Data feature extraction method and system for aero-engine component characteristic map

Abstract

This invention relates to the field of performance analysis technology for aero-engine components, specifically to a method and system for extracting data features from characteristic maps of aero-engine components. The method includes: representing the characteristic map as a combination of a set of characteristic curves and metadata, and setting category labels and physical indices for the characteristic curves; obtaining geometric relationship components, trend relationship components, curvature relationship components, and performance variable relationship components between the characteristic curves; obtaining the node feature matrix, curve relationship matrix, multi-channel relationship tensor, adjacency weight matrix, and similarity of the characteristic map, and constructing a dataset; constructing a relationship feature learning model and training it; and extracting the curve relationship matrix, adjacency weight matrix, multi-channel relationship tensor, and similarity of the target characteristic map. This invention enables the measurement, comparison, classification, and retrieval of similarity between different characteristic maps, and provides a unified input for subsequent intelligent processing.

Description

Technical Field

[0001] This invention relates to the field of performance analysis technology for aero-engine components, and specifically to a method and system for extracting data features from characteristic maps of aero-engine components. Background Technology

[0002] The performance characteristics of aero-engine components are typically described in the form of characteristic diagrams. Figure 1 A typical aerodynamic performance characteristic curve consists of multiple characteristic curves, used to characterize the performance changes of a component under different operating conditions or different operating parameters. For example, in compressor or fan components, multiple converted speed lines, efficiency characteristic curves, or boundary characteristic curves are often used to describe the aerodynamic performance characteristics of the component.

[0003] Existing component characteristic analysis methods are mostly based on single characteristic curves or single characteristic diagrams, typically analyzing characteristic curves through manual experience or comparison of local parameters. While these methods are applicable to processing small amounts of characteristic data, they struggle to provide a unified description and quantitative comparison of the overall structure of a component characteristic diagram when the diagram contains multiple characteristic curves and these curves differ in number, distribution, and shape. Furthermore, in engineering practice, component characteristic diagrams from different sources often suffer from inconsistent parameter scales, different coordinate definitions, and varying data resolutions, making direct alignment and comparison of characteristic curves difficult. Even within the same characteristic diagram, there is a lack of systematic modeling and characterization methods for the relative positions, trends, and overall distribution of different characteristic curves.

[0004] With the development of data-driven methods and intelligent computing technologies, analyzing component characteristic maps as a whole and quantifying their structural information through feature extraction methods has become an important development direction for improving the efficiency and consistency of component characteristic analysis. However, existing data-driven methods mostly focus on modeling single performance parameters or discrete data points, lacking effective technical solutions for relationship modeling and feature learning for characteristic map structures composed of multiple characteristic curves.

[0005] Therefore, there is a need to provide a data feature extraction method and system for aero-engine component characteristic maps to solve the above problems. Summary of the Invention

[0006] To address the problem that existing data-driven methods often focus on modeling single performance parameters or discrete data points, lacking effective technical solutions for relationship modeling and feature learning of "characteristic diagram structures composed of multiple characteristic curves," this invention provides a data feature extraction method and system for aero-engine component characteristic diagrams to solve the existing problems.

[0007] The first aspect of this invention provides a method for extracting data features from characteristic maps of aero-engine components, employing the following technical solution, including: Obtain the characteristic map corresponding to each component of the aero-engine, and represent the characteristic map as a combination of a set of characteristic curves and metadata; set the category labels and physical indexes of the characteristic curves; The discrete points in the characteristic curves are standardized to obtain standardized discrete points. Based on the three-dimensional parameters of the standardized discrete points, each characteristic curve is reparameterized, resampled, and aligned to obtain the correspondence between the characteristic curves. The three-dimensional parameters include: a first-dimensional performance variable, a second-dimensional performance variable, and a third-dimensional performance variable. Based on the correspondence between the characteristic curves, the average geometric distance between two characteristic curves is taken as the geometric relationship component between the two characteristic curves. The trend relationship component between the two characteristic curves is obtained based on the discrete first-order difference vector of the characteristic curves, the discrete curvature is obtained based on the discrete second-order difference vector of the two characteristic curves, and the curvature relationship component between the curves is obtained based on the discrete curvature. The performance variable relationship component between the characteristic curves is obtained based on the third-dimensional performance variable. Based on the geometric, trend, curvature, and performance variable relationships between two characteristic curves, the curve relationship matrix, multi-channel relationship tensor, and adjacency weight matrix of the characteristic graph are obtained. Based on the number of characteristic curves, physical index, and curve relationship matrix of the two characteristic graphs, the structural difference between the two characteristic graphs is obtained, and the similarity between the two characteristic graphs is obtained based on the structural difference. Based on the three-dimensional parameters, discrete first-order difference vector, and discrete curvature of each resampling point in the characteristic curve, the node feature vector of each characteristic curve and the node feature matrix of the characteristic graph are constructed. The dataset is constructed with the node feature matrix as input parameter and the curve relationship matrix, adjacency weight matrix, multi-channel relationship tensor, and similarity of the characteristic graph as output parameters. Construct a relation feature learning model, and train the relation feature learning model on the dataset to obtain a trained target relation feature learning model; The node feature matrix of the target feature map to be extracted is input into the target relation feature learning model to predict the curve relation matrix, adjacency weight matrix, multi-channel relation tensor, and similarity of the target feature map.

[0008] A further technical solution of the present invention is that the step of representing the characteristic map as a combination of a set of characteristic curves and metadata is as follows:

[0009] In the formula, Representing the characteristic map, This represents the set of characteristic curves in the characteristic diagram. A collection of metadata related to the feature graph; For the first Characteristic curves.

[0010] A further technical solution of the present invention is as follows: the steps of reparameterizing, resampling, and aligning the three-dimensional parameters of the standardized discrete points in the characteristic curves to obtain the correspondence between the characteristic curves are as follows: The reparameterized characteristic curve is obtained by reparameterizing each characteristic curve based on the three-dimensional parameters of the standardized discrete points in the characteristic curve; Set a uniform number of resampling points and perform interpolation resampling on the reparameterized characteristic curve to obtain resampling points; The geometric distance between two resampled points on the two characteristic curves is obtained based on the three-dimensional parameters of the resampled points of the two reparameterized characteristic curves, and the geometric distance is used as the alignment cost. Define the alignment path of the two characteristic curves as a sequence of physical index pairs; The alignment cost of two characteristic curves is obtained based on the physical index, sequence, and alignment cost. The alignment path with the minimum alignment cost is taken as the optimal alignment path. Obtain the optimal physical index pair sequence between characteristic curves based on the optimal alignment path; The optimal physical index pair sequence between characteristic curves is used as the correspondence between characteristic curves.

[0011] A further technical solution of the present invention is as follows: based on the correspondence between characteristic curves, the average geometric distance between two characteristic curves is used as the geometric relationship component between the two characteristic curves; the trend relationship component between the two characteristic curves is obtained based on the discrete first-order difference vector of the characteristic curves; the discrete curvature is obtained based on the discrete second-order difference vector between the two characteristic curves; the curvature relationship component between the curves is obtained based on the discrete curvature; the step of obtaining the performance variable relationship component between the curves based on the third-dimensional performance variable is as follows: The expression for the geometric relationship components between the two characteristic curves is:

[0012] In the formula, Characteristic curves With characteristic curve Geometric relationship components between them; Indicates the total number of resampling points; Characteristic curves Chinese Physical Index The corresponding three-dimensional parameters at the resampling point; Characteristic curves Neutral and characteristic curves Same physical index The corresponding three-dimensional parameters at the resampling point; This indicates a request for distance; Indicates the characteristic curve after alignment mapping. Physical index of resampling points In the characteristic curve The corresponding physical index on; The expression for the trend relationship component between the two characteristic curves is:

[0013] In the formula, Characteristic curves With characteristic curve The trend relationship between the components; Characteristic curves Chinese Physical Index The discrete first-order difference vector at the corresponding resampling point; Characteristic curves Neutral and characteristic curves Same physical index The discrete first-order difference vector at the corresponding resampling point; The expression for the curvature relationship components between the two characteristic curves is:

[0014]

[0015] In the formula, Characteristic curves With characteristic curve The curvature relationship components between them; Characteristic curves Chinese Physical Index The discrete curvature at the corresponding resampling point; Characteristic curves Neutral and characteristic curves Same physical index The discrete curvature at the corresponding resampling point; This represents the first-order difference of the first-dimensional performance variable; This represents the second difference of the first-dimensional performance variable; This represents the first difference of the second-dimensional performance variable; This represents the second-order difference of the second-dimensional performance variable; This represents a non-negative constant, used to avoid the denominator being zero; The expression for the performance variable relationship components between the two characteristic curves is as follows:

[0016] In the formula, Characteristic curves With characteristic curve The performance variable relationships between components; Characteristic curves Chinese Physical Index The third-dimensional performance variable at the corresponding resampling point; Characteristic curves Neutral and characteristic curves Same physical index The third-dimensional performance variable at the corresponding resampling point.

