Method for evaluating performance of multiple components in gas path of aero-engine

Abstract

The present application relates to the technical field of aero-engine operation performance monitoring, and in particular to an aero-engine gas path multi-component performance evaluation method capable of significantly improving the operation performance monitoring efficiency and accuracy of an aero-fan engine, wherein the performance degradation of each gas path component is regarded as a hidden variable, first, the difference between the deviation value of the gas path parameters after preprocessing and standardization and the deviation value of the gas path parameters of the engine of the same type is taken as the model input; then, with the aid of a nonlinear encoder constructed by a 1DCNN and an MLP, the performance degradation of each gas path component hidden in the deviation value of the gas path parameters is decoupled; then, the influence of the initial unit performance degradation of the gas path component on the deviation value of the gas path parameters is taken as the weight to construct a linear decoder based on the fingerprint map extraction, and the input is reconstructed with the aid of the linear decoder; and finally, the neural network model is trained to obtain the optimal decoder parameters with the minimization of the difference between the input and output as the target.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of aero-engine operation performance monitoring, in particular to an aero-engine gas path multi-component performance evaluation method capable of significantly improving the aero-engine operation performance monitoring efficiency and accuracy. BACKGROUND

[0002] Domestic researchers regard each component of the aero-engine gas path as a whole working cooperatively, and on this basis, combining a large number of operation monitoring parameters, they have proposed a large number of engine gas path component performance evaluation methods from the aspects of mechanism knowledge analysis and data information mining. However, these methods have problems in actual application, such as the unavailability of key design parameters of the engine, the difficulty in obtaining performance monitoring data, the lack of unified performance evaluation criteria, and the fact that the performance evaluation results of the same type of engine gas path do not have comparative reference significance, etc. Therefore, it is difficult to land research results and solve the problem of aero-engine gas path component performance evaluation and provide guidance for actual engine operation and maintenance management.

[0003] Taking the CFM56-5B2 / 3 aero-engine as an example, it adopts modular design and contains three main unit bodies and six main gas path components. The specific structure of the engine is shown in Figure 1 The names of the main gas path components are shown in Table 1, and it is a typical double-rotor turbofan aero-engine. The power device of the aero-engine mainly relies on the air system to provide thrust, and the air system mainly includes a first-stage gas path (inner duct) and a second-stage gas path (outer duct). The first-stage gas path (inner duct): air is pressurized by the 1st fan, 5th low-pressure compressor and 9th high-pressure compressor, sent into the combustion chamber, mixed with aviation fuel, expanded and combusted, high-temperature and high-pressure gas is injected to drive the 1st high-pressure turbine to drive the high-pressure compressor to continue to work, and the gas is discharged after the high-pressure turbine; the second-stage gas path (outer duct): air is pushed by the fan and mixed with the exhaust gas of the first-stage gas path, and the first-stage gas path provides 80% of the required thrust of the engine. The remaining accessory systems include: fuel system, oil system and accessory gearbox system. In terms of operation and maintenance management, the aero-engine of this type mainly adopts condition-based maintenance measures, such as life parts reaching the end of life or health monitoring system warning, etc.

[0004] The attributes and functions based on the fingerprint map are important theoretical basis and diagnostic tools in the process of engine gas path component performance evaluation. For example: the friction between the HPT blade and the casing leads to an increase in the tip clearance, which in turn leads to a decrease in the working efficiency of the HPT, and finally reflects in the abnormal increase of the engine exhaust temperature deviation value Since the fingerprint map can be used to explain the small deviation of the engine gas path parameter deviation value under the working efficiency decline or failure of different gas path components, the fingerprint map can be used as mechanism knowledge to evaluate the aging or damage of the engine gas path components.

[0005] If the working efficiency reduction of the gas path components is regarded as the performance degradation of the gas path components, according to the fingerprint map, the performance degradation of the CFM56-5B2 / 3 type aero-engine is different in the form of the gas path parameter deviation value. When the performance of the gas path components is evaluated, the deviation value of the gas path parameter at the time to be evaluated is usually used to evaluate the performance of the gas path components and the deviation value of the gas path parameter at the reference time . The linear relationship between the unit working efficiency reduction of each gas path component shown in the fingerprint map and the influence amount of the gas path parameter deviation value , as shown in equation (1), can be used to determine the performance degradation of the engine gas path components to be evaluated.

[0006] (1),

[0007] In the equation, is the original gas path parameter deviation value at the time to be evaluated; is the unit opening angle of VSV and BVB or the unit bleed air leakage amount at the time to be evaluated; is the offset amount of the gas path parameter deviation value caused by the unit working efficiency reduction of each gas path component in the fingerprint map. When , it represents that the offset amount of the gas path parameter deviation value caused by the performance degradation of the gas path components to be evaluated is consistent with the offset amount of the gas path parameter deviation value caused by the theoretical unit performance degradation, that is, the performance of the gas path components is degraded by 1%; when

[0008] or , it respectively represents that the actual performance degradation is lighter or more severe than the theoretical unit performance degradation. Table 1 CFM56-5B2 / 3 type aero-engine gas path components

[0009]

[0010] In fact, during the actual operation of the engine, the working states of the engine gas path components and the components and the components and the whole engine are nonlinearly coupled, and the performance degradation of the multiple components and the performance degradation of the single component may be consistent in the form . At this time, if the fingerprint map is directly used for the performance evaluation of the gas path components, the evaluation failure will occur when the abnormality is detected or the fault is diagnosed. In addition, due to the cooperative working characteristics of each gas path component of the engine, the degradation process of each gas path component will maintain a slow degradation feature of dynamic balance during the service period, and the feature is hidden in the gas path parameter deviation value with one-dimensional time sequence characteristics.

[0011] SUMMARY

[0012] ​​This invention addresses the problem of assessment failure risks in existing assessment techniques by proposing a method for assessing the performance of multiple components in the air path of aero-engines that can significantly improve the efficiency and accuracy of monitoring the operational performance of aero-turbofan engines.

