An engine mechanism feature extraction method based on parameter nonlinear mining layer

Abstract

The application relates to the technical field of aero-engine gas path fault diagnosis, and discloses an engine mechanism feature extraction method based on a parameter nonlinear mining layer, which comprises the following steps: acquiring engine sensor measurement parameters of features to be extracted and standard parameters output by an engine reference model; loading an engine mechanism feature extractor which is generated based on parameter nonlinear mining layer training and is used for realizing engine mechanism feature extraction; inputting the engine sensor measurement parameters of the features to be extracted and the standard parameters output by the engine reference model into the engine mechanism feature extractor, acquiring feature residuals output by the engine mechanism feature extractor, and completing engine mechanism feature extraction. The engine mechanism feature extractor based on parameter nonlinear correlation is constructed, the nonlinear proportional relationship of engine measurement parameters is effectively extracted, the engine gas path fault features are accurately extracted, and the engine gas path fault can be diagnosed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of aero-engine gas path fault diagnosis, and particularly relates to an engine mechanism feature extraction method based on a parameter nonlinear mining layer. BACKGROUND

[0002] An aero-engine is a whole composed of complex systems. In the full envelope range, it is difficult to reflect the gas path characteristics and performance of the engine by the size of a single measured parameter value, and it is necessary to analyze the ratio between different cross-section parameters to obtain the gas path characteristics and performance changes of the engine. For example, the engine pressure ratio (EPR) is the ratio of the total pressure Pt6 at the low-pressure turbine outlet to the total pressure Pt2 at the fan inlet. The parameter is one of the key target parameters in the engine control system, and is linearly related, which can indirectly reflect the current performance of the engine.

[0003] However, the correlation between different cross-section parameters is not necessarily linear. For example, in the process of accelerating along the engine acceleration working line, the correlation between the converted fuel and the converted speed flow is nonlinear, which can be seen from the following calculation expression:

[0004] ;

[0005] In the formula:

[0006] -- main fuel flow;

[0007] -- converted fuel flow;

[0008] -- high-pressure rotor physical speed;

[0009] -- high-pressure rotor converted speed;

[0010] -- total temperature at the fan inlet;

[0011] -- total pressure at the fan inlet;

[0012] -- constant, usually 0.5-0.8.

[0013] In the prior art, when a neural network algorithm is used to diagnose the gas path fault of an aero-engine, the nonlinear proportional relationship between parameters in the engine needs to be captured first. However, when the traditional graph neural network calculates the correlation between adjacent nodes, it uses a weighted summation method, which cannot accurately capture the nonlinear proportional relationship between the parameters of adjacent nodes, and therefore cannot accurately extract the engine gas path fault features, ultimately leading to the need to improve the explainability and accuracy of gas path fault diagnosis.

[0014] As disclosed in the patent document CN120850204A, a kind of aero-engine gas path performance parameter multi-modal pulse fusion network prediction framework is disclosed, belonging to aero-engine gas path performance parameter prediction and fault diagnosis technical field. First, aero-engine flight data is obtained, and each sensor data is preprocessed;Second, a multi-modal pulse fusion neural network model including time series data pulse processing module, image data pulse processing module, multi-modal feature fusion module and prediction output module is constructed;Finally, the multi-modal pulse fusion neural network model is trained, and the trained model is used to predict the gas path performance parameter. Deep feature extraction and fusion of time series data set and image data set are realized, and complex nonlinear correlation between data can be captured, but in the scheme, weighted summation method is still used for calculation, and the accuracy of captured nonlinear proportional relationship needs to be improved.

[0015] Therefore, in order to solve this problem, a technical scheme that can accurately capture the nonlinear proportional relationship between the parameters of adjacent nodes and accurately extract the engine gas path fault features is needed. SUMMARY

[0016] To solve the technical problems that the prior art cannot accurately capture the nonlinear proportional relationship between the parameters of adjacent nodes and cannot accurately extract the engine gas path fault features, an engine mechanism feature extraction method based on parameter nonlinear mining layer is provided, comprising:

[0017] Obtaining engine sensor measurement parameters and engine reference model output standard parameters to be extracted features ;

[0018] Loading an engine mechanism feature extractor, which is trained and generated based on a parameter nonlinear mining layer, to realize engine mechanism feature extraction. The engine mechanism features are features that can reflect the working law, performance and state of the engine according to the principle of the engine, including total pressure ratio, converted fuel flow and high-pressure converted speed.

[0019] Inputting the engine sensor measurement parameters and engine reference model output standard parameters to be extracted features into the engine mechanism feature extractor , obtain the feature residual output by the engine mechanism feature extractor , complete the engine mechanism feature extraction.