[0017] A further technical solution of the present invention is that the steps for obtaining the curve relationship matrix, multi-channel relationship tensor, and adjacency weight matrix of the feature map are as follows: The expression for the curve relationship matrix of the characteristic map is:

[0018]

[0019] In the formula, The matrix representing the curve relationships in the characteristic plot; A comprehensive relational matrix representing the characteristic map; Characteristic curves With characteristic curve The performance variable relationships between components; Characteristic curves With characteristic curve The curvature relationship components between them; Characteristic curves With characteristic curve The trend relationship between the components; Characteristic curves With characteristic curve Geometric relationship components between them; Non-negative weighting coefficients representing geometric relational components; Non-negative weighting coefficients representing trend relationship components; Non-negative weighting coefficients representing curvature relation components; Represents the non-negative weighting coefficients of the relational components of performance variables; The expression for the multichannel relation tensor of the feature map is:

[0020]

[0021] In the formula, The multichannel relation tensor representing the feature map; Characteristic curves With characteristic curve The multichannel relationship tensor between them; This indicates the number of characteristic curves in the characteristic diagram; The expression for the adjacency weight matrix of the feature graph is:

[0022]

[0023] In the formula, The adjacency weight matrix represents the characteristic graph; Characteristic curves With characteristic curve The adjacency weights between them; This represents the kernel width parameter.

[0024] A further technical solution of the present invention is to obtain the structural difference degree of two characteristic maps based on the number of characteristic curves, physical index, and curve relationship matrix of the two characteristic maps, and the step of obtaining the similarity of the two characteristic maps based on the structural difference degree is as follows: When the number of characteristic curves in two characteristic maps is the same and their indices are aligned, the structural difference between the two characteristic maps is obtained based on the curve relationship matrix; and the similarity between the two characteristic maps is obtained based on the structural kernel scale parameter and the structural difference. When the number of characteristic curves in two characteristic maps is inconsistent, the spectral signature vector is obtained based on the adjacency weight matrix, and the similarity between the two characteristic maps is obtained based on the spectral signature vector.

[0025] A further technical solution of the present invention is that the expression for the nodal feature vector of the characteristic curve is:

[0026] In the formula, Characteristic curves Node feature vectors; Characteristic curves In physical index The discrete first-order difference vector at the resampling point; Characteristic curves Chinese Physical Index The first-dimensional performance parameter at the corresponding resampling point; Characteristic curves Chinese Physical Index The second-dimensional performance parameter at the corresponding resampling point; Characteristic curves Chinese Physical Index The third-dimensional performance parameter at the corresponding resampling point; Characteristic curves Chinese Physical Index The discrete curvature at the corresponding resampling point; This indicates the number of resampling points.

[0027] A further technical solution of the present invention is that the relation feature learning model includes an attention weight module and a graph-level embedding module. The attention weight module is used to introduce attention weights into the update rules of adjacent layers of the relation feature learning model to obtain the target update rule. The graph-level embedding module is used to standardize each output parameter in the dataset and concatenate the standardization results to obtain a graph-level embedding vector. The target update rule is as follows:

[0028] In the formula, Represents a non-linear activation function; Indicates the first The learnable parameter matrix used by the layer for transforming the features of its own nodes; Indicates the first The first in the layer The hidden representation values ​​corresponding to the feature curves; Indicates attention weight; Indicates the first The set of neighbors of each characteristic curve on the characteristic graph; Indicates the first The layer is a learnable parameter matrix used for aggregating neighbor node information; Indicates the number of model layers; Characteristic curves With characteristic curve The multichannel relationship tensor between them; Indicates the updated result of the first... The first in the layer The hidden representation values ​​corresponding to the feature curves; Indicates the first The first in the layer The hidden representation values ​​corresponding to the feature curves; The graph-level embedding vector is:

[0029] In the formula, Represents a graph-level embedding vector; Indicates the first Weighted operators for characteristic curves; Indicates the first in the final layer The hidden representation values ​​of the characteristic curves; This indicates the number of characteristic curves in the characteristic graph.

[0030] A further technical solution of the present invention includes: setting a total loss function, using the total loss function as the loss function of the relation feature learning model, wherein the expression of the total loss function is:

[0031]

[0032]

[0033]

[0034] In the formula, This is the total loss function; To measure the learning loss function; The structural reconstruction loss function; The consistency constraint loss function; The weights are used to measure the learning loss function; The weights for the structural reconstruction loss function; The weights of the consistency constraint loss function; For each anchor sample, the graph-level embedding vector is used. The graph-level embedding vector corresponding to the positive sample; The graph-level embedding vector corresponding to the negative sample; For interval parameters; For the Sigmoid function; For the final layer Hidden representation of the characteristic curves; For the final layer Hidden representation of the characteristic curves; For the reconstructed characteristic curve With characteristic curve The adjacency weights between them; Characteristic curves With characteristic curve The adjacency weights between them; The number of characteristic curves involved in modeling relational features in the current component characteristic diagram; It is a symmetric consistency constraint function; This is a hierarchical consistency constraint function; This is the cross-graph distance consistency constraint function; The weights are the symmetric consistency constraint functions. The weights of the hierarchical consistency constraint function; The weights of the cross-graph distance consistency constraint function; The expression for the symmetric consistency constraint function is:

[0035] In the formula, Characteristic curve For characteristic curves Relationship matrix; Characteristic curve For characteristic curves Relationship matrix; Characteristic curve Its relation matrix to itself; The expression for the hierarchical consistency constraint function is:

[0036]

[0037] In the formula, For the first The physical index of each characteristic curve; For the first The physical index of each characteristic curve; For the first The projection scalar of the characteristic curve; For the first The projection scalar of the characteristic curve; For the final layer Hidden representation of the characteristic curves; This is the sorting interval parameter; for Learnable projection vectors under the category; The expression for the cross-graph distance consistency constraint function is:

[0038]

[0039] In the formula, For characteristic map and characteristic diagram Teacher structural distance; For scale matching function; For a set of sample pairs; For characteristic map and characteristic diagram Student distance; For characteristic map ; For characteristic map ; For characteristic map The corresponding graph-level embedding vector; For characteristic map The corresponding graph-level embedding vector.

[0040] A second aspect of the present invention provides a data feature extraction system for characteristic maps of aero-engine components, comprising: The parameter preprocessing module is used to obtain the characteristic map corresponding to each component of the aero-engine, and represent the characteristic map as a combination of a set of characteristic curves and metadata; and to set the category labels and physical indexes of the characteristic curves. The relation component acquisition module is used to standardize the discrete points in the characteristic curves to obtain standardized discrete points. Based on the three-dimensional parameters of the standardized discrete points in the characteristic curves, each characteristic curve is reparameterized, resampled, and aligned for mapping to obtain the correspondence between the characteristic curves. The three-dimensional parameters include: a first-dimensional performance variable, a second-dimensional performance variable, and a third-dimensional performance variable. Based on the correspondence between the characteristic curves, the average geometric distance between two characteristic curves is used as the geometric relation component between them. The module also obtains the trend relation component between the two characteristic curves based on the discrete first-order difference vector, the discrete curvature based on the discrete second-order difference vector, and the curvature relation component between the curves based on the discrete curvature. Finally, the module obtains the performance variable relation component between the two characteristic curves based on the third-dimensional performance variable. The dataset construction module is used to obtain the curve relationship matrix, multi-channel relationship tensor, and adjacency weight matrix of the characteristic map based on the geometric relationship components, trend relationship components, curvature relationship components, and performance variable relationship components between two characteristic curves; to obtain the structural difference between the two characteristic maps based on the number of characteristic curves, physical index, and curve relationship matrix; and to obtain the similarity between the two characteristic maps based on the structural difference. Based on the three-dimensional parameters, discrete first-order difference vector, and discrete curvature of each resampling point in the characteristic curve, the module constructs the node feature vector of each characteristic curve and the node feature matrix of the characteristic map. The module uses the node feature matrix as input parameters and the curve relationship matrix, adjacency weight matrix, multi-channel relationship tensor, and similarity of the characteristic map as output parameters to construct the dataset. The relation feature learning model module is used to build a relation feature learning model. The relation feature learning model is trained on the dataset to obtain a trained target relation feature learning model. The system also includes a feature extraction module, which inputs the node feature matrix of the target feature map to be extracted into the target relation feature learning model to predict the curve relation matrix, adjacency weight matrix, multi-channel relation tensor, and similarity of the target feature map.

