A dynamic thrust estimation method for turbofan engines based on recurrent neural network

Abstract

The present invention discloses a method for dynamic thrust estimation of a turbofan engine based on a recurrent neural network, comprising the following steps: generating a full-envelope dynamic thrust data set based on a turbofan engine thermodynamic model; designing a full-envelope dynamic thrust estimation evaluation index for a turbofan engine, training and optimizing a turbofan engine dynamic thrust estimator; estimating the thrust output level of the turbofan engine during use according to the obtained dynamic thrust estimator. The method for generating a dynamic thrust data set of the present invention improves the diversity of engine dynamic samples, solves the problems of difficulty in obtaining training data and single dynamic characteristics; the established dynamic thrust estimation evaluation index can intuitively evaluate the dynamic estimation performance of the thrust estimator within the full envelope, and effectively improves the optimization effect and efficiency of the structural parameters and training parameters during the training of the estimator.

Description

Technical Field

[0001] The invention belongs to the technical field of aviation engine control, and in particular relates to a method for estimating dynamic thrust of a turbofan engine based on a recurrent neural network. Background Art

[0002] As aircraft engines develop towards high thrust-to-weight ratio / power-to-weight, high performance and high economy, traditional indirect thrust control has greatly restricted the development of high-performance aircraft engines due to problems such as imprecise thrust control, nonlinear throttle angle, mismatched dual-engine thrust, and conservative control plan. Therefore, direct thrust control technology has become necessary for research. The United States' Intelligent Engine (IEC) is one of the top four advanced control concepts, and one of its core contents is direct thrust control. The Aviation Technology Development Seminar initiated by NASA also proposed the need to reduce the workload of pilots and improve the autonomy of aircraft engine operation. Research on direct thrust control is receiving widespread attention around the world.

[0003] In order to achieve direct thrust control, the thrust estimator needs to provide accurate thrust feedback information. Existing thrust estimation methods can be roughly divided into two categories: one is model-based thrust estimation methods, and the other is data-driven thrust estimation methods. The model-based thrust estimation methods can be further divided into thrust estimation methods based on aerodynamic thermodynamic component-level models and thrust estimation methods based on linear models.

[0004] Thrust estimation based on component-level models can greatly improve the accuracy of thrust estimation and also has a good estimation effect on the dynamic process of thrust change. However, this method has the problems of large computational complexity and poor real-time performance, making it difficult to apply onboard. Most thrust estimation methods based on linear models use Kalman filters to estimate thrust on the basis of linear models. Although this greatly reduces the amount of computation, there is a large model uncertainty and the thrust estimation accuracy is low. In addition, the characteristics of the Kalman filtering method itself will cause large noise in the thrust estimation process, which is not conducive to the stability of the control system.

[0005] Most data-driven thrust estimation methods use various neural networks to fit the relationship between engine state and thrust output. This type of method has good fitting ability for known data sets, has good thrust estimation effect, and has small computational complexity when used online, which can meet the real-time requirements of the control system. However, it places extremely high demands on the diversity of data sets, and the model has poor estimation ability for dynamic characteristics not included in the data set. Moreover, since the engine has strong time series characteristics, most neural network structures have poor processing capabilities for time series, which also reduces the estimation accuracy of the estimator to a certain extent.

[0006] In general, although a lot of research has been carried out in the field of aero-engine thrust estimation, so far no thrust estimation method can be applied to engineering practice.

