CN117556310B

CN117556310B - Spacecraft residual life prediction method

Info

Publication number: CN117556310B
Application number: CN202410035054.5A
Authority: CN
Inventors: 杨志; 邓绍建; 刘威
Original assignee: Beijing Zhongke Spaceflight Talent Service Co ltd
Current assignee: Beijing Zhongke Spaceflight Talent Service Co ltd
Priority date: 2024-01-10
Filing date: 2024-01-10
Publication date: 2024-03-26
Anticipated expiration: 2044-01-10
Also published as: CN117556310A

Abstract

The invention discloses a spacecraft residual life prediction method, which belongs to the technical field of electric digital data processing (G06F), and comprises the following steps: acquiring telemetry data of a spacecraft; performing convolution operation on time sequence data of each parameter by using a convolution neural network model, and extracting time domain features and frequency domain features; constructing a spacecraft part connection topology structure diagram; aligning and fusing the time domain features, the frequency domain features and the structural features on a time axis; dynamically clustering the fusion characteristics through a clustering algorithm; constructing a stage prediction model; identifying the current degradation stage of the newly acquired telemetry data of the spacecraft, and inputting the degradation stage into a stage prediction model for online prediction; and carrying out weighted fusion on the prediction results of the prediction models of all the stages to obtain the residual life prediction result of the spacecraft. The invention combines various data processing and feature extraction technologies to comprehensively and deeply understand the working state of the spacecraft and predict the possible failure modes of the spacecraft.

Description

Spacecraft residual life prediction method

Technical Field

The invention relates to the technical field of electric digital data processing (G06F), in particular to a method for predicting the residual life of a spacecraft.

Background

Spacecraft are key tools for human research and exploration of the universe, each of which is a vast structure composed of numerous complex subsystems. With advances in technology, aerospace technology has also evolved dramatically over the last decades. More and more countries begin to recognize the value of space exploration, with a large number of spacecraft being launched into space each year to perform specific scientific tasks. When the spacecraft is designed and manufactured, a certain margin is reserved in order to ensure that the spacecraft can work normally in a remote space environment. Thus, even if the spacecraft has exceeded a preset lifetime, they can continue to provide service for a long time. Therefore, how to perform health management on the out-of-service spacecraft becomes an important and significant task.

In the health management research of the over-service spacecraft, predicting the residual life of the spacecraft is a key link. This requires extensive analysis of the spacecraft telemetry data collected by the ground measurement and control center to determine the current health of the on-orbit spacecraft, predicting the likely occurrence of the fault and the time at which the fault occurs. The prediction can not only provide the basis for the ground measurement and control personnel to replace equipment or adjust working state in time, but also greatly improve the use efficiency and safety of the spacecraft.

In the prior art, prediction schemes for the residual life of a spacecraft exist:

the Chinese patent with the bulletin number of CN115563888B discloses a method, a system, electronic equipment and a medium for predicting the residual life of a spacecraft; specifically disclosed is: the method relates to the technical field of spacecraft residual life prediction, and comprises the following steps: acquiring real-time telemetry time sequence data of a spacecraft; constructing a space-time self-encoder model; the space-time self-encoder model comprises a space-time encoder, a decoder and a residual life prediction module; the space-time encoder comprises a time domain encoder, a space domain encoder and a feature fusion module; and inputting the real-time telemetry time sequence data into a trained space-time self-encoder model to obtain the residual life of the spacecraft. The method and the device can improve the prediction accuracy of the residual life of the spacecraft.

The Chinese patent with the bulletin number of CN116011109B discloses a spacecraft service life prediction method, a spacecraft service life prediction device, electronic equipment and a storage medium; specifically disclosed is: the method comprises the following steps: acquiring telemetry data of a spacecraft, wherein the telemetry data comprises data of a plurality of working parameters of the spacecraft; sequentially performing outlier processing, feature extraction processing and standardization processing on the data of each working parameter to obtain target features of each working parameter; inputting target characteristics of all working parameters into a trained spacecraft health evolution model to obtain health factors corresponding to spacecraft telemetry data; inputting health factors corresponding to the spacecraft telemetry data into a trained spacecraft life prediction model to obtain predicted life and confidence intervals corresponding to the spacecraft telemetry data. The invention can synthesize multidimensional data to predict the residual life of the spacecraft and give out a confidence interval, thereby playing an important role in improving the accuracy of the life prediction of the spacecraft.

