A method for constructing analytical margin of gas turbine based on variable weight neural network

Abstract

The application provides a variable weight neural network-based gas turbine analytical redundancy construction method, and belongs to the field of intelligent aero-engine control.The application comprises the following steps: collecting gas turbine operation process data;making input data graphs;constructing a variable weight neural network;training the network by using the data graphs;and based on the trained variable weight neural network, realizing real-time and accurate estimation of the analytical redundancy.The analytical redundancy construction method provided by the application has the advantages of strong nonlinear expression capacity and fast calculation speed, and is suitable for gas turbine fault diagnosis and fault-tolerant control.

Description

TECHNICAL FIELD

[0001] The application provides a variable weight neural network-based gas turbine analytical redundancy construction method, and belongs to the field of intelligent aero-engine control. BACKGROUND

[0002] The sensor fault diagnosis and fault-tolerant control of an aero-engine control system usually use a hardware redundancy method, but as the amount of hardware redundancy increases, the weight and volume of the engine will inevitably increase, and even the overall performance of the system will be reduced. In order to avoid the above problems, the analytical redundancy is usually used as the redundancy information, that is, according to the internal relationship between various variables of the engine, the measured value of a certain parameter can be estimated by using the remaining sensors, and the estimated value is the analytical redundancy. This can provide a reference value for sensor fault diagnosis, and according to the residual error between the estimated value and the measured value, it can be judged whether a fault occurs. For sensors without hardware redundancy, the analytical redundancy can also be used to replace the failed sensors to ensure the stable operation of the system. Therefore, the construction of the analytical redundancy of the gas turbine is the basis for the fault diagnosis and fault-tolerant control of the gas turbine, and has important engineering value.

[0003] Machine learning methods can automatically adjust weights and biases to make the nonlinear characteristics of the model consistent with the actual aero-engine. Machine learning mainly includes shallow learning methods and deep learning algorithms. The nonlinear expression ability of shallow learning methods is weak, and the nonlinear expression ability of deep learning algorithms is strong. The estimation accuracy of the analytical redundancy is positively correlated with the nonlinear expression ability. Therefore, in order to obtain high-precision analytical redundancy, it is necessary to improve the nonlinear expression ability. Typical shallow learning methods include support vector machines. The patent with publication number CN112052871A discloses a rocket engine health diagnosis method and system based on a support vector machine. The support vector machine is used to diagnose the real-time working parameters of the rocket engine. This method belongs to a shallow learning method. The parameters of the shallow learning method must be updated to adapt to the nonlinear characteristics of the engine at different operating points. Generally, it needs to be established on the common operating line or the boundary. The nonlinear expression ability is low, and it cannot meet the strong nonlinear characteristics of the gas turbine, especially the variable cycle engine.

[0004] Deep learning algorithms can solve strong nonlinear problems, typical of which are long short-term memory neural networks (LSTM) and convolutional neural networks (CNN). The patent with publication number CN111639467A discloses an aero-engine life prediction method based on a convolutional neural network and a long short-term memory neural network. The convolutional neural network and the long short-term memory neural network are used to predict the life of the aero-engine. This method belongs to a deep learning algorithm. Although deep learning algorithms can enhance nonlinear expression ability, a large number of parameters and slow computing speed are unacceptable for gas turbines with high real-time requirements.

[0005] In summary, there is no good method to reduce the number of parameters while improving the nonlinear expression ability of the algorithm, so as to improve the accuracy and calculation speed of the analytical margin and apply it to the gas turbine, a new gas turbine analytical margin construction method based on variable weight neural network can effectively solve this problem. SUMMARY

[0006] In view of the problems existing in the existing shallow learning method and deep learning algorithm, the present application proposes a gas turbine analytical margin construction method based on variable weight neural network, which realizes real-time and accurate estimation of the analytical margin of the gas turbine.

[0007] To achieve the above-mentioned purpose, the concept and technical solution of the present application are realized as follows:

[0008] The basic concept of the present application is that by the method of variable weight, the weight value becomes a function of the input, so that the weight value can change with the input, while the weight value of the traditional neural network is a constant value, so for the same number of parameters, the nonlinear expression ability of the variable weight neural network is stronger. For the same difficulty task, because the nonlinear expression ability of the variable weight neural network is stronger, fewer operation parameters are needed, and the calculation speed is faster, which can overcome the problems of slow calculation speed and poor real-time performance of deep learning algorithm; using the variable weight neural network model for analytical margin estimation can solve the problem of slow calculation speed of the gas turbine analytical margin.