[0041] The beneficial effects of this invention are: This invention takes curve standardization, curve alignment, relation feature learning model, and feature learning as the main line, transforms the feature graph composed of multiple feature curves into a learnable structured representation, and outputs curve relation matrix, adjacency weight matrix, multi-channel relation tensor, and similarity that can reflect the overall structural features of the feature graph. This enables the similarity measurement, comparison, classification, and retrieval between different feature graphs, and provides a unified input for subsequent intelligent processing. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a flowchart illustrating a method for extracting data features from a characteristic map of an aero-engine component according to the present invention. Figure 2 This is a system block diagram of a data feature extraction system for aero-engine component characteristic maps according to the present invention; Figure 3 This is a schematic diagram illustrating the composition of the characteristic curves in the characteristic graph of an embodiment of the present invention; Figure 4 This is a diagram showing the separability of the relation matrix of similar characteristic graphs in an embodiment of the present invention; Figure 5 This is a diagram showing the separability of the heterogeneous characteristic graph relation matrix in an embodiment of the present invention; Figure 6 This is a schematic diagram of the feature learning training convergence curve of the relation feature learning model in an embodiment of the present invention; Figure 7 This is a clustering partitioning result diagram of the characteristic diagram of the compression component in an embodiment of the present invention; Figure 8 This is a clustering partitioning result diagram of the characteristic diagram of the expansion component in an embodiment of the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] An embodiment of the data feature extraction method for characteristic maps of aero-engine components according to the present invention is as follows: Figure 1As shown, it includes: S1. Represent the feature graph as a combination of a set of feature curves and metadata, and set the category labels and physical indexes of the feature curves; Specifically, the characteristic map corresponding to each component of the aero-engine is obtained, and the characteristic map is represented as a combination of a set of characteristic curves and metadata; the category labels and physical indexes of the characteristic curves are set.

[0046] For example, in one specific embodiment, the steps of representing a feature map as a combination of a set of feature curves and metadata are as follows:

[0047] In the formula, Representing the characteristic map, This represents the set of characteristic curves in the characteristic diagram. A collection of metadata related to the feature graph; For the first Characteristic curves.

[0048] For example, in one specific embodiment, the steps of setting the category label and physical index of the characteristic curve are as follows: To preserve the structural information of the curve family, a curve identifier vector is introduced. Set the category label and physical index of the characteristic curve to the curve, that is:

[0049] In the formula, This indicates the curve category label, used to distinguish curve types, including converted speed lines, efficiency lines, and boundary lines. This represents the curve's physical index, used to characterize the curve's sorting or physical value (e.g., converted speed value or efficiency level) within its category.

[0050] S2. Obtain the geometric relationship components, trend relationship components, curvature relationship components, and performance variable relationship components between characteristic curves; Specifically, S21, standardize the discrete points in the characteristic curves to obtain standardized discrete points; S22, based on the three-dimensional parameters of the standardized discrete points in the characteristic curves, reparameterize, resample, and align the mapping of each characteristic curve to obtain the correspondence between the characteristic curves; the three-dimensional parameters include: the first-dimensional performance variable, the second-dimensional performance variable, and the third-dimensional performance variable; S23, based on the correspondence between the characteristic curves, take the average geometric distance between two characteristic curves as the geometric relationship component between the two characteristic curves; obtain the trend relationship component between the two characteristic curves based on the discrete first-order difference vector of the characteristic curves; obtain the discrete curvature based on the discrete second-order difference vector between the two characteristic curves; obtain the curvature relationship component between the curves based on the discrete curvature; obtain the performance variable relationship component between the curves based on the third-dimensional performance variable.

[0051] For example, in one specific embodiment, the process of standardizing the discrete points in the characteristic curve to obtain standardized discrete points in step S21 is as follows: Any characteristic curve can be represented as a parameterized expression of discrete points, that is:

[0052] In the formula, Characteristic curves The number of discrete points; Characteristic curves The first Three-dimensional parameters of a discrete point; Characteristic curves ; This represents the first dimension of the performance variable; This represents the second dimension of the performance variable; This represents the third-dimensional performance variable; in this embodiment, the first-dimensional performance variable is the converted flow rate. or equivalent flow The second performance variable is either the pressure ratio or the expansion ratio. The third performance variable is efficiency. .

[0053] For example, in this embodiment, considering the differences in coordinate scale, unit, and range among various components and characteristic maps of the aero-engine, a standardized mapping operator is constructed. Based on standardized mapping operators The three-dimensional parameters of the discrete points are standardized, i.e., the standardized three-dimensional parameters of the discrete points are:

[0054] In the formula, Indicates the characteristic curve The first The three-dimensional parameters after standardization of discrete points; This represents the first dimension of the performance variable after standardization. This represents the second dimension of the performance variable after standardization. This represents the third-dimensional performance variable after standardization. Among them, the normalized mapping operator Component normalization or Z-score standardization is employed; the steps for obtaining the three-dimensional parameters of the standardized discrete points using the component normalization method are as follows:

[0055] in, This represents the minimum value of the first-dimensional performance variable within the statistical range of the characteristic plot; This represents the minimum value of the second-dimensional performance variable within the statistical range of the characteristic plot; This represents the minimum value of the third-dimensional performance variable within the statistical range of the characteristic plot; This represents the maximum value of the first dimension performance variable within the statistical range of the characteristic plot; This represents the maximum value of the second-dimensional performance variable within the statistical range of the characteristic plot; This represents the maximum value of the third-dimensional performance variable within the statistical range of the characteristic plot; The steps for obtaining the standardized three-dimensional parameters of discrete points using the Z-score standardization method are as follows:

[0056] In the formula, This represents the mean of the first dimension of the performance variable within the statistical range of the characteristic plot; The standard deviation of the first dimension of the performance variable within the statistical range of the characteristic plot; This represents the mean of the second-dimensional performance variable within the statistical range of the characteristic plot; The standard deviation of the second-dimensional performance variable within the statistical range of the characteristic plot; This represents the mean of the third-dimensional performance variable within the statistical range of the characteristic plot; It represents the standard deviation of the third-dimensional performance variable within the statistical range of the characteristic plot.

[0057] For example, in a specific embodiment, the step S22, which involves reparameterizing, resampling, and aligning each characteristic curve based on the three-dimensional parameters of the standardized discrete points in the characteristic curves to obtain the correspondence between characteristic curves, is as follows: S221, reparameterizing each characteristic curve based on the three-dimensional parameters of the standardized discrete points in the characteristic curves to obtain reparameterized characteristic curves; S222, setting a uniform resampling number and interpolating and resampling the reparameterized characteristic curves to obtain resampling points; S223, obtaining the geometric distance between two resampling points on the two characteristic curves based on the three-dimensional parameters of the resampling points of the two reparameterized characteristic curves, and using the geometric distance as the alignment cost; defining the alignment path of the two characteristic curves as a physical index pair sequence; obtaining the alignment cost of the two characteristic curves based on the physical index pair sequence and the alignment cost; taking the alignment path corresponding to the minimum alignment cost as the optimal alignment path; obtaining the optimal physical index pair sequence between the characteristic curves based on the optimal alignment path; and using the optimal physical index pair sequence between the characteristic curves as the correspondence between the characteristic curves.