[0007] Recurrent Neural Network (RNN) is a type of recursive neural network that takes sequence data as input, recurses in the direction of sequence evolution, and all nodes are connected in a chain. It has the characteristics of memory, parameter sharing, and Turing completeness, so it has certain advantages when learning the nonlinear characteristics of the sequence. The Long Short-Term Memory Networks (LSTM) developed on the basis of RNN not only have the advantages of RNN for sequence data processing, but also can effectively solve the gradient explosion and gradient disappearance problems existing in RNN training, and are very suitable for the thrust estimation problem of turbofan engines with strong nonlinearity. Therefore, in order to improve the shortcomings of the existing thrust estimation method, the present invention proposes a dynamic thrust estimation method based on a recurrent neural network for turbofan engines, focusing on improving the dynamic thrust estimation capability of the estimator, solving the data acquisition problem, and improving the full envelope thrust estimation accuracy. Summary of the invention

[0008] Purpose of the invention: In order to solve the problem that the thrust estimation accuracy of the conventional turbofan engine thrust estimation method is not high for the random dynamic characteristics, the present invention proposes a turbofan engine dynamic thrust estimation method based on a recurrent neural network. It is intended to improve the diversity of engine dynamic samples through a dynamic thrust data set generation method; through the dynamic thrust estimation evaluation index, the optimization effect and optimization efficiency of the structural parameters and training parameters in the estimator training process are improved, and finally the dynamic thrust estimation accuracy of the turbofan engine thrust estimator in the full envelope range is improved, and the high-precision estimation of the thrust without measuring the full envelope range of the turbofan engine is realized.

[0009] Technical solution: To achieve the above purpose, the technical solution adopted by the present invention is:

[0010] A method for estimating dynamic thrust of a turbofan engine based on a recurrent neural network comprises the following steps:

[0011] Step 1), based on the turbofan engine thermodynamic model, generate a full envelope dynamic thrust data set;

[0012] Step 2), designing a turbofan engine full envelope dynamic thrust estimation evaluation index, training and optimizing a turbofan engine dynamic thrust estimator;

[0013] Step 3) estimating the thrust output level of the turbofan engine during use based on the obtained dynamic thrust estimator.

[0014] Furthermore, the specific steps of the method for generating the dynamic thrust data set in step 1) are as follows:

[0015] Step 1.1), in order to train the random dynamic thrust model of the full flight envelope of the selected aircraft engine, select the atmospheric environment parameters and the measurable state parameters of the engine, the state x of the engine at time t t Select the current and previous n-time state of the engine. Step 1.1) In order to train the selected aircraft engine full flight envelope random dynamic thrust model, select the atmospheric environment parameters and the engine measurable state parameters to form the engine state x at time t. t ; Select the current and previous n-time states of the engine to form the engine state group X at time t t =[x t-n ,...,x t ], as the input of the LSTM RNN neural network; select the current moment thrust F t As the neural network output, written as Y t =[y t ]; by X t =[x t-n ,...,x t ]→Y t =[y t ] The input-output pairs constitute the basic structure of the dynamic thrust data set;

[0016] Step 1.2), randomly select the operating point [H,Ma] and the initial fuel flow rate W within the full envelope. fb and the initial tail nozzle area A8, the final thrust command r and the PI control parameter k p and k i ; After initializing the engine according to the initial conditions, according to the thrust command r F Calculate the thrust error e from the component and model thrust output F F =r F -F, and perform thrust error feedback control to obtain the random dynamic characteristics of the engine;

[0017] Step 1.3), in the dynamic process, collecting engine dynamic data based on the data set structure in step 1.1), and stopping data collection when the engine enters a steady state;

[0018] Step 1.4), repeat the steps from step 1.2) to step 1.3) for o times to complete the dynamic thrust data collection.

[0019] Furthermore, the specific steps of the full flight envelope dynamic thrust estimator training method in step 2) are as follows:

[0020] Step 2.1), based on experience and trial and error, select the number of LSTM recurrent neural network layers, number of nodes and time series length;

[0021] Step 2.2), using the Adam optimizer, training is performed in batches at a given learning rate based on the dynamic thrust dataset;

[0022] Step 2.3), designing the evaluation index of the dynamic thrust estimation of the turbofan engine in the full envelope, and evaluating the comprehensive performance of the dynamic thrust estimator of the turbofan engine in the full envelope;

[0023] In step 2.4), based on the dynamic thrust evaluation index designed in step 2.3), the estimation performance of the dynamic thrust estimator trained in step 2 is evaluated within the full envelope. According to the evaluation results, the number of network layers, number of nodes, time series length, and learning rate in step 2.1) and step 2.2) are adjusted.

