Variable cycle engine parameter estimation method based on variable activation function neural network

Abstract

The application provides a variable cycle engine parameter estimation method based on a variable activation function neural network, and belongs to the field of intelligent aero-engine control.The application comprises the following steps: collecting data of a variable cycle engine during operation; preparing a two-dimensional input data atlas according to a time sequence; constructing a variable activation function neural network; training the network by using the data atlas, and establishing a nonlinear mapping relationship between the two-dimensional input data atlas and engine parameters; and based on the trained variable activation function neural network, realizing real-time and accurate estimation of the parameters.The parameter estimation method provided by the application has the advantages of high accuracy and fast calculation speed, and is suitable for thrust control and fault-tolerant control of the variable cycle engine.

Description

TECHNICAL FIELD

[0001] The application provides a variable cycle engine parameter estimation method based on an appropriate amount of information fusion convolutional neural network, and belongs to the field of intelligent aero-engine control. BACKGROUND

[0002] An aero-engine is the source power of an aircraft, and its stability directly determines the safety of the aircraft. However, due to the complex structure and harsh working environment, the failure rate is usually high. Engine fault detection and diagnosis technology is an important technical measure to ensure the safe operation of the engine and prolong the service life of the engine, and has very important significance for enhancing the airworthiness of the aircraft and ensuring flight safety. At present, the methods of aero-engine parameter estimation and fault diagnosis mainly include three types: a method based on a physical model, a data-driven method, and a knowledge-based method. In actual problems, due to the complex structure of the aero-engine, the multiple fault categories, and other reasons, it is difficult to accurately and quickly collect data reflecting the characteristics of the engine components, and it is easy to misclassify the first failure. It is difficult to establish an accurate physical model. The prediction complexity of the knowledge-based method is high, and the expert system has the problems of incomplete knowledge and poor adaptability, and is also difficult to implement. The measured data is easy to obtain, so the data-based method is most commonly used in engine parameter estimation. Among them, the fault prediction technology based on machine learning and deep learning model is most commonly used, especially the technology of constructing a neural network model for parameter estimation based on deep learning, which does not require accurate engine component characteristics, and can continuously learn from the historical operation data of the engine, gradually improving the prediction and estimation accuracy during the data accumulation process.

[0003] The patent with publication number CN112131673A discloses an engine surge fault prediction system and method based on a fusion neural network model, which improves the time series prediction method, realizes the generation of a specified length of prediction time series and the extraction of corresponding local features and semantic relationships between data, and further realizes more accurate and rapid advance prediction of engine surge failure, which belongs to a data-driven method. The patent with publication number CN112610339A discloses a variable cycle engine parameter estimation method based on an appropriate amount of information fusion convolutional neural network, which realizes the prediction of the working mode of the variable cycle engine, and belongs to a data-driven method based on a deep neural network, and has excellent feature extraction and nonlinear expression capability. However, most of the current aero-engine parameter estimation methods including the above two patents are based on model or knowledge prediction, and there are very few schemes for predicting based on data to establish a deep learning model or using a machine learning algorithm. Even if there are, the activation functions used are fixed formulas, which are not sufficient to meet the nonlinear demand of a large number of parameter processing and expression.

[0004] In summary, this invention proposes a method for estimating parameters of a variable-cycle engine based on a convolutional neural network with appropriate information fusion. Summary of the Invention

[0005] To address the problems existing in current deep learning algorithms and engine thrust engineering calculation methods, this invention proposes a method for estimating parameters such as thrust of a variable cycle engine based on a neural network with a variable activation function. This method enables real-time and accurate estimation of the parameters to be estimated for a variable cycle engine under multiple operating modes.