The Chinese patent with the bulletin number of CN116305531B discloses a modeling method, a device, equipment and a medium for a spacecraft health evolution model; specifically disclosed is: the method comprises the following steps: acquiring telemetry data of spacecrafts at different times and preprocessing the telemetry data; performing feature extraction based on the preprocessed spacecraft telemetry data; based on all the characteristic quantities, carrying out fusion processing to obtain one-dimensional characteristics corresponding to time; performing curve fitting on the one-dimensional characteristics according to the time sequence to obtain a fitted curve, and taking the fitting value of the fitted curve as a health factor corresponding to time; screening all the characteristic quantities based on the correlation degree between the health factors and the characteristic quantities, and selecting the characteristic quantities with high correlation degree; and constructing a fuzzy neural network, taking the selected characteristic quantity as input and the selected health factor as output, and training the fuzzy neural network to obtain the spacecraft health evolution model. The invention can obtain the evolution model for evaluating the health of the spacecraft.

The Chinese patent with publication number of CN113761722A discloses a spacecraft multi-working-condition life prediction method based on PCA; specifically disclosed is: the method comprises the following steps: s1, acquiring a data set, clustering the data under the working condition, and carrying out data standardization processing; s2, respectively carrying out PCA decomposition on each cluster data set and carrying out principal component analysis; s3, constructing a health index and a residual service life model based on main components, and predicting the residual service life of the spacecraft; s4, inputting sensor data, and predicting and estimating the residual life of the spacecraft by using a training model; the scheme provides a multi-working condition life prediction method based on PCA for spacecraft life prediction, can well process multi-working condition characteristics in the spacecraft operation process, and solves the problems that a plurality of telemetry parameters cannot objectively describe importance of the telemetry parameters to life prediction and the like.

Corresponding disclosures are also made in the relevant art, including CN113420624A, CN115982621A, CN116150901a, etc.

However, in the above-mentioned prior art, analysis of a single-factor (usually time-domain parameter) regression model is generally performed, and there is a problem that the prediction range is narrow and the prediction universality is not high.

The existing telemetry data processing method of the spacecraft may not fully utilize all available information, such as various parameters of positions, running states and the like, which are critical to understanding the working state of the spacecraft and predicting possible failure modes, and the existing processing method may not fully extract time domain features and frequency domain features related to the working performance and possible failure modes of the spacecraft, which are critical to fully understanding the working state of the spacecraft and possible changes thereof. In addition, in a complex spacecraft system, the performance of other components may be affected by the failure of one component, which may be captured by structural features, but the existing processing method may not fully take this into consideration, and at the same time, the existing processing method may not fully utilize the advantages of the density-based clustering algorithm DBSCAN, which may implement dynamic clustering of spacecraft telemetry data, so as to identify the current degradation stage, and has good adaptability.

For the problems in the related art, no effective solution has been proposed at present.

Disclosure of Invention

In order to overcome the above problems, the present invention aims to provide a method for predicting the residual life of a spacecraft, which aims to solve the problem that in a complex spacecraft system, the failure of one part may affect the performance of other parts, and the effect may be captured by structural features, but the existing processing method may not fully consider the problem.

For this purpose, the invention adopts the following specific technical scheme:

a method for predicting the remaining life of a spacecraft, the method comprising the steps of:

s1, acquiring telemetry data of a spacecraft, wherein the telemetry data of the spacecraft at least comprises time sequence data of each parameter;

s2, performing convolution operation on time sequence data of each parameter by using a convolution neural network model, and extracting time domain features and frequency domain features;

s3, constructing a spacecraft part connection topological structure diagram, and learning structural features among parts by adopting a graph convolution network model;

s4, aligning and fusing time domain features, frequency domain features and structural features on a time axis by adopting a splicing and transferring time sequence model to obtain fusion features;

S5, dynamically clustering the fusion characteristics through a clustering algorithm, and identifying the degradation stage of the spacecraft telemetry data at present;

s6, modeling a degradation stage where spacecraft telemetry data are currently located by adopting an LSTM network to obtain a stage prediction model;

s7, repeatedly executing the steps S2 to S5 on the newly acquired spacecraft telemetry data, identifying the current degradation stage of the newly acquired spacecraft telemetry data, and inputting the degradation stage into a stage prediction model for online prediction;

and S8, carrying out weighted fusion on the prediction results of the prediction models of all the stages to obtain the residual life prediction result of the spacecraft.

