Convolutional neural network-based turbofan engine individual deviation bias correction method

Abstract

The application discloses a turbofan engine individual difference deviation correction method based on a convolutional neural network, and comprises the following steps: based on a nominal turbofan engine component level model, an individual difference turbofan engine component level model is established, and an individual difference deviation correction training data set is generated; a turbofan engine full envelope individual difference deviation correction precision evaluation index is designed, and a turbofan engine individual difference deviation correction model network structure and parameters are optimized; and according to the obtained individual difference deviation correction model, the output of the on-board model of the nominal turbofan engine is corrected. The application generates a training data set based on a thermodynamic component level model, solves the problem of difficult acquisition of real machine data, and the established individual difference deviation correction model can effectively reduce the data dimension, extract feature information and improve the dynamic precision of individual difference deviation correction in the full envelope range.

Description

Technical Field

[0001] This invention belongs to the field of aero-engine control technology, specifically relating to a method for correcting individual differences in turbofan engines based on convolutional neural networks. Background Technology

[0002] Aero engines are highly complex and precise thermodynamic machines; even minute manufacturing errors can alter their thermodynamic properties. Furthermore, the extremely harsh operating environment makes aero engine components prone to deformation and damage, inevitably resulting in individual differences between each aero engine. These individual differences stem primarily from two factors: first, manufacturing defects leading to dimensional deviations and assembly errors; and second, environmental factors and mechanical fatigue during service causing performance degradation of engine components.

[0003] Individual differences not only significantly impact engine performance but also lead to control quality degradation in offline-designed controllers during online operation. To avoid this degradation, controller retraining is necessary. However, the training process of intelligent controllers is highly dependent on the onboard model, with the agent's input consisting of multiple engine states at various time points. Individual differences cause significant deviations in these states, greatly influencing the final decision-making. The most accurate method for correcting these deviations is remodeling based on real data; however, component-level modeling is extremely labor-intensive and typically requires a long period, while individual differences change very rapidly, making component-level modeling for individual deviations impractical. If an onboard model for individual deviation correction could be established based on existing nominal engine component-level models, it would greatly facilitate the online retraining of intelligent controllers.

[0004] To establish a widely applicable and highly accurate individual difference correction model, it is first necessary to study the mechanism of individual differences in aero-engines, analyze the main reasons for individual differences in various components, and explore the impact of various individual differences on the performance of each component. The impact of individual differences is attributed to specific parameters in the component-level model, and simulation of various differences can be achieved by modifying these parameters. Then, under specific individual differences, an individual difference correction method based on convolutional neural networks (CNNs) is proposed. CNNs are a type of feedforward neural network with a deep structure that includes convolutional computation. They contain a feature extractor composed of convolutional and pooling layers, possessing strong feature extraction capabilities. By extracting difference features from multi-dimensional data using CNNs and combining them with the nominal engine component-level model, real-time difference correction of airborne engines can be achieved. Summary of the Invention

[0005] Objective: To address the issue of control quality degradation in offline-designed controllers during online operation due to individual variations in aero-engines, this invention proposes a method for correcting individual variation deviations in turbofan engines based on convolutional neural networks. The aim is to collect engine data under various individual variations using a method for generating individual variation deviation label datasets, thereby increasing dataset diversity; to improve the efficiency of parameter adjustment and structural optimization during the training process of the individual variation deviation correction model by establishing an evaluation index for the accuracy of individual variation deviation correction within the full envelope; and finally, to use the trained individual variation deviation correction model to correct the nominal engine airborne model.

[0006] Technical solution: To achieve the above objectives, the technical solution adopted by this invention is as follows:

[0007] A method for correcting individual differences in turbofan engines based on convolutional neural networks includes the following steps:

[0008] Step 1) Based on the nominal turbofan engine component-level model, establish an individual-difference turbofan engine component-level model and generate an individual-difference deviation correction training dataset.