[0006] To achieve the above objectives, the concept and technical solution of the present invention are as follows:

[0007] The basic concept of this invention is to use a variable activation function method, making the activation function a function of the input, so that the activation function can change with the input. Traditional neural networks use a fixed activation function, therefore, for the same number of parameters, the variable activation function neural network has a stronger nonlinear expressive capability. For tasks of similar difficulty, because the variable activation function neural network has a stronger nonlinear expressive capability, it requires fewer computational parameters and has a faster computation speed, thus overcoming the problems of slow computation speed and poor real-time performance of deep learning algorithms. Using a variable activation function neural network model for thrust and other parameter estimation can solve the problem of low accuracy in existing engine thrust engineering calculation methods.

[0008] Based on the above basic concept, the technical solution proposed in this invention is a method for estimating parameters of a variable cycle engine based on a variable activation function neural network, comprising the following steps:

[0009] Step 1: Collect data on the operation of the variable cycle engine;

[0010] Step 2: Create a 3D input data atlas based on the time series;

[0011] Step 3: Construct a neural network with variable activation functions;

[0012] Step 4: Train the network using the data atlas to establish a nonlinear mapping relationship between the 3D input data atlas and engine parameters;

[0013] Step 5: Based on the trained variable activation function neural network, achieve real-time and accurate estimation of parameters.

[0014] Furthermore, the data collected in step 1 during the operation of the gas turbine includes engine gas path parameters, environmental variables, control variables, and health parameters.

[0015] Further, the data set in step 2 is m×n×l, where m>1, n>1, l>1, m represents the number of rows in a data set, and the physical meaning of m is: m measurable parameters of the variable cycle aero-engine; n represents the number of columns in a data set, denoted as time t, corresponding to the 1st column, and the pth column corresponding to time t-τ×(p-1), where p is an integer greater than or equal to 1 and less than or equal to n, τ is the time interval for each data acquisition; l represents the number of data sets. The estimated parameters are the parameters at time t.

[0016] Furthermore, the variable activation function layer in step 3 is a function algorithm that uses the activation function as the input of this layer. It is the core algorithm of this patent. Considering that the input between each layer in a neural network structure is generally a matrix or vector, this algorithm is general and can be applied to other neural networks.

[0017] The specific formula for calculating the variable activation function is as follows:

[0018]

[0019] W f This represents the activation function weight matrix, with size (m, n, z), which is the input X. in The function f(x) is related to the weights W. f1 is the leaky RelU function, f2 is the cosine function, f3 is the y = -x function, and so on, indicating that more activation functions can be used. The number is unlimited, but excessive use is not recommended to avoid unnecessary computation. Activation functions do not change the size of the matrix, and n activation functions form a three-dimensional matrix F(X). in ), of size (m, n, z), represents the matrix obtained by applying an activation function to z input matrices of size (m, n). (Symbol) This indicates digit-by-digit multiplication.

[0020] The specific formula for calculating its backpropagation is as follows:

[0021] W(s+1)=W(s)-L out *F(X in )*X in ×W(s) (1-8)

[0022] Its backpropagation formula is as follows:

[0023] L in =X loss *[F(X in )*W+F′(X in )*W(X in (1-8)

[0024] Wherein, F(X) in) is the form of a three-dimensional matrix in equation (1-6) where f1, f2, f3, ...

[0025] Furthermore, step 5, based on the trained variable activation function neural network, achieves real-time and accurate estimation of the parameters of the variable cycle engine. The specific steps are as follows:

[0026] Step 5.1: Collect data of m measurable parameters of the variable cycle engine at the current time t and at times tT, ..., tT*8;

[0027] Step 5.2: Create a two-dimensional input data graph with dimensions m×n based on the time series;

[0028] Step 5.3: Input the two-dimensional data graph into the trained variable activation function neural network to achieve real-time and accurate estimation of the variable cycle engine.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0030] In this invention, a variable activation function method is used, making the activation function a function of the input. This allows the activation function to change with the input, whereas traditional neural networks use a fixed activation function. Therefore, for the same number of parameters, the variable activation function neural network has a stronger nonlinear expression capability. For tasks of similar difficulty, because the variable activation function neural network has a stronger nonlinear expression capability and faster computation speed, it can process more parameters in the same amount of time, thus solving the problem of low accuracy in existing engine thrust engineering calculation methods. This invention achieves a significant improvement in the nonlinear expression capability of the variable activation function neural network while reducing the number of parameters. Furthermore, due to its faster computation speed, the variable activation function neural network can process more parameters in the same amount of time, ultimately achieving real-time and accurate estimation of parameters such as thrust of variable cycle engines. Attached Figure Description

