CN-121658871-B - Low-cost aircraft operation supporting method based on attention characterization integrated clustering

CN121658871BCN 121658871 BCN121658871 BCN 121658871BCN-121658871-B

Abstract

The invention belongs to the field of aerospace engineering and data mining, and discloses a low-cost aircraft operation support method based on attention characterization integrated clustering, which comprises the steps of acquiring sensor data of aircraft history, preprocessing, obtaining a feature matrix and constructing a sample; aiming at the feature matrix, a plurality of base partitions are generated by using a base clustering algorithm, a global high-order associated matrix is constructed, groups are established by using the global high-order associated matrix, centroid feature vectors of each group are determined, a deep neural network is constructed, the deep neural network comprises an encoder, a decoder and an output unit, the feature matrix is used for training, the total objective function during training comprises clustering loss and reconstruction loss, the reconstruction loss is constructed on the basis of the reconstruction adjacent matrix and the adjacent matrix corresponding to the base partitions, real-time sensor data of an aircraft are acquired, and after the sensor data are divided according to time slices, the sensor data are input into the trained deep neural network to obtain corresponding prediction results.

Inventors

LUO SHENG
LI GUOXU
LIU JUNSHENG
ZHAO LIRAN
DONG YUE
LIU YIXIN

Assignees

西安现代控制技术研究所

Dates

Publication Date: 20260508
Application Date: 20260206

Claims (8)

