Low-cost aircraft operation support method based on attention representation integrated clustering

Abstract

The application belongs to the field of aerospace engineering and data mining, and discloses a low-cost aircraft operation support method based on attention representation integrated clustering, which comprises the following steps: obtaining historical sensor data of an aircraft and performing preprocessing to obtain a feature matrix and construct a sample; for the feature matrix, a plurality of base divisions are generated by using a base clustering algorithm; a global high-order correlation matrix is constructed; a group is established by using the global high-order correlation matrix, and a centroid feature vector of each group is determined; a deep neural network is constructed, including an encoder, a decoder and an output unit; the deep neural network is trained by using the feature matrix; a total objective function during the training comprises a clustering loss and a reconstruction loss; the reconstruction loss is constructed based on a reconstructed adjacency matrix and an adjacency matrix corresponding to the base division; after real-time sensor data of the aircraft is obtained and is divided according to time slicing, the data is input into the trained deep neural network to obtain a corresponding prediction result.

Description

Technical Field

[0001] This invention belongs to the fields of aerospace engineering and data mining, and specifically relates to a low-cost aircraft operation support method based on attention representation ensemble clustering. Background Technology

[0002] With the rapid development of modern aviation technology, the complexity and integration of aircraft systems are increasing, placing extremely high demands on flight safety and maintenance capabilities. During operation, fixed-wing aircraft, helicopters, and drones continuously generate massive amounts of condition monitoring data from key components such as engines, hydraulic systems, and avionics systems. For example, a modern wide-body passenger aircraft can generate several gigabytes of data per flight, covering hundreds or even thousands of sensor parameters (such as exhaust temperature, rotor speed, fuel flow, vibration levels, flight altitude, and airspeed). This data contains rich information about aircraft performance evolution, operating habits, and potential warning signs. How to effectively utilize this "big data" to shift from traditional "reactive maintenance" or "scheduled maintenance" to "condition-based maintenance" and "predictive maintenance" is a current focus of attention in the aviation industry and academia.

[0003] Currently, aircraft operational status monitoring technologies are mainly divided into three categories:

[0004] Physical model-based methods simulate system behavior by establishing accurate mathematical models of aircraft components (such as thermodynamic cycle models and dynamic differential equations) and performing feature detection through residual analysis. However, the actual flight environment is complex and variable, and component aging leads to parameter drift. It is extremely difficult to establish a high-precision physical model that is adaptable to the entire life cycle, and the computational complexity is high, making it difficult to meet real-time requirements.

[0005] Knowledge-based methods rely on the experience of domain experts, such as expert systems and feature tree analysis. These methods are effective for known feature patterns, but they struggle to discover unknown, novel, or complex coupled feature patterns, and the maintenance and updating of the knowledge base is time-consuming and labor-intensive.

[0006] Data-driven approaches utilize machine learning algorithms to directly extract patterns from monitoring data. Supervised learning methods (such as support vector machines and convolutional neural networks) have achieved significant results, but they heavily rely on large amounts of high-quality labeled feature data. In the aviation field, high-level feature data is extremely scarce, and manual annotation is costly and subjective, limiting the application of supervised learning methods. Summary of the Invention

[0007] The purpose of this invention is to provide a low-cost aircraft operation support method based on attention representation ensemble clustering, which aims to solve the problems of high computational complexity, reliance on expert experience, and reliance on labeled data in existing low-cost aircraft operation support technologies.

[0008] To achieve the above objectives, the present invention employs the following technical solution:

[0009] Low-cost aircraft operation support methods based on attention representation ensemble clustering include:

[0010] Historical sensor data of the aircraft is acquired and preprocessed to obtain a feature matrix; samples are constructed based on time-sliced ​​division of the feature matrix.

[0011] For the feature matrix, multiple basis partitions are generated using a basis clustering algorithm; based on the indicator matrix corresponding to each basis partition, a global high-order correlation matrix is ​​constructed.

[0012] Groups are established using the global high-order correlation matrix, and the centroid eigenvectors of each group are determined.

