train the aeroengine model based on a deep neural network built in S3; S5: If there is aeroengine test run data or flight data, further using the test run data or flight data to train the aeroengine model based on a deep neural network built in S3; The specific steps of S3 are as follows: S31: Taking the inputs of each component of the aeroengine thermodynamic model as inputs and the outputs as outputs, determining the number of neural network layers and the number of hidden nodes in each layer, and constructing the neural network structure of each component of the aeroengine; among them, the network input is the total temperature, total pressure, Mach number, air flow rate at the component inlet and other component-related input parameters; the number of neural network layers is 4, and the number of hidden nodes in each layer is xx; S32: Connecting the neural network structures of each component built in S31 head to tail in the gas path flow order of the aeroengine to form the neural network model of the aeroengine; the finally formed neural network model of the aeroengine; S33: Determining that the input of the neural network model of the aeroengine is the input of the aerothermodynamic model, and the output is the aerothermodynamic parameters of each section; the input of the neural network model of the aeroengine is: total temperature, total pressure, flow rate, ambient pressure, high-pressure rotational speed, low-pressure rotational speed, fuel flow rate at the engine inlet; the output is: total temperature, total pressure, Mach number, air / gas flow rate at each section; The specific steps of S4 are as follows: S41: Determining the training step size, batch training capacity, and weight initialization method during the training process of the aeroengine model; S42: Using the section parameters output by the aeroengine neural network model constructed in S3 to construct the balance equation satisfied by the co-operation of each component of the aeroengine, and taking the difference between both sides of the balance equation as the thermodynamic loss function for neural network training; for the co-operation balance equation of a dual-rotor mixed-exhaust turbofan engine, the obtained thermodynamic loss function for neural network training is: loss d = loss w + loss p loss w = ∑(w in - w ex - w cl ) loss p = ∑(P t - P c - P ex ) where loss w is the flow balance loss: W in is the inlet flow of the network component unit, W ex is the outlet flow of the network component unit, W cl is the cooling and other air extraction flow. When the engine is operating stably, the inlet and outlet flows of the component should be balanced, i.e., W in -W ex -W cl = 0. Use loss w to measure the approximation degree of the network output to the flow balance; loss p is the power balance loss, P t is the turbine output power calculated by the network, P c is the compressor required power calculated by the network, P ex is the power extraction. When the engine is operating stably, the power on the same rotor is balanced, i.e., P t -P c -P ex = 0. Use loss p to measure the approximation degree of the network to the power balance; the flow rates of each section and the powers of the high and low pressure rotors in the above formula are all represented by the aerodynamic and thermodynamic parameters of each section output by the aeroengine neural network model; S43: Using a certain error metric to define the similarity loss function between the model output and the output of the simulation data set generated in S2 during the training process; Using the mean square error as the similarity loss function: Among them, loss m is the similarity loss, y i is the target value vector of each cross-section parameter in the training samples, is the output estimated value vector of each cross-section parameter of the aero-engine neural network model; S44: Combine the thermodynamic loss defined in S42 and the similarity loss defined in S43 with a certain weight as the loss function for the final neural network training process. Using the simulation training set constructed in S24 and the training process parameters given in S41, train the aero-engine neural network model constructed in S32; the final loss function is: loss = loss m + loss d ; On the simulation training set constructed in S24, use the Adam optimization method to train the aero-engine neural network model; during the training process, the maximum error of each generation T5 is defined as: S45: Verify the accuracy and efficiency of the aero-engine neural network model on the simulation validation set constructed in S24; The specific steps of S5 are as follows: S51: Perform data preprocessing on the obtained aero-engine test run data or flight data to form an aero-engine test run data set or flight data set, and divide the above data sets into non-overlapping training sets and validation sets; the training set contains flight data of 2000 steady-state points, and the validation set contains flight data of 6970 steady-state points; S52: Define the thermodynamic loss function for training in the same way as S42; S53: For the specific measurement cross-section data in the data set, select the corresponding aerodynamic and thermodynamic parameters from the aero-engine neural network model, and define the similarity loss function using the S43 method; there is only the turbine outlet temperature cross-section parameter in this flight data, and the similarity loss function is defined as: where loss m is the similarity loss, T5 i is the target value of the temperature after the turbine in the training samples, is the estimated value of the temperature after the turbine output by the aero-engine neural network model; S54: Train the aero-engine neural network model in the same way as S44; S55: Verify the model accuracy and efficiency in the same way as S45; the model accuracy is measured by the mean absolute error and the maximum absolute error:

2. The aero-engine modeling method based on thermodynamic processes and deep neural networks according to claim 1, characterized in that The specific steps of S1 are as follows: S11: Determine the flight envelope of the aero-engine and the engine operating state interval; S12: Randomly generate multiple aero-engine flight data points within the flight envelope and the engine operating state interval; S13: Convert the data points generated in S12 into the input variables of the aero-engine thermodynamic model; the input variables of the aero-engine thermodynamic model are the total temperature at the engine inlet, the total pressure at the inlet, the ambient atmospheric pressure, and the high-pressure rotor speed; using the general calculation method for the total pressure recovery coefficient of the inlet, convert the flight altitude and Mach number of the aero-engine in S12 into the total temperature and total pressure at the inlet to obtain the simulation input data set Ω in : Ω in = {ω i = (T 2i , P 2i , Pamb 2i , N 2i ) | i = 1, 2, …, 16384}.

3. The aero-engine modeling method based on thermodynamic processes and deep neural networks according to claim 1, characterized in that The specific steps of S2 are as follows: S21: Select the aero-engine thermodynamic model; select the dual-rotor mixed-exhaust aero-engine thermodynamic model of the Gasturb aero-engine simulation software; S22: Input the input quantities generated in S1 into the thermodynamic model selected in S21, and calculate the output quantities of the aerodynamic and thermodynamic parameters of each cross-section; input the 15360 aero-engine operating points given by the obtained simulation input data set into the thermodynamic model selected in S21, and calculate the aerodynamic and thermodynamic parameters of 15360 aero-engine cross-sections; including: cross-section total temperature, cross-section total pressure, cross-section static temperature, cross-section static pressure, cross-section Mach number, cross-section flow rate; S23: Establish a mapping from each input quantity in S1 to the output quantity in S22 to form a training sample, and combine all the training samples to form a simulation data set; correspond the cross-section parameters obtained in S22 with the input to form a simulation data set Ω: Ω = {ω i = (T 2i , P 2i , Pamb 2i , N 2i , sttn i ) | i = 1, 2, …, 15360} wherein, sttn i represents the cross-sectional parameters of each thermodynamic model corresponding to the i-th simulation input; S24: Divide the simulation data set constructed in S23 into non - overlapping simulation training set Ω tr and simulation verification set Ω val , which are respectively used for the training and verification of the aero - engine neural network model; Among them, the simulation training set Ω tr contains 12,288 pieces of data, and the simulation verification set Ω val contains 3,072 pieces of data.

4. An aero-engine modeling system based on thermodynamic processes and deep neural networks for implementing the aero-engine modeling method based on thermodynamic processes and deep neural networks according to any one of claims 1 to 3, characterized in that, the aero-engine modeling system based on thermodynamic processes and deep neural networks includes: a model input quantity module for simulating the input quantities of the thermodynamic model; a data set construction module for constructing a simulation data set under the thermodynamic model; an engine model structure construction module for constructing the aero-engine model structure based on deep neural networks; an engine model training and verification module for training and verifying the aero-engine model based on deep neural networks; an engine model retraining and verification module for the retraining and verification process of the aero-engine model based on deep neural networks using test run data or flight data.

5. A computer-readable storage medium storing a computer program, which when executed by a processor causes the processor to execute the steps of the aero-engine modeling method based on thermodynamic processes and deep neural networks according to any one of claims 1 to 3.

6. An aero-engine, characterized in that, the aero-engine executes the aero-engine modeling method based on thermodynamic processes and deep neural networks according to any one of claims 1 to 3.

Images