A method for describing process characteristics of a turbofan engine by using a multi-feature fusion residual network

Abstract

The application discloses a method for describing the process characteristics of a turbofan engine by using a multi-feature fusion residual network, and relates to the following technical scheme: feature selection is completed by using a correlation analysis method based on collected engine ground test data; data preprocessing is performed by using a wavelet transform threshold filtering method and standardization; each feature data after filtering is fused into a three-dimensional matrix; and the three-dimensional matrix is input into a Resnet residual network added with a wavelet transform filtering structure to obtain final data. The application realizes the elimination of redundant variables in the test data and the coarse analysis of variable correlation, and uses the correlation coefficient between each group of variables as the basis and the maximum mutual information coefficient as the supplement, so that the problems of narrow application range and low modeling precision of a traditional aerodynamic thermodynamic model are avoided. By using the feature extraction capability of the residual network, a complete and accurate nonlinear multivariable prediction model is established, and the model has a strong noise suppression capability, which effectively assists the subsequent multivariable controller design and fault diagnosis of the aero-engine.

Description

Technical Field

[0001] This invention belongs to the field of high-precision modeling technology for key parameters of turbofan engines, and specifically relates to a method for describing the process characteristics of turbofan engines using a multi-feature fusion residual network. Background Technology

[0002] Turbofan engines, as the heart of aircraft, hold a pivotal position in the aviation industry. Research on turbofan engines is inseparable from the description of their process characteristics. A good model is the foundation for subsequent design and manufacturing of aero-engines, effectively reducing unnecessary experiments and enabling more accurate and reliable controller design.

[0003] Existing methods for describing the process characteristics of turbofan engines mainly fall into two categories: mechanistic description methods based on thermodynamics and kinetics, and identification description methods based on experimental data. The first method decomposes the turbofan engine according to the direction of airflow, describes the process characteristics of each component individually, and then uses the common working equations of each component to connect them according to the engine structure diagram to establish an overall process characteristic description of the engine. Since the equations are all nonlinear implicit equations, although this method is applicable to the entire working process of a turbofan engine, it cannot obtain accurate analytical solutions, and changes in the characteristic curves of each component under different operating conditions will affect the model accuracy.

[0004] The characterization of the identification process based on experimental data, without considering the internal structural principles of aero-engines, can be categorized into two types based on their underlying principles: identification description methods based on system identification algorithms and artificial intelligence description methods based on machine learning algorithms. The former utilizes the Volterra series model and recursive least squares to establish a third-order single-variable Volterra series model for the non-afterburning main control system of an aero-engine. However, as the number of variables, system order, and memory length increase, the number of parameters to be identified increases exponentially. Although a sparse model of the Volterra series can be obtained using kernel symmetry, some nonlinear characteristics are still lost. Machine learning methods include artificial neural networks, but due to the limited accuracy of traditional BP networks, they cannot be used for strongly nonlinear systems, and they suffer from the problem of gradient vanishing with large computational loads and deep layers. Furthermore, when considering the data memory length, the number of input layer nodes is relatively large. Therefore, convolutional neural networks are considered to solve the memory length problem by using a multidimensional matrix approach. However, traditional convolutional neural networks do not solve the gradient vanishing problem, which not only leads to slow training speed and low utilization of network weights, but also results in large errors in feature description results, which is not conducive to subsequent research work. For example, the design of multivariable controllers for engines and engine fault detection both require the good and accurate process characteristics of turbofan engines as a basis. However, the existing process characteristic descriptions restrict the design of multivariable controllers for engines and engine fault detection. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, the present invention aims to provide a data analysis method for turbofan engines using a multi-feature fusion residual deep learning network, thereby improving the modeling accuracy of multivariables in aero-engines, reducing the errors of key engine parameters (such as high-pressure shaft speed) under operating conditions, establishing a complete and accurate nonlinear multivariable prediction model, and expecting the average relative error of engine high-pressure speed to be no greater than 0.40%, the average error of low-pressure turbine inlet temperature to be no greater than 8K, the average relative error of engine thrust to be no greater than 1.5%, and the average relative errors of nozzle position indication (NPI) and low-pressure speed to be no greater than 0.8%, ultimately achieving design guidance for engine multivariable controllers and accurate engine fault detection.

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

[0007] A method for describing the process characteristics of a turbofan engine using a multi-feature fusion residual network includes the following steps:

[0008] Step 1: Correlation analysis. Collect engine ground test data and use correlation analysis to select 9 input variables and 5 output variables.