[0013] A deep neural network is constructed, including an encoder, a decoder, and an output unit. The encoder obtains a low-dimensional embedding representation based on the centroid feature vector and the global high-order correlation matrix. The decoder reconstructs the adjacency matrix based on the low-dimensional embedding representation. The output unit determines the predicted probability of a sample belonging to different preset categories based on the low-dimensional embedding representation.

[0014] The deep neural network is trained using the feature matrix; the overall objective function during training includes clustering loss and reconstruction loss; the reconstruction loss is constructed based on the reconstruction adjacency matrix and the adjacency matrix corresponding to the basis partition.

[0015] The real-time sensor data of the aircraft is acquired and divided into time slices, and then input into a trained deep neural network to obtain the corresponding prediction results.

[0016] Furthermore, for the historical sensor data of the aircraft, each type of sensor data is subjected to physical boundary filtering, statistical filtering and normalization to obtain a standardized feature matrix; a sliding window is used to divide the time interval of the sensor data, each time window is a time slice, and a sample is all types of sensor data in a time slice.

[0017] Furthermore, for the feature matrix, multiple basis partitions are generated using a basis clustering algorithm; based on the indicator matrix corresponding to each basis partition, a global high-order correlation matrix is ​​constructed, including:

[0018] The K-means algorithm is used to analyze the feature matrix. run Next; in the first The next run starts from the interval The number of clusters is randomly set in the middle. ,in The minimum and maximum number of state categories are preset.

[0019] No. The base partitioning result obtained from the second run is: Then the final generation Basis partitioning results , ;

[0020] For the Divide into bases and construct a binary indicator matrix. If the first The sample belongs to the first The first in the basis partition Each cluster, then The Middle Line number Column elements Otherwise, it is 0; where , ; Indicates the number of samples;

[0021] Concatenate all indicator matrices to form a global high-order correlation matrix. ,in The total number of clusters across all basis partitions.

[0022] Furthermore, groups are established using the global high-order correlation matrix, and the centroid eigenvectors of each group are determined, including:

[0023] For global high-order incidence matrix Treat each column as a group; global higher-order correlation matrix The Centroid eigenvectors of each group as follows:

[0024] ;

[0025] in, Indicates the first One sample; Represents the global higher-order incidence matrix The Middle Line number The elements of the column are used to indicate the sample. Does it belong to the first One group, When indicating sample Belongs to the When a group ,otherwise ; The total number of clusters across all basis partitions; Indicates the number of samples.

[0026] Furthermore, the encoder employs L fully connected layers; each fully connected layer performs attention coefficient calculation, normalization, and feature aggregation on the sample features corresponding to each sample in its input features to obtain updated features; all updated features serve as input to the next fully connected layer; wherein, the input features of the first fully connected layer are the global high-order correlation matrix. The aggregated features of all the updated features output by the last fully connected layer are used as the low-dimensional embedding representation Z.

[0027] Furthermore, the attention coefficient in each fully connected layer is calculated according to the following formula:

[0028] ;

[0029] in, For the first The first fully connected layer Sample In the global high-order correlation matrix The corresponding number in Row vector, take it as the first row vector. Sample Corresponding sample features; ; For the first Sample In the In the fully connected layer, relative to the first Attention coefficient of each group; , The first Learnable attention weights and learnable weight matrices for fully connected layers; superscript Indicates transpose; This represents a vector concatenation operation; Represents a linear rectified function; Global high-order incidence matrix The The centroid feature vectors of each group;

[0030] The attention coefficient was normalized.

[0031] ;

[0032] in, The normalized attention coefficient. For the first Sample In the In the fully connected layer, relative to the first Attention coefficient of each group , Represents the global higher-order incidence matrix The Middle Line number Column elements, Indicates sample Belongs to the One group; It is a natural exponential function;

[0033] Using the normalized attention coefficients, weighted aggregation groups are used to obtain updated features:

[0034] ;

[0035] in, For the first Sample The updated features are the input features of the next fully connected layer; This is the balance coefficient, with a value ranging from 0.1 to 1. The total number of clusters across all basis partitions; Represents the global higher-order incidence matrix The Middle Line number Column elements,

[0036] Update features for all samples output from the last fully connected layer Aggregation is performed to obtain a low-dimensional embedding representation. .