[0013] This invention achieves its purpose through the following measures:

[0014] A method for evaluating the performance of multiple components in the air path of an aero-engine, characterized by comprising the following steps:

[0015] Step 1: Establish a nonlinear encoder model and perform feature extraction. The nonlinear encoder consists of a one-dimensional convolutional neural network and a traditional encoder. A traditional 1D-CNN typically contains three basic units: convolutional layers, nonlinear activation layers, and pooling layers. The convolutional layers and nonlinear activation layers work together to extract parametric features, while the pooling layers are used to downsample the model parameters. Since an autoencoder is used to reconstruct the input and output, it is necessary to ensure that the data before and after reconstruction is consistent. Therefore, the pooling layers used to reduce the model parameters need to be removed. The feature extraction process of the 1D-CNN after removing the pooling layers is as follows:

[0016] Based on the length of time The feature parameter is 3 Constructing the input feature surface of 1D-CNN By adopting indivual The convolution kernel performs a convolution operation on the input feature surface according to equation (5), and then inputs the result of the convolution operation into the nonlinear activation layer to output the result. Convolutional feature surfaces ;

[0017] Because of the adoption The convolution kernel is such that, without data padding, the output feature surface is obtained.

[0018] It will be relative to the original input feature surface To reduce the number of time steps by two, and to ensure consistent data length before and after input, it is necessary to modify the input feature surface. The beginning and end of the feature face are filled with zeros. ;

[0019] (5),

[0020] In the formula, It is the first The first convolutional feature surface 1 eigenvalue, , is the first row of the convolution region corresponding to the i-th row of the input value, ; is the weight value of the i-th row and j-th column of the i-th convolution kernel to be learned; is a nonlinear activation function;

[0021] After the convolution, a nonlinear activation function is introduced, which can map the original linear mapping feature surface to a nonlinear space. In the mapped nonlinear space, the linearly inseparable linear mapping features will have the opportunity to find a linearly separable path, thereby decoupling the performance degradation of each gas path component in the nonlinear coupled working state . The ELU function is used as the nonlinear activation function of the network, as shown in equation (6).

[0022] (6),

[0023] wherein, is a combination weight coefficient, and ;

[0024] Step two: nonlinear encoding:

[0025] After extracting the features by 1D-CNN, the input of the nonlinear encoder is constructed with the convolution feature surface with a time length of T and a feature number of n , which is mapped to the hidden layer for representing the performance degradation of each gas path component after decoupling by a nonlinear encoder composed of a multilayer perceptron (MLP), and the nonlinear mapping relationship thereof is as shown in equations (7) and (8):

[0026] (7), (8),

[0027] wherein, is the nonlinear output of the encoder input layer, ; is the output of the encoder, i.e., the performance degradation of the gas path component after decoupling, ; is the encoder input layer weight to be learned, ; is the nonlinear activation function of the encoder; is the encoder input layer bias to be learned, ; ​​​​​​is the encoder output layer weight to be learned, ; is the encoder output layer bias to be learned, ;

[0028] Step three: linear decoding process is carried out:

[0029] The performance degradation of the decoupled air path components is obtained through feature extraction and nonlinear coding , and the influence of the unit performance degradation of each air path component on the air path parameter deviation value in the fingerprint map is used as the weight to construct a linear decoder to reconstruct the air path parameter deviation value , and the linear decoding method is shown in equation (9).

[0030] (9),

[0031] Step four: parameter optimization: since the nonlinear encoder model aims to minimize the reconstruction error of the air path parameter deviation value before and after reconstruction, the mean square error function shown in equation (10) can be used as one of the optimization targets,

[0032] (10),

[0033] wherein, is the mean square error of the air path parameter deviation value before and after reconstruction; is the number of samples; is the model parameter to be learned, ; is the difference between the reconstructed th air path parameter deviation value and the reference value; is the difference between the th air path parameter deviation value before reconstruction and the reference value;

[0034] In addition, since the aero-engine is a high-reliability thermal mechanical device, the performance degradation of each component in the non-failure state should be slow and smooth, and combined with the cooperative working characteristics of each air path component, the performance degradation of the air path component during service should also have cooperativity, that is, the performance degradation of each air path component should be roughly in the same order of magnitude, but since the influence of the unit performance degradation of each air path component on the aero-engine is not consistent, that is, the disturbance degree of the air path parameter deviation value is different, in order to eliminate the inconsistency of the influence of each air path component and make the performance fluctuation index not the same, the influence of the unit performance degradation of each air path component on the air path parameter deviation value is used as the weight to eliminate the dimensional influence of the column vector, so the loss function shown in equation (11) can be constructed to narrow the search range of the optimal parameters and reduce the difficulty of model training:

[0035] (11),

[0036] wherein, is the performance degradation amount loss function of each gas path component; is the performance degradation amount of the i-th gas path component of the j-th sample; is the performance degradation amount of the i-th gas path component of the j-th sample; is the initial performance degradation amount of the gas path component, is the initial performance degradation amount of the gas path component, is the initial performance degradation amount of the gas path component, is the Euclidean norm of the i-th column vector of the weight matrix is the Euclidean norm of the i-th column vector of the weight matrix is the Euclidean norm of the i-th column vector of the weight matrix

[0037] The loss function of the engine multi-gas path component performance evaluation model is shown in formula (12):

[0038] (12).