[0009] Based on the above basic concept, the technical solution proposed by the present application is a gas turbine analytical margin construction method based on variable weight neural network, comprising the following steps:

[0010] Step 1: Collecting gas turbine operation process data;

[0011] Step 2: Making input data graph;

[0012] Step 3: Constructing variable weight neural network;

[0013] Step 4: Training the network using the data graph;

[0014] Step 5: Based on the trained variable weight neural network, realizing real-time and accurate estimation of the analytical margin.

[0015] Further, the collecting of gas turbine operation process data in step 1 includes engine gas path parameters, environmental variables, control variables and health parameters.

[0016] Further, the data graph in step 2 has a dimension of m*n, m>1, n>1, m represents the total number of rows, n represents the total number of columns, each row represents different kinds of running data, and each column represents different time points; when the current time point is t, the nth column corresponds to the t time point, and the pth column corresponds to the t-n+p time point, p is an integer greater than or equal to 1 and less than or equal to n, and the constructed analytic margin is the analytic margin at the t time point.

[0017] Further, the variable weight neural network in step 3 comprises an input layer, a feature separation layer, a variable weight layer, an activation layer and an output layer.

[0018] wherein,

[0019] The input dimension of the feature separation layer is m*n, and the output dimension is m*1.

[0020] wherein,

[0021] The output y of the variable weight layer is W(x)*x+o, that is, the weight W is a nonlinear mapping of the input x, without loss of generality, the nonlinear mapping can be sigmoid, leaky relu, sin(x), cos(x), e x Any one. The mapping relationship of sigmoid and leaky relu is as follows:

[0022]

[0023] a is a constant much smaller than 1, without loss of generality, which can be 0.01.

[0024] Further, step 4 utilizes the data graph to train the network.

[0025] Further, step 5 realizes real-time and accurate estimation of the analytic margin based on the trained variable weight neural network.

[0026] Compared with the prior art, the beneficial effects of the present application are:

[0027] In the present application, by constructing a variable weight neural network model, using a variable weight method, by making the weight value a function of the input, the weight value can change with the input, while the weight value of the traditional neural network is a constant value, so for the same number of parameters, the variable weight neural network has stronger nonlinear expression ability. For tasks of the same difficulty, because the variable weight neural network has stronger nonlinear expression ability, it requires fewer operation parameters and faster calculation speed, which can overcome the problem of slow calculation speed and poor real-time performance of deep learning algorithms. The nonlinear expression ability of the variable weight neural network can be significantly improved while reducing the number of parameters, and because the variable weight neural network has fewer calculation parameters and faster calculation speed than other algorithms, the real-time and accurate estimation of the analytical margin of the gas turbine is finally realized. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 is a flow chart of the gas turbine analytical margin estimation method provided by the present application;

[0029] Figure 2 is a structural schematic diagram of the variable weight neural network provided by the present application;

[0030] Figure 3 is a variable weight layer l x a graph back propagation diagram;

[0031] Figure 4 is a T5 analytical margin based on a mobile network, a variable weight network and a dense network provided by the present application; DETAILED DESCRIPTION

[0032] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0033] Please refer to Figure 1 which shows a flow chart of a variable weight neural network-based gas turbine analytical margin construction method provided by the present application, specifically including the following steps:

[0034] Step 1: collect gas turbine operation data, including engine gas path parameters, environmental variables, control variables, health parameters, select a more complex variable cycle engine as the research object, which includes 21 measurable parameters and 5 to be estimated parameters, specifically: flight altitude H, flight Mach number Ma, fuel flow Wf, tail nozzle area A8, front adjustable area bypass ejector opening FVBE, rear adjustable area bypass 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, high pressure compressor inlet total pressure P21, outer bypass outlet total temperature T22, outer bypass outlet total pressure P22, high pressure turbine outlet total temperature T5, high pressure turbine outlet total pressure P5, mixing chamber outlet total temperature T7, mixing chamber outlet total pressure P7, tail nozzle outlet total temperature T8, tail nozzle outlet total pressure P8 and to be estimated parameters low pressure rotor speed Nl, high pressure rotor speed Nh, high pressure compressor outlet total pressure P3, high pressure turbine outlet total temperature T5, low pressure turbine outlet total pressure P6, wherein the collection time is 2000 seconds, and the collection period is 0.02 seconds;

[0035] Step 2: make an input data graph, the data graph has dimensions m*n, m>1, n>1, m represents the total number of rows, n represents the total number of columns, each row represents different types of operation data, and each column represents different time, when the current time is t time, the nth column corresponds to the t-n+p time, p is an integer greater than or equal to 1 and less than or equal to n, and the analytical margin constructed is the analytical margin of the t time.