[0037] Furthermore, the input to the decoder is a low-dimensional embedding representation. The connection strength between samples is calculated using the following formula:

[0038] ;

[0039] in, They represent the first Sample In low-dimensional embedding representation The corresponding vector in; ; superscript Indicates transpose; Indicates sample The strength of the connection between them; This represents the Sigmoid activation function;

[0040] Connection strength between all samples Together they form a reconstructed adjacency matrix .

[0041] Furthermore, for low-dimensional embedding representation The K-means algorithm is used for clustering. Different preset categories are set during clustering to obtain multiple clusters corresponding to the preset categories and the cluster center of each cluster.

[0042] Calculate the predicted probability that a sample belongs to different preset categories:

[0043] ;

[0044] in, Indicates the first Sample In low-dimensional embedding representation The corresponding vector in Indicates the first Cluster centers for each predefined category, The number of preset categories; Indicates the first Sample Belongs to the The predicted probability of each preset category. ; express Norm;

[0045] After obtaining the predicted probability of a sample belonging to each preset category, a judgment is made:

[0046] Set confidence threshold Determine the probability of prediction Are they all less than If all values ​​are less than 0, it indicates that the prediction result is unreliable, and the output is "Unknown state".

[0047] If not all are less than Then, among all predicted probabilities, the preset category corresponding to the highest predicted probability is selected as the sample. The prediction results are output.

[0048] Furthermore, the overall objective function The settings are as follows:

[0049] ;

[0050] in, The clustering loss weights have a range of values. ; The reconstruction loss is used to measure the reconstructed adjacency matrix. The differences from the base partitioning results are expressed as follows:

[0051] ;

[0052] in, Indicates the first The adjacency matrix corresponding to each basis partition, if the sample In the same cluster within this base partition, then its th line, number Column elements Otherwise, it is 0; For the first Each basis partitions the weights, satisfying the constraints. ; Denotes the Frobenius norm; ; The number of base partitions;

[0053] The clustering loss is expressed as follows:

[0054] ;

[0055] in, Indicates the number of samples; The number of preset categories; Indicates the first Sample Belongs to the The predicted probability of each preset category. Indicates the first Sample Compared to the first The distribution of auxiliary targets for each preset category is represented as follows:

[0056] ;

[0057] in, Indicates the first Cluster frequencies corresponding to each preset category; Indicates the first Cluster frequencies corresponding to each preset category Indicates the first Sample Belongs to the The predicted probability of each preset category. .

[0058] A terminal device includes a processor, a memory, and a computer program stored in the memory; when the processor executes the computer program, it implements the low-cost aircraft operation support method based on attention representation ensemble clustering.

[0059] A computer-readable storage medium storing a computer program; when executed by a processor, the computer program implements the low-cost aircraft operation support method based on attention representation ensemble clustering.

[0060] Compared with the prior art, the present invention has the following technical features:

[0061] 1. High robustness and noise resistance: By integrating multiple base partitions and introducing an adaptive weighting mechanism, this invention can automatically identify and suppress poor clustering results caused by sensor transient interference or improper initialization. This feature is crucial in high-noise environments such as flight data, effectively reducing the false alarm rate and computational complexity.

[0062] 2. Deep Feature Mining Capability: Utilizing an attention encoder, this invention not only considers the similarity between samples but also delves into the nonlinear correlation between samples and groups (latent feature patterns). This enables the method to discover complex coupled features (such as the nonlinear deviation between exhaust temperature and rotor speed) that traditional linear methods cannot identify, without relying on expert experience.

[0063] 3. No need for large amounts of labeled data: As an unsupervised / self-supervised learning method, this invention does not require expensive feature label data to work, making it very suitable for the current situation of scarce feature samples in the aviation field; it can automatically discover new flight patterns or unknown feature types.