[0009] Step 2) Design an evaluation index for the accuracy of individual difference deviation correction of the full envelope of turbofan engine, and optimize the network structure and parameters of the individual difference deviation correction model of turbofan engine;

[0010] Step 3) Correct the output of the nominal turbofan engine airborne model based on the obtained individual difference deviation correction model.

[0011] Furthermore, the specific steps of the method for generating the training dataset to correct for individual differences in step 1) are as follows:

[0012] Step 1.1) Select a nominal component-level model of the turbofan engine to simulate the nominal engine; by adjusting the component efficiency coefficient and flow coefficient of the nominal component-level model of the turbofan engine, simulate the engine with individual differences.

[0013] Step 1.2) Select the nominal engine measurable section state parameters, engine operating atmospheric environment parameters, engine control variables, and parameters characterizing safe operating conditions to form the engine state x at time t. t The input X of the deviation correction network is formed by selecting the engine's states at n time points. t =[x t-n ,x t-n-1 ,...,x t-1 ,x t ];

[0014] Step 1.3) Select the state quantity error e of the nominal engine and the individually differentiated engine under the same atmospheric environment and input conditions.s,t Y is the output label of the bias correction network at time t. t =[e s,t ]; X is determined by input and output data. t =[x t-n ,x t-n-1 ,...,x t-1 ,x t →Y t =[e s,t The basic structure of the individual difference bias correction network;

[0015] Step 1.4): Randomly select i operating points within the full envelope to control the nominal engine model and the individual difference engine model, and collect training data during the control process according to the input-output structure in steps 1.2)-1.3).

[0016] Furthermore, the specific steps for training data acquisition in step 1.4) are as follows:

[0017] Step 1.4.1): Randomly select the operating point [H, Ma] and the initial fuel flow rate W within the entire envelope range. fb,ini Tail nozzle area A 8,ini Initialize the nominal engine model and the individual engine model; randomly generate the reference thrust command r, and pass it through the PI controller and a set of random PI parameters k. p and k i To control the nominal engine, set the control cycle and the upper limit of the control task time;

[0018] Step 1.4.2): During the control process, the nominal control quantity W of the engine at each moment, calculated by the PI controller, is... fb,t and A 8,t As an input to the individual engine, the control process is immediately stopped when the nominal engine enters a steady state;

[0019] In step 1.4.3), during the control process of steps 1.4.1)-1.4.2), the nominal engine state x is recorded based on the input data structure described in steps 1.2)-1.3). t The input label data X of the bias correction network is obtained. t Based on the output structure described in steps 1.2)-1.3), the state and unmeasurable thrust errors of the nominal engine and the individually differentiated engine under the same atmospheric conditions and input are calculated, and the output label data Y of the deviation correction network is obtained. t Record X at each moment during the control process. t →Y t The data pairs are stored in the training dataset;

[0020] Step 1.4.4), repeat steps 1.4.1)-1.4.3) i times to complete the collection of training data for the individual bias correction model.

[0021] Furthermore, the specific steps for designing and optimizing the individual difference deviation correction model for the turbofan engine in step 2) are as follows:

[0022] Step 2.1): Based on experience and the relationship between the parameters, select the initial number of input and output channels, convolution kernel and stride, input and output dimensions and number of convolutional layers for the convolutional neural network, and use average pooling layers for pooling.

[0023] Step 2.2) The Adam optimizer is used to train the dataset based on individual difference bias labels in batches at a given learning rate;

[0024] Step 2.3) Based on the full envelope individual difference correction accuracy evaluation index, evaluate the correction accuracy of the trained individual difference bias correction model under the full envelope. Based on the evaluation results, adjust the number of input and output channels, convolution kernel and stride, input and output dimensions, number of convolutional layers and learning rate in steps 2.1) and 2.2).