[0058] S221. The steps for reparameterizing each characteristic curve by standardizing the three-dimensional parameters of discrete points in the characteristic curve to obtain the reparameterized characteristic curve are as follows: To achieve a comparable parameterized representation between characteristic curves, the three-dimensional coordinate vector of the standardized discrete points is represented as:

[0059] In the formula, Characteristic curves The Middle A three-dimensional coordinate vector of a standardized discrete point; Define the cumulative arc length of the characteristic curve as:

[0060] In the formula, Characteristic curves From the first standardized discrete point to the... The cumulative arc length of a standardized discrete point; Represents the L2 norm; Characteristic curves The total number of standardized discrete points; Characteristic curves The Middle A three-dimensional coordinate vector of a standardized discrete point; Characteristic curves The Middle A three-dimensional coordinate vector of a standardized discrete point; The three-dimensional parameters of the standardized discrete points on the reparameterized characteristic curve are then expressed as:

[0061] In the formula, The first parameter on the reparameterized characteristic curve Three-dimensional parameters of a standardized discrete point; Characteristic curves From the first standardized discrete point to the... The cumulative arc length of a standardized discrete point; S222. The steps for setting a uniform resampling number and interpolating and resampling the reparameterized characteristic curve to obtain resampling points are as follows: Set a uniform number of resampling points as Then we have:

[0062] In the formula, Represents the first in the uniform resampling parameter grid. The normalized arc length parameter has a range of values. ; Indicates the number of resampling points. Indicates the resampling parameter number; Interpolation and resampling are performed on each reparameterized characteristic curve to obtain:

[0063] In the formula, Represents the reparameterized characteristic curve The Middle Three-dimensional parameters of each resampling point; Represents the reparameterized characteristic curve The Middle The first dimension of the performance variable at each resampling point; Represents the reparameterized characteristic curve The Middle The second-dimensional performance variable of each resampling point; Represents the reparameterized characteristic curve The Middle The third-dimensional performance variable of each resampling point; Represents the reparameterized characteristic curve The interpolation sampling function in the standardized space; wherein the interpolation sampling function adopts one of linear interpolation, spline interpolation or piecewise polynomial interpolation, and this embodiment adopts linear interpolation.

[0064] The three-dimensional coordinate vector of the resampling point is then represented as:

[0065] In the formula, Represents the reparameterized characteristic curve The Middle The three-dimensional coordinate vector of each resampled point.

[0066] S223. The steps to obtain the correspondence between characteristic curves are as follows: To establish the correspondence between curves, characteristic curves are defined. With characteristic curve The alignment cost matrix elements are:

[0067] In the formula, Characteristic curves The first in Individual sampling points and characteristic curves The first in The geometric distance between resampled points; in this embodiment, the geometric distance between two resampled points is used as the performance curve. With characteristic curve The alignment cost matrix elements; Define from characteristic curves With characteristic curve The alignment path is a sequence of physical index pairs, that is:

[0068] In the formula, Characteristic curves With characteristic curve Alignment path; Indicates the path length; Indicates the first position in the alignment path For each corresponding physical index pair, the alignment path satisfies the index monotonicity and continuity constraints, ensuring that the resampled point sequence in the characteristic curve does not undergo reverse matching during the alignment process.

[0069] Based on this, the alignment cost can be defined as:

[0070] In the formula, Characteristic curves With characteristic curve Minimum alignment cost between; This indicates the penalty coefficient for inconsistency in curve categories; Indicates an indicator function, when When the condition inside the parentheses is true Take 1, otherwise Take 0.

[0071] The alignment path corresponding to the minimum alignment cost is taken as the optimal alignment path. An alignment mapping function can be constructed from the optimal alignment path, and the alignment mapping function is used to indicate the characteristic curve. resampling index In the characteristic curve For the corresponding index, the alignment mapping function is:

[0072] In the formula, Indicates the characteristic curve after alignment mapping. Physical index of resampling points In the characteristic curve The corresponding physical index on the graph can then be used to obtain the characteristic curve. With characteristic curve The corresponding optimal physical index pair sequence; the optimal physical index pair sequence between characteristic curves is the correspondence between characteristic curves.

[0073] For example, in a specific embodiment, S23, based on the correspondence between characteristic curves, the average geometric distance between two characteristic curves is taken as the geometric relationship component between the two characteristic curves; the trend relationship component between the two characteristic curves is obtained based on the discrete first-order difference vector of the characteristic curves; the discrete curvature is obtained based on the discrete second-order difference vector between the two characteristic curves; the curvature relationship component between the curves is obtained based on the discrete curvature; the step of obtaining the performance variable relationship component between the curves based on the third-dimensional performance variable is as follows: In one specific embodiment, the step of obtaining the geometric relationship components between the characteristic curves is as follows: Based on the correspondence between the characteristic curves in step S22, that is, based on the optimal physical index pair sequence of the two characteristic curves, the average geometric distance between the two characteristic curves is:

[0074] In the formula, Characteristic curves With characteristic curve Geometric relationship components between them; Indicates the total number of resampling points; Characteristic curves Chinese Physical Index The corresponding three-dimensional parameters at the resampling point; Characteristic curves Neutral and characteristic curves Same physical index The corresponding three-dimensional parameters at the resampling point; This indicates a request for distance; Indicates the characteristic curve after alignment mapping. Physical index of resampling points In the characteristic curve The corresponding physical index is used in this embodiment, where the average geometric distance between the two characteristic curves is taken as the geometric relationship component between the characteristic curves.

[0075] In one specific embodiment, the step of obtaining the trend relationship component between the two characteristic curves is as follows: To characterize the differences in curve shape trends, the discrete first-order difference vector is defined as:

[0076] In the formula, Characteristic curves Chinese Physical Index The discrete first-order difference vector at the corresponding resampling point; Characteristic curves Chinese Physical Index The corresponding three-dimensional parameters at the resampling point; Characteristic curves Chinese Physical Index The corresponding three-dimensional parameters at the resampling point; Represents the first in the uniform resampling parameter grid. A normalized arc length parameter value; The expression for the trend relationship component between the two characteristic curves is:

[0077] In the formula, Characteristic curves With characteristic curve The trend relationship between the components; Characteristic curves Chinese Physical Index The discrete first-order difference vector at the corresponding resampling point; Characteristic curves Neutral and characteristic curves Same physical index The discrete first-order difference vector at the corresponding resampling point; In one specific embodiment, the step of obtaining the curvature relationship components between the two characteristic curves is as follows: The expression for the discrete second-order difference vector is:

[0078] In the formula, Represents a discrete second-order difference vector; make , The discrete curvature is:

[0079] In the formula, This represents the first-order difference of the first-dimensional performance variable; This represents the second difference of the first-dimensional performance variable; This represents the first difference of the second-dimensional performance variable; This represents the second-order difference of the second-dimensional performance variable; This represents a non-negative constant, used to avoid the denominator being zero. Therefore, the expression for the curvature relationship components between the two characteristic curves is:

[0080] In the formula, Characteristic curves With characteristic curve The curvature relationship components between them; Characteristic curves Chinese Physical Index The discrete curvature at the corresponding resampling point; Characteristic curves Neutral and characteristic curves Same physical index The discrete curvature at the corresponding resampling point.

[0081] In one specific embodiment, the expression for the performance variable relationship components between the two characteristic curves is as follows:

[0082] In the formula, Characteristic curves With characteristic curve The performance variable relationships between components; Characteristic curves Chinese Physical Index The third-dimensional performance variable at the corresponding resampling point; Characteristic curves Neutral and characteristic curves Same physical index The third-dimensional performance variable at the corresponding resampling point.

[0083] At this point, the geometric relationship components, trend relationship components, curvature relationship components, and performance variable relationship components between the two characteristic curves can be obtained.

[0084] S3. Obtain the node feature matrix, curve relationship matrix, multi-channel relationship tensor, adjacency weight matrix, and similarity of the feature graph, and construct the dataset; Specifically, S31, based on the geometric relationship components, trend relationship components, curvature relationship components, and performance variable relationship components between curves, obtain the curve relationship matrix and multi-channel relationship tensor of the characteristic map; obtain the adjacency weight matrix of the characteristic map based on the curve relationship matrix; S32, based on the number of characteristic curves, physical index, and curve relationship matrix of the two characteristic maps, obtain the structural difference degree of the two characteristic maps, and obtain the similarity of the two characteristic maps based on the structural difference degree; S33, based on the three-dimensional parameters, discrete first-order difference vector, and discrete curvature of each resampling point in the characteristic curve, construct the node feature vector of each characteristic curve and the node feature matrix of each characteristic map; S34, using the node feature matrix as input parameters and the curve relationship matrix, adjacency weight matrix, multi-channel relationship tensor, and similarity of the characteristic map as output parameters, construct the dataset.

[0085] For example, in one specific embodiment, S31, based on the geometric relationship components, trend relationship components, curvature relationship components, and performance variable relationship components between curves, the curve relationship matrix and multi-channel relationship tensor of the characteristic map are obtained; the step of obtaining the adjacency weight matrix of the characteristic map based on the curve relationship matrix is ​​as follows: The comprehensive relationship matrix is ​​obtained based on the geometric relationship components, trend relationship components, curvature relationship components, and performance variable relationship components between curves. The expression for the comprehensive relationship matrix is ​​as follows:

[0086] The expression for the curve relationship matrix of the characteristic map is:

[0087] In the formula, The matrix representing the curve relationships in the characteristic plot; A comprehensive relational matrix representing the characteristic map; Characteristic curves With characteristic curve The performance variable relationships between components; Characteristic curves With characteristic curve The curvature relationship components between them; Characteristic curves With characteristic curve The trend relationship between the components; Characteristic curves With characteristic curve Geometric relationship components between them; Non-negative weighting coefficients representing geometric relational components; Non-negative weighting coefficients representing trend relationship components; Non-negative weighting coefficients representing curvature relation components; The non-negative weighting coefficients represent the relational components of performance variables.