[0064] 4. End-to-end optimization framework: Feature extraction (embedding learning), graph structure reconstruction and cluster assignment are unified in a loss function for joint optimization, avoiding the error accumulation caused by step-by-step processing, so that the learned features are specifically used for feature classification tasks, thus improving diagnostic accuracy.

[0065] 5. Enhanced interpretability: Attention coefficients provide an intuitive explanation, indicating which base clustering perspectives contribute most to the sample decision, helping maintenance personnel understand the basis for feature determination. Attached Figure Description

[0066] Figure 1 This is a schematic flowchart of the method of the present invention;

[0067] Figure 2 This is a schematic diagram of the deep neural network structure in this invention;

[0068] Figure 3 This is a schematic diagram illustrating the construction of a high-order global correlation matrix in this invention;

[0069] Figure 4 The figure shows the output of a deep neural network in an embodiment of the present invention; each color in the figure represents a preset category. Detailed Implementation

[0070] This invention provides a low-cost aircraft operation support method based on attention-based ensemble clustering. It combines the nonlinear mapping capability of deep neural networks with the robustness of ensemble clustering, fuses high-order group information through an attention mechanism, and employs an adaptive weighting strategy to optimize the ensemble effect. See also Figure 1 The specific steps of this invention are as follows:

[0071] Step 1: Acquire historical sensor data of the aircraft and preprocess it to obtain a feature matrix; construct samples based on time slices of the feature matrix.

[0072] The sensor data includes categories such as exhaust temperature, rotor speed, fuel flow rate, pressure parameters, and auxiliary parameters; the auxiliary parameters include fuel-air ratio, cooling air flow rate, and bleed air enthalpy.

[0073] The preprocessing method is as follows:

[0074] For each type of sensor data, physical limit filtering, statistical filtering, and normalization are performed. Physical limit filtering means that each type of sensor data must not exceed the corresponding physical limit; if it does, it is considered noise and is either zeroed out or removed. Statistical filtering means that for each type of sensor data, the mean and standard deviation are determined; if the absolute value of the difference between a sensor data and the mean is greater than 3 times the standard deviation, it is judged as noise and is filled using linear interpolation. Normalization means that each sensor data is normalized to the range [0,1].

[0075] Therefore, the standardized feature matrix is ​​denoted as ;in express The real space of dimension 1 This indicates the dimension of the sensor data, i.e., the number of categories of sensor data; It is the space of real numbers; This indicates the number of samples; one sample is sensor data of all categories within a time slice; the time slice refers to dividing the time interval of sensor data using a sliding window, with each time window being a time slice.

[0076] Step 2, for the feature matrix Multiple basis partitions are generated using a basis clustering algorithm; based on the indicator matrix corresponding to each basis partition, a global high-order correlation matrix is ​​constructed. .

[0077] To capture the multi-view structure of sensor data, the base clustering algorithm uses the K-means algorithm as the base clusterer, from the feature matrix... Generated in The basis partitioning is as follows:

[0078] Step 2.1: Use the K-means algorithm to process the feature matrix. run Next; in the The next run starts from the interval The number of clusters is randomly set in the middle. ,in The preset minimum and maximum number of state categories; minimum number of state categories It should be greater than the actual number of state categories; for example, the actual state categories of sensor data include three categories: normal, degraded, and faulty. .

[0079] Step 2.2, the The base partitioning result obtained from the second run is: ,in Indicates the first The first in the basis partition There are several clusters; then the final generation will be... Basis partitioning results .

[0080] Step 2.3, in order to Basis partitioning results The input is fed into the deep neural network, for the first... Divide into bases and construct a binary indicator matrix. If the first The sample belongs to the first The first in the basis partition Each cluster, then The Middle Line number Column elements Otherwise, it is 0; where , .

[0081] Step 2.4, as follows Figure 3 As shown, all indicator matrices are concatenated to form a global high-order correlation matrix. ,in This represents the total number of clusters across all base partitions; for example, if running... In each instance, with a tie of 20 clusters, the global higher-order correlation matrix is... have List.