[0025] Furthermore, the specific steps for estimating the accuracy of the turbofan engine's full envelope individual difference deviation correction in step 2.3) include:

[0026] Step 2.3.1) Within the entire envelope, divide the area into p points at height intervals and l points at Mach number intervals, forming p×l working points; at each working point, randomly generate k sets of 1×7 test arrays. composition This serves as the initial state and reference command for the engine; where W fb,j Let A represent the initial state of the fuel in the j-th test group. 8,j This represents the initial state of the tail nozzle area in the j-th test group. This represents the proportional gain of the fuel flow PI controller in the j-th test group. This represents the proportional coefficient of the PI controller for the tail nozzle area in the j-th test group. This represents the integral coefficient of the fuel flow PI controller in the j-th test group. R represents the integral coefficient of the PI controller for the nozzle area in the j-th test group. j This indicates the command for the thrust in the j-th test group;

[0027] Step 2.3.2): At each operating point, initialize the nominal engine and the individual engine based on the test array, initialize the PI controller based on the set PI parameters, perform thrust differential feedback control on the nominal engine, and directly use the control quantity calculated by the nominal engine PI controller as the input of the individual engine to control the individual engine, and set the control cycle and control task time limit.

[0028] Step 2.3.3) Record the nominal engine and individual engine state quantity error e at each moment during the test. s,t Calculate the average error e of each state variable in k sets of tests at a single operating point. mean and maximum instantaneous error e max and the overall average error of all state variables at all operating points within the full envelope. and maximum instantaneous error This is used to evaluate the correction accuracy of the individual difference deviation correction model for turbofan engines within the full envelope.

[0029] Furthermore, the specific steps for the airborne model correction in step 3) are as follows:

[0030] Step 3.1): In actual use, the trained individual deviation model is uploaded. The atmospheric environment where the engine is located and the control quantities given by the airborne controller are used as inputs to characterize the airborne component-level model. After calculation, the airborne component-level model obtains the measurable cross-sectional state parameters of the engine and the parameters characterizing the safe operating state. The atmospheric environment parameters and the engine control quantities are combined to obtain the engine state x at time t. t ;

[0031] Step 3.2): Select the current data set and the data set from the previous n time points of the engine to form the input X of the individual difference deviation correction model at the current time t. t =[x t-n ,x t-n-1 ,...,x t-1 ,x t ];

[0032] Step 3.3): After obtaining the input data set, the individual difference deviation correction model calculates the error e of the state variables to be used in the control process. s,t This is used to correct the nominal engine condition.

[0033] S o,t =S d,t +e s,t

[0034] In the formula, S o,t S represents the corrected state variable. d,t This represents the state quantity of the nominal engine output.

[0035] Beneficial Effects: The present invention provides a method for correcting individual differences in turbofan engines based on convolutional neural networks. Compared with existing technologies, the above technical solution has the following technical advantages:

[0036] (1) The individual difference deviation correction training dataset generation method proposed in this invention simulates the impact of individual differences through a component-level model, which solves the problem of difficulty in obtaining real machine data during the design process;

[0037] (2) The individual difference deviation correction model established by the present invention can effectively reduce the data dimension and automatically extract feature information, thus solving the problem of difficulty in extracting key individual difference features due to data redundancy.

[0038] The established individual difference bias correction model can effectively reduce data dimensionality, extract feature information, and improve the dynamic accuracy of individual difference bias correction within the entire envelope range.

[0039] (3) The present invention proposes a method for correcting individual differences in turbofan engines based on convolutional neural networks, which can correct airborne model deviations with high precision within the entire envelope range. This method helps to solve the model accuracy problem of retraining airborne intelligent controllers and is of great significance for improving the overall performance of advanced turbofan engines. Attached Figure Description

[0040] Figure 1 This is the design process for an airborne model individual difference bias correction method based on convolutional neural networks.

[0041] Figure 2 This is a schematic diagram of the individual difference bias correction model.

[0042] Figure 3 This is a schematic diagram of a turbofan engine.

[0043] Figure 4 This is the flight envelope diagram of a turbofan engine.