[0020] Further, before loading the engine mechanism feature extractor, the engine mechanism feature extractor is constructed, including the following contents:

[0021] obtain the existing engine sensor measurement parameters and the standard parameters output by the engine reference model , form a training sample set;

[0022] Based on the parameter nonlinear correlation mining layer, the mechanism feature parameters of the engine are correlated and mined, combined with the graph attention mechanism layer, two convolution layers and the filter response normalization layer FRN, to form an initial engine mechanism feature extractor;

[0023] Train the initial engine mechanism feature extractor with the training sample set to generate the engine mechanism feature extractor.

[0024] Further, the training process of the initial engine mechanism feature extractor includes the following contents:

[0025] Set the fault label t, classify and number the fault features;

[0026] Substitute the training sample set into the initial engine mechanism feature extractor to extract the engine sensor measurement parameter features and the engine reference model output parameter features , obtain the feature residual ;

[0027] According to whether it is a fault feature, the feature residual is separated into non-fault feature residual and fault feature residual , calculate the difference between the non-fault feature residual and the fault feature residual ;

[0028] After full connection layer FC and probability normalization Softmax processing are performed on the feature residual , according to the fault label t, the cross-entropy loss function is used to determine the fault category;

[0029] After spherical clustering analysis is performed on the difference , according to the fault label t, the hyperspherical feature clustering loss function is used to quantify the features;

[0030] Carry out feature extraction training until the feature residual satisfying the requirement of the cross-entropy loss function and the difference satisfying the requirement of the hyperspherical feature clustering loss function, the training of the initial engine mechanism feature extractor is completed, and the engine mechanism feature extractor is generated.

[0031] wherein n is the fault type number.

[0032] Further, the requirement of the cross-entropy loss function is to ensure correct fault classification.

[0033] Further, the requirement of the hyperspherical feature clustering loss function is to evenly distribute features of different classes on a hypersphere, control the distance from the center of the sphere, and make the two-norm of the fault-free feature residual as the center of the sphere tend to 0.

[0034] Further, the formation process of the initial engine mechanism feature extractor includes the following contents:

[0035] The activation value q of the parameter nonlinear correlation mining layer is obtained, and after the activation value q is processed by the graph attention mechanism, it passes through two convolution layers and a filter response normalization layer FRN to form an initial engine mechanism feature extractor, thereby realizing the feature extraction function.

[0036] Further, the calculation expression of the activation value q is:

[0037] ;

[0038] wherein represents a column vector of all 1s with a length of m, and the superscript T represents the transpose operation; represents the Hadamard product; represents the element-wise power operation; y is the input parameter, including engine sensor measurement parameters and engine reference model output parameters .

[0039] The beneficial effects of the present application: through the new engine mechanism feature extraction method based on parameter nonlinear mining layer, the graph attention mechanism and the parameter nonlinear mining layer are used to construct the engine mechanism feature extractor, the features of the sensor measurement parameters and the engine benchmark model output parameters are extracted, the feature residuals, the fault feature residuals and the difference between the fault feature residuals and the non-fault feature residuals are analyzed, the super-spherical feature clustering loss and the cross-entropy loss function are analyzed, and the engine mechanism feature extractor with parameter nonlinear correlation is formed through continuous training iteration, which can effectively extract the nonlinear proportional relationship of the engine measurement parameters, accurately extract the engine gas path fault features, and realize the effective diagnosis of the engine gas path fault. The problems that the traditional graph neural network cannot capture the nonlinear proportional relationship between the parameters of adjacent nodes when calculating the correlation between adjacent nodes, and it is difficult to accurately extract the engine gas path fault features are solved. The aero-engine gas path fault can be accurately positioned, the accuracy and robustness of the gas path fault diagnosis are improved, and help is provided for engine health management and gas path fault analysis. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 The present application provides a parameter nonlinear mining layer based engine mechanism feature extraction method schematic diagram;

[0041] Figure 2 The present application provides a training framework diagram of the initial engine mechanism feature extractor. DETAILED DESCRIPTION

[0042] The technical solutions of the present application are further described below, but the scope of protection is not limited to the description.