[0082] Step 3: Use the global high-order correlation matrix to establish groups and determine the centroid eigenvector of each group.

[0083] For the global high-order incidence matrix Each column is treated as a group; using the concept of graph attention networks, the association strength between samples and groups is learned: a global high-order association matrix. The Centroid eigenvectors of each group as follows:

[0084] ;

[0085] in, Indicates the first One sample; Represents the global higher-order incidence matrix The Middle Line number The elements of the column are used to indicate the sample. Does it belong to the first Groups When it belongs to ,otherwise .

[0086] Step 4: Construct a deep neural network, including an encoder, decoder, and output unit; for example... Figure 2 As shown, the encoder obtains a low-dimensional embedding representation based on the centroid feature vector and the global high-order correlation matrix; the decoder outputs a reconstructed adjacency matrix based on the low-dimensional embedding representation; and the output unit determines the predicted probability of a sample belonging to different preset categories based on the low-dimensional embedding representation.

[0087] Step 4.1: The encoder employs an L-layer fully connected architecture. Each fully connected layer performs attention coefficient calculation, normalization, and feature aggregation on the sample features corresponding to each sample in its input features to obtain updated features. All updated features serve as input to the next fully connected layer. The input features of the first fully connected layer are the global high-order correlation matrix. The aggregated features of all the updated features output by the last fully connected layer are used as the low-dimensional embedding representation Z.

[0088] The attention coefficient is calculated in each fully connected layer according to the following formula:

[0089] ;

[0090] in, For the first In a fully connected layer, the feature matrix The Sample In the global high-order correlation matrix The corresponding number in Row vector, take it as the first row vector. Sample Corresponding sample features; ; For the first Sample In the In the fully connected layer, relative to the first Attention coefficient of each group; , The first Learnable attention weights and learnable weight matrices for fully connected layers (mapping sample features to a high-dimensional space to capture non-linear relationships); superscript Indicates transpose; This represents a vector concatenation operation; Represents a linear rectified function; Global high-order incidence matrix The The centroid eigenvectors of each group.

[0091] Step 4.2: Normalize the attention coefficients:

[0092] ;

[0093] in, The normalized attention coefficient. For the first Sample In the In the fully connected layer, relative to the first Attention coefficient of each group , Represents the global higher-order incidence matrix The Middle Line number Column elements, Indicates sample Belongs to the One group; It is a natural exponential function.

[0094] Step 4.3, Feature aggregation.

[0095] Using the normalized attention coefficients, weighted aggregation groups are used to obtain updated features:

[0096] ;

[0097] in, For the first Sample The updated features are the input features of the next fully connected layer; The balancing coefficient has a value range of 0.1 to 1; the first term of the updated features retains the individual characteristics of the sample itself; the second term introduces global topological information.

[0098] After the above processing, the updated features of all samples output by the last fully connected layer are obtained. Aggregation is performed by concatenating the updated features of all samples to obtain a low-dimensional embedding representation. .

[0099] Step 4.4, Construct a decoder for use based on the low-dimensional embedding representation. Output the reconstructed adjacency matrix.

[0100] In this scheme, the input to the decoder is a low-dimensional embedding representation. The connection strength between samples is calculated using the following formula:

[0101] ;

[0102] in, They represent the first Sample In low-dimensional embedding representation The corresponding vector in the middle, i.e., the low-dimensional embedding representation The OK; ; superscript Indicates transpose; Indicates sample The strength of the connection between them; This represents the Sigmoid activation function.

[0103] Finally, the connection strength between all samples is determined. Together they form a reconstructed adjacency matrix .

[0104] Step 4.5, construct the output unit; the output unit is based on the low-dimensional embedding representation. The predicted probabilities of a sample belonging to different preset categories are determined as follows:

[0105] First, for low-dimensional embedding representation The K-means algorithm is used for clustering. Different preset categories are set during clustering, resulting in multiple clusters corresponding to the preset categories and the cluster center of each cluster. For example, in one embodiment, there are three preset categories: normal, deteriorated, and faulty, which are used to characterize different health states of the parts collected by the sensor.