[0031] Figure 1 This is a flowchart of the variable cycle engine parameter estimation method provided by the present invention;

[0032] Figure 2 This is a schematic diagram of a variable activation function neural network provided by the present invention;

[0033] Figure 3 This invention provides parameter estimation based on mobile networks, variable activation function networks, and dense networks. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] Please see Figure 1 The flowchart illustrates a method for estimating parameters of a variable-cycle engine based on a variable activation function neural network, provided by the present invention, which specifically includes the following steps:

[0036] Step 1: Collect data on the operation of the variable cycle engine, including 21 measurable parameters and 5 parameters to be estimated, specifically: flight altitude H, flight Mach number Ma, fuel flow rate Wf, exhaust nozzle area A8, forward adjustable area duct ejector opening FVBE, rear adjustable area duct ejector opening BVBE, mode selection valve opening MSV, fan inlet total temperature T1, fan inlet total pressure P1, CDFS inlet total temperature T2, CDFS inlet total pressure P2, high-pressure compressor inlet total temperature T21, and high-pressure compressor total pressure P21. The following parameters are collected: total inlet pressure of the turbine P21, total outlet temperature of the bypass tunnel T22, total outlet pressure of the bypass tunnel P22, total outlet temperature of the high-pressure turbine T5, total outlet pressure of the high-pressure turbine P5, total outlet temperature of the mixing chamber T7, total outlet pressure of the mixing chamber P7, total outlet temperature of the tail nozzle T8, total outlet pressure of the tail nozzle P8, and the parameters to be estimated: low-pressure rotor speed N1, high-pressure rotor speed Nh, total outlet pressure of the high-pressure compressor P3, total outlet temperature of the high-pressure turbine T5, and total outlet pressure of the low-pressure turbine P6. The data acquisition time is 2000 seconds, and the acquisition period is 0.02 seconds.

[0037] Step 2: Create the output data graphs. The data graph set is m×n×l, where m>1, n>1, l>1. m represents the number of rows in a data graph, and the physical meaning of m is: m measurable parameters of the variable cycle aero-engine; n represents the number of columns in a data graph, where the current time is t, corresponding to the 1st column, and the pth column corresponds to t-τ×p, where p is an integer greater than or equal to 1 and less than or equal to n; τ is the time interval for each data acquisition; l represents the number of data graphs. The estimated parameters are those at time t+τ.

[0038] Step 3: Constructing a variable activation function neural network refers to building a neural network capable of using variable activation functions based on the coupling relationships between parameters of a variable-cycle engine. Figure 2 This provides a schematic diagram of a variable activation function neural network, including an input layer, a feature separation layer, a variable activation function layer, a fully connected layer, and an output layer. For a graphical representation of the backpropagation of the variable activation function layer, please refer to [link to diagram]. Figure 3 The specific descriptions of each layer of this neural network and its forward and backward propagation are as follows;

[0039] 3.1 Input Layer

[0040] The input layer maps and standardizes the input data; the mapping standardization uses the standard deviation standardization method, and the specific calculation formula is as follows:

[0041]

[0042] In the formula, the denominator is the sample standard deviation of the i-th row of the input data;

[0043] 3.2 Output Layer

[0044] The output layer determines the specific form of the loss function during backpropagation and performs the inverse operation of mapping and standardization on the output data of the previous layer before outputting it. The specific formula for calculating the loss function is as follows:

[0045]

[0046] The specific formula for the inverse operation is as follows:

[0047]

[0048] 3.3 Feature Separation Layer

[0049] The feature separation layer is used to extract one-dimensional horizontal features, vertical features, and two-dimensional features from the input feature map. The specific calculation formula for feature separation is as follows:

[0050] X out =sum(X) in .*W,2)+b (1-4)

[0051] The specific calculation formula for its parameter update is as follows:

[0052]

[0053] In the formula, (s+1) in W(s+1) represents the weight W in the (s+1)th training iteration, W(s) represents the weight W in the th training iteration, η is the learning rate, which can be specified manually, and L out It is the input loss of the next layer.

[0054] Its backpropagation formula is as follows:

[0055] L in =W′(S)×L out (1-6)

[0056] 3.4 Variable Activation Function Layer

[0057] The variable activation function layer is a function algorithm that uses the activation function as the input of this layer. It is the core algorithm of this patent. Considering that the input between each layer in a neural network structure is generally a matrix or vector, this algorithm has generality and can be applied to other neural networks.

[0058] The specific formula for calculating the variable activation function is as follows:

[0059]

[0060] Among them W f This represents the activation function weight matrix, with size (m, n, z), which is the input X. in The function f(x) is related to the weights W. f1 is the leaky RelU function, f2 is the cosine function, f3 is the y = -x function, and so on, indicating that more activation functions can be used. The number is unlimited, but excessive use is not recommended to avoid unnecessary computation. Activation functions do not change the size of the matrix, and n activation functions form a three-dimensional matrix F(X). in ), of size (m, n, z), represents the matrix obtained by applying an activation function to z input matrices of size (m, n). (Symbol) This indicates digit-by-digit multiplication.

[0061] The specific formula for calculating its backpropagation is as follows:

[0062] W(s+1)=W(s)-L out *F(X in )*X in ×W(s) (1-8)

[0063] Its backpropagation formula is as follows:

[0064] L in =X loss *[F(X in )*W+F′(X in )*W(X in (1-8)

[0065] Wherein, F(X) in ) is the form of a three-dimensional matrix in equation (1-6) where f1, f2, f3, ...

[0066] 3.5 Fully Connected Layer

[0067] Fully connected layers in this neural network serve to integrate weights. The specific calculation formula for a fully connected layer is:

[0068] X out =W×X in +b (1-9)

[0069] Where W represents the weight matrix, and its number of rows is X. out The number of rows and columns is X. in The number of columns; where b represents the bias vector, having a box X out The same number of rows, one column.

[0070] The specific calculation formula for its parameter update is as follows:

[0071]

[0072] Its backpropagation formula is as follows:

[0073] L in =W′(S)×L out (1-11)

[0074] Step 5: Based on the trained variable activation function neural network, achieve real-time and accurate estimation of the parameters of the variable cycle engine. The specific steps are as follows:

[0075] Step 5.1: Collect data on 21 measurable parameters of the variable cycle engine at the current time t and at times t-0.02, ..., t-0.02*8;

[0076] Step 5.2: Create a two-dimensional input data plot with a dimension of 21×9 based on the time series;

[0077] Step 5.3: Input the two-dimensional data graph into the trained variable activation function neural network. The data passes through an input layer, a feature separation layer, a variable activation function layer, a fully connected layer, and an output layer to achieve real-time and accurate estimation of the variable cycle engine. The output layer outputs the estimated parameters at time t.

[0078] This invention compares the dynamic performance of parameter estimation based on mobile networks, variable activation function neural networks, and dense networks with the test dataset. The simulation results are as follows: Figure 3 As shown.