[0024] Furthermore, the specific steps of the full envelope dynamic thrust estimation evaluation index design method in step 2.3) are as follows:

[0025] Step 2.3.1), randomly generate l groups of 1×5 random arrays c j =[W fb,j ,A 8,j ,k p,j ,k i,j ,r j ](j=1,...,l), composed of Where W fb,j represents the normalized initial state of the fuel in the jth test group, A 8,j represents the normalized initial state of the tail nozzle area in the jth group of tests, k p,j represents the proportionality coefficient in the jth group of experiments, k i,j represents the integral coefficient in the jth group of experiments, r j represents the normalized thrust command in the jth group of trials, and W fb,j ∈[0,1],A 8,j ∈[0,1], k p,j ∈[0,1], k i,j ∈[0,1],r j ∈[0,1];

[0026] Step 2.3.2), at the same working point (H, Ma), obtain the maximum thrust F of the current working point respectively max and minimum thrust F min , maximum fuel flow W fb,max and minimum fuel flow W fb,min , maximum tail nozzle area A 8,max and the minimum tail nozzle area, according to step 2.3.1) Perform inverse normalization to obtain the initial fuel, initial tail nozzle area and thrust command value, establish l test samples, and the inverse normalization process is expressed as

[0027] W fb,ini,j =W fb,j ·(W fb,max -W fb,min )+W fb,min

[0028] A 8,ini,j =A 8,j ·(A 8,max -A 8,min )+A 8,min

[0029] r F,j =r j ·(F max -F min )+F min

[0030] On these l test samples, by presetting the coefficient k p,j and k i,j Based on the PI controller, the thrust is controlled by error feedback. The error value between the thrust estimation and the engine thrust output is recorded at each moment during the test process. The average thrust estimation error e at the working point is calculated. F,mean and single point maximum instantaneous thrust estimation error e F,max ;

[0031] Step 2.3.3), divide n points at altitude intervals within the full envelope, divide m points at Mach number intervals, and form n×m working points within the full envelope. Perform single-point tests at these working points according to steps 2.3.2) to 2.3.3), and record the single-point average thrust estimation error e of n×m grid points F,mean and single point maximum instantaneous thrust estimation error e F,max ;

[0032] Step 2.3.4) Estimate the error e based on the single point average thrust F,mean and single point maximum instantaneous thrust estimation error e F,max , calculate the overall average error of thrust estimation in the entire envelope Thrust estimation error variance var(e F,mean ), and the maximum thrust estimation error within the envelope

[0033] Further, the specific steps of step 3) are as follows:

[0034] Step 3.1), in actual use, based on the sensor measurement information, record the measurable state parameters of the engine to form the engine state x at time t t ;

[0035] Step 3.2), select the current and previous n-time states of the engine to form the current engine state group X at time t t =[x t-n ,...,x t ];

[0036] Step 3.3), group the engine status into X t As the input of the dynamic thrust estimator, it estimates the current thrust of the engine

[0037] Beneficial effects: The present invention provides a method for estimating dynamic thrust of a turbofan engine based on a recurrent neural network. Compared with the prior art, the above technical solution has the following technical effects:

[0038] (1) The method for generating a dynamic thrust data set proposed in the present invention improves the diversity of engine dynamic samples and solves the problem of difficulty in obtaining training data and single dynamic characteristics;

[0039] (2) The dynamic thrust estimation evaluation index established by the present invention can intuitively evaluate the dynamic estimation performance of the thrust estimator within the full envelope, effectively improving the optimization effect and efficiency of the structural parameters and training parameters during the estimator training process;

[0040] (3) The present invention proposes a method for dynamic thrust estimation of a turbofan engine based on a recurrent neural network, which greatly improves the thrust estimation capability of the turbofan engine in the random dynamic process of the full envelope, and plays a positive role in promoting the application of direct thrust control in turbofan engines and improving the overall performance of the engine. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a design flow of dynamic thrust estimator based on recurrent neural network.