[0009] Step 2: Data preprocessing. First, use wavelet threshold filtering to reduce noise in the data, and then perform standardization.

[0010] Step 3: Feature fusion, fusing the data obtained in Step 2 into a 9*4*4 three-dimensional data matrix;

[0011] Step 4: Network learning. The matrix obtained in Step 3 is used as the input data of the ResNet residual network. A wavelet transform filter structure is added after the network output layer to obtain the final result after filtering.

[0012] Step 5: Use the final results to describe the process characteristics of the turbofan engine, and based on these process characteristics, guide the design of the engine's multivariable controller or perform fault diagnosis.

[0013] Compared with existing technologies, the multivariate correlation analysis method proposed in this invention combines the advantages of Spearman correlation analysis and maximum mutual information coefficient analysis, achieving refined analysis of correlations among multiple variables. Furthermore, a multi-feature fusion residual deep network is constructed based on the characteristics of turbofan engine test data. The three-dimensional matrix input format solves the problem of excessive input nodes when considering data memory length in traditional networks. The noise reduction method used in the multi-feature fusion residual deep network proposed in this invention, employing a ResNet residual network, not only improves the gradient vanishing problem but also accelerates training speed while maintaining accuracy. The signal is decomposed and reconstructed using the Mallet algorithm under a 3-layer Sym5 wavelet basis. The minimax thresholding method and Garrote thresholding function are used for wavelet coefficients, making it more suitable for practical engineering applications. In addition, when describing the process characteristics of turbofan engines, this invention presents a mapping relationship between the numerical values ​​and relative changes of multiple parameters and various engine states. This is beneficial for clarifying the relationship between fault states and signals in subsequent fault diagnosis and for selecting control parameter combinations in control law design. Attached Figure Description

[0014] Figure 1 This is a flowchart of the overall method of the present invention.

[0015] Figure 2 This is a flowchart illustrating the specific implementation method of the present invention.

[0016] Figure 3 The flowchart shows the calculation process for the maximum mutual information coefficient.

[0017] Figure 4 The diagram illustrates the principle of three-level wavelet decomposition using the Mallat decomposition algorithm.

[0018] Figure 5 This is a graph of the Sym5 wavelet basis functions.

[0019] Figure 6 The data denoising effect diagrams are as follows: a is the wavelet denoising effect diagram of engine test data variable P3, b is the wavelet denoising detail diagram of engine test data variable P3, c is the wavelet denoising effect diagram of engine test data variable T6, and d is the wavelet denoising detail diagram of engine test data variable T6.

[0020] Figure 7 This is a flowchart of the feature fusion process for step three.

[0021] Figure 8 This is a diagram of the residual block structure.

[0022] Figure 9 This is a diagram showing the overall structure of the residual network for a turbofan engine.

[0023] Figure 10The modeling results for various engine parameters are as follows: a is the modeling result for low-pressure shaft speed, b is the modeling result for high-pressure shaft speed, c is the modeling result for low-pressure turbine after-temperature, d is the modeling result for engine thrust, and e is the modeling result for nozzle position indication. Detailed Implementation

[0024] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings and examples.

[0025] refer to Figure 1 and Figure 2 This invention provides a method for describing the process characteristics of a turbofan engine using a multi-feature fusion residual network, comprising the following steps:

[0026] Step 1: Collect engine ground test data and complete feature selection through correlation analysis.

[0027] Specifically, in this invention, Spearman correlation analysis and maximum information correlation analysis were used to perform correlation analysis on engine ground test data and complete feature selection, ultimately selecting 9 input variables and 5 output variables.

[0028] Spearman correlation analysis introduces relative rank differences, which can be used to calculate the Spearman rank correlation coefficient between groups of variables. For non-monotonic data with complex monotonicity, the maximum mutual information coefficient analysis method is needed for supplementary analysis. The maximum mutual information coefficient correlation analysis method, based on mutual information, measures the degree to which a change in one variable causes a change in another variable. The formula is as follows:

[0029]

[0030] Where I(X,Y) represents the mutual information between two sets of variables, p(x,y) is the joint probability distribution of the two sets of variables, and p(x) and p(y) are the marginal probability distributions of each set of variables. Using the concept of a two-dimensional scatter plot, a non-equidistant optimization method is employed to calculate the mutual information between variables, and then normalization is performed to obtain the maximum mutual information coefficient between the variables. The calculation steps are as follows: Figure 3 As shown.