[0043] The present application provides a parameter nonlinear mining layer based engine mechanism feature extraction method, as shown in Figure 1 The present application provides a parameter nonlinear mining layer based engine mechanism feature extraction method, as shown in

[0044] Step S100, obtaining engine sensor measurement parameters and engine benchmark model output standard parameters to be extracted features ; the engine sensor measurement parameters and the engine benchmark model output standard parameters to be extracted features contain the same number of data, the same working condition; the sensor includes a temperature sensor, a pressure sensor and a flow sensor that can reflect the engine cross-section parameter;

[0045] Step S200, loading the engine mechanism feature extractor, the engine mechanism feature extractor is generated based on parameter nonlinear mining layer training, used for realizing engine mechanism feature extraction, the engine mechanism feature is the feature that can reflect the engine working rule, performance and state according to the engine principle, including total pressure ratio, converted fuel flow and high pressure converted speed;

[0046] Before the loading engine mechanism feature extractor, an engine mechanism feature extractor is constructed, including the following contents:

[0047] Obtain existing engine sensor measurement parameters and standard parameters output by the engine benchmark model , form a training sample set;

[0048] The existing engine sensor measurement parameters and standard parameters output by the engine benchmark model contain the same number of data under the same working condition.

[0049] Based on the parameter nonlinear correlation mining layer, the mechanism characteristic parameters of the engine are correlated and mined, combined with the graph attention mechanism layer, two convolution layers and the filter response normalization layer FRN, to form an initial engine mechanism feature extractor.

[0050] Train the initial engine mechanism feature extractor with the training sample set to generate the engine mechanism feature extractor.

[0051] The training framework of the initial engine mechanism feature extractor is as shown in Figure 2 The training process includes the following contents:

[0052] Set the fault label t, classify and number the fault features;

[0053] Substitute the training sample set into the initial engine mechanism feature extractor to extract engine sensor measurement parameter features and engine benchmark model output parameter features , obtain feature residuals ;

[0054] According to whether it is a fault feature, the feature residuals are separated into non-fault feature residuals and fault feature residuals , and the difference between the non-fault feature residuals and the fault feature residuals is calculated;

[0055] After the feature residuals are processed by the full connection layer FC and the probability normalization Softmax, according to the fault label t, the cross-entropy loss function is used to determine the fault category;

[0056] After the difference is subjected to spherical clustering analysis, according to the fault label t, the hyperspherical feature clustering loss function is used to quantify the features;

[0057] carrying out feature extraction training until the feature residuals meet the requirement of the cross-entropy loss function and the difference meet the requirement of the hyperspherical feature clustering loss function, the training of the initial engine mechanism feature extractor is completed to generate the engine mechanism feature extractor;

[0058] wherein n is the fault type number.

[0059] And in order to ensure the consistency of different classes in working conditions and quantities, in each training iteration period, the existing engine sensor measurement parameters and standard parameters output by the engine reference model should include the same number of data of the same working condition.

[0060] The requirement of the cross-entropy loss function is to ensure the correctness of fault classification.

[0061] The requirement of the hyperspherical feature clustering loss function is to distribute the features of different classes on a hypersphere, control the distance from the center of the sphere, and make the two-norm of the fault-free feature residual as the center of the sphere tend to 0.

[0062] The formation process of the initial engine mechanism feature extractor includes the following contents:

[0063] The activation value q of the parameter nonlinear correlation mining layer is obtained, and after the activation value q is processed by the graph attention mechanism, it passes through two convolution layers and a filter response normalization layer FRN to form an initial engine mechanism feature extractor, realizing the function of feature extraction.

[0064] In the embodiment, the calculation expression of the activation value q is:

[0065] ;

[0066] wherein, represents a column vector of all 1s with a length of m, and the superscript T represents the transpose operation; represents Hadamard product (corresponding position multiplication); represents the element-wise power operation; y is the input parameter, including engine sensor measurement parameters and engine reference model output parameters .

[0067] Step S300, input the engine sensor measurement parameters and standard parameters output by the engine reference model of the features to be extracted to the engine mechanism feature extractor , the feature residual output by the engine mechanism feature extractor is obtained, and the engine mechanism feature extraction is completed; the feature residual characterizes the difference between the engine sensor measurement parameter feature and the engine reference model output parameter feature , and the feature residual is different under different conditions (with or without faults or different types of faults); therefore, the feature residual corresponds to the engine mechanism feature under different conditions of the engine. The feature residual is obtained, that is, the engine mechanism feature extraction is completed.

[0068] After full connection layer FC and probability normalization Softmax processing are performed on the feature residual output by the engine mechanism feature extractor , the feature residual can be applied to the gas path fault diagnosis of an aero-engine.

[0069] The engine mechanism feature extraction method based on the parameter nonlinear mining layer can effectively extract the nonlinear proportion of the measurement parameters of the engine, and can be used for effectively diagnosing the gas path fault of the engine. The method solves the problem that the traditional graph neural network cannot capture the nonlinear proportional relationship between the parameters of adjacent nodes when calculating the correlation between adjacent nodes, and further cannot accurately extract the gas path fault feature of the engine, and can accurately locate the gas path fault of the aero-engine, improves the accuracy and robustness of the gas path fault diagnosis, and provides help for the health management and gas path fault analysis of the engine.