[0042] Figure 2 This is a schematic diagram of the turbofan engine structure.

[0043] Figure 3 It is the flight envelope diagram of the turbofan engine.

[0044] Figure 4 This is the thrust estimation test result of test sample 1.

[0045] Figure 5 This is the thrust estimation test result of test sample 2.

[0046] Figure 6It is a flight path diagram for full envelope simulation flight test.

[0047] Figure 7 It is the test result of full envelope simulated flight test. DETAILED DESCRIPTION

[0048] The present invention discloses a method for dynamic thrust estimation of a turbofan engine based on a recurrent neural network, comprising the following steps: step 1), based on a turbofan engine thermodynamic model, generating a full envelope dynamic thrust data set; step 2), designing a turbofan engine full envelope dynamic thrust estimation evaluation index, training and optimizing a turbofan engine dynamic thrust estimator; step 3), estimating the thrust output level of the turbofan engine during use according to the obtained dynamic thrust estimator. The present invention effectively improves the diversity of engine dynamic samples; the established dynamic thrust estimation evaluation index can intuitively evaluate the dynamic estimation performance of the thrust estimator within the full envelope, effectively improving the optimization effect and optimization efficiency of the structural parameters and training parameters during the training of the estimator; greatly improving the thrust estimation capability of the turbofan engine in the random dynamic process of the full envelope, and playing a positive role in promoting the application of direct thrust control in turbofan engines and improving the overall performance of the engine.

[0049] A dynamic thrust estimation method for turbofan engines based on recurrent neural networks, the design process is as follows Figure 1 As shown, the following steps are included:

[0050] Step 1), based on the turbofan engine thermodynamic model, generate a full envelope dynamic thrust data set;

[0051] Step 1.1), in order to train the random dynamic thrust model of the full flight envelope of the selected aircraft engine, the atmospheric environment parameters and the measurable state parameters of the engine are selected to form the engine state x at time t t ; Select the current and previous n-time states of the engine to form the engine state group X at time t t =[x t-n ,...,x t ], as the input of the LSTM RNN neural network; select the current moment thrust F t As the neural network output, written as Y t =[y t ]; by X t =[x t-n ,…,x t ]→Y t =[y t ] The input-output pairs constitute the basic structure of the dynamic thrust data set;

[0052] Step 1.2), randomly select the operating point [H,Ma] and the initial fuel flow rate W within the full envelope.fb and the initial tail nozzle area A8, the final thrust command r and the PI control parameter k p and k i ; After initializing the engine according to the initial conditions, according to the thrust command r F Calculate the thrust error e from the component and model thrust output F F =r F -F, and perform thrust error feedback control to obtain the random dynamic characteristics of the engine;

[0053] Step 1.3), in the dynamic process, collecting engine dynamic data based on the data set structure in step 1.1), and stopping data collection when the engine enters a steady state;

[0054] Step 1.4), repeat the steps from step 1.2) to step 1.3) for o times to complete the dynamic thrust data collection.

[0055] Step 2), designing a turbofan engine full envelope dynamic thrust estimation evaluation index, training and optimizing a turbofan engine dynamic thrust estimator;

[0056] Step 2.1), based on experience and trial and error, select the number of LSTM recurrent neural network layers, number of nodes and time series length;

[0057] Step 2.2), using the Adam optimizer, training is performed in batches at a given learning rate based on the dynamic thrust dataset;

[0058] Step 2.3), designing the evaluation index of the dynamic thrust estimation of the turbofan engine in the full envelope, and evaluating the comprehensive performance of the dynamic thrust estimator of the turbofan engine in the full envelope;