[0039] The step one also includes pre-processing the data before feature extraction, using the Grubbs criterion and EEMD-SG to pre-process the gas path parameter deviation value, effectively eliminating the influence of the engine non-steady state working condition on the gas path parameter deviation value, removing the gross error in the data and reducing the data noise, and specifically:

[0040] Step 1: For the gas path parameter deviation value , the correction formula based on the fingerprint map is as follows:

[0041] (13),

[0042] wherein, is the uncorrected original gas path parameter deviation value at the moment; is the unit opening angle or unit bleed air leakage of VSV and VBV at the moment; is the offset caused by the fingerprint map to the gas path parameter deviation value;

[0043] Step 2: The Grubbs criterion is used to remove the gross error of the grouped gas path parameter deviation value, and the specific steps of step 2 are as follows:

[0044] Step 2-1: For a group of corrected gas path parameter deviation values , the mean and the standard deviation of the data can be calculated according to formula (14) (15),

[0045] (14), ​​​​ (15)

[0046] Step 2-2: The group of data is arranged in ascending order according to formula (16): (16);

[0047] Step 2-3: The sorted group of data is calculated according to formula (17):

[0048] (17), If the data point corresponds to the statistical quantity , the point is a gross error and needs to be removed from the group of data, wherein is the number of repeated measurements , and the statistical quantity critical value when the confidence probability is is obtained by consulting the Grubbs critical value table;

[0049] Step 2-4: In order to ensure the integrity of the data, the removed data is filled in by using the interpolation method of formula (18): (18);

[0050] Step 3: The gas path parameter deviation value after gross error removal is filtered and denoised by using EEMD-SG: the specific steps are as follows:

[0051] Step 3-1: The gas path parameter deviation value is decomposed based on EEMD, and step 3-1 specifically includes the following steps:

[0052] Step 3-1-1: Set the number of repetitions of superimposed Gaussian white noise and the amplitude of white noise relative to the standard deviation of the original signal ,

[0053] Step 3-1-2: Superimpose Gaussian white noise with amplitude on the gas path parameter deviation value after gross error removal according to formula (19) to obtain a new signal , (19),

[0054] Step 3-1-3: Decompose the new signal by using EMD, and decompose it into intrinsic mode components and one residual component according to formula (20)​​​ : (20),

[0055] Step 3-1-4: Repeat sub-step 3-1-2 and step 3-1-3, calculate the mean of each IMF component and residual component obtained as the air path parameter deviation value according to formula (21) and (22) the decomposed intrinsic modal component and residual component :

[0056] (21), (22),

[0057] Data noise is mainly contained in the high-frequency intrinsic modal component , which is regarded as a noisy component, while signal trends are mainly reflected in the low-frequency intrinsic modal component and residual component , which can be regarded as evolution components, so data denoising mainly depends on the processing of the high-frequency intrinsic modal component ;

[0058] Step 3-2: Perform 0 mean value test on the cumulative sum of IMF components using single-sample T test to distinguish high and low frequency IMFs, and the passed test is high frequency IMF and the failed test is low frequency IMF, the specific steps are as follows:

[0059] Step 3-2-1: Calculate the cumulative IMF component according to formula (23):

[0060] (23),

[0061] Step 3-2-2: Establish test hypothesis and determine test level: set test mean , null hypothesis H0: , alternative hypothesis H1: , significance level , where is the mean of the cumulative IMF component ;

[0062] Step 3-2-3: Select test method and calculate test statistic: use single-sample T test formula (24) to calculate the statistic, and perform two-sided test:

[0063] (24),

[0064] where is the sample size of ; is the standard deviation; step 3-2-4: determine P value and make inference conclusion: calculate the obtained value according to formula (24), consult the T distribution table to obtain the P value, if accept H0, accumulate the IMF component through 0 mean value T test; if accept H1, accumulate the IMF component not pass 0 mean value T test, when the accumulated IMF component passing the test is , the IMF component of in the IMF component is high frequency signal, the IMF component of is low frequency signal.

[0065] Step 4: smooth the high frequency IMF signal by using Savitzky-Golay filter. The specific steps are as follows:

[0066] Step 4-1: for the high frequency IMF component , a equal length overlapping sliding window is set in , the corresponding , the corresponding , a polynomial of order is constructed by fitting this group of data according to formula (25).

[0067] (25),

[0068] In the formula, is the undetermined coefficient of the polynomial;

[0069] Step 4-2: determine the undetermined coefficient of the polynomial, which can be fitted by least square method according to formula (26) through the residual between the data and the original data:

[0070] (26);

[0071] Step 4-3: in order to make the residual minimum, the reciprocal of each undetermined coefficient of the polynomial should be 0,

[0072] (27), Solve:

[0073]

[0074] (28),

[0075] Use the fitted polynomial to calculate the center point estimate value of the window, and use the equal length overlapping sliding window to traverse the high frequency intrinsic mode component of any time step​ The signal smoothing process can be completed in this way to obtain the reconstructed high-frequency IMF signal. ;

[0076] Step 5: Reconstruct the deviation value of the gas path parameters after noise reduction: According to formula (29), the high-frequency IMF component obtained after smoothing by the SG filter is... Low-frequency IMF components and residual components By summing the values, we can obtain the deviation values ​​of the reconstructed gas path parameters. ,

[0077] (29)

[0078] In the formula, These are the high-frequency IMF components processed by the SG filter; The low-frequency IMF components are those not processed by the SG filter. These are the residual components obtained from EEMD decomposition.

[0079] This invention addresses the performance degradation assessment needs of multiple airflow components in aircraft engines. Based on fingerprint images and airflow parameter deviations, and combined with deep learning methods, a performance assessment model for multiple airflow components of aircraft engines is proposed. This model treats the performance degradation of each airflow component as a latent variable. First, the difference between the pre-processed and standardized airflow parameter deviations and the benchmark airflow parameter deviations for this engine model is used as the model input. Then, a nonlinear encoder constructed using 1DCNN and MLP decouples the performance degradation of each airflow component hidden in the airflow parameter deviations. Next, a linear decoder is constructed using the impact of unit performance degradation of the airflow component on the airflow parameter deviations extracted from the fingerprint image as weights. This linear decoder is used to reconstruct the input. The neural network model is then trained with the goal of minimizing the difference between input and output to obtain the optimal decoder parameters. Finally, the trained decoder is used to assess the performance degradation of airflow components. Experiments verify that the airflow component with the most severe performance degradation in the assessment results is consistent with the disassembly and re-engine reasons recorded in the engine's repair work kit, indirectly verifying the effectiveness of the model's assessment results. Attached Figure Description

[0080] Appendix Figure 1 This is a structural schematic diagram of the CFM56-5B2 / 3 aircraft engine.

[0081] Appendix Figure 2 This is a schematic diagram of the automatic encoder model in this invention.

[0082] Appendix Figure 3 This is a schematic diagram of the performance evaluation model for multiple components of the engine air passage in this invention.