[0031] The overall correlation analysis formula can be obtained as follows:

[0032]

[0033] set up The formula for calculating the correlation coefficient is as follows:

[0034]

[0035] Where, x i (t), xj R(x) represents two different variables at time t, respectively. i (t)), R(x j (t) represent the position of the variable, respectively. The value represents the average rank of the variable; `sort()` is the sorting function; `r()` is the function to calculate the relative rank difference between any two distinct variables at different times. i ||1=||x j ||1 = n, where n represents the size of the variable, and M represents the total number of variables in the dataset. Represents variable x i x j The updated correlation coefficient, where G represents different meshing schemes for the plane, p(x i (t),x j Let p(x) be the joint probability distribution of the two sets of variables. i (t)) and p(x j (t) represents the marginal probability distribution of each group of variables, and ε represents the set high correlation threshold parameter.

[0036] First, based on the knowledge of turbofan engine mechanism modeling and analysis, the high-pressure shaft speed (NH), low-pressure shaft speed (NL), engine thrust (XG), nozzle position indication (NPI), and low-pressure turbine outlet temperature (T6) are respectively used as variables x. i The remaining 34 sets of variables (such as throttle lever angle, high-pressure compressor inlet guide vane angle position, low-pressure compressor afterburner pressure, low-pressure compressor afterburner temperature, high-pressure compressor afterburner pressure, high-pressure compressor afterburner temperature, fuel flow rate, lubricating oil pressure, etc.) are used as variable x. j Substituting these values ​​into the aforementioned correlation analysis method (i.e., Spearman correlation analysis), we calculate the correlation coefficient (in this example, the Spearman rank correlation coefficient) and the average of their absolute values. We then sort the data variables in descending order of their average values ​​and select the first 8 groups. Next, we analyze the variable x... i For variables with relatively low correlation coefficients, the maximum mutual information (MIC) coefficient correlation analysis method is used. Based on the magnitude of the correlation coefficient and MIC coefficient between the two variables, conclusions about moderate or high correlation between the variables are obtained.

[0037] Finally, nine input variables were extracted based on the results: throttle lever angle (PLA), high-pressure compressor inlet guide vane angle position (IGV), low-pressure compressor afterburner pressure (P2), low-pressure compressor afterburner temperature (T2), high-pressure compressor afterburner pressure (P3), high-pressure compressor afterburner temperature (T3), fuel flow rate (FE_small), lubricating oil pressure (P_slippery), and intake manifold airflow meter static pressure (PQ2). Five output variables were extracted: high-pressure shaft speed (NH), low-pressure shaft speed (NL), engine thrust (XG), nozzle position indication (NPI), and low-pressure turbine outlet temperature (T6).

[0038] Step 2: Data preprocessing. First, use wavelet threshold filtering to reduce noise in the data, and then perform standardization.

[0039] Specifically, in this step, outlier removal is first performed on the data. This invention uses the method of replacing the median of adjacent data to modify outliers into normal values. Then, the second-order spline interpolation method in the piecewise interpolation method is used to perform data interpolation. This increases the amount of data and improves the modeling accuracy without destroying the original data distribution characteristics.

[0040] Next, wavelet threshold filtering is used to denoise the data. The specific method is described as follows:

[0041] First, the noisy data to be processed is decomposed using wavelet transform. This invention utilizes the Mallet decomposition algorithm to perform multi-level decomposition of the noisy signal, i.e., each variable, with a decomposition level of 3. The principle is as follows: Figure 4 As shown. The specific decomposition results of each layer are as follows:

[0042]

[0043]

[0044] Where the decomposition levels l = 1, 2, 3, t = 1, 2, ..., N, and N represents the number of points in the wavelet transform sequence; x l,L [t] represents the low-frequency component of the l-th layer at time t, x l,H [t] represents the high-frequency component of the l-th layer at time t; x[t] represents the original signal; h and g represent a pair of orthogonal filter banks; h[2k-t] and g[2k-t] are the k-th coefficients of the low-pass filter and the high-pass filter, respectively, and K represents the filter length.

[0045] This invention primarily employs Symlet wavelet basis functions to denoise the data. Figure 5 The curves of the Sym5 wavelet basis are given. Based on the data characteristics and denoising requirements, an appropriate threshold function and corresponding threshold are used to process the high-frequency coefficients after wavelet decomposition according to the threshold function, so as to obtain the denoised wavelet coefficients of each frequency band.

[0046] The engine test data used in this invention has a high signal-to-noise ratio for most variables, and adopts a maximum-minimum threshold setting method, with the calculation formula as follows.

[0047]

[0048] Where N is the data length.