[0059] Step 2.3.1), randomly generate l groups of 1×5 random arrays c j =[W fb,j ,A 8,j ,k p,j ,k i,j ,r j ](j=1,...,l), composed of Where W fb,j represents the normalized initial state of the fuel in the jth test group, A 8,j represents the normalized initial state of the tail nozzle area in the jth group of tests, k p,j represents the proportionality coefficient in the jth group of experiments, k i,j represents the integral coefficient in the jth group of experiments, r j represents the normalized thrust command in the jth group of trials, and W fb,j ∈[0,1],A 8,j ∈[0,1], k p,j∈[0,1], k i,j ∈[0,1],r j ∈[0,1];

[0060] Step 2.3.2), at the same working point (H, Ma), obtain the maximum thrust F of the current working point respectively max and minimum thrust F min , maximum fuel flow W fb,max and minimum fuel flow W fb,min , maximum tail nozzle area A 8,max and the minimum tail nozzle area, according to step 2.3.1) Perform inverse normalization to obtain the initial fuel, initial tail nozzle area and thrust command value, establish l test samples, and the inverse normalization process is expressed as

[0061] W fb,ini,j =W fb,j ·(W fb,max -W fb,min )+W fb,min

[0062] A 8,ini,j =A 8,j ·(A 8,max -A 8,min )+A 8,min

[0063] r F,j =r j ·(F max -F min )+F min

[0064] On these l test samples, by presetting the coefficient k p,j and k i,j Based on the PI controller, the thrust is controlled by error feedback. The error value between the thrust estimation and the engine thrust output is recorded at each moment during the test process. The average thrust estimation error e at the working point is calculated. F,mean and single point maximum instantaneous thrust estimation error e F,max ;

[0065] Step 2.3.3), divide n points at altitude intervals within the full envelope, divide m points at Mach number intervals, and form n×m working points within the full envelope. Perform single-point tests at these working points according to steps 2.3.2) to 2.3.3), and record the single-point average thrust estimation error e of n×m grid points F,mean and single point maximum instantaneous thrust estimation error e F,max ;

[0066] Step 2.3.4) Estimate the error e based on the single point average thrust F,mean and single point maximum instantaneous thrust estimation error e F,max , calculate the overall average error of thrust estimation in the entire envelope Thrust estimation error variance var(e F,mean ), and the maximum thrust estimation error within the envelope

[0067] In step 2.4), based on the dynamic thrust evaluation index designed in step 2.3), the estimation performance of the dynamic thrust estimator trained in step 2 is evaluated within the full envelope. According to the evaluation results, the number of network layers, number of nodes, time series length, and learning rate in step 2.1) and step 2.2) are adjusted.

[0068] Step 3) estimating the thrust output level of the turbofan engine during use based on the obtained dynamic thrust estimator.

[0069] Step 3.1), in actual use, based on the sensor measurement information, record the measurable state parameters of the engine to form the engine state x at time t t ;

[0070] Step 3.2), select the current and previous n-time states of the engine to form the current engine state group X at time t t =[x t-n ,...,x t ];

[0071] Step 3.3), group the engine status into X t As the input of the dynamic thrust estimator, it estimates the current thrust of the engine

[0072] Example

[0073] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0074] The present invention can be better understood according to the following examples. However, it is easy for those skilled in the art to understand that the specific material ratios, process conditions and results described in the examples are only used to illustrate the present invention, and should not and will not limit the present invention described in detail in the claims.

[0075] The object described in the embodiment is a certain type of dual-shaft military turbofan engine, which contains two rotor parts, a high-pressure shaft and a low-pressure shaft. Figure 2 In this embodiment, the control quantity of the turbofan engine is the fuel flow rate W fb And the tail nozzle area A8, the controlled quantity is thrust F. Figure 3 This is a schematic diagram of the object's flight envelope.

[0076] Step A: Establish an engine dynamic thrust data set.