[0088] The expression for the multi-channel relation tensor of the feature map is:

[0089]

[0090] In the formula, The multichannel relation tensor representing the feature map; Characteristic curves With characteristic curve The multichannel relationship tensor between them; This indicates the number of characteristic curves in the characteristic diagram; To facilitate relation learning based on the computational model, the relation matrix is ​​mapped to adjacency weights using a kernel function. The expression for the adjacency weight matrix of the feature graph is as follows:

[0091]

[0092] In the formula, The adjacency weight matrix represents the characteristic graph; Characteristic curves With characteristic curve The adjacency weights between them; This represents the kernel width parameter.

[0093] For example, in a specific embodiment, step S32, obtaining the structural difference degree of the two characteristic maps based on the number of characteristic curves, physical index, and curve relationship matrix of the two characteristic maps, and the specific process of obtaining the similarity of the two characteristic maps based on the structural difference degree is as follows: when the number of characteristic curves of the two characteristic maps is the same and the index is aligned, the structural difference degree of the two characteristic maps is obtained based on the curve relationship matrix; and the similarity of the two characteristic maps is obtained based on the structural kernel scale parameter and the structural difference degree; when the number of characteristic curves of the two characteristic maps is inconsistent, the structural difference degree of the two characteristic maps is obtained based on the adjacency weight matrix, the spectral signature vector is obtained based on the structural difference degree, and the similarity of the two characteristic maps is obtained based on the spectral signature vector.

[0094] In this embodiment, when the number of characteristic curves in two characteristic maps is the same and their indices are aligned, the structural difference between the two characteristic maps is obtained based on the curve relationship matrix; and the similarity between the two characteristic maps is obtained based on the structural kernel scale parameter and the structural difference. When two characteristic maps have the same number of characteristic curves and their indices are aligned, the expression for the structural difference between the two characteristic maps is:

[0095] In the formula, express Feature diagram and Feature diagram Structural differences; express Feature diagram The curve relationship matrix; express Feature diagram The curve relationship matrix; It is the Frobenius norm.

[0096] The similarity between the two feature maps is:

[0097] In the formula, express Feature diagram and Feature diagram Similarity; For the structural kernel scale parameter.

[0098] In this embodiment, when the number of characteristic curves in two characteristic maps is inconsistent, the spectral signature vector is obtained based on the adjacency weight matrix, and the similarity between the two characteristic maps is obtained based on the spectral signature vector as follows: When two feature maps have different numbers of feature curves, it is difficult to directly calculate the matrix difference due to the different dimensions of the curve relationship matrix. This embodiment uses the feature map structural features constructed from the adjacency weight matrix to form a dimension-independent, comparable representation, defining the degree matrix as:

[0099] In the formula, Characteristic curves With characteristic curve The degree matrix between them; Characteristic curves With characteristic curve The adjacency weights between them; The normalized Laplace matrix is ​​then defined as:

[0100] In the formula, This represents the normalized Laplacian matrix, used to characterize the overall structural features of the feature map; Represents the identity matrix, whose dimensions and adjacency weight matrix are... same; The adjacency weight matrix represents the characteristic graph; The inverse square root matrix of the degree matrix of the adjacency weight matrix; Take before The smallest eigenvalues ​​form the spectral signature vector as follows:

[0101] In the formula, Represents the spectral signature vector; Represents the eigenvector; Indicates spectral dimension; express; Representing spectral dimension The corresponding spectral signature; the spectral distance is:

[0102] In the formula, express Feature diagram and Feature diagram spectral distance; express Feature diagram Spectral signature vector; express Feature diagram The spectral signature vector.

[0103] When the number of characteristic curves in two characteristic maps is inconsistent, the matrix difference cannot be directly calculated due to the different dimensions of their relation matrices. Therefore, the aforementioned spectral distance is used to compare the relation map structures of the two characteristic maps, and the similarity is further defined as follows:

[0104] In the formula, express Feature diagram and Feature diagram Similarity; This represents the spectral kernel scale parameter, used to adjust the effect of spectral distance on the degree of similarity decay; For example, in one specific embodiment, step S33, the process of constructing the node feature vector of each characteristic curve and the node feature matrix of each characteristic map based on the three-dimensional parameters, discrete first-order difference vector, and discrete curvature of each resampling point in the characteristic curve, is as follows: The expression for the nodal eigenvectors of each characteristic curve is:

[0105] In the formula, Characteristic curves Node feature vectors; Characteristic curves In physical index The discrete first-order difference vector at the resampling point; Characteristic curves Chinese Physical Index The first-dimensional performance parameter at the corresponding resampling point; Characteristic curves Chinese Physical Index The second-dimensional performance parameter at the corresponding resampling point; Characteristic curves Chinese Physical Index The third-dimensional performance parameter at the corresponding resampling point; Characteristic curves Chinese Physical Index The discrete curvature at the corresponding resampling point; This indicates the number of resampling points.

[0106] The node feature matrix of the feature graph is then:

[0107] In the formula, The node feature matrix represents the feature map. , For node feature dimensions; It is the field of real numbers, that is, the set of all real numbers.

[0108] For example, in one specific embodiment, S34, a dataset is constructed using the node feature matrix as input parameters and the curve relationship matrix, adjacency weight matrix, multi-channel relationship tensor, and similarity of the feature graph as output parameters.

[0109] S4. Construct a relation feature learning model and train it; Specifically, a relation feature learning model is constructed, and the model is trained on the dataset to obtain a trained target relation feature learning model.

[0110] For example, in one specific embodiment, the relation feature learning model includes: an attention weight module and a graph-level embedding module. The attention weight module is used to introduce attention weights into the update rules of adjacent layers of the relation feature learning model to obtain the target update rule; the graph-level embedding module is used to standardize each output parameter in the dataset and concatenate the standardization results to obtain a graph-level embedding vector.

[0111] In this embodiment, the initialization of the hidden representation is as follows:

[0112] In the formula, For the 0th layer hidden representation; then the 1st layer hidden representation... layer to the first The target update rule for the layer is:

[0113] In the formula, Represents a non-linear activation function; Indicates the first The learnable parameter matrix used by the layer for transforming the features of its own nodes; Indicates the first The first in the layer The hidden representation values ​​corresponding to the feature curves; Indicates attention weight; Indicates the first The set of neighbors of each characteristic curve on the characteristic graph; Indicates the first The layer is a learnable parameter matrix used for aggregating neighbor node information; Indicates the number of model layers; Characteristic curves With characteristic curve The multichannel relationship tensor between them; Indicates the updated result of the first... The first in the layer The hidden representation values ​​corresponding to the feature curves; Indicates the first The first in the layer The hidden representation values ​​corresponding to the feature curves; To enable the model to utilize different types of relational information, an attention score is defined:

[0114] In the formula, This indicates the score for attention. Characteristic curves With characteristic curve The multichannel relationship tensor between them; This is the attention scoring vector, used to map the concatenated feature representations to scalar attention scores; The linear transformation matrix representing the hidden representation of the current characteristic curve is used to extract the characteristic curve. Query characteristics; The linear transformation matrix representing the hidden representation of adjacent characteristic curves is used to extract characteristic curves. Key features; The linear transformation matrix representing the multi-channel relation tensor is used to transform the characteristic curves. With characteristic curve Multi-channel relationship information, including geometric, trend, curvature, and performance variable relationships, is mapped to the attention scoring space; attention scores are then assigned to these relationships. Normalization is performed to obtain a normalized attention score, which is then used as the attention weight. .

[0115] In this embodiment, the graph-level embedding vector is:

[0116] In the formula, Represents a graph-level embedding vector; Indicates the first Weighted operators for characteristic curves; Indicates the first in the final layer The hidden representation values ​​of the characteristic curves; This indicates the number of characteristic curves in the characteristic graph.