Claims

1. A method for estimating parameters of a variable-cycle engine based on a neural network with variable activation functions, characterized in that, Specifically, the following steps are included: Step 1: Collect data on the operation of the variable cycle engine; Step 2: Create a 3D input data atlas based on the time series; Step 3: Construct a variable activation function neural network. The variable activation function neural network refers to a neural network that includes an input layer, a feature separation layer, a variable activation function layer, a fully connected layer, and an output layer. The variable activation function layer is a function algorithm that uses the activation function as the input of this layer. Considering that the input between each layer in the neural network structure is generally a matrix or vector, this algorithm has generality and can be applied to other neural networks. The specific formula for calculating the variable activation function is as follows: W f This represents the activation function weight matrix, with size (m, n, z), which is the input X. in The function f1 is the leaky_RelU function, f2 is the cosine function, f3 is the y = -x function, ... indicates that more activation functions can be used here, with no limit on the number, but it is not recommended to use too many to avoid unnecessary calculations. The activation functions do not change the size of the matrix, and n activation functions form a three-dimensional matrix F(X). in ), of size (m, n, z), represents the matrix obtained by the activation function of z input matrices of size (m, n), with symbol . This indicates digit-by-digit multiplication; The specific formula for calculating its backpropagation is as follows: W(s+1)=W(s)-L out *F(X in )*X in ×W(s) (1-8) Its backpropagation formula is as follows: L in =X loss *[F(X in )*W+F′(X in )*W(X in )] (1-8) Wherein, F(X) in ) is the form of a three-dimensional matrix representing f1, f2, f3, ... in equation (1-6); Step 4: Train the network using the data atlas to establish a nonlinear mapping relationship between the 3D input data atlas and engine parameters; Step 5: Based on the trained variable activation function neural network, achieve real-time and accurate estimation of parameters.

2. The method for estimating parameters of a variable-cycle engine based on a variable activation function neural network according to claim 1, characterized in that, Step 1 involves collecting data on the operation of the variable cycle engine, including 21 measurable parameters and 1 parameter to be estimated. Specifically, these parameters are: flight altitude H, flight Mach number Ma, fuel flow rate Wf, exhaust nozzle area A8, mode selection valve opening MSV, forward adjustable area duct ejector opening FVBE, aft adjustable area duct ejector opening BVBE, fan inlet total temperature T1, fan inlet total pressure P1, core engine drive fan CDFS inlet total temperature T2, and core engine drive fan CDF. The parameters are: total inlet pressure P2, total inlet temperature T21 of high-pressure compressor, total inlet pressure P21 of high-pressure compressor, total outlet temperature T22 of bypass, total outlet pressure P22 of bypass, total outlet temperature T5 of high-pressure turbine, total outlet pressure P5 of high-pressure turbine, total outlet temperature T7 of mixing chamber, total outlet pressure P7 of mixing chamber, total outlet temperature T8 of tailpipe, total outlet pressure P8 of tailpipe, and parameter K to be estimated. The number of data acquisitions is N, the acquisition period is T seconds, N is an integer greater than 41, and T is a real number greater than 0.

3. The method for estimating parameters of a variable-cycle engine based on a variable activation function neural network according to claim 1, characterized in that, The data set in step 2 is m×n×l, where m>1, n>1, l>1, m represents the number of rows in a data set, and the physical meaning of m is: m measurable parameters of a variable cycle aero-engine; n represents the number of columns in a data set, where the current time is t, corresponding to the 1st column, and the pth column corresponds to t-τ×p, where p is an integer greater than or equal to 1 and less than or equal to n, and τ is the time interval for each data collection; l represents the number of data sets, and the estimated parameters are the parameters at time t+τ.

4. The method for estimating parameters of a variable-cycle engine based on a variable activation function neural network according to claim 1, characterized in that, Step 5, based on the trained variable activation function neural network, achieves real-time and accurate estimation of the parameters of the variable cycle engine. The specific steps are as follows: Step 5.1: Collect data of m measurable parameters of the variable cycle engine at the current time t and at times tT, ..., tT*8; Step 5.2: Create a two-dimensional input data graph with dimensions m×n based on the time series; Step 5.3: Input the two-dimensional data graph into the trained variable activation function neural network to achieve real-time and accurate estimation of the variable cycle engine.

Images