[0083] Figure 8 is a graph of engine flight mission frequency statistics. Figure 4

[0084] Figure 9 is a graph of engine flight mission frequency statistics. Figure 5 Figure 10 is a graph of mean square error of different hyperparameters on the test set.

[0085] Figure 11 is a graph of engine flight mission frequency statistics. Figure 6 Figure 12 is a graph of engine flight mission frequency statistics.

[0086] Figure 7 Figure 13 is a graph of engine flight mission frequency statistics. DETAILED DESCRIPTION

[0087] The application will be further described below in conjunction with the accompanying drawings and examples.

[0088] Example 1

[0089] The research object of this example is a CFM56-5B2 / 3 aero-engine, which adopts modular design and includes three main unit bodies and six main air path components. The specific structure of the engine is shown in FIG. 1, and the names of the main air path components are shown in Table 1. It is a typical dual-rotor turbofan aero-engine. Figure 1 Based on the attributes and functions of the fingerprint map, it is an important theoretical basis and diagnostic tool in the performance evaluation process of the engine air path components. For example, the friction between the HPT blade and the casing causes the tip clearance to increase, which in turn causes the HPT working efficiency to decrease, and finally reflects in the abnormal increase of the engine exhaust temperature deviation value

[0090] . Since the fingerprint map can be used to explain the small deviation of the engine air path parameter deviation value under the working efficiency decline or failure of different air path components, the fingerprint map can be used as mechanism knowledge to evaluate the aging or damage of the engine air path components. If the working efficiency decline of the air path component is regarded as the performance decline of the air path component, according to the fingerprint map, the performance forms of different failures or performance declines of the CFM56-5B2 / 3 aero-engine are different on the air path parameter deviation value. When evaluating the performance of the air path component, usually only the linear relationship between the difference between the air path parameter deviation value at the evaluation time and a certain reference air path parameter deviation value

[0091] and the unit working efficiency decline of each air path component shown on the fingerprint map is needed, and the performance decline of the engine air path component to be evaluated can be judged according to formula (1).

[0092]

[0093] (1),​​​​​​

[0094] In the formula, is the original gas path parameter deviation value at the moment; is the unit opening angle or unit bleed air leakage of VSV and BVB at the moment; refers to the printout the offset caused by the gas path parameter deviation value.

[0095] When , it represents that the gas path parameter deviation value offset caused by the performance degradation of the gas path component to be evaluated is consistent with the gas path parameter deviation value offset caused by the theoretical unit performance degradation, i.e., the performance degradation of the gas path component is 1%; when or , it respectively represents that the actual performance degradation is lighter or more severe than the theoretical unit performance degradation.

[0096] In fact, during the actual operation of the engine, the working states of the engine gas path components and components, and the components and the whole machine are nonlinearly coupled. The performance degradation of multiple components and the performance degradation of a single component may be consistent in form. At this time, if the printout is directly used for performance evaluation of the gas path single component, abnormality detection or fault diagnosis, evaluation failure will occur.

[0097] In addition, when using the printout to evaluate the performance of multiple gas path components, if the gas path parameter deviation value is regarded as the collection of the impact caused by the performance degradation of each gas path component, according to the printout, an underdetermined equation set shown in formula (2) can be constructed.

[0098] (2),

[0099] In the formula, is the bias coefficient, ; is the performance degradation amount of the engine gas path component, ; is the weight of the impact of the unit performance degradation of each gas path component on the gas path parameter deviation value based on the printout, as shown in formula (3);

[0100] (3),

[0101] During the solving of the equation set, there are two difficulties: 1) Since it is not a square matrix, it is difficult to solve the inverse matrix of the matrix. 2) The unknowns of the underdetermined equation system (the performance degradation of each gas path component) are 5 greater than 3 known quantities (gas path parameter deviation values), and it is impossible to solve them by introducing other relevant conditions to construct 2 constraint equations. Therefore, it is impossible to directly use the fingerprint diagram and the gas path parameter deviation values ​​to realize the performance evaluation of multiple gas path components under the collaborative working state.

[0102] Therefore, a data-driven approach can be used, employing deep learning technology to learn the performance of multiple components in the gas path.

[0103] Deviation value from pre-treated air circuit parameters of the engine Deviation value of air circuit parameters of this engine model as a benchmark Difference The nonlinear mapping relationship between them is shown in Equation (4). The learned performance degradation of the gas path components is then verified by combining the fingerprint diagram. Whether it is accurate, that is, whether the offset of the gas path parameter deviation value can be restored by using the fingerprint diagram.

[0104] (4),

[0105] In the formula, It is the offset of the gas path parameter deviation value. Performance degradation of multiple components in the gas path Nonlinear mapping function between them.

[0106] As mentioned above, the deviation value of the pre-processed gas path parameters can be considered as... Deviation of gas path parameters from a certain reference time The difference between The data fluctuations are caused by the performance degradation of various airflow components under the nonlinear coupling operating condition of the engine. As shown in the fingerprint diagram, if the performance degradation of each airflow component is known... It can be linearly reconstructed from the fingerprint image. Therefore, this example uses, for example... Figure 2 The automatic encoder shown proposes, as Figure 3 The engine air passage multi-component performance evaluation model is shown.

[0107] Figure 3 The model will measure the performance degradation of each pneumatic component. Treated as an intermediate latent variable, i.e., the coding target, using the preprocessed gas path parameter deviation value. Relative to the benchmark Inter-difference As model input, the nonlinear mapping capability of the encoder is used to mine the performance degradation of each decoupled pneumatic component. Then, the impact of the unit performance degradation of each pneumatic component on the deviation value of the pneumatic parameters in the fingerprint is used as the weight. The constructed linear decoder (Decoder) approximates to reconstruct the difference value For , the loss function of the reconstructed difference value before and after and is minimized The network model parameters of the encoder and decoder are trained by a back propagation algorithm , so as to realize the performance evaluation of the engine gas path components by means of the trained encoder.