[0044] Figure 5 This is the model correction result for test sample 1.

[0045] Figure 6 This is the model correction result for test sample 2. Detailed Implementation

[0046] This invention discloses a method for correcting individual differences in turbofan engines based on convolutional neural networks, the design process of which is as follows: Figure 1 As shown, it includes the following steps:

[0047] Step 1) Based on the nominal turbofan engine component-level model, establish an individual-difference turbofan engine component-level model and generate an individual-difference deviation correction training dataset.

[0048] Step 1.1) Select a nominal component-level model of the turbofan engine to simulate the nominal engine; by adjusting the component efficiency coefficient and flow coefficient of the nominal component-level model of the turbofan engine, simulate the engine with individual differences.

[0049] Step 1.2) Select the nominal engine measurable section state parameters, engine operating atmospheric environment parameters, engine control variables, and parameters characterizing safe operating conditions to form the engine state x at time t. t The input X of the deviation correction network is formed by selecting the engine's states at n time points. t =[x t-n ,x t-n-1 ,...,x t-1 ,x t ];

[0050] Step 1.3) Select the state quantity error e of the nominal engine and the individually differentiated engine under the same atmospheric environment and input conditions. s,t Y is the output label of the bias correction network at time t. t =[e s,t ]; X is determined by input and output data. t =[x t-n ,x t-n-1 ,...,x t-1 ,x t →Y t =[e s,t The basic structure of the individual difference bias correction network;

[0051] Step 1.4): Randomly select i operating points within the full envelope to control the nominal engine model and the individual difference engine model, and collect training data during the control process according to the input-output structure in steps 1.2)-1.3).

[0052] Step 1.4.1): Randomly select the operating point [H, Ma] and the initial fuel flow rate W within the entire envelope range. fb,ini Tail nozzle area A 8,ini Initialize the nominal engine model and the individual engine model; randomly generate the reference thrust command r, and pass it through the PI controller and a set of random PI parameters k. p and k i To control the nominal engine, set the control cycle and the upper limit of the control task time;

[0053] Step 1.4.2): During the control process, the nominal control quantity W of the engine at each moment, calculated by the PI controller, is... fb,t and A 8,t As an input to the individual engine, the control process is immediately stopped when the nominal engine enters a steady state;

[0054] In step 1.4.3), during the control process of steps 1.4.1)-1.4.2), the nominal engine state x is recorded based on the input data structure described in steps 1.2)-1.3). t The input label data X of the bias correction network is obtained. t Based on the output structure described in steps 1.2)-1.3), the state and unmeasurable thrust errors of the nominal engine and the individually differentiated engine under the same atmospheric conditions and input are calculated, and the output label data Y of the deviation correction network is obtained. t Record X at each moment during the control process. t →Y t The data pairs are stored in the training dataset;

[0055] Step 1.4.4), repeat steps 1.4.1)-1.4.3) i times to complete the collection of training data for the individual bias correction model.

[0056] Step 2) Design an evaluation index for the accuracy of individual difference deviation correction of the full envelope of turbofan engine, and optimize the network structure and parameters of the individual difference deviation correction model of turbofan engine;

[0057] Step 2.1): Based on experience and the relationship between the parameters, select the initial number of input and output channels, convolution kernel and stride, input and output dimensions and number of convolutional layers for the convolutional neural network, and use average pooling layers for pooling.

[0058] Step 2.2) The Adam optimizer is used to train the dataset based on individual difference bias labels in batches at a given learning rate;

[0059] Step 2.3) Based on the full envelope individual difference correction accuracy evaluation index, evaluate the correction accuracy of the trained individual difference bias correction model under the full envelope. Based on the evaluation results, adjust the number of input and output channels, convolution kernel and stride, input and output dimensions, number of convolutional layers and learning rate in steps 2.1) and 2.2).