[0077] First, the data set structure is established based on step 1.1). The atmospheric environment parameters altitude and Mach number are selected, and the engine measurable state parameters low-pressure shaft speed n are selected. L , high pressure shaft speed n H , turbine pressure drop ratio EPR, high pressure compressor outlet atmospheric pressure P3, low pressure turbine outlet temperature T6 constitute the engine state x at time t t =[H t ,Ma t ,n L,t ,n H,t ,EPR t ,P 3,t ,T 6,t ], based on the control period T = 0.02s, the engine status at the current and previous n moments is selected to form the engine status group X at time t t =[x t-n ,...,x t [, as the input of the LSTM RNN neural network, in order to test the influence of different time series lengths on the estimation network accuracy, n = 3 and n = 6 are selected for exploration. Select the current moment thrust F t As the output of the neural network, it can be written as Y t =[y t ].

[0078] Next, a random operating point [H,Ma] and an initial fuel flow rate W are selected within the full envelope. fb and the initial tail nozzle area A8, thrust command r F and PI control parameters kp and k i The engine component level model is set to enter a stable operation state according to the initial atmospheric conditions and initial control quantities, and the thrust error e F =r F -F performs thrust error feedback PI control.

[0079] Then, the engine dynamic process data is collected. t =[x t-n ,...,x t ]→Y t =[y t ] is used as a set of data in the data set. In order to prevent the steady-state data from occupying too large a proportion of the data set and affecting the estimation accuracy of the dynamic process by the estimator, data collection is stopped when the engine enters the steady state. The basis for judging whether the engine enters the steady state is: the engine has ten consecutive control cycles (i.e., 0.2s in a row), and the thrust error e F<0.01%.

[0080] 40,000 working points were selected within the envelope and the data collection work was repeated to complete the generation of the dynamic thrust data set.

[0081] Step B: Select different network structures and training parameters based on experience to form five schemes, as shown in Table 1. In the table, scheme A is the reference group, and schemes BE are obtained by changing one of the parameters on the basis of scheme A through the control variable method to explore the influence of each parameter on the estimation accuracy.

[0082] Table 1 Thrust estimator structure and training parameter scheme

[0083]

[0084] The Adam optimizer is used to train the model in batches at a given learning rate based on the dynamic thrust dataset.

[0085] Step C: In order to facilitate the debugging of network structure parameters and training parameters, the dynamic thrust estimation evaluation index of the turbofan engine full envelope is designed.

[0086] Select l = 100 and randomly generate one hundred 1×5 random arrays c j =[W fb,j ,A 8,j ,k p,j ,k i,j ,r j ](j=1,...,100), composed of Where W fb,j represents the normalized initial state of the fuel in the jth test group, A 8,j represents the normalized initial state of the tail nozzle area in the jth group of tests, k p,j represents the proportionality coefficient in the jth group of experiments, k i,j represents the integral coefficient in the jth group of experiments, r j represents the normalized thrust command in the jth group of trials, and W fb,j ∈[0,1],A 8,j ∈[0,1], k p,j ∈[0,1], k i,j ∈[0,1],r j ∈[0,1].

[0087] At the same working point (H, Ma), obtain the maximum thrust F of the current working point respectively max and minimum thrust F min , maximum fuel flow W fb,max and minimum fuel flow W fb,min , maximum tail nozzle area A 8,maxand the minimum tail nozzle area, according to step 2.3.1) After reverse normalization, the initial fuel, initial nozzle area and thrust command values ​​are obtained. One hundred test cases are established. The reverse normalization process can be expressed as

[0088] W fb,ini,j =W fb,j ·(W fb,max -W fb,min )+W fb,min

[0089] A 8,ini,j =A 8,j ·(A 8,max -A 8,min )+A 8,min

[0090] r F,j =r j ·(F max -F min )+F min

[0091] On these test cases, the thrust estimation performance of the estimator AE trained in step C is tested over the full envelope.