[0049] To ensure a smooth transition of the threshold function at the threshold point while minimizing bias caused by threshold processing, this invention employs a classic and concise improved threshold function, the Garrote function. Its function expression is as follows:

[0050]

[0051] Where ω is the wavelet coefficient before processing, and γ is the set threshold.

[0052] Based on the above, the formula for applying the minima threshold to wavelet coefficients at each level can be obtained as follows:

[0053]

[0054] in

[0055] Data noise reduction effect such as Figure 6 As shown in the diagram, a) shows the wavelet denoising effect of engine test data variable P3, b) shows the detailed wavelet denoising effect of engine test data variable P3, c) shows the wavelet denoising effect of engine test data variable T6, and d) shows the detailed wavelet denoising effect of engine test data variable T6. It can be seen that the denoised data is smoother, and the signal-to-noise ratio is significantly reduced.

[0056] The denoised wavelet coefficients are then reconstructed using wavelet coefficients to obtain the denoised data. Furthermore, to increase the comparability between data of different variables, data standardization is required after denoising. This invention employs a standardization method, transforming the data into a standard normal distribution. The mathematical relationship between the relative error and the directly calculated result and the expected value after standardization is as follows.

[0057]

[0058] Where Y s The variables are the standardized results of the calculation after inverse normalization.

[0059] Step 3: Feature fusion. The data obtained in Step 2 is fused into a 9*4*4 three-dimensional data matrix.

[0060] In this step, the feature variables obtained after processing in step two are fused into a three-dimensional matrix. First, each set of input variable data is truncated using a sampling window of 16 data points. Then, each set of 16 data points is transformed into a 4x4 two-dimensional data matrix. Finally, the 4x4 two-dimensional data matrices obtained from a total of 9 sets of input variables are fused into a 9x4x4 three-dimensional data matrix, and the output value corresponding to each matrix is ​​determined. The feature fusion process is as follows: Figure 7 As shown.

[0061] Step 4: Network learning. The matrix obtained in Step 3 is used as the input data of the ResNet residual network. A wavelet transform filter structure is added after the network output layer to obtain the final result after filtering.

[0062] The ResNet residual network of this invention can be described as follows:

[0063] For the turbofan engine ground test dataset, a multi-feature fusion residual deep network is constructed using a residual network structure. High and low layer features are fully extracted and combined through a 4-layer residual structure. A wavelet transform filter structure is added after the network output layer, and the filtered result is used as the final calculation result. The input memory length is determined by the input characteristics of the turbofan engine ground test dataset. The ReLU function is used as the activation function, and the convolutional layer adopts a 3×3 convolutional kernel structure. The network is trained using mini-batch stochastic gradient descent (MBGD), and the L1 loss function is selected as the absolute value error function.

[0064] Specifically, in this step, the ResNet residual network algorithm is used to model the ground test data of the aero-engine, and its input characteristics are used to determine the input memory length. The residual block structure is as follows: Figure 8 As shown, a basic unit of a residual block typically includes two convolutional layers, two ReLU activation layers, two batch normalization (BN) layers, and one shortcut. The input is x, the output is H(x), and the non-short-cut part calculates the output F(x), so H(x) = F(x) + x.

[0065] Since the data preprocessing in step two results in negative values, and the ReLU function is less sensitive to negative values, this invention improves upon it by using the hyperbolic tangent function, as shown in the following formula.

[0066]

[0067] Batch Normalization (BN) performs a standardization operation on the variables of each intermediate layer to make them conform to a standard normal distribution with a mean of 0 and a variance of 1. Then, the network's expressive power is improved through scaling and other operations on the results. The formula is as follows:

[0068]

[0069] y i =αp i +β (12)

[0070] To address the high-accuracy modeling requirement of high-pressure speed in the output variables, a separate network was designed for it. A four-layer residual structure was used to fully extract and combine features from both high and low layers. A wavelet transform filter was added after the network output layer, and the filtered result was used as the final calculation result. The overall structure of the turbofan engine residual network is as follows: Figure 9 As shown.

[0071] To ensure both training speed and result accuracy, mini-batch stochastic gradient descent (MBGD) is used, and the L1Loss loss function is selected for network training, as defined below.

[0072]

[0073] Where y(i) is the estimated value and Y(i) is the true value.