[0106] Secondly, calculate the predicted probability that the sample belongs to different preset categories.

[0107] ;

[0108] in, Indicates the first Sample In low-dimensional embedding representation The corresponding vector in Indicates the first Cluster centers for each predefined category, The number of preset categories; Indicates that the sample belongs to the first The predicted probability of each preset category. ; express Norm.

[0109] Step 4.6: After obtaining the predicted probability of a sample belonging to each preset category, a judgment is made:

[0110] Set the confidence threshold according to actual needs. It is generally set to be no less than 0.6, such as 0.7.

[0111] Determine each prediction probability Are they all less than If all values ​​are less than 0, it indicates that the prediction result is unreliable, and an "unknown state" is output to remind human intervention for analysis; this mechanism can effectively reduce the false alarm rate.

[0112] If not all are less than Then, among all predicted probabilities, the preset category corresponding to the highest predicted probability is selected as the sample. The prediction results are output, and these prediction results are the aircraft health status diagnosis results.

[0113] Step 5, using the feature matrix Train the deep neural network; the overall objective function during training is... It includes clustering loss and reconstruction loss; the reconstruction loss is constructed based on the reconstructed adjacency matrix and the adjacency matrix corresponding to the basis partition.

[0114] Overall objective function The settings are as follows:

[0115] ;

[0116] in, The clustering loss weights have a range of values. ; The reconstruction loss is used to measure the reconstructed adjacency matrix. The differences from the base partitioning results are expressed as follows:

[0117] ;

[0118] in, Indicates the first The adjacency matrix corresponding to each basis partition, if the sample In the same cluster within this basis partition, then its th line, number Column elements Otherwise, it is 0; We use a base to partition the weights; this is a variable to be optimized through training, satisfying constraints. ; Denotes the Frobenius norm; .

[0119] The clustering loss is expressed as follows:

[0120] ;

[0121] in, Indicates the first Sample Belongs to the The predicted probability of each preset category. Indicates the first Sample Compared to the first The distribution of auxiliary targets for each preset category is represented as follows:

[0122] ;

[0123] in, Indicates the first Cluster frequencies corresponding to each preset category; Indicates the first Cluster frequencies corresponding to each preset category Indicates the first Sample Belongs to the The predicted probability of each preset category. .

[0124] The auxiliary target distribution in the above formula can enhance high-confidence prediction. During the optimization process, samples located at the edge of the cluster can be continuously pushed towards the center of the cluster, thereby strengthening the compactness of the cluster.

[0125] In this invention, an alternating optimization strategy is used during the training of the deep neural network, and the weights are divided for the basis. and network parameters (the first) Learnable attention weights of fully connected layers Learnable weight matrix Alternate optimization is performed; ultimately, when the overall objective function... Upon convergence, a well-trained deep neural network is obtained.

[0126] Step 6: In practical applications, acquire real-time sensor data from the aircraft and preprocess it to obtain the corresponding feature matrix; divide the feature matrix into time slices using a sliding window; input the sensor data corresponding to each time slice into the trained deep neural network in sequence to obtain the corresponding prediction result.

[0127] Example

[0128] Fourteen sensor data categories strongly correlated with performance degradation were selected for a certain aircraft, specifically: total outlet temperature of the low-pressure compressor; total outlet temperature of the high-pressure compressor; total outlet temperature of the low-pressure turbine; total outlet pressure of the high-pressure compressor; physical fan speed; physical core engine speed; fuel flow rate; static pressure at the outlet of the high-pressure compressor; fuel-air ratio; corrected fan speed; corrected core engine speed; bypass ratio; bleed air enthalpy; and cooling air flow rate.

[0129] After constructing the feature matrix, a total of Basis partitioning, minimum number of state categories It is much larger than the real state categories (the real state categories include three categories: normal, degraded, and faulty).