[0117] For example, in one specific embodiment, a total loss function is defined and used as the loss function of the relation feature learning model, wherein the expression of the total loss function is:

[0118]

[0119]

[0120]

[0121] In the formula, This is the total loss function; To measure the learning loss function; The structural reconstruction loss function; The consistency constraint loss function; The weights are used to measure the learning loss function; The weights for the structural reconstruction loss function; The weights of the consistency constraint loss function; For each anchor sample, the graph-level embedding vector is used. The graph-level embedding vector corresponding to the positive sample; The graph-level embedding vector corresponding to the negative sample; For interval parameters; For the Sigmoid function; For the final layer Hidden representation of the characteristic curves; For the final layer Hidden representation of the characteristic curves; For the reconstructed characteristic curve With characteristic curve The adjacency weights between them; Characteristic curves With characteristic curve The adjacency weights between them; The number of characteristic curves involved in modeling relational features in the current component characteristic diagram; It is a symmetric consistency constraint function; This is a hierarchical consistency constraint function; This is the cross-graph distance consistency constraint function; The weights are the symmetric consistency constraint functions. The weights of the hierarchical consistency constraint function; The weights of the cross-graph distance consistency constraint function; The expression for the symmetric consistency constraint function is:

[0122] In the formula, Characteristic curve For characteristic curves Relationship matrix; Characteristic curve For characteristic curves Relationship matrix; Characteristic curve Its relation matrix to itself; The expression for the hierarchical consistency constraint function is:

[0123]

[0124] In the formula, For the first The physical index of each characteristic curve; For the first The physical index of each characteristic curve; For the first The projection scalar of the characteristic curve; For the first The projection scalar of the characteristic curve; For the final layer Hidden representation of the characteristic curves; This is the sorting interval parameter; for Learnable projection vectors under the category; The expression for the cross-graph distance consistency constraint function is:

[0125]

[0126] In the formula, For characteristic map and characteristic diagram Teacher structural distance; For scale matching function; For a set of sample pairs; For characteristic map and characteristic diagram Student distance; For characteristic map ; For characteristic map ; For characteristic map The corresponding graph-level embedding vector; For characteristic map The corresponding graph-level embedding vector.

[0127] It should be noted that consistency constraint inference through the total loss function ensures that the feature representation meets the requirements of structural self-consistency and hierarchical consistency, thereby enabling the overall structure of the feature map to be quantitatively described and used for comparison, classification and subsequent intelligent analysis.

[0128] S5. Extract the curve relationship matrix, adjacency weight matrix, multi-channel relationship tensor, and similarity from the target feature map; Specifically, the node feature matrix of the target feature map from which features are to be extracted is input into the target relation feature learning model to predict the curve relation matrix, adjacency weight matrix, multi-channel relation tensor, and similarity of the target feature map.

[0129] This invention proposes an embodiment of a data feature extraction system for aero-engine component characteristic maps, such as... Figure 2As shown, this embodiment includes: a parameter preprocessing module, a relation component acquisition module, a dataset construction module, a relation feature learning model module, and a feature extraction module. The parameter preprocessing module is used to acquire the characteristic map corresponding to each component of the aero-engine and represent the characteristic map as a combination of a set of characteristic curves and metadata; it sets the category labels and physical indexes of the characteristic curves. The relation component acquisition module is used to standardize the discrete points in the characteristic curves to obtain standardized discrete points, and based on the three-dimensional parameters of the standardized discrete points in the characteristic curves, it sequentially performs reparameterization, resampling, and alignment mapping on each characteristic curve to obtain the correspondence between the characteristic curves. The three-dimensional parameters include: a first-dimensional performance variable, a second-dimensional performance variable, and a third-dimensional performance variable. Based on the correspondence between the characteristic curves, the average geometric distance between two characteristic curves is used as the geometric relationship component between the two characteristic curves. The trend relationship component between the two characteristic curves is obtained based on the discrete first-order difference vector of the characteristic curves, the discrete curvature is obtained based on the discrete second-order difference vector between the two characteristic curves, and the curvature relationship component between the curves is obtained based on the discrete curvature. The performance variable relationship between the two characteristic curves is obtained based on the third-dimensional performance variable. The dataset construction module is used to obtain the curve relationship matrix, multi-channel relationship tensor, and adjacency weight matrix of the characteristic map based on the geometric relationship components, trend relationship components, curvature relationship components, and performance variable relationship components between two characteristic curves. Based on the number of characteristic curves, physical index, and curve relationship matrix of the two characteristic maps, it obtains the structural difference between the two characteristic maps and the similarity between them. Based on the three-dimensional parameters, discrete first-order difference vector, and discrete curvature of each resampling point in the characteristic curves, it constructs the node feature vector of each characteristic curve and the node feature matrix of the characteristic map. Using the node feature matrix as input parameters and the curve relationship matrix, adjacency weight matrix, multi-channel relationship tensor, and similarity of the characteristic map as output parameters, it constructs the dataset. The relationship feature learning model module is used to construct a relationship feature learning model. Based on the dataset, it trains the relationship feature learning model to obtain the trained target relationship feature learning model. The feature extraction module inputs the node feature matrix of the target characteristic map from which features to be extracted into the target relationship feature learning model to predict the curve relationship matrix, adjacency weight matrix, multi-channel relationship tensor, and similarity of the target characteristic map.

[0130] The invention will be described below with reference to simulation data and accompanying drawings: like Figure 3 As shown, the characteristic diagram of the component consists of a set of curves. Composition, in which each curve Each has a corresponding curve category label. and physical index The isochronous rotation lines in the diagram, (Efficiency) contour lines, surge boundary lines, and boundary lines correspond to different types of characteristic curves. Iso-rotational speed lines are used to characterize the performance variation relationship under the same equivalent rotational speed or relative rotational speed. Contour lines are used to represent lines connecting operating points with the same efficiency; surge boundaries are used to define the stable operating region of the component; boundary lines are used to represent other operating range boundaries of the characteristic map. These curves together constitute the multi-curve structure of the component characteristic map, which is the foundation for subsequent relationship modeling and feature extraction.

[0131] like Figure 4 As shown, after completing the characteristic curve standardization and alignment, this embodiment constructs a relationship matrix between the characteristic curves to characterize the pairwise similarity between them. For similar characteristic graph pairs, the relationship matrix exhibits a clear block distribution structure, meaning that the relationship strength between characteristic curves within a class is high, while the relationship strength between characteristic curves between classes is low, thus forming a clear partitioning pattern. This result demonstrates that the method of this invention can explicitly express the category structure originally implicit in multiple curves, enabling similar characteristic graphs to have good separability in the relationship structure space. Figure 4 This demonstrates the effectiveness of the present invention in structural identification and relational modeling, illustrating that the method can truly reflect the consistency characteristics within a family of curves.

[0132] like Figure 5 As shown, when the input consists of heterogeneous feature graph pairs with significant structural differences, their relationship matrix no longer exhibits a clear block structure, but rather a discrete distribution lacking obvious partitioning features. This result demonstrates that the method of this invention does not artificially construct non-existent structural patterns, but objectively reflects the actual differences between curves, thereby ensuring the authenticity and robustness of the relationship modeling results. Figure 4 Used with Figure 5 In comparison, this further illustrates that the present invention is effective in identifying both structural consistency and structural differences, providing a reliable foundation for subsequent classification or clustering based on embedded features.

[0133] like Figure 6 As shown, after inputting the relational structure representation into the relational feature learning model for feature learning, the training loss gradually decreases and tends to stabilize with increasing training epochs. Specifically, the final loss value of the basic method is approximately 0.042, while the final loss value of the method in this invention is approximately 0.015, a reduction of about 64.3% compared to the basic method. Furthermore, the method in this invention essentially enters a stable convergence interval after 150 training epochs, with relatively small curve oscillation amplitude, indicating that the feature learning process based on the curve relational structure has better convergence stability and lower final error. These results demonstrate that the constructed curve relational matrix, multi-channel relational tensor, and adjacency weight matrix can provide the model with a more discriminative structural input, thereby improving the effectiveness and engineering feasibility of feature graph structural feature extraction.

[0134] like Figure 7 and Figure 8 As shown, this embodiment uses the graph-level embedding representation obtained in step S4 (feature learning stage) to perform two-dimensional visualization and clustering partitioning analysis on characteristic graph samples of different components. Among them, Figure 7 The clustering and partitioning results of the component characteristic map are for compression. Figure 8 This is the clustering partitioning result of the expansion component characteristic map. In the figure, "Feature 1" and "Feature 2" are two visual feature components obtained by two-dimensional projection or dimensionality reduction of the component characteristic map structural feature embedding vectors "pressure ratio / expansion ratio" and "flow rate". They are used to show the distribution relationship of different characteristic map samples in the structural feature space.