1. A low-cost aircraft operation support method based on attention-characterization integrated clustering, comprising: acquiring historical sensor data of an aircraft, preprocessing the sensor data to obtain a feature matrix, and dividing time slices of the feature matrix to construct a sample; Constructing a global high-order incidence matrix based on an indication matrix corresponding to each base division; establishing groups by using a global high-order incidence matrix, and determining a centroid feature vector of each group; the encoder obtains a low-dimensional embedded representation based on a centroid feature vector and a global high-order incidence matrix, and comprises: The encoder adopts L-layer full-connection layers, each full-connection layer calculates attention coefficients, normalizes and aggregates the characteristics according to the sample characteristics corresponding to each sample in the input characteristics to obtain updated characteristics, all the updated characteristics are used as the input of the next full-connection layer, wherein the input characteristics of the first full-connection layer are global high-order correlation matrix All the updated features output by the last full connection layer are aggregated and then used as a low-dimensional embedded representation Z; the attention coefficient is calculated in each fully connected layer according to the following formula: ; Wherein, the Is the first Layer full connection layer of the first Individual samples In global high-order incidence matrix Corresponding first of (a) Line vector, take it as the first Individual samples Corresponding sample characteristics; ; Is the first Individual samples In the first place Relative first of all-layer connection layers Attention coefficients of the individual groups; 、 Respectively the first Learning attention weight and learning weight matrix of layer full-connection layer Representing a transpose; Representing a vector concatenation operation; representing a linear rectification function; for global higher-order incidence matrices Is the first of (2) Centroid feature vectors for the individual groups; Normalization processing is carried out on the attention coefficient: ; Wherein, the For the normalized attention coefficient, Is the first Individual samples In the first place Relative first of all-layer connection layers The attention coefficients of the individual groups, , Representing a global higher-order correlation matrix Middle (f) Line 1 The elements of the column are arranged such that, Representing a sample Belonging to the first Groups; Is a natural exponential function; weighting the aggregate group with the normalized attention coefficients to obtain updated features: ; Wherein, the Is the first Individual samples The updated characteristics of the next full connection layer are input characteristics; the value range is 0.1-1 for the balance coefficient; Dividing the total number of clusters in the base group; representing a global higher-order correlation matrix Middle (f) Line 1 The elements of the column are arranged such that, Update features for all samples output by last full connection layer Aggregation to obtain a low-dimensional embedded representation ; The decoder outputs a reconstructed adjacency matrix according to the low-dimensional embedded representation, and the output unit determines the prediction probability of the samples belonging to different preset categories according to the low-dimensional embedded representation; training the deep neural network by utilizing a feature matrix, wherein the total objective function during training comprises clustering loss and reconstruction loss, and the reconstruction loss is constructed based on a reconstruction adjacent matrix and an adjacent matrix corresponding to base division; And acquiring real-time sensor data of the aircraft, dividing the sensor data according to time slices, and inputting the sensor data into a trained deep neural network to obtain a corresponding prediction result.
2. The method for supporting low-cost aircraft operation based on attention-characterization integrated clustering according to claim 1, wherein physical boundary filtering, statistical filtering and normalization processing are performed on each type of sensor data according to the aircraft history to obtain a normalized feature matrix, and a sliding window is adopted to divide time intervals of the sensor data, wherein each time window is a time slice, and one sample is sensor data of all types under one time slice.
3. The method for supporting low-cost aircraft operation based on attention-characterization integrated clustering of claim 1, wherein generating a plurality of base partitions by using a base clustering algorithm for the feature matrix, constructing a global higher-order correlation matrix based on an indication matrix corresponding to each base partition, comprises: the K-means algorithm is adopted for the feature matrix Operation Next, in the first place Secondary run-time slave interval The number of clusters is randomly set Wherein The method comprises the steps of obtaining a preset minimum state class number and a preset maximum state class number; First, the The base division result obtained by the secondary operation is Then finally generate Radix division results , ; For the first Dividing the base to construct a binary indicating matrix And, if you get The samples belong to The first of the base partitions Clusters of, then Middle (f) Line 1 Elements of columns Otherwise 0, wherein , ; Representing the number of samples; All the indication matrixes are spliced to form a global high-order incidence matrix Wherein The total number of clusters in the partition is for all the bases.
4. The method of claim 1, wherein establishing groups using a global higher-order correlation matrix, determining centroid feature vectors for each group, comprises: For global higher-order incidence matrix Each column is used as a group, and the global high-order incidence matrix Is the first of (2) Centroid feature vector of individual group The following are provided: ; Wherein, the Represent the first A sample number; representing a global higher-order correlation matrix Middle (f) Line 1 Column elements for indicating samples Whether or not it is the first The number of groups of the device is, When the sample is indicated Belonging to the first When each group is Otherwise ; Dividing the total number of clusters in the base group; Representing the number of samples.
5. The attention-token-integrated-clustering-based low-cost aircraft operation support method of claim 1, wherein the decoder input is a low-dimensional embedded representation And the connection strength between the samples was calculated according to the following formula: ; Wherein, the Respectively represent the first Individual samples Embedding representations in low dimensions Corresponding vector in (a); superscript (I) Representing a transpose; Representing a sample The connection strength between the two; representing a Sigmoid activation function; the connection strength between all samples Together form a reconstructed adjacency matrix 。
6. The method for low-cost aircraft operation support based on attention-characterization integrated clustering of claim 1, wherein the representation is embedded in low dimensions Adopting a K-means algorithm to perform clustering, and setting different preset categories during clustering, so as to obtain a plurality of clusters corresponding to the preset categories and a clustering center of each cluster; Calculating the prediction probability of the samples belonging to different preset categories: ; Wherein, the Represent the first Individual samples Embedding representations in low dimensions Is used to determine the vector of the corresponding vector, Represent the first The cluster centers of the preset categories are selected, Is the number of preset categories; Represent the first Individual samples Belonging to the first The probability of prediction for a pre-set category, ; Representation of A norm; After obtaining the prediction probability that the sample belongs to each preset category, judging: setting confidence threshold Judging the prediction probability Whether or not both are smaller than If the prediction results are smaller than the preset value, the prediction results are not credible, and an unknown state is output; if not all smaller than Selecting a preset category corresponding to the maximum prediction probability from all the prediction probabilities as a sample And outputting the prediction result of the (c).
7. The method for low-cost aircraft operation support based on attention-characterization integrated clustering of claim 1, wherein the total objective function The settings were as follows: ; Wherein, the The clustering loss weight is in the value range of ; Reconstruction loss for weighting reconstruction adjacency matrix The difference from the base partition result is expressed as follows: ; Wherein, the Represent the first Dividing adjacent matrix corresponding to each base, if the sample The same cluster in the base partition, then its first Line 1 Elements of columns Otherwise, 0; Is the first The weight is divided by the base to satisfy the constraint ; Representing the Frobenius norm; ; Number of base partitions; The clustering loss is represented as follows: ; Wherein, the Representing the number of samples; Is the number of preset categories; Represent the first Individual samples Belonging to the first The probability of prediction for a pre-set category, Represent the first Individual samples Relative to the first Auxiliary target distribution of each preset category is expressed as follows: ; Wherein, the Represent the first Cluster frequencies corresponding to preset categories; Represent the first Cluster frequencies corresponding to the respective preset categories, Represent the first Individual samples Belonging to the first The probability of prediction for a pre-set category, 。
8. Terminal device comprising a processor, a memory and a computer program stored in said memory, which processor, when executing the computer program, is characterized in that a low-cost aircraft operation support method based on attention-characterizing integrated clusters as defined in any one of claims 1-7 is implemented.