Optionally, the convolution neural network model is used for carrying out convolution operation on the time sequence data of each parameter, and the extracting of the time domain features and the frequency domain features comprises the following steps:

s21, inputting time sequence data of each parameter into a convolutional neural network model, performing convolutional operation on the time sequence data by utilizing a one-dimensional convolutional layer, and extracting local features in the time sequence data;

s22, increasing nonlinearity of a convolutional neural network model by adopting a nonlinear activation function, reducing the dimension of time sequence data by using a pooling layer, and reserving time domain features in the time sequence data;

S23, performing fast Fourier transform on the time sequence data, and converting the time sequence data into frequency domain data;

s24, carrying out convolution operation on the frequency domain data by using a one-dimensional convolution layer, extracting frequency domain features in the frequency domain data, and reducing the dimension of the frequency domain features by using a pooling layer.

Optionally, the constructing a spacecraft component connection topological structure diagram and adopting a graph convolution network model to learn structural features between components comprises the following steps:

s31, collecting parameter data of spacecraft parts, wherein the parameter data of the spacecraft parts at least comprise the types, functions and position information of the spacecraft parts and the connection modes among the spacecraft parts;

s32, constructing a connection topology structure diagram of the spacecraft component according to parameter data of the spacecraft component, wherein each node in the connection topology structure diagram represents a component, each side represents connection between the components, and simultaneously, the weights of the nodes and the sides and the directions of the sides are determined;

s33, inputting a connection topological structure diagram into a graph convolution network model, and initializing characteristics of nodes and edges;

s34, learning structural characteristics among spacecraft components by using a graph roll-up network model.

Optionally, step S4 includes:

S41, constructing a multidimensional characteristic parameter setThe method comprises the steps of carrying out a first treatment on the surface of the The method meets the following conditions:

wherein,、/>and->Respectively multidimensional characteristic parameter sets ++>At time->A time domain subset, a frequency domain subset, and a topology subset; />Is->、/>Is->、/>Is->The method comprises the steps of respectively obtaining a time domain parameter sequence number, a time domain parameter dimension, a frequency domain parameter sequence number, a frequency domain parameter dimension, a topological structure parameter sequence number and a topological structure parameter dimension;

s42, constructing a moment feature vector; constructing time domain feature vectors based on the time domain subset, the frequency domain subset and the topological structure subset respectivelyFrequency domain feature vector->And topology feature vector->The method comprises the following steps:

s43, inputting the feature vectors into a transducer model, and learning the dependency relationship between the feature vector data through a self-attention mechanism.

Optionally, the inputting the feature vector data into the transducer model and learning the dependency relationship between the feature vector data through the self-attention mechanism includes the steps of:

s431, setting parameters of a transducer model, wherein the parameters of the transducer model at least comprise the number of layers of an encoder, the number of layers of a decoder and the number of heads of multi-head attention;

s432, taking the feature vector data as input of a transducer model, and processing through a word vector layer and a position coding layer to obtain an input tensor;

S433, in the encoder, the input tensor learns the dependency relationship between the input features through the multi-head self-attention layer;

s434, processing the output characteristics of the encoder through a full connection layer to obtain the output of the encoder;

s435, in the decoder, the target feature interacts with the encoder output through a multi-head self-attention layer to obtain encoder information;

s436, the dependency relationship among target features is also learned through the multi-head self-attention layer in the decoder information;

s437, the decoder outputs target characteristics through full-connection layer processing, and generates a prediction sequence through linear transformation and softmax;

s438, repeatedly executing the steps of S431-S437, performing iterative optimization on the transducer model, completing sequence modeling of the features, calculating a cross entropy loss function, and updating parameters of the transducer model through back propagation.

Optionally, the calculation formula for calculating the cross entropy loss function is:

；

in the method, in the process of the invention,as a loss function;

is a mean square error loss function;

is a root mean square error loss function;

is a weight coefficient.

Optionally, the dynamic clustering of the fusion features by a clustering algorithm and the identification