[0130] In this embodiment, the encoder of the deep neural network has two fully connected layers. The dimension of the updated features output by the first layer is 128, and that of the second layer is 10. The activation function is a linear rectified function, and the balance coefficients are... Clustering loss weights When training the deep neural network, the learning rate is set to 0.001, and the Adam optimizer is used.

[0131] The deep neural network of this invention is used to predict the category of the sample, and the result is as follows: Figure 4 As shown, the final prediction results of the sample clearly show multiple categories, proving the effectiveness of the technical solution of the present invention.

[0132] Compared to the single K-means algorithm, this invention improves classification accuracy by approximately 15% and normalized mutual information (NMI) by approximately 20%.

[0133] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A low-cost aircraft operation support method based on attention representation ensemble clustering, characterized in that, include: Acquire historical sensor data of the aircraft and preprocess it to obtain a feature matrix; Samples are constructed by dividing the feature matrix into time slices; For the feature matrix, multiple basis partitions are generated using a basis clustering algorithm; based on the indicator matrix corresponding to each basis partition, a global high-order correlation matrix is ​​constructed. Groups are established using the global high-order correlation matrix, and the centroid eigenvectors of each group are determined. Construct a deep neural network, including an encoder, a decoder, and an output unit; The encoder, based on the centroid feature vector and the global high-order correlation matrix, obtains a low-dimensional embedding representation, including: The encoder employs an L-layer fully connected structure. Each fully connected layer performs attention coefficient calculation, normalization, and feature aggregation on the sample features corresponding to each sample in its input features to obtain updated features. All updated features serve as the input to the next fully connected layer. The input features of the first fully connected layer are the global high-order correlation matrix. The aggregated features of all the updated features output by the last fully connected layer are used as the low-dimensional embedding representation Z. The attention coefficient is calculated in each fully connected layer according to the following formula: ； in, For the first The first fully connected layer Sample In the global high-order correlation matrix The corresponding number in Row vector, take it as the first row vector. Sample Corresponding sample features; ; For the first Sample In the In the fully connected layer, relative to the first Attention coefficient of each group; , The first Learnable attention weights and learnable weight matrices for fully connected layers; superscript Indicates transpose; This represents a vector concatenation operation; Represents a linear rectified function; Global high-order incidence matrix The The centroid feature vectors of each group; The attention coefficient is normalized: ； in, The normalized attention coefficient. For the first Sample In the In the fully connected layer, relative to the first Attention coefficient of each group , Represents the global higher-order incidence matrix The Middle Line number Column elements, Indicates sample Belongs to the One group; It is a natural exponential function; Using the normalized attention coefficients, weighted aggregation groups are used to obtain updated features: ； in, For the first Sample The updated features are the input features of the next fully connected layer; This is the balance coefficient, with a value ranging from 0.1 to 1. The total number of clusters across all basis partitions; Represents the global higher-order incidence matrix The Middle Line number Column elements, Update features for all samples output from the last fully connected layer Aggregation is performed to obtain a low-dimensional embedding representation. ; The decoder outputs a reconstructed adjacency matrix based on the low-dimensional embedding representation; the output unit determines the predicted probability of a sample belonging to different preset categories based on the low-dimensional embedding representation. The deep neural network is trained using the feature matrix; the overall objective function during training includes clustering loss and reconstruction loss; the reconstruction loss is constructed based on the reconstruction adjacency matrix and the adjacency matrix corresponding to the basis partition. The real-time sensor data of the aircraft is acquired and divided into time slices, and then input into a trained deep neural network to obtain the corresponding prediction results.

2. The low-cost aircraft operation support method based on attention representation ensemble clustering according to claim 1, characterized in that, For the historical sensor data of the aircraft, each type of sensor data is subjected to physical boundary filtering, statistical filtering and normalization to obtain a standardized feature matrix. A sliding window is used to divide the time interval of sensor data. Each time window is a time slice, and a sample is the sensor data of all categories in a time slice.