[0036] Step 3: build a variable weight neural network, which is a neural network capable of variable weight built according to the coupling relationship between the gas turbine parameters, Figure 2 A structural diagram of the variable weight neural network is provided, which includes an input layer, a feature separation layer, a variable weight layer, an activation layer and an output layer, the variable weight layer l x The graph back propagation is described in Figure 3 ;

[0037] Among them,

[0038] The variable weight layer is a function algorithm taking the weight as input; the variable weight layer output of the variable weight neural network is y=W(x)*x+o, that is, the weight W is a nonlinear mapping of the input x; the variable weight algorithm is a mapping relationship from input to weight, which can be used for fully connected layers, long short term memory neural network (LSTM) layers and convolutional layers. Considering that any input array can be converted into a vector, without loss of generality, the input and output are both vectors. The forward propagation specific calculation formula is:

[0039]

[0040] where x is an input vector (m, 1), is a variable weight tensor with size (m, n, m), W is a weight matrix (m, n), B is a bias matrix (m, n), o is a bias vector (n, 1), and T represents transposition. The first dimension of all arrays is row, the second dimension is column, and the third dimension is page. +, -, *, / , ^ represent element-wise operations, and the specific calculation formula is:

[0041]

[0042] where if the dimensions of two arrays do not match, the lower-dimensional array will be expanded to a higher dimension by copying the missing dimension. X is a matrix of m rows and n columns, denoted as (m, n), and after sum(X, 2), X is converted into an x vector with size (m, 1). sum(array, k) is a function that adds array elements after the dimension, satisfying k th The dimensions. If the array of flip() is a vector or matrix, the missing dimension will be assigned to 1;

[0043] W does not need to be updated because the weight is determined by the variable weight tensor. The specific calculation formula for back propagation is:

[0044]

[0045] o = o - η · d o

[0046]

[0047] where, is the variable weight tensor gradient, is the bias gradient, d o is the offset bias gradient;

[0048] The input layer maps and normalizes the input data, and the specific calculation formula is:

[0049] Y = (x - μ) / σ

[0050] where the input data mapping size of X is (m, n), μ is a constant column vector whose row elements are the average of each row, and σ is a column vector whose row elements are the average of each row and no element is equal to zero. The output data mapping size of Y is (m, n);

[0051] The activation layer plays a role in introducing activation functions. In this paper, the LeakyReLU function is considered to have good performance and is used as the activation function. The specific calculation formula for forward propagation is:

[0052] y = act(x)

[0053]

[0054] The specific calculation formula of back propagation is:

[0055]

[0056] Wherein, l y is the output loss, l x is the input loss;

[0057] The input dimension of the feature separation layer is m x n, and the output dimension is m x 1, and the specific calculation formula is:

[0058] y = sum(X.*W, 2) + b

[0059] Wherein, X is the input data mapping, the size is (m, n). W is the weight matrix, the size is (m, n), b is a bias vector, the size is (m, 1), y is the output data, the size is (m, 1);

[0060] The specific calculation formula of back propagation of feature separation layer is:

[0061] L x = l y .*W

[0062] D W = l y .*X

[0063] d b = l y

[0064] W = W - η·D W

[0065] b = b - η·d b

[0066] Wherein, l y is the output loss, L X is the input loss, X is the input matrix of the feature input layer, D W is the gradient of weight, d b is the gradient of bias, and η is the learning rate;

[0067] Step 4: train the network by using data graph;

[0068] Step 5: input two-dimensional data graph into the trained variable weight neural network to realize real-time and accurate estimation of gas turbine analytical margin.

[0069] In order to verify the effect of the variable weight neural network-based gas turbine analytical margin construction method provided by the application, two-dimensional input data graphs are made by using 21 measurable parameter real-time data of the variable cycle engine, the variable weight neural network trained is input, ten steady state points are randomly selected to perform real-time estimation on the analytical margins of N l , P3, T5 of the variable cycle engine, and the analytical margin estimation effect is as shown in Table 1:

[0070] Table 1 Variable weight neural network-based N1, P3, T5 analytical margins of ten steady state points

[0071]

[0072] Among them, AV represents the actual value, EV represents the estimated value, ARE represents the average of all relative errors, and RE represents the relative error, and the specific calculation formula is:

[0073]

[0074] Table 1 shows the variable weight neural network-based N1, P3, T5 analytical margins, and the ARE of the three is less than 1.5%, so it can be concluded that the analytical margin based on the variable weight neural network meets the requirements.