[0070] The above disclosure is only a specific embodiment of the present application, but the present application is not limited thereto, and any changes that can be thought of by those skilled in the art should fall within the protection scope of the present application.

Claims

1. An engine mechanism feature extraction method based on parameter nonlinear mining layer, characterized in that, The application relates to an engine mechanism feature extractor, which is generated based on a graph attention mechanism and a parameter nonlinear mining layer training and is used for realizing engine mechanism feature extraction, wherein the engine mechanism feature is a feature reflecting engine working rules, performance and states according to engine principles, and the engine mechanism feature includes a total pressure ratio, a converted fuel flow and a high-pressure converted rotating speed. obtaining engine sensor measurement parameters and standard parameters of engine reference model outputs of a feature to be extracted ; The engine mechanism feature extractor is generated by training an initial engine mechanism feature extractor, wherein the initial engine mechanism feature extractor determines a fault category by using a cross-entropy loss function and quantifies features by using a hyperspherical feature clustering loss function; and the formation process of the initial engine mechanism feature extractor includes the following contents: inputting the engine sensor measurement parameters and the standard parameters of the engine reference model output of the features to be extracted into the engine mechanism feature extractor , obtaining the feature residuals output by the engine mechanism feature extractor , completing the engine mechanism feature extraction; An activation value q of the parameter nonlinear correlation mining layer is obtained, the activation value q is processed by the graph attention mechanism, passes through two convolution layers and a filter response normalization layer FRN, forms an initial engine mechanism feature extractor, and realizes a feature extraction function. Before the engine mechanism feature extractor is loaded, an engine mechanism feature extractor is constructed, 2. The parameter-based nonlinear mining layer-based engine mechanism feature extraction method according to claim 1, characterized in that, which includes the following contents: The parameter nonlinear correlation mining layer is used to correlate and mine mechanism feature parameters of an engine, a graph attention mechanism layer, two convolution layers and a filter response normalization layer FRN are combined to form an initial engine mechanism feature extractor; acquiring existing engine sensor measurement parameters and standard parameters output by an engine reference model to form a training sample set; The initial engine mechanism feature extractor is trained by using the training sample set to generate the engine mechanism feature extractor. The training process of the initial engine mechanism feature extractor includes the following contents:

3. The parameter-based nonlinear mining layer-based engine mechanism feature extraction method of claim 2, wherein, A fault label t is set to classify and number fault features. Wherein, n is the fault type number. applying the training sample set to the initial engine mechanism feature extractor to extract engine sensor measurement parameter features and engine reference model output parameter features , obtaining feature residuals ; calculating a difference between the non-faulty feature residuals and the faulty feature residuals separating into non-faulty feature residuals and faulty feature residuals calculating a difference between the non-faulty feature residuals and the faulty feature residuals and faulty feature residuals ;​ For the characteristic residual After performing fully connected layer (FC) and probability normalized Softmax processing, the fault category is determined using the cross-entropy loss function based on the fault label t. on the difference After spherical clustering analysis, the feature is quantified according to the fault label t using the hyperspherical feature clustering loss function; performing feature extraction training until the feature residuals satisfy the requirements of the cross-entropy loss function and the difference satisfy the requirements of the hyperspherical feature clustering loss function, the training of the initial engine mechanism feature extractor is completed, and the engine mechanism feature extractor is generated; The cross-entropy loss function is required to ensure correct fault classification.

4. The parameter-based nonlinear mining layer-based engine mechanism feature extraction method of claim 3, wherein, The calculation expression of the activation value q is:

5. The parameter-based nonlinear mining layer-based engine mechanism feature extraction method of claim 3, wherein, The requirement of the hyperspherical feature clustering loss function is to distribute the features of different classes on a hypersphere, control the distance from the center of the sphere, and make the two-norm of the fault-free feature residual as the center of the sphere approach 0. The requirement of the hyperspherical feature clustering loss function is to distribute the features of different classes on a hypersphere, control the distance from the center of the sphere, and make the two-norm of the fault-free feature residual as the center of the sphere approach 0.

6. The parameter-based nonlinear mining layer-based engine mechanism feature extraction method of claim 1, wherein, ​ ； wherein, represents a column vector of all ones of length m, and the superscript T represents the transpose operation; represents the Hadamard product; represents the element-wise power operation; y is an input parameter including engine sensor measured parameters and engine reference model output parameters .

Images