[0074] The experimental training epoch length was set to 1500, the learning rate to 0.0001, the weight decay to 8e-4, and the batch size for batch normalization to 256. The learning rate was set to decrease stepwise to 0.1 times its original value at 5, 10, and 15 epochs. The experimental equipment used was an MSI GE75 laptop with an Intel Core i7-8750 CPU and an NVIDIA RTX 2070 GPU. The experimental results showed an average MSE of 4.0e-04 and an average relative error of 0.8%. The results are as follows... Figure 10 As shown, where, Figure 10 'a' represents the modeling results for the low-pressure shaft speed. Figure 10 b represents the modeling result of the high-pressure shaft speed. Figure 10 c represents the low-pressure turbine after-temperature modeling result. Figure 10 d represents the thrust modeling result. Figure 10 e represents the modeling result for the nozzle position indication.

[0075] After inverse normalizing the results data, the average relative error of each index was calculated. The average error for NL was 0.54%, for NH 0.36%, for T6 0.3%, for XG 1.64K 1.37%, and for nozzle position indication 0.78%.

[0076] Step 5: Real-time operating parameters from airborne sensors are processed by a ground station system embedding a multi-feature fusion residual network to obtain process characteristic parameters. These parameters are compared with the standard performance parameters provided by the engine manufacturer to identify deviations. Through deviation analysis and trend analysis, the engine process characteristics are described. The predicted variables accurately describe the nonlinear dynamic characteristics during engine testing and exhibit strong noise suppression capabilities, avoiding the narrow application range and low modeling accuracy issues of traditional aero-thermodynamic models. Therefore, based on these process characteristics, different controlled parameters can be controlled in different operating segments according to engine operating conditions. A control law is selected that achieves the thermodynamic and strength characteristics specified in the engine design, ensuring aerodynamic stability during operation. The deviations between the measured engine parameter values ​​and baseline values, representing different trends in measured parameter changes, are selected as input data for the intelligent fault diagnosis model. The corresponding engine fault modes are taken as the output of the diagnostic model, thereby establishing a multi-class classifier.

[0077] In practice, the system for describing the process characteristics of this invention mainly includes various sensors for collecting engine test data, a network model, and an output section. The network model can execute steps two to four as described above, while the output section can provide the described process characteristics to the fault intelligent diagnosis model, which outputs a fault model or outputs it to a processor. The model is then compared with the standard performance parameters given by the engine manufacturer, and deviation data is output. The deviation data can be represented in the form of numbers, curves, images, etc. This deviation data can be input into the parameter module of the engine controller to guide and adjust the design of the engine controller.

[0078] In summary, this invention achieves the removal of redundant variables and coarse correlation analysis of variables in test data. It also provides a relatively accurate description of the dynamic process of the non-afterburning main control system of a certain type of turbofan engine. Verification was conducted using identification data and non-identification data from similar stages of the same engine. The comparison shows that, in both small and large deviation dynamic processes with different throttle thrust and yield rates, the predicted values ​​of each variable in this invention's model achieve good accuracy compared to actual ground test data.

Claims

1. A method for describing the process characteristics of a turbofan engine using a multi-feature fusion residual network, characterized in that, Includes the following steps: Step 1: Correlation analysis. Ground test data of the engine was collected, and correlation analysis was used to identify 9 input variables and 5 output variables. The correlation analysis method was as follows: set up The formula for calculating the correlation coefficient is as follows: in, , They represent t Two different variables at time, , These represent the position of the variable, , Indicates the average rank of the variable. For sorting functions, A function to calculate the relative rank difference between any two different variables at different times; , n Indicates the size of the variable. M This represents the total number of variables in the dataset. Representing variables , Updated correlation coefficient G Represents different meshing schemes for a plane. The joint probability distribution of the two sets of variables. and These are the marginal probability distributions for each group of variables. This represents the set high correlation threshold parameter; Step 2: Data preprocessing. First, use wavelet threshold filtering to reduce noise in the data, and then perform standardization. Step 3: Feature fusion, fusing the data obtained in Step 2 into a 9*4*4 three-dimensional data matrix; Step 4: Network learning. The matrix obtained in Step 3 is used as the input data of the ResNet residual network. A wavelet transform filter structure is added after the network output layer to obtain the final result after filtering. Step 5: Use the final results to describe the process characteristics of the turbofan engine, and based on these process characteristics, guide the design of the engine's multivariable controller or perform fault diagnosis.