[0060] Step 2.3.1) Within the entire envelope, divide the area into p points at height intervals and l points at Mach number intervals, forming p×l working points; at each working point, randomly generate k sets of 1×7 test arrays. composition This serves as the initial state and reference command for the engine; where Wfb,j Let A represent the initial state of the fuel in the j-th test group. 8,j This represents the initial state of the tail nozzle area in the j-th test group. This represents the proportional gain of the fuel flow PI controller in the j-th test group. This represents the proportional coefficient of the PI controller for the tail nozzle area in the j-th test group. This represents the integral coefficient of the fuel flow PI controller in the j-th test group. R represents the integral coefficient of the PI controller for the nozzle area in the j-th test group. j This indicates the command for the thrust in the j-th test group;

[0061] Step 2.3.2): At each operating point, initialize the nominal engine and the individual engine based on the test array, initialize the PI controller based on the set PI parameters, perform thrust differential feedback control on the nominal engine, and directly use the control quantity calculated by the nominal engine PI controller as the input of the individual engine to control the individual engine, and set the control cycle and control task time limit.

[0062] Step 2.3.3) Record the nominal engine and individual engine state quantity error e at each moment during the test. s,t Calculate the average error e of each state variable in k sets of tests at a single operating point. mean and maximum instantaneous error e max and the overall average error of all state variables at all operating points within the full envelope. and maximum instantaneous error This is used to evaluate the correction accuracy of the individual difference deviation correction model for turbofan engines within the full envelope.

[0063] Step 3) Correct the output of the nominal turbofan engine airborne model based on the obtained individual difference deviation correction model;

[0064] Step 3.1): In actual use, the trained individual deviation model is uploaded. The atmospheric environment where the engine is located and the control quantities given by the airborne controller are used as inputs to characterize the airborne component-level model. After calculation, the airborne component-level model obtains the measurable cross-sectional state parameters of the engine and the parameters characterizing the safe operating state. The atmospheric environment parameters and the engine control quantities are combined to obtain the engine state x at time t. t ;

[0065] Step 3.2): Select the current data set and the data set from the previous n time points of the engine to form the input X of the individual difference deviation correction model at the current time t. t =[x t-n ,x t-n-1 ,...,x t-1 ,xt ];

[0066] Step 3.3): After obtaining the input data set, the individual difference deviation correction model calculates the error e of the state variables to be used in the control process. s,t This is used to correct the nominal engine condition.

[0067] S o,t =S d,t +e s,t

[0068] In the formula, S o,t S represents the corrected state variable. d,t This represents the state quantity of the nominal engine output.

[0069] Example

[0070] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0071] The present invention can be better understood from the following embodiments. However, those skilled in the art will readily understand that the specific material ratios, process conditions, and results described in the embodiments are for illustrative purposes only and should not, and will not, limit the invention as described in detail in the claims.

[0072] The object described in this embodiment is a certain type of twin-shaft military turbofan engine, which contains two rotor components: a high-pressure shaft and a low-pressure shaft. Its structural schematic diagram is shown below. Figure 3 As shown. In this embodiment, the control variable for the turbofan engine is the fuel flow rate W. fb The tail nozzle area is A8, and the controlled variable is thrust F. Figure 4 This is a schematic diagram of the flight envelope of the object.

[0073] Step A: Establish an engine individual difference deviation training dataset.

[0074] First, establish the dataset structure based on step 1.1). Select the total temperature T at the engine intake manifold outlet. 2,t and total pressure P 2,t Total fan outlet pressure T 21,t and total temperature P 21,t Total compressor outlet temperature T 3,t and total pressure P 3,t Low-pressure turbine outlet total temperature T 5,t and total pressure P 5,t Total temperature T at the outlet of the mixing chamber 65,t and total pressure P 65,t The altitude H of the engine's operating environment t Mach number Ma t The fuel flow rate W input to the engine at the current moment fb,t Tail nozzle area A8,t And the fan surge margin SM, which characterizes the engine's safe operating condition. f,t Compressor surge margin SM c,t The bias correction network consists of the data set at time t:

[0075] x t =[H t Ma t W fb,t A 8,t ,T 2,t ,P 2,t ,T 21,t ,P 21,t ,T 3,t ,P 3,t ,T 5,t ,P 5,t ,T 65,t ,P 65,t SM f,t SM c,t ]

[0076] The current data set and the previous 5 time steps of the engine are selected as the input data for the current time step of the convolutional neural network (CNN).