[0108] In addition, due to the synergistic working characteristics of each gas path component of the engine, the degradation process of each gas path component will maintain a slow degradation feature of dynamic balance during its service period, and the feature is hidden in the gas path parameter deviation value with one-dimensional time sequence characteristics. In the construction process of the nonlinear encoder, the nonlinear encoder constructed by using the traditional encoder often cannot fully extract the hidden features. Considering that the convolutional neural network (CNN) with weight sharing and local connection as the characteristics has strong data feature extraction capability, a one-dimensional convolutional neural network (1D-CNN) can be used to identify the degradation features of the engine, and the degradation features mentioned above are input into the traditional encoder, so as to construct a more powerful nonlinear mapping encoder for mining the performance degradation of each gas path component hidden in the engine gas path parameter deviation value.

[0109] The detailed steps of the engine gas path multi-component performance evaluation in this example are as follows:

[0110] (1) Feature extraction

[0111] The nonlinear encoder of the model mainly consists of a one-dimensional convolutional neural network and a traditional encoder.

[0112] The traditional 1D-CNN usually contains three basic units of convolutional layers, nonlinear activation layers and pooling layers. The convolutional layers and the nonlinear activation layers work together to extract parameter features, and the pooling layer is used to realize the down-sampling of model parameters. Since the input and output of the auto-encoder are reconstructed in the present application, it is necessary to ensure that the data before and after reconstruction are consistent, so it is necessary to remove the pooling layer for reducing the model parameters. The feature extraction process of the 1D-CNN after removing the pooling layer is as follows:

[0113] The input feature surface of the 1D-CNN is constructed with a time length of and a feature parameter of 3 , by using a ​The convolution kernel is used to perform convolution operation on the input feature surface according to formula (3-5), and then the result of the convolution operation is input into a nonlinear activation layer, and output

[0114] Due to the use of the convolution kernel, the output feature surface

[0115] will be reduced by 2 time steps relative to the original input feature surface In order to ensure that the data length before and after input is consistent, it is necessary to perform 0 padding at the beginning and end of the input feature surface , and the padded feature surface .

[0116] (5),

[0117] In the formula, is the (i, j) th feature value of the i th convolution feature surface, , is the input value of the i th row and j th column of the convolution region corresponding to the i th row, is the weight value of the i th row and j th column of the i th convolution kernel to be learned; is a nonlinear activation function; After convolution, the nonlinear activation function is introduced, which can map the original linear mapping feature surface

[0118] to a nonlinear space. In the mapped nonlinear space, the original linearly inseparable linear mapping features will have the opportunity to find a linearly separable path, so as to decouple the performance degradation of each gas path component in the nonlinear coupling working state . Common nonlinear activation functions include Sigmoid function, Tanh function and ReLU function. In the use process, the Sigmoid function and the Tanh function have the problem of gradient dispersion, and the ReLU function will cause the neuron to "die" due to the negative gradient being 0. Therefore, the present application adopts the ELU function as the nonlinear activation function of the network, as shown in formula (6).

[0119] (6),

[0120] ​​​​​​​​​​​​The function fuses the Sigmoid function and the ReLU function, so that the network has strong robustness to input features.

[0121] (2) Nonlinear encoding

[0122] After extracting the features by 1D-CNN, the input of the nonlinear encoder is constructed with the convolutional feature surface with a time length of T and n features , which is mapped to the hidden layer for representing the performance degradation of each gas path component after decoupling by a nonlinear encoder composed of a multilayer perceptron (MLP) , and the nonlinear mapping relationship thereof is as shown in equations (7) and (8).

[0123] (7),

[0124] (8),

[0125] In the formula, is the nonlinear output of the input layer of the encoder, ; is the output of the encoder, i.e., the performance degradation of the gas path component after decoupling, ; is the input layer weight of the encoder to be learned, ; is the nonlinear activation function of the encoder; is the input layer bias of the encoder to be learned, ; is the output layer weight of the encoder to be learned, ; is the input layer bias of the encoder to be learned, ;

[0126] (3) Linear decoding

[0127] The performance degradation of the gas path component after decoupling is obtained through feature extraction and nonlinear encoding , and then a linear decoder is constructed by using the influence of the unit performance degradation of each gas path component in the fingerprint map on the gas path parameter deviation value as the weight to reconstruct the gas path parameter deviation value , and the linear decoding method is as shown in equation (9).

[0128] (9),

[0129] In the formula is the bias of the decoder to be learned, .

[0130] (4) Parameter optimization

[0131] Since the model aims to minimize the reconstruction error of the gas path parameter deviation before and after reconstruction, the mean square error function as shown in equation (10) can be used as one of the optimization objectives.

[0132] (10)

[0133] In the formula, It is the mean square error of the deviation values ​​of the gas path parameters before and after reconstruction; It is the number of samples; These are the model parameters to be learned. ; It is the reconstructed first The difference between the deviation value of each gas path parameter and the reference value; It is the first before the reconstruction The difference between the deviation value of each gas path parameter and the reference value;

[0134] Furthermore, as aero-engines are highly reliable thermodynamic machines, the performance degradation of each component under non-fault conditions should be slow and stable. Considering the collaborative working characteristics of each gas path component, the performance degradation of these components during service should also be synergistic, meaning the performance degradation of each component should be roughly within the same order of magnitude. However, since the impact of unit performance degradation of each gas path component on the aero-engine is not uniform—that is, the degree of disturbance to the gas path parameter deviations differs—to eliminate the inconsistency in performance fluctuation indicators caused by the inconsistent force dimensions of each gas path component, the impact of unit performance degradation of each gas path component on the gas path parameter deviations can be used as a reference. The norm of the column vector eliminates the influence of dimensions, so the loss function shown in equation (11) can be constructed. This reduces the difficulty of model training by narrowing the search range for optimal parameters.

[0135] (11),

[0136] In the formula, It is a function of the performance degradation loss of each gas path component; It is the first The first sample Performance degradation of individual pneumatic components; It is a pneumatic component The initial performance degradation amount, take ; It is a weight matrix No. The Euclidean norm of the column vector; In summary, the loss function of the engine multi-airway component performance evaluation model is shown in equation (12).

[0137] (12).