2. The method for describing the process characteristics of a turbofan engine using a multi-feature fusion residual network according to claim 1, characterized in that, The nine input variables are: throttle lever angle PLA, high-pressure compressor inlet guide vane angle position IGV, low-pressure compressor afterpressure P2, low-pressure compressor aftertemperature T2, high-pressure compressor afterpressure P3, high-pressure compressor aftertemperature T3, fuel flow rate, lubricating oil pressure, and intake manifold airflow meter static pressure PQ2; the five output variables are: high-pressure shaft speed NH, low-pressure shaft speed NL, engine thrust XG, nozzle position indication NPI, and low-pressure turbine outlet temperature T6.

3. The method for describing the process characteristics of a turbofan engine using a multi-feature fusion residual network according to claim 1, characterized in that, The residual network ResNet is: For the turbofan engine ground test dataset, a multi-feature fusion residual deep network is constructed using a residual network structure. High and low layer features are fully extracted and combined through a 4-layer residual structure. A wavelet transform filter structure is added after the network output layer, and the filtered result is used as the final calculation result. The input memory length is determined by the input characteristics of the turbofan engine ground test dataset. The ReLU function is used as the activation function, and the convolutional layer adopts a 3×3 convolutional kernel structure. The network is trained using the mini-batch stochastic gradient descent (MBGD) method, and the L1 loss function is selected as the absolute value error function.

4. The method for describing the process characteristics of a turbofan engine using a multi-feature fusion residual network according to claim 1, characterized in that, In step one, the high-pressure shaft speed NH, low-pressure shaft speed NL, engine thrust XG, nozzle position indicator NPI, and low-pressure turbine outlet temperature T6 are respectively used as variables. The remaining 34 groups of variables are used as variables. Substitute these values ​​into the correlation analysis method to calculate the correlation coefficient and the average of their absolute values. Sort the data variables in descending order of their average values, select the top 8 groups, and then analyze the variables... For variables with a correlation coefficient less than the set conditions, the maximum mutual information (MIC) coefficient correlation analysis method is used to determine whether the correlation between the two variables is moderate or high based on the magnitude of the correlation coefficient and the MIC coefficient.

5. The method for describing the process characteristics of a turbofan engine using a multi-feature fusion residual network according to claim 1, characterized in that, In step two, the wavelet threshold filtering method for noise reduction is as follows: The Mallet algorithm is used to perform three-level wavelet decomposition and reconstruction on each variable, as shown in the following formula: The number of decomposition layers l =1,2,3 t =1,2,..., N, N Indicates the number of points in the wavelet transform sequence; for t Time of the first l Low-frequency components of the layer, for t Time of the first l Layer high-frequency components; x [ t ] represents the original signal; h , g This represents a pair of orthogonal filter banks; h [ 2k-t ]、 g [ 2k-t [The first part is a description of the low-pass and high-pass filters.] k One coefficient, K Indicates the filter length; The Garrote function is used to perform minimax thresholding on the wavelet coefficients, as shown in the following formula: The threshold value is set.

6. The method for describing the process characteristics of a turbofan engine using a multi-feature fusion residual network according to claim 1, characterized in that, The standardization process involves transforming the data into a standard normal distribution.

7. The method for describing the process characteristics of a turbofan engine using a multi-feature fusion residual network according to claim 1, characterized in that, In step three, the data of each set of input variables is first extracted using 16 data points as the sampling window; then, the 16 data points of each set of input variables are transformed into a 4*4 two-dimensional data matrix; finally, the 4*4 two-dimensional data matrices obtained from a total of 9 sets of input variables are merged into a 9*4*4 three-dimensional data matrix, and the output value corresponding to each matrix is ​​determined.

8. The method for describing the process characteristics of a turbofan engine using a multi-feature fusion residual network according to claim 1, characterized in that, In step five, the real-time operating parameters of the airborne sensors are processed by a ground station system with an embedded multi-feature fusion residual network to obtain process characteristic parameters. These parameters are then compared with the standard performance parameters given by the engine manufacturer to identify the changes in deviation. Through deviation analysis and deviation trend analysis, the engine process characteristics are described.

9. The method for describing the process characteristics of a turbofan engine using a multi-feature fusion residual network according to claim 1, characterized in that, Step five, the specific rules guiding the design of the engine multivariable controller, are as follows: based on the established turbofan engine model and engine operating conditions, different controlled parameters are controlled in different operating segments, and a control law that can realize the thermodynamic and strength characteristics specified in the engine design and ensure aerodynamic stability during operation is selected; the specific rules for fault diagnosis are as follows: the deviation between the engine measured parameter values ​​and the baseline values, which characterize different trends of measured parameters, is selected as the input data of the intelligent fault diagnosis model, and the corresponding engine fault mode is taken as the output of the diagnostic model, thereby establishing a multi-class classifier.

Images