3. The low-cost aircraft operation support method based on attention representation ensemble clustering according to claim 1, characterized in that, For the feature matrix, multiple basis partitions are generated using a basis clustering algorithm; based on the indicator matrix corresponding to each basis partition, a global high-order correlation matrix is ​​constructed, including: The K-means algorithm is used to analyze the feature matrix. run Next; in the The next run starts from the interval The number of clusters is randomly set in the middle. ,in The minimum and maximum number of state categories are preset. No. The base partitioning result obtained from the second run is: Then the final generation Basis partitioning results , ; For the Divide into bases and construct a binary indicator matrix. If the first The sample belongs to the first The first in the basis partition Each cluster, then The Middle Line number Column elements Otherwise, it is 0; where , ; Indicates the number of samples; Concatenate all indicator matrices to form a global high-order correlation matrix. ,in The total number of clusters across all basis partitions.

4. The low-cost aircraft operation support method based on attention representation ensemble clustering according to claim 1, characterized in that, Groups are established using the global high-order correlation matrix, and the centroid eigenvectors of each group are determined, including: For the global high-order incidence matrix Treat each column as a group; global higher-order correlation matrix The Centroid eigenvectors of each group as follows: ； in, Indicates the first One sample; Represents the global higher-order incidence matrix The Middle Line number The elements of the column are used to indicate the sample. Does it belong to the first One group, When indicating sample Belongs to the When a group ,otherwise ; The total number of clusters across all basis partitions; Indicates the number of samples.

5. The low-cost aircraft operation support method based on attention representation ensemble clustering according to claim 1, characterized in that, The decoder's input is a low-dimensional embedding representation. The connection strength between samples is calculated using the following formula: ； in, They represent the first Sample In low-dimensional embedding representation The corresponding vector in; ; superscript Indicates transpose; Indicates sample The strength of the connection between them; This represents the Sigmoid activation function; Connection strength between all samples Together they form a reconstructed adjacency matrix .

6. The low-cost aircraft operation support method based on attention representation ensemble clustering according to claim 1, characterized in that, For low-dimensional embedding representation The K-means algorithm is used for clustering. Different preset categories are set during clustering to obtain multiple clusters corresponding to the preset categories and the cluster center of each cluster. Calculate the predicted probability of a sample belonging to different preset categories: ； in, Indicates the first Sample In low-dimensional embedding representation The corresponding vector in Indicates the first Cluster centers for each preset category, The number of preset categories; Indicates the first Sample Belongs to the The predicted probability of each preset category. ; express Norm; After obtaining the predicted probability of a sample belonging to each preset category, a judgment is made: Set confidence threshold Determine the probability of prediction Are they all less than If all values ​​are less than 0, it indicates that the prediction result is unreliable, and the output is "Unknown state". If not all are less than Then, among all predicted probabilities, the preset category corresponding to the highest predicted probability is selected as the sample. The prediction results are output.

7. The low-cost aircraft operation support method based on attention representation ensemble clustering according to claim 1, characterized in that, Overall objective function The settings are as follows: ； in, The clustering loss weights have a range of values. ; The reconstruction loss is used to measure the reconstructed adjacency matrix. The differences from the base partitioning results are expressed as follows: ； in, Indicates the first The adjacency matrix corresponding to each basis partition, if the sample In the same cluster within this basis partition, then its th line, number Column elements Otherwise, it is 0; For the first Each basis partitions the weights, satisfying the constraints. ; Denotes the Frobenius norm; ; The number of base partitions; The clustering loss is expressed as follows: ； in, Indicates the number of samples; The number of preset categories; Indicates the first Sample Belongs to the The predicted probability of each preset category. Indicates the first Sample Compared to the first The distribution of auxiliary targets for each preset category is represented as follows: ； in, Indicates the first Cluster frequencies corresponding to each preset category; Indicates the first Cluster frequencies corresponding to each preset category Indicates the first Sample Belongs to the The predicted probability of each preset category. .

8. A terminal device, comprising a processor, a memory, and a computer program stored in the memory; wherein, when the processor executes the computer program, it is characterized in that, Implement the low-cost aircraft operation support method based on attention representation ensemble clustering as described in any one of claims 1-7.

Images