[0138] In this example, the loss function calculated by model forward propagation is optimized by using the adaptive momentum estimation (Adam) algorithm The parameters are updated using backpropagation (BP) , so as to obtain the optimal multi-gas path component performance evaluation model.

[0139] Experimental verification:

[0140] First, data selection and pretreatment:

[0141] The 76 CFM56-5B2 / 3 aero-engine full-life running monitoring data collected from an airline, totaling more than 510,000, were used as the data basis for model verification. Through analysis, it was found that the collected running data included engine takeoff and cruise data. The cruise data collection required the aircraft and engine to operate stably for more than 5 minutes, and the on-board sensor had no data change within 12 seconds. Therefore, this chapter used this monitoring data with certain stability to analyze the long-term and slow performance degradation of the engine gas path components.

[0142] In view of the various "flaws" in the original running data, the gas path parameter deviation values were pretreated, and the Grubbs criterion and EEMD-SG were used to pretreat the gas path parameter deviation values, effectively eliminating the influence of engine non-steady state conditions on the gas path parameter deviation values, removing the coarse errors in the data and reducing the data noise. The pretreated part of the gas path parameter deviation values is shown in Table 2.

[0143] Table 2 CFM56-5B2 / 3 aero-engine gas path parameter deviation values

[0144]

[0145] In addition, the calculation methods of each gas path parameter deviation value are not the same, so that the dimensions of each gas path parameter deviation value after pretreatment are different. To this end, the Euclidean norm of column vector can be used to standardize it according to formula (30).

[0146] (30),

[0147] In the formula, is the unit dimension gas path parameter deviation value; is the pretreated gas path parameter deviation value; is the weight matrix The Euclidean norm of the first row vector.

[0148] Performance evaluation model benchmark established:

[0149] Due to differences in manufacturing levels, even engines of the same model may have slightly different factory performance. However, as mentioned in the current research, traditional comprehensive evaluation methods can only reflect the performance differences of the air passage components at different times within the same engine. They cannot provide a unified evaluation benchmark for all engines of that model. That is, they cannot show the changes in component performance at different times for different engine numbers under a certain model relative to a certain benchmark performance of that model. Therefore, it is necessary to reasonably select a set of air passage parameter deviation values ​​for that model of engine as the benchmark for evaluating the component performance of that engine type. .

[0150] Therefore, statistical analysis can be performed on the deviation values ​​of the air path parameters generated by new engines of this model within a certain period of service after leaving the factory, and a set of air path parameter deviation values ​​can be determined as the benchmark for evaluating the performance of engine components based on actual needs.

[0151] Statistical analysis of uncorrected engine operating data revealed the daily flight mission frequency of the 76 engines as follows: Figures 3-4 As shown, different color layers correspond to different engine numbers under the same model. It can be seen that for a single engine, the daily flight mission frequency is generally less than 6 times, with 2 flight missions per day being typical.

[0152] Since the task has the highest frequency, the basic data is grouped into sets of two points.

[0153] To avoid the impact of control parameter fine-tuning on the deviation of the gas path parameters during the initial service period of a new engine, 2156 data points were selected from the first two weeks (14 days) of service for each new engine. The worst initial performance point of the gas path parameter deviation of this type of engine was calculated according to formulas (31) and (32). With average initial performance point As a benchmark for evaluating the performance of engine components .

[0154] (31),

[0155] (32),

[0156] In the formula, It refers to the amount of sample data. .

[0157] The worst initial performance point of the CFM56-5B2 / 3 aero-engine's air path parameter deviation value was determined by solving the problem. With average initial performance point The corresponding deviation values ​​of the gas path parameters are shown in equations (33) and (34).

[0158] as shown.

[0159] (33),

[0160] (34),

[0161] Model training and hyperparameter adjustment:

[0162] The engine air path multi-component performance evaluation model based on the fingerprint map and air path parameter deviation value is trained as shown below:

[0163] (1) Data preprocessing and standardization processing are performed on the air path parameter deviation value

[0164] (2) Select an initial performance point as a reference, and construct a model data set with the difference of the air path parameter deviation value relative to the reference

[0165] (3) Combine 1DCNN and AE to construct an aero-engine air path multi-component performance evaluation model based on the fingerprint map and air path parameter deviation value;

[0166] (4) Divide the model training set and test set in a ratio of 7:3;

[0167] (5) Set the model hyperparameters according to Table 3, wherein the time step and the number of convolution kernels are the hyperparameters to be adjusted;

[0168] (6) Use the Adam algorithm combined with the loss function to train the model, and search for the optimal value of the hyperparameters to be adjusted;

[0169] (7) Determine the hyperparameters and train the final model;

[0170] (8) Use the trained nonlinear encoder to evaluate the performance of the air path multi-component;

[0171] (9) Compare the air path component performance evaluation results with the maintenance report to verify the accuracy of the model evaluation.

[0172] The time step and the number of convolution kernels are the model hyperparameters to be adjusted, and the specific steps for parameter search are as follows:

[0173] Let the time step value set be , and the number of convolution kernels be ​​, the mean square error on the test set after the model training is completed is used to determine the value of the hyperparameters. In order to avoid the interference of neural network randomness and different initialization methods on the model training results, the parameters determined for each experiment are trained 10 times, and the mean of the mean square error of the 10 times is taken as the mean square error of the group of parameters. The group of parameters with the smallest mean square error is selected as the optimal value of the adjusted hyperparameters. The grid search result is shown in Figure 5 , considering the network model parameters and training accuracy, the optimal time step , the number of convolution kernels .

[0174] Table 3 Model hyperparameters

[0175]

[0176] Multi-air-path component performance evaluation application example:

[0177] Limited by the length of the article, only the air path parameter deviation value of a certain engine of CFM56-5B2 / 3 type from the factory service to the first overhaul under the whole life cruise state is taken as an example for multi-air-path component performance evaluation. The performance degradation of each air-path component during service decoupled by the nonlinear encoder of the trained model is as follows Figure 6 The reconstructed model output and input based on the air-path component performance evaluation results are as follows Figure 7 Wherein, the superscript a represents that the data is preprocessed and standardized. It can be seen that the performance of each decoupled air-path component can be used to approximate the input parameters of the model by means of the fingerprint map.