[0092] In the full envelope, 32 points are divided by altitude intervals, and 11 points are divided by Mach number intervals, forming 32×11=352 working points in the full envelope. At the same working point (H, Ma), based on the 100 established test samples, the preset coefficient k p,j and k i,j Based on the PI controller, the thrust is controlled by error feedback. The error value between the thrust estimation and the engine thrust output is recorded at each moment in the control process, and the average thrust estimation error e at the working point is calculated. F,mean and single point maximum instantaneous thrust estimation error e F,max .

[0093] Estimation error e based on the single point average thrust F,mean and single point maximum instantaneous thrust estimation error e F,max , calculate the overall average error of thrust estimation in the entire envelope Thrust estimation error variance var(e F,mean ), and the maximum thrust estimation error within the envelope The results are shown in Table 2.

[0094] Table 2 Full envelope accuracy test results of Scheme A-Scheme E

[0095]

[0096] Step D: Based on the test results in Table 2, adjust the network structure parameters and training parameters. Table 3 shows the network structure and training parameters of Scheme F. Based on the Adam optimizer, the estimator of Scheme F is retrained. Table 4 shows the thrust estimation accuracy of the thrust estimator of Scheme F within the full envelope.

[0097] Table 3 Final thrust estimator structure and training parameters

[0098]

[0099] Table 4 Scheme G full envelope accuracy test results

[0100]

[0101] The designed thrust estimator F is tested at randomly selected points within the full envelope to test the thrust estimation accuracy. Figure 4 The thrust estimation test diagram of test 1 is shown, which was conducted at an altitude of 387.0m and a Mach number of 0.56. The dotted line is the thrust command, the solid line is the actual thrust output of the engine, and the long dashed line is the thrust output of the estimator. During the entire control process, the average estimation error is 0.041% and the maximum thrust estimation error is 0.173%.

[0102] Figure 5 The thrust estimation test diagram of test 2 is shown, which was conducted at an altitude of 5615.0m and a Mach number of 0.82. The dotted line is the thrust command, the solid line is the actual thrust output of the engine, and the long dashed line is the thrust output of the estimator. During the entire control process, the average estimation error is 0.083% and the maximum thrust estimation error is 0.351%.

[0103] Figure 6 The figure shows the flight path of the simulated flight within the full envelope. The thrust during the simulated flight is estimated as follows: Figure 7 The results are shown. Figure 7 In the figure, the first picture from top to bottom is the thrust estimation result, the second to fourth pictures are the flight environment and engine input changes, and the fifth picture is the estimation error. In the simulated flight test, the average estimation error is 0.085% and the maximum estimation error is 0.309%.