[0135] like Figure 7 As shown, this embodiment performs two-dimensional visualization and clustering partitioning analysis on compressed component characteristic map samples based on the graph-level embedding representation obtained in step S4, the feature learning stage. "Feature 1" and "Feature 2" in the figure are two visualized feature components obtained by two-dimensional projection or dimensionality reduction of the component characteristic map structural feature embedding vectors "compression ratio" and "flow rate". Sample points of different colors represent compressed component characteristic map samples of different categories or partitions; solid lines represent decision boundaries used to distinguish different sample groups; dashed lines represent the interval boundaries on both sides of the decision boundaries; and hollow circles represent support vector samples that play a major constraining role on the partition boundaries. Figure 7 As can be seen, the characteristic map samples of compression components exhibit relatively obvious linear separability in the structural feature space. Low-pressure component characteristic samples are mainly distributed in the upper right region of the feature space, while high-pressure component characteristic samples are mainly distributed in the lower left to middle region. The two types of samples can be distinguished by an approximately linear decision boundary. This result indicates that, for compression components, low-pressure and high-pressure components exhibit differences that can be captured by structural features in terms of curve family distribution, pressure ratio variation trends, and flow matching relationships. The structural features extracted in this invention can effectively reflect the category differences and overall structural differences between the characteristic maps of compression components.

[0136] like Figure 8 As shown, this embodiment performs two-dimensional visualization and clustering partitioning analysis on the expansion component characteristic map samples based on the graph-level embedding representation obtained in step S4, the feature learning stage. "Feature 1" and "Feature 2" in the figure are two visualized feature components obtained by two-dimensional projection or dimensionality reduction of the component characteristic map structural feature embedding vectors "expansion ratio" and "flow rate". Sample points of different colors represent expansion component characteristic map samples of different categories or partitions, black curves represent nonlinear partition boundaries, and hollow circles represent support vector samples. Figure 7 compared to, Figure 8The characteristic maps of medium-expansion components exhibit a more pronounced nonlinear distribution in the structural feature space. Low-pressure component characteristic maps are mainly concentrated in the upper left and middle regions of the feature space, while high-pressure component characteristic maps are mainly distributed in the lower middle to right-side regions. These two types of samples need to be distinguished by nonlinear partition boundaries. Figure 7 Compared to the compression components in the middle, Figure 8 The more curved boundary morphology of the expansion component samples indicates a stronger nonlinear coupling relationship between the expansion ratio, flow rate, and curve morphology in the expansion component characteristic diagram. The structural features extracted in this invention can still form a discriminative sample distribution in a low-dimensional feature space, which can be used to support component characteristic diagram classification, zonal modeling, and subsequent engineering analysis tasks.

[0137] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method of data feature extraction for an aeroengine component signature map, the method comprising: include: Obtain the characteristic map corresponding to each component of the aero-engine, and represent the characteristic map as a combination of a set of characteristic curves and metadata; Set the category label and physical index of the characteristic curve; The discrete points in the characteristic curves are standardized to obtain standardized discrete points. Based on the three-dimensional parameters of the standardized discrete points, each characteristic curve is reparameterized, resampled, and aligned to obtain the correspondence between the characteristic curves. The three-dimensional parameters include: a first-dimensional performance variable, a second-dimensional performance variable, and a third-dimensional performance variable. Based on the correspondence between the characteristic curves, the average geometric distance between two characteristic curves is taken as the geometric relationship component between the two characteristic curves. The trend relationship component between the two characteristic curves is obtained based on the discrete first-order difference vector of the characteristic curves, the discrete curvature is obtained based on the discrete second-order difference vector of the two characteristic curves, and the curvature relationship component between the curves is obtained based on the discrete curvature. The performance variable relationship component between the characteristic curves is obtained based on the third-dimensional performance variable. Based on the geometric, trend, curvature, and performance variable relationships between two characteristic curves, the curve relationship matrix, multi-channel relationship tensor, and adjacency weight matrix of the characteristic graph are obtained. Based on the number of characteristic curves, physical index, and curve relationship matrix of the two characteristic graphs, the structural difference between the two characteristic graphs is obtained, and the similarity between the two characteristic graphs is obtained based on the structural difference. Based on the three-dimensional parameters, discrete first-order difference vector, and discrete curvature of each resampling point in the characteristic curve, the node feature vector of each characteristic curve and the node feature matrix of the characteristic graph are constructed. The dataset is constructed with the node feature matrix as input parameter and the curve relationship matrix, adjacency weight matrix, multi-channel relationship tensor, and similarity of the characteristic graph as output parameters. Construct a relation feature learning model, and train the relation feature learning model on the dataset to obtain a trained target relation feature learning model; The node feature matrix of the target feature map to be extracted is input into the target relation feature learning model to predict the curve relation matrix, adjacency weight matrix, multi-channel relation tensor, and similarity of the target feature map.

2. The data feature extraction method for aero-engine component characteristic diagrams according to claim 1, characterized in that, The steps to represent a characteristic graph as a combination of a set of characteristic curves and metadata are as follows: In the formula, Representing the characteristic map, This represents the set of characteristic curves in the characteristic diagram. A collection of metadata related to the feature graph; For the first Characteristic curves.

3. The data feature extraction method for aero-engine component characteristic diagrams according to claim 1, characterized in that, The steps for obtaining the correspondence between characteristic curves by sequentially reparameterizing, resampling, and aligning the three-dimensional parameters of the standardized discrete points in the characteristic curves are as follows: The reparameterized characteristic curve is obtained by reparameterizing each characteristic curve based on the three-dimensional parameters of the standardized discrete points in the characteristic curve; Set a uniform number of resampling points and perform interpolation resampling on the reparameterized characteristic curve to obtain resampling points; The geometric distance between two resampled points on the two characteristic curves is obtained based on the three-dimensional parameters of the resampled points of the two reparameterized characteristic curves, and the geometric distance is used as the alignment cost. Define the alignment path of the two characteristic curves as a sequence of physical index pairs; The alignment cost of two characteristic curves is obtained based on the physical index, sequence, and alignment cost. The alignment path with the minimum alignment cost is taken as the optimal alignment path. Obtain the optimal physical index pair sequence between characteristic curves based on the optimal alignment path; The optimal physical index pair sequence between characteristic curves is used as the correspondence between characteristic curves.

4. The method for extracting data features from a characteristic map of an aero-engine component according to claim 1, characterized in that, Based on the correspondence between characteristic curves, the average geometric distance between two characteristic curves is taken as the geometric relationship component between the two characteristic curves; the trend relationship component between the two characteristic curves is obtained based on the discrete first-order difference vector of the characteristic curves; the discrete curvature is obtained based on the discrete second-order difference vector between the two characteristic curves; and the curvature relationship component between the curves is obtained based on the discrete curvature. The steps to obtain the performance variable relationship components between curves based on the third-dimensional performance variable are as follows: The expression for the geometric relationship components between the two characteristic curves is: In the formula, Characteristic curves With characteristic curve Geometric relationship components between them; Indicates the total number of resampling points; Characteristic curves Chinese Physical Index The corresponding three-dimensional parameters at the resampling point; Characteristic curves Neutral and characteristic curves Same physical index The corresponding three-dimensional parameters at the resampling point; This indicates a request for distance; Indicates the characteristic curve after alignment mapping. Physical index of resampling points In the characteristic curve The corresponding physical index on; The expression for the trend relationship component between the two characteristic curves is: In the formula, Characteristic curves With characteristic curve The trend relationship between the components; Characteristic curves Chinese Physical Index The discrete first-order difference vector at the corresponding resampling point; Characteristic curves Neutral and characteristic curves Same physical index The discrete first-order difference vector at the corresponding resampling point; The expression for the curvature relationship components between the two characteristic curves is: In the formula, Characteristic curves With characteristic curve The curvature relationship components between them; Characteristic curves Chinese Physical Index The discrete curvature at the corresponding resampling point; Characteristic curves Neutral and characteristic curves Same physical index The discrete curvature at the corresponding resampling point; This represents the first-order difference of the first-dimensional performance variable; This represents the second difference of the first-dimensional performance variable; This represents the first difference of the second-dimensional performance variable; This represents the second-order difference of the second-dimensional performance variable; This represents a non-negative constant, used to avoid the denominator being zero; The expression for the performance variable relationship components between the two characteristic curves is as follows: In the formula, Characteristic curves With characteristic curve The performance variable relationships between components; Characteristic curves Chinese Physical Index The third-dimensional performance variable at the corresponding resampling point; Characteristic curves Neutral and characteristic curves Same physical index The third-dimensional performance variable at the corresponding resampling point.