[0178] The performance of the air-path components of the engine at the factory, before overhaul, and the performance of each air-path component are shown in Tables 3-4. It can be seen that the performance degradation of the HPT is the most serious before overhaul, which is 0.6150%.

[0179] Table 4 Performance of air-path components of the engine

[0180]

[0181] Since the actual engine air-path component performance degradation needs to be measured by physical experiments, and the research process of the present application does not have the experimental conditions, in order to verify the accuracy of the evaluation results, it is found through consulting the diagnosis results of the engine repair work package (Engine Repair Work-scope) that the engine disassembly reason is HPT leading edge ablation. This is consistent with the evaluation results of the model, which proves that the model can better evaluate the performance degradation of the engine air-path components.

[0182] The application faces the demand of airlines for performance degradation evaluation of engine air path multi-component, based on fingerprint map and air path parameter deviation value, combined with deep learning method, an aero-engine air path multi-component performance evaluation model is proposed. The model regards the performance degradation of each air path component as a hidden variable, firstly, the difference between the pre-processed and standardized air path parameter deviation value and the air path parameter deviation value benchmark of the engine is taken as the input of the model; then with the help of the nonlinear encoder constructed by 1DCNN and MLP, the performance degradation of each air path component hidden in the air path parameter deviation value is decoupled; then the influence of the initial air path component unit performance degradation on the air path parameter deviation value is extracted based on the fingerprint map to construct a linear decoder, and the input is reconstructed by means of the linear decoder; then the neural network model is trained to obtain the optimal decoder parameters by minimizing the difference between the input and output; finally, the performance degradation of the air path component is evaluated by the trained decoder, and the performance degradation of the air path component with the most serious performance degradation in the performance evaluation result is consistent with the recorded disassembly reason in the engine repair work package, which verifies the effectiveness of the model evaluation result of the model from the side.

Claims

1. An aeroengine gas path multi-component performance assessment method, characterized in that, Comprising the following steps: Step one: establish a nonlinear encoder model, feature extraction, the nonlinear encoder consists of a one-dimensional convolutional neural network (1D-CNN) and traditional encoder, 1D-CNN contains convolution layer, nonlinear activation layer and pooling layer 3 basic units; Because the automatic encoder needs to ensure that the data before and after reconstruction are consistent, the pooling layer used to reduce the model parameters needs to be removed, and the feature extraction process of 1D-CNN after removing the pooling layer is as follows: The input feature surface of the 1D-CNN is constructed with a time length of T and a characteristic parameter of 3 , By adopting n 3x3 convolution kernels to perform convolution operation on the input feature surface according to formula (5), and then inputting the result of the convolution operation into a nonlinear activation layer, n convolution feature surfaces Y are output i In order to ensure that the data length before and after input is consistent, 0 padding needs to be performed at the beginning and end of the input feature surface X, and the padded feature surface ； （5）， wherein Y i,j — the jth feature value of the ith convolutional feature surface, i ∈ (1, n); - the input value of the u-th row and v-th column of the convolution region corresponding to the j-th row, j e (1, T); - the weight of the u-th row and v-th column of the i-th convolution kernel to be learned; σ c — a non-linear activation function; After convolution, the original linear mapping feature surface K i *x j is mapped to a nonlinear space, in which the original linearly inseparable linear mapping features will have the opportunity to find a linearly separable path, thereby decoupling the performance degradation H of each gas path component in the nonlinear coupled working state. The ELU function is used as the nonlinear activation function of the network. Step two: nonlinear coding: After the feature is extracted by the 1D-CNN, the input of the nonlinear encoder is constructed by the convolution feature surface with the time length T and the feature number n , which is mapped to the hidden layer for representing the performance degradation of each gas path component after decoupling by the nonlinear encoder composed of a multilayer perception mechanism Step three: linear decoding process: after obtaining the decoupled air path component performance degradation H through feature extraction and nonlinear encoding, the influence of the unit performance degradation of each air path component in the fingerprint diagram on the air path parameter deviation value is used as the weight W d The linear decoder is constructed to reconstruct the air path parameter deviation value The linear decoding method is shown in formula (9): ， wherein b d - decoder bias to be learned, ； Step four: parameter optimization: the nonlinear encoder model takes the reconstruction error of the parameter deviation value of the air path before and after reconstruction as the target, so the mean square deviation function as formula (10) is used as one of the optimization targets, ， In the formula, L1 is the mean square deviation of the air path parameter deviation value before and after reconstruction; n is the sample number; Step one also includes pre-processing the data before feature extraction, using the Grubbs criterion and EEMD-SG to pre-process the air path parameter deviation value, effectively eliminating the influence of engine non-steady state conditions on the air path parameter deviation value, removing large errors in the data and reducing data noise, which is as follows: θ - model parameters to be learned, θ = {K i W' e b' e W e b e b d} — the difference between the reconstructed i-th air path parameter deviation value and the reference; Δd i - the difference between the i-th air path parameter deviation value before reconstruction and the reference.