[0077] X t =[x t ,x t-1 ,x t-2 ,x t-3 ,x t-4 ,x t-5 ]

[0078] The error of the state variables needed in the control process is selected as the output of the deviation correction network at the current moment:

[0079] Y t =[e nL,t ,e nH,t ,e EPR,t ,e P3,t ,e T6,t ,e F,t ]

[0080] In the formula, e represents the deviation between the nominal engine and the engine state variables with individual differences, and n L For the low-pressure shaft speed of the engine, n H is the engine high-pressure shaft speed, EPR is the engine pressure ratio, P3 is the compressor outlet total pressure, T6 is the mixing chamber outlet total temperature, and F is the thrust.

[0081] Next, within the entire envelope range, a working point [H, Ma] is randomly selected, and the control quantity, fuel flow rate W, is initialized. fb,t Tail nozzle area A 8,tReference thrust command r F and the parameter k of the PI controller p and k i Furthermore, the component-level model is initialized based on the above conditions and made to run smoothly, based on the thrust error e. F =r F -F performs thrust error feedback PI control on the nominal engine. Data acquisition is performed during the control process. Input and output data are collected at each sampling time to compare with X. t →Y t This data was used as training data. 20,000 working points were selected within the data envelope, and the data collection process was repeated to generate the individual difference bias training dataset.

[0082] Step B: Based on experience, select the convolutional neural network parameters as shown in Scheme A. Building upon Scheme A, use the controlled variable method to investigate the effects of variations in the number of input / output channels, convolutional kernel and stride, and the number of convolutional layers on the network training performance. Groups B1 and B2 use the number of input / output channels as variables, groups C1 and C2 use the convolutional kernel and stride as variables, and groups D1 and D2 use the number of convolutional layers as variables. In groups D1 and D2, since the network's input and output dimensions are always integers, the convolutional kernel and stride are fine-tuned appropriately. The final scheme is shown in the table below:

[0083] Table 1 Network Parameter Adjustment Scheme A

[0084]

[0085] Table 2 Network Parameter Adjustment Scheme B1

[0086]

[0087] Table 3 Network Parameter Adjustment Scheme B2

[0088]

[0089] Table 4 Network Parameter Adjustment Scheme C1

[0090]

[0091] Table 5 Network Parameter Adjustment Scheme C2

[0092]

[0093] Table 6 Network Parameter Adjustment Scheme D1

[0094]

[0095] Table 7 Network Parameter Adjustment Scheme D2

[0096]

[0097] The Adam optimizer is used to train the dataset on individual difference bias labels in batches at a given learning rate.

[0098] Step C: To facilitate the optimization of network structure parameters, design an evaluation index for the individual difference deviation correction accuracy of the full envelope of the turbofan engine.

[0099] Within the entire envelope, 51 points are divided at height intervals, and 9 points are divided at Mach number intervals, forming 51 × 9 = 459 working points. At each working point, 100 sets of 1 × 7 random arrays are randomly generated. composition This serves as the initial state of the engine, reference command, and parameters of the PI controller in each operating point test group.

[0100] Subsequently, PI feedback control is applied to the nominal engine based on the thrust error. The state output of the nominal engine is corrected using an individual deviation correction model. Simultaneously, the control input of the nominal engine at each moment is used as the control input for the individual-specific engine, and the individual-specific engine is controlled accordingly. During the process, the error between the state input of the individual-specific engine and the corrected state input of the nominal engine at each moment is recorded, and the average correction error e for each state input at each operating point is calculated. mean and maximum correction error e max and the average correction error within the full envelope. and maximum correction error This is used to evaluate the accuracy of the individual difference bias correction model within the entire envelope.