Description

Low-cost aircraft operation supporting method based on attention characterization integrated clustering Technical Field The invention belongs to the field of aerospace engineering and data mining, and particularly relates to a low-cost aircraft operation support method based on attention characterization integrated clustering. Background With the rapid development of modern aviation technology, the complexity and integration degree of an aircraft system are increasingly improved, and extremely high requirements on flight safety and maintenance guarantee capability are put forward. In the running process of aircrafts such as fixed wing aircrafts, helicopters, unmanned planes and the like, key components such as engines, hydraulic systems, avionics systems and the like can continuously generate massive state monitoring data. For example, a modern wide-body passenger aircraft may generate several gigabytes of data per flight, covering hundreds to thousands of sensor parameters (e.g., exhaust temperature, rotor speed, fuel flow, vibration values, altitude, airspeed, etc.). These data are rich in the evolution of the performance, operating habits and potential characterization of the aircraft. How to effectively use these "big data" to realize the transition from traditional "post-maintenance" or "timing maintenance" to "optionality maintenance" and "predictive maintenance" is the focus of current aviation industry and academia. Currently, aircraft operation state monitoring technologies are mainly divided into three categories: The method based on the physical model simulates the system behavior by establishing an accurate mathematical model (such as a thermodynamic cycle model and a kinetic differential equation) of the aircraft component and performs feature detection by residual analysis, however, the actual flight environment is complex and changeable, parameter drift is caused by component aging, the establishment of the physical model with high precision and adaptation to the whole life cycle is extremely difficult, the calculation complexity is high, and the real-time requirement is difficult to meet. Knowledge-based methods utilize expert systems, feature tree analysis, and other domain-dependent expert experiences. Such methods are effective for known feature patterns, but it is difficult to find unknown, new or complex coupled feature patterns, and maintenance and updating of the knowledge base is time consuming and labor intensive. The data-driven based method utilizes a machine learning algorithm to mine patterns directly from the monitored data. Among other things, supervised learning methods (e.g., support vector machines, convolutional neural networks) have achieved significant results, but they rely heavily on large amounts of high quality tagged feature data. In the aviation field, high-grade characteristic data are extremely scarce, and manual labeling cost is high and subjectivity is strong, so that the application of the supervised learning method is limited. Disclosure of Invention The invention aims to provide a low-cost aircraft operation support method based on attention-characterization integrated clustering, which aims to solve the problems of high computational complexity, dependence on expert experience, dependence on tagged data and the like in the existing low-cost aircraft operation support technology. In order to realize the tasks, the invention adopts the following technical scheme: a low cost aircraft operation support method based on attention-characterization integrated clustering, comprising: acquiring historical sensor data of an aircraft, preprocessing the sensor data to obtain a feature matrix, and dividing time slices of the feature matrix to construct a sample; Constructing a global high-order incidence matrix based on an indication matrix corresponding to each base division; establishing groups by using a global high-order incidence matrix, and determining a centroid feature vector of each group; the method comprises the steps of constructing a deep neural network, wherein the deep neural network comprises an encoder, a decoder and an output unit, the encoder obtains a low-dimensional embedded representation based on a centroid feature vector and a global high-order incidence matrix, the decoder outputs a reconstructed adjacent matrix according to the low-dimensional embedded representation, and the output unit determines prediction probabilities of samples belonging to different preset categories according to the low-dimensional embedded representation; training the deep neural network by utilizing a feature matrix, wherein the total objective function during training comprises clustering loss and reconstruction loss, and the reconstruction loss is constructed based on a reconstruction adjacent matrix and an adjacent matrix corresponding to base division; And acquiring real-time sensor data of the aircraft, dividing the sensor data according to time slices, and inputt