[0104] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for estimating dynamic thrust of a turbofan engine based on a recurrent neural network, characterized in that: The following steps are involved: Step 1), based on the turbofan engine thermodynamic model, generate a full envelope dynamic thrust data set; Step 2), designing a turbofan engine full envelope dynamic thrust estimation evaluation index, training and optimizing a turbofan engine dynamic thrust estimator; Step 3), estimating the thrust output level of the turbofan engine during use according to the obtained dynamic thrust estimator; The specific steps of the method for generating the dynamic thrust data set in step 1) are as follows: Step 1.1), in order to train the random dynamic thrust model of the full flight envelope of the selected aircraft engine, the atmospheric environment parameters and the measurable state parameters of the engine are selected to form the engine state x at time t t ; Select the current and previous n-time states of the engine to form the engine state group X at time t t =[x t-n ,...,x t ], as the input of the LSTM RNN neural network; select the current moment thrust F t As the neural network output, written as Y t =[y t ]; by X t =[x t-n ,...,x t ]→Y t =[y t ] The input-output pairs constitute the basic structure of the dynamic thrust data set; Step 1.2), randomly select the operating point [H,Ma] and the initial fuel flow rate W within the full envelope. fb and the initial tail nozzle area A8, the final thrust command r and the PI control parameter k p and k i ; After initializing the engine according to the initial conditions, according to the thrust command r F Calculate the thrust error e from the component and model thrust output F F =r F -F, and perform thrust error feedback control to obtain the random dynamic characteristics of the engine; Step 1.3), in the dynamic process, collecting engine dynamic data based on the data set structure in step 1.1), and stopping data collection when the engine enters a steady state; Step 1.4), repeating steps 1.2) to 1.3) for o times to complete dynamic thrust data collection; The specific steps of the full flight envelope dynamic thrust estimator training method in step 2) are as follows: Step 2.1), based on experience and trial and error, select the number of LSTM recurrent neural network layers, number of nodes and time series length; Step 2.2), using the Adam optimizer, training is performed in batches at a given learning rate based on the dynamic thrust dataset; Step 2.3), designing the evaluation index of the dynamic thrust estimation of the turbofan engine in the full envelope, and evaluating the comprehensive performance of the dynamic thrust estimator of the turbofan engine in the full envelope; Step 2.4), based on the dynamic thrust evaluation index designed in step 2.3), the estimation performance of the dynamic thrust estimator trained in step 2.2) within the full envelope is evaluated, and according to the evaluation results, the number of network layers, number of nodes, time series length, and learning rate in step 2.1) and step 2.2) are adjusted; The specific steps of the full envelope dynamic thrust estimation evaluation index design method in step 2.3) are as follows: Step 2.3.1), randomly generate l groups of 1×5 random arrays c j =[W fb,j ,A 8,j ,k p,j ,k i,j ,r j ], j = 1, ..., l, composed of Where W fb,j represents the normalized initial state of the fuel in the jth test group, A 8,j represents the normalized initial state of the tail nozzle area in the jth group of tests, k p,j represents the proportionality coefficient in the jth group of experiments, k i,j represents the integral coefficient in the jth group of experiments, r j represents the normalized thrust command in the jth group of trials, and W fb,j ∈[0,1],A 8,j ∈[0,1], k p,j ∈[0,1], k i,j ∈[0,1],r j ∈[0,1]; Step 2.3.2), at the same working point [H, Ma], obtain the maximum thrust F of the current working point respectively max and minimum thrust F min , maximum fuel flow W fb,max and minimum fuel flow W fb,min , maximum tail nozzle area A 8,max and the minimum tail nozzle area, according to step 2.3.1) Perform inverse normalization to obtain the initial fuel, initial tail nozzle area and thrust command value, establish l test samples, and the inverse normalization process is expressed as IN fb,ini,j =In fb,j ·(IN fb,max -IN fb,min )+W fb,min A 8,ini,j =A 8,j ·(A 8,max -A 8,min )+A 8,min r F,j =r j ·(F max -F min )+F min On these l test samples, by presetting the coefficient k p,j and k i,j Based on the PI controller, the thrust is controlled by error feedback. The error value between the thrust estimation and the engine thrust output is recorded at each moment during the test process. The average thrust estimation error e at the working point is calculated. F,mean and single point maximum instantaneous thrust estimation error e F,max ; Step 2.3.3), divide n points at altitude intervals within the full envelope, divide m points at Mach number intervals, and form n×m working points within the full envelope. Perform single-point tests at these working points according to steps 2.3.2) to 2.3.3), and record the single-point average thrust estimation error e of n×m grid points F,mean and single point maximum instantaneous thrust estimation error e F,max ; Step 2.3.4) Estimate the error e based on the single point average thrust F,mean and single point maximum instantaneous thrust estimation error e F,max , calculate the overall average error of thrust estimation in the entire envelope Thrust estimation error variance var(e F,mean ), and the maximum thrust estimation error within the envelope 2. The method for dynamic thrust estimation of a turbofan engine based on a recurrent neural network according to claim 1, characterized in that: The specific steps of step 3) are as follows: Step 3.1), in actual use, based on the sensor measurement information, record the measurable state parameters of the engine to form the engine state x at time t t ; Step 3.2), select the current and previous n-time states of the engine to form the current engine state group X at time t t =[x t-n ,…,x t ]; Step 3.3), group the engine status into X t As the input of the dynamic thrust estimator, it estimates the current thrust of the engine

Images