2. The method of claim 1, wherein, Step 2: use Grubbs criterion to remove large errors from the grouped air path parameter deviation value, and the specific steps of step 2 are as follows: Step 1 : For air path parameter bias values The correction formula based on the fingerprint map is as follows: （13）， In the formula - i time uncorrected original gas path parameter deviation value; x ij - i the unit opening angle of VSV, B V or the unit bleed air leakage at time instant w j — the x in the fingerprint map j the offset caused by the air path parameter deviation value; Step 3: use EEMD-SG to filter and denoise the air path parameter deviation value D after removing large errors: the specific steps are as follows: Step 2-1 : For a set of modified airpath parameter bias values that follow a Gaussian distribution The mean μ and standard deviation σ of the set of data are calculated as per equations (14) and (15), （14）， （15）， Step 2-2: Sort the set of data In ascending order according to formula (16): d (1) <d (2) <... <d (n) (16); Step 2-3: Calculate the sorted data set according to formula (17) The corresponding statistic g (i) : （17）， If the data point d (i) The corresponding statistical quantity g (i) > g (0) (n, α), the point is a gross error, which needs to be removed from the group data, wherein g (0) (n, α) is the statistical critical value when the number of repeated measurements is n and the confidence probability is α, which is obtained by consulting the Grubbs critical value table. Step 2-4: To ensure the integrity of the data, the data d is rejected i The data is filled in by the interpolation method of (18) （18）； Step 3-1: decompose the air path parameter deviation value based on EEMD, which specifically includes the following steps: Step 3-1-1: set the number of repeated superimposed Gaussian white noise m and the amplitude α of white noise relative to the standard deviation of the original signal, Step 3-2: use single-sample T test to test the cumulative sum of IMF components to 0 to distinguish high and low frequency IMF, and the high frequency IMF is tested, and the low frequency IMF is not tested, and the specific steps are as follows: Step 3-1-2: superimpose a Gaussian white noise N of amplitude a in the air path parameter deviation value D after coarse error removal according to formula (19) i , and the new signal obtained is denoted as , （19）， Step 3-1-3: Using EMD on the new signal decomposed into n intrinsic mode components I i,j and 1 residual component R i : （20）， Step 3-1-4: Repeat steps 3-1-2 and 3-1-3 m times, and calculate the mean of each IMF component and the residual component obtained as the decomposed eigenmode component I of the air path parameter deviation value D according to formulas (21) and (22) j and the residual component R: （21）， （22）， Data noise is mainly contained in high-frequency intrinsic mode component I h , which is regarded as noise component, while signal trend is mainly embodied in low-frequency intrinsic mode component I l and residual component R, which can be regarded as evolution component, so data noise reduction mainly depends on the processing of high-frequency intrinsic mode component I h . Step 3-2-3: select test method and calculate test statistic: use single-sample T test formula (24) to calculate the statistic, and perform two-sided test: Step 3-2-1: Calculate cumulative IMF component I according to formula (23) s : （23）， Step 3-2-2: Establish test hypothesis and determine test level: Let test mean μ = 0, null hypothesis H0: μ Is = μ, alternative hypothesis H1: μ Is ≠ μ, significance level α = 0.05, where μ Is is the mean of the cumulative IMF component I s ​ Step 4: use Savitzky-Golay filter to smooth high frequency IMF signal, and the specific steps are as follows: （24）， where k — number of samples; and s of samples; σ Is — I s standard deviation of Step 3-2-4: Determine P value and make inference: Calculate t value from equation (24) α,k-1 value, look up T distribution table to get P value, if P > a, accept H0, accumulate IMF component I s by 0 mean value T test; if P < a, accept H1, accumulate IMF component I s not pass 0 mean value T test, when the accumulated IMF component passing the test is I d , the IMF components with j < d in the IMF components are high frequency signals, and the IMF components with j > d are low frequency signals; Step 4-3: in order to minimize the residual E, the reciprocal of each polynomial undetermined coefficient E is 0, Step 4-1: For high frequency IMF component I h corresponding 2m+1 consecutive values x within an equal length overlapping sliding window i , i e (-m, m), by constructing a kth order polynomial fitting this set of data according to equation (25), where k < 2m+1. （25）， wherein a j - polynomial pending coefficients; Step 4-2: Determine the polynomial pending coefficient a j The data and the residual E of the original data can be fitted by least square method according to formula (26): （26）； The solution is: （27）， R is the residual component obtained by EEMD decomposition. （28）， The center point estimate value of the window is calculated by using the fitted polynomial, and the equal-length overlapping sliding window is used to traverse the high-frequency intrinsic modal component I of any time step h The signal smoothing processing is completed, and the reconstructed high-frequency IMF signal I' is obtained h ; Step 5: Reconstruct the denoised air path parameter deviation value: according to formula (29), the high-frequency IMF component I' obtained after smoothing by the SG filter is added to the low-frequency IMF component I h , and the residual component R, to obtain the reconstructed air path parameter deviation value D'; l . （29）， wherein I' i - high frequency IMF components processed by the SG filter; I i - low frequency IMF components without SG filter processing; The ELU function in step one is shown in formula (6):

3. The method of claim 1, wherein, In the formula, α is the combination weight coefficient, and α=1. ， In step two, the nonlinear encoder composed of multiple layers of perception is mapped to the hidden layer for representing the performance degradation of each air path component after decoupling, and the nonlinear mapping relationship is as follows:

4. The method of claim 1, wherein, ​ Z = σ e (YW′ e +b′ e ) (7), H = ZW e +b e (8), wherein is a non-linear output of the encoder input layer, ; is an encoder input layer weight to be learned, ; is a non-linear activation function of the encoder; is an encoder input layer bias to be learned, ; is an encoder output layer weight to be learned, ; is an encoder input layer bias to be learned, .

5. The method of claim 1, wherein, Since the aero-engine is a high-reliability thermal machine, the performance degradation of each component in the non-failure state should be slow and smooth. Combined with the cooperative working characteristics of each gas path component, the performance degradation of the gas path component during service should also have synergy, that is, the performance degradation of each gas path component should be roughly in the same order of magnitude. However, the influence of the unit performance degradation of each gas path component on the aero-engine is not consistent, that is, the disturbance degree of the gas path parameter deviation value is different. In order to eliminate the inconsistency of the performance fluctuation index caused by the influence of each gas path component, the influence W of the unit performance degradation of each gas path component on the gas path parameter deviation value can be used to eliminate the influence of the unit performance degradation of each gas path component on the aero-engine. d The norm of the column vector eliminates the influence of the dimension, so the loss function L2 shown in equation (11) can be constructed to reduce the optimal parameter search range and reduce the model training difficulty: ， In the formula, It is a function of the performance degradation loss of each gas path component; It is the first The first sample Performance degradation of individual pneumatic components; It is a pneumatic component The initial performance degradation amount, take ; It is a weight matrix No. The Euclidean norm of the column vector; then the loss function of the engine multi-airway component performance evaluation model is shown in equation (12): （12）。

Images