[0075] In recent years, neural networks have been applied to many fields due to their strong learning ability, among which typical ones are mobile networks and dense networks. The dense network improves the accuracy by increasing the number of parameters at the expense of the calculation speed, and the mobile network focuses on the real-time performance of operation and seeks a balance between the number of parameters and the calculation accuracy. Both kinds of networks have advantages and disadvantages, so the application compares the dynamic performance of the T5 analytical margin based on the mobile network, the variable weight neural network and the dense network with the test data set, and the simulation results are as shown in Figure 4 It can be seen that for the estimation of the T5 analytical margin, the accuracy of the dense network and the variable weight neural network is similar, and both are better than that of the mobile network. Therefore, it can be proved that the variable weight neural network has high estimation accuracy in the analytical margin estimation task, and in some cases, the accuracy is higher than that of the mobile network.

[0076] The training time, the number of forward propagation layers and the calculation time of the three networks with a batch size of 128 for 6 periods are as shown in Table 2:

[0077] Table 2 Training time, number of layers and calculation time of the variable weight network, the mobile network and the dense network

[0078]

[0079] From Table 2, it can be seen that the variable weight neural network is the most compact algorithm, and the training time, the number of forward propagation layers and the calculation time are all much lower than those of the other two networks.

[0080] The average value of the relative error of the analytical margin and the calculation efficiency of N1, P3 and T5 are shown in Table 3:

[0081] Table 3 The average value of the relative error of the analytical margin and the calculation efficiency of N1, P3 and T5

[0082]

[0083] wherein, the calculation efficiency E c The specific calculation formula is:

[0084]

[0085] From Table 3, it can be seen that in terms of analytical margin estimation, the mobile network and the variable weight neural network have similar accuracy, and the dense network is not accurate in N1 analytical margin estimation. Although the dense network is relatively accurate in P3 and T5 analytical margin estimation, the high accuracy comes at the cost of a large number of parameters and a long calculation time. Obviously, the analytical margin estimation method based on the variable weight neural network proposed in the present application can meet the accuracy requirement of the analytical margin estimation of the gas turbine, and the calculation speed is significantly better than that of the mobile network and the dense network.

[0086] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can make equivalent replacement or change according to the technical solution and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered in the protection scope of the present application.

Claims

1. A method for constructing analytical margins of a gas turbine based on a variable weight neural network, characterized in that, Specifically comprising the following steps Step 1: Collecting m operating data of the gas turbine at time t, t-1,..., t-n+1, including engine gas path parameters, environmental variables, control variables, and health parameters; Step 2: Making a two-dimensional input data graph with a dimension size of m*n, the data graph has a dimension of m*n, m>1, n>1, m represents the total number of rows, n represents the total number of columns, each row represents a different type of operating data, and each column represents a different time, with the current time being t, the nth column corresponds to t, and the pth column corresponds to t-n+p, p is an integer greater than or equal to 1 and less than or equal to n, and the constructed analytical margin is the analytical margin at time t; Step 3: Constructing a variable weight neural network, including an input layer, a feature separation layer, a variable weight layer, an activation layer, and an output layer; the variable weight layer of the variable weight neural network outputs y=W(x)*x+o, that is, the weight W is a nonlinear mapping of the input x, and o is an offset vector; the variable weight neural network is a mapping relationship from input to weight, used for fully connected layers, long short-term memory neural network LSTM layers, and convolutional layers; the input and output are both vectors; Step 4: Training the network using the data graph; Step 5: Based on the trained variable weight neural network, real-time and accurate estimation of the analytical margin is realized.

2. The method of claim 1, wherein the method is based on a variable weight neural network. The forward propagation specific calculation formula is: where x is an input vector (m, 1), is a variable weight tensor with size (m, n, m), W is a weight matrix (m, n), B is a bias matrix (m, n), o is a bias vector (n, 1), and T represents transposition. The first dimension of all arrays is row, the second dimension is column, and the third dimension is page. * represents an element-wise operation, and the specific calculation formula is: where if the dimensions of the two arrays do not match, the lower dimensional array will be expanded to the higher dimension by replicating the missing dimensions; X is a matrix of size m rows, n columns, denoted as (m, n), after sum(X, 2), X is converted to an x vector of size (m, 1); sum(array, k) is a function that adds the elements of an array after the dimension, satisfying k th dimensions; if the array of flip() is a vector or matrix, the missing dimension will be assigned 1.

Images