5. The data feature extraction method for aero-engine component characteristic diagrams according to claim 1, characterized in that, The steps to obtain the curve relationship matrix, multi-channel relationship tensor, and adjacency weight matrix of the feature map are as follows: The expression for the curve relationship matrix of the characteristic map is: In the formula, The matrix representing the curve relationships in the characteristic plot; A comprehensive relational matrix representing the characteristic map; Characteristic curves With characteristic curve The performance variable relationships between components; Characteristic curves With characteristic curve The curvature relationship components between them; Characteristic curves With characteristic curve The trend relationship between the components; Characteristic curves With characteristic curve Geometric relationship components between them; Non-negative weighting coefficients representing geometric relational components; Non-negative weighting coefficients representing trend relationship components; Non-negative weighting coefficients representing curvature relation components; Represents the non-negative weighting coefficients of the relational components of performance variables; The expression for the multichannel relation tensor of the feature map is: In the formula, The multichannel relation tensor representing the feature map; Characteristic curves With characteristic curve The multi-channel relationship tensor between them; This indicates the number of characteristic curves in the characteristic diagram; The expression for the adjacency weight matrix of the feature graph is: In the formula, The adjacency weight matrix represents the characteristic graph; Characteristic curves With characteristic curve The adjacency weights between them; Indicates the kernel width parameter; This indicates the number of characteristic curves in the characteristic graph.

6. The method for extracting data features from a characteristic map of an aero-engine component according to claim 1, characterized in that, Based on the number of characteristic curves, physical index, and curve relationship matrix of two characteristic maps, the structural difference between the two characteristic maps is obtained. The steps to obtain the similarity between the two characteristic maps based on the structural difference are as follows: When the number of characteristic curves in two characteristic maps is the same and their indices are aligned, the structural difference between the two characteristic maps is obtained based on the curve relationship matrix. The similarity between the two characteristic maps is obtained based on the structural kernel scale parameters and structural difference degree; When the number of characteristic curves in two characteristic maps is inconsistent, the spectral signature vector is obtained based on the adjacency weight matrix, and the similarity between the two characteristic maps is obtained based on the spectral signature vector.

7. The method for extracting data features from a characteristic map of an aero-engine component according to claim 1, characterized in that, The expression for the nodal eigenvectors of the characteristic curve is: In the formula, Characteristic curves Node feature vectors; Characteristic curves In physical index The discrete first-order difference vector at the resampling point; Characteristic curves Chinese Physical Index The first-dimensional performance parameter at the corresponding resampling point; Characteristic curves Chinese Physical Index The second-dimensional performance parameter at the corresponding resampling point; Characteristic curves Chinese Physical Index The third-dimensional performance parameter at the corresponding resampling point; Characteristic curves Chinese Physical Index The discrete curvature at the corresponding resampling point; This indicates the number of resampling points.

8. The method for extracting data features from a characteristic map of an aero-engine component according to claim 1, characterized in that, The relation feature learning model includes an attention weight module and a graph-level embedding module. The attention weight module is used to introduce attention weights into the update rules of adjacent layers of the relation feature learning model to obtain the target update rule; the graph-level embedding module is used to standardize each output parameter in the dataset and concatenate the standardization results to obtain the graph-level embedding vector. The target update rule is as follows: In the formula, Represents a nonlinear activation function; Indicates the first The learnable parameter matrix used by the layer for transforming the features of its own nodes; Indicates the first The first in the layer The hidden representation values ​​corresponding to the feature curves; Indicates attention weight; Indicates the first The set of neighbors of each characteristic curve on the characteristic graph; Indicates the first The layer is a learnable parameter matrix used for aggregating neighbor node information; Indicates the number of model layers; Characteristic curves With characteristic curve The multi-channel relationship tensor between them; Indicates the updated result of the first... The first in the layer The hidden representation values ​​corresponding to the feature curves; Indicates the first The first in the layer The hidden representation values ​​corresponding to the feature curves; The graph-level embedding vector is: In the formula, Represents a graph-level embedding vector; Indicates the first Weighted operators for characteristic curves; Indicates the first in the final layer The hidden representation values ​​of the characteristic curves; This indicates the number of characteristic curves in the characteristic graph.

9. The method for extracting data features from a characteristic map of an aero-engine component according to claim 1, characterized in that, Also includes: Define a total loss function, which is used as the loss function for the relation feature learning model. The expression for the total loss function is: In the formula, This is the total loss function; To measure the learning loss function; The structural reconstruction loss function; The consistency constraint loss function; The weights are used to measure the learning loss function; The weights for the structural reconstruction loss function; The weights of the consistency constraint loss function; For each anchor sample, the graph-level embedding vector is used. The graph-level embedding vector corresponding to the positive sample; The graph-level embedding vector corresponding to the negative sample; For interval parameters; For the Sigmoid function; For the final layer Hidden representation of the characteristic curves; For the final layer Hidden representation of the characteristic curves; For the reconstructed characteristic curve With characteristic curve The adjacency weights between them; Characteristic curves With characteristic curve The adjacency weights between them; The number of characteristic curves involved in modeling relational features in the current component characteristic diagram; It is a symmetric consistency constraint function; This is a hierarchical consistency constraint function; This is the cross-graph distance consistency constraint function; The weights are the weights of the symmetric consistency constraint function; The weights of the hierarchical consistency constraint function; The weights of the cross-graph distance consistency constraint function; The expression for the symmetric consistency constraint function is: In the formula, Characteristic curve For characteristic curves Relationship matrix; Characteristic curve For characteristic curves Relationship matrix; Characteristic curve Its relation matrix to itself; The expression for the hierarchical consistency constraint function is: In the formula, For the first The physical index of each characteristic curve; For the first The physical index of each characteristic curve; For the first The projected scalar of the characteristic curve; For the first The projected scalar of the characteristic curve; For the final layer Hidden representation of the characteristic curves; This is the sorting interval parameter; for Learnable projection vectors under the category; The expression for the cross-graph distance consistency constraint function is: In the formula, For characteristic map and characteristic diagram Teacher structural distance; For scale matching function; For a set of sample pairs; For characteristic map and characteristic diagram Student distance; For characteristic map ; For characteristic map ; For characteristic map The corresponding graph-level embedding vector; For characteristic map The corresponding graph-level embedding vector.

10. A data feature extraction system for characteristic maps of aero-engine components, characterized in that, include: The parameter preprocessing module is used to obtain the characteristic map corresponding to each component of the aero-engine and represent the characteristic map as a combination of a set of characteristic curves and metadata. Set the category label and physical index of the characteristic curve; The relation component acquisition module is used to standardize the discrete points in the characteristic curves to obtain standardized discrete points. Based on the three-dimensional parameters of the standardized discrete points in the characteristic curves, each characteristic curve is reparameterized, resampled, and aligned for mapping to obtain the correspondence between the characteristic curves. The three-dimensional parameters include: a first-dimensional performance variable, a second-dimensional performance variable, and a third-dimensional performance variable. Based on the correspondence between the characteristic curves, the average geometric distance between two characteristic curves is used as the geometric relation component between them. The module also obtains the trend relation component between the two characteristic curves based on the discrete first-order difference vector, the discrete curvature based on the discrete second-order difference vector, and the curvature relation component between the curves based on the discrete curvature. Finally, the module obtains the performance variable relation component between the two characteristic curves based on the third-dimensional performance variable. The dataset construction module is used to obtain the curve relationship matrix, multi-channel relationship tensor, and adjacency weight matrix of the characteristic map based on the geometric relationship components, trend relationship components, curvature relationship components, and performance variable relationship components between two characteristic curves; to obtain the structural difference between the two characteristic maps based on the number of characteristic curves, physical index, and curve relationship matrix; and to obtain the similarity between the two characteristic maps based on the structural difference. Based on the three-dimensional parameters, discrete first-order difference vector, and discrete curvature of each resampling point in the characteristic curve, the module constructs the node feature vector of each characteristic curve and the node feature matrix of the characteristic map. The module uses the node feature matrix as input parameters and the curve relationship matrix, adjacency weight matrix, multi-channel relationship tensor, and similarity of the characteristic map as output parameters to construct the dataset. The relation feature learning model module is used to build a relation feature learning model. The relation feature learning model is trained on the dataset to obtain a trained target relation feature learning model. The system also includes a feature extraction module, which inputs the node feature matrix of the target feature map to be extracted into the target relation feature learning model to predict the curve relation matrix, adjacency weight matrix, multi-channel relation tensor, and similarity of the target feature map.

Images