[0101] Based on the aforementioned individual difference correction accuracy evaluation index within the full envelope, the correction accuracy of the trained individual difference bias correction model under the full envelope is evaluated. Based on the evaluation results, the number of input and output channels, convolution kernel and stride, input and output dimensions, number of convolutional layers and learning rate are adjusted.

[0102] Step D: Based on the test results from Step C, adjust the network structure and parameters. Table 8 shows the final network structure and parameters used.

[0103] Table 8 shows the adopted schemes.

[0104]

[0105] Typical operating point tests were conducted on the individual difference bias correction model constructed from the selected network structure. Figure 5The diagram shows the corrected results for Test 1, which was conducted at an altitude of 0m and a Mach number of 0. The dashed line represents the state variable output of the engine with individual differences, the dotted line represents the output of the nominal engine, and the solid line represents the result after correcting the nominal engine's airborne model using the designed individual difference correction model. The errors of each state variable throughout the control process are shown in the table below:

[0106] Table 9. Correction Table for Individual Differences when H=0m, Ma=0

[0107]

[0108] It can be seen that the average correction error of each state quantity at the ground working point is less than 0.09%, and the maximum error is less than 0.28%.

[0109] Figure 6 The image shown is a corrected result diagram of Test 2, which was conducted at an altitude of 2000m and a Mach number of 0.1.

[0110] Table 10. Error Correction Table for Individual Differences (H = 2000m, Ma = 0.1)

[0111]

[0112] It can be seen that at a height of 2000m and a Mach number of 0.1, the average correction error for each state quantity is less than 0.11%, and the maximum error is less than 0.36%.

[0113] Therefore, the proposed individual difference bias correction method based on convolutional neural networks has good correction accuracy.

[0114] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for correcting individual differences in turbofan engines based on convolutional neural networks, characterized in that: Includes the following steps: Step 1) Based on the nominal turbofan engine component-level model, establish an individual-difference turbofan engine component-level model and generate an individual-difference deviation correction training dataset; Step 2) Design an evaluation index for the accuracy of individual difference deviation correction of the full envelope of turbofan engine, and optimize the network structure and parameters of the individual difference deviation correction model of turbofan engine; Step 3) Correct the output of the nominal turbofan engine airborne model based on the obtained individual difference deviation correction model; The specific steps of the method for generating the training dataset to correct for individual differences in step 1) are as follows: Step 1.1) Select a nominal component-level model of the turbofan engine to simulate the nominal engine; by adjusting the component efficiency coefficient and flow coefficient of the nominal component-level model of the turbofan engine, simulate the engine with individual differences; Step 1.2) Select the nominal engine measurable section state parameters, engine operating atmospheric environment parameters, engine control variables, and parameters characterizing safe operating conditions to form the engine state at time t. The input to the deviation correction network is formed by selecting the engine's states at n time points. ; Step 1.3) Select the state quantity error of the nominal engine and the individually differentiated engine under the same atmospheric environment and input conditions. As the output label of the deviation correction network at time t ; by input / output structure The basic structure of the individual difference bias correction network; Step 1.4): Randomly select i operating points within the full envelope to control the nominal engine model and the individual difference engine model, and collect training data during the control process according to the input-output structure in Step 1.3). The specific steps for designing and optimizing the individual difference deviation correction model for the turbofan engine in step 2) are as follows: Step 2.1) Based on experience and the relationship between the parameters, select the initial number of input and output channels, convolution kernel and stride, input and output dimensions and number of convolutional layers for the convolutional neural network, and use average pooling layers for pooling layers; Step 2.2) uses the Adam optimizer to train in batches based on the individual difference bias label dataset at a given learning rate; Step 2.3) Based on the full envelope individual difference correction accuracy evaluation index, evaluate the correction accuracy of the trained individual difference bias correction model under the full envelope. Based on the evaluation results, adjust the number of input and output channels, convolution kernel and stride, input and output dimensions, number of convolutional layers and learning rate in Step 2.1) and Step 2.2). The specific steps for estimating the accuracy of individual difference deviation correction for the full envelope of the turbofan engine in step 2.3) include: Step 2.3.1) Within the entire envelope, divide the area into p points at height intervals and l points at Mach number intervals, forming a... There are 1 working point; at each working point, k groups are randomly generated. test array ,composition This serves as the initial state and reference command for the engine; among which This represents the initial state of the fuel in the j-th test group. This represents the initial state of the tail nozzle area in the j-th test group. This represents the proportional gain of the fuel flow PI controller in the j-th test group. This represents the proportional coefficient of the PI controller for the tail nozzle area in the j-th test group. This represents the integral coefficient of the fuel flow PI controller in the j-th test group. This represents the integral coefficient of the PI controller for the tail nozzle area in the j-th test group. This indicates the command for the thrust in the j-th test group; Step 2.3.2): At each operating point, initialize the nominal engine and the individual engine based on the test array, initialize the PI controller based on the set PI parameters, perform thrust differential feedback control on the nominal engine, and directly use the control quantity calculated by the nominal engine PI controller as the input of the individual engine to control the individual engine, and set the control cycle and control task time limit. Step 2.3.3) Record the nominal engine and individual engine state quantity errors at each moment during the test. Calculate the average error of each state variable in k sets of tests at a single operating point. and maximum instantaneous error and the overall average error of all state variables at all operating points within the full envelope. and maximum instantaneous error This is used to evaluate the correction accuracy of the individual difference deviation correction model for turbofan engines within the full envelope.

2. The method for correcting individual differences in turbofan engines based on convolutional neural networks according to claim 1, characterized in that: The specific steps for training data acquisition in step 1.4) are as follows: Step 1.4.1): Randomly select a working point within the entire envelope. Initial fuel flow Tail nozzle area Initialize the nominal engine model and the individual engine model; randomly generate the reference thrust command r, and pass it through the PI controller and a set of random PI parameters k. p and k i To control the nominal engine, set the control cycle and the upper limit of the control task time; Step 1.4.2): During the control process, the nominal engine control quantity calculated by the PI controller at each moment is... and As an input to the individual engine, the control process is immediately stopped when the nominal engine enters a steady state; In step 1.4.3), during the control process of steps 1.4.1)-1.4.2), the nominal engine state is recorded based on the input / output structure described in step 1.3). The input label data for the bias correction network is obtained. ; Based on the input-output structure described in step 1.3), the state and unmeasurable thrust errors of the nominal engine and the individually differentiated engine under the same atmospheric conditions and input are calculated, and the output label data of the deviation correction network is obtained. Record every moment during the control process. The data pairs are stored in the training dataset; Step 1.4.4), repeat steps 1.4.1) to 1.4.3) i times to complete the collection of training data for the individual bias correction model.

3. The method for correcting individual differences in turbofan engines based on convolutional neural networks according to claim 1, characterized in that: The specific steps for correcting the airborne model in step 3) are as follows: Step 3.1): In actual use, the trained individual deviation model is uploaded. The atmospheric environment where the engine is located and the control quantities given by the airborne controller are used as inputs to characterize the airborne component-level model. After calculation, the airborne component-level model obtains the measurable cross-sectional state parameters of the engine and the parameters characterizing the safe operating state. The atmospheric environment parameters and the engine control quantities are combined to obtain the engine state at time t. ; Step 3.2) Select the current data set and the data set from the previous n time steps of the engine to form the input of the individual difference bias correction model at time t. ; Step 3.3): After obtaining the input data set, the individual difference deviation correction model calculates the error of the state variables to be used in the control process. This is used to correct the nominal engine condition. ； In the formula, This represents the corrected state variable. This represents the state quantity of the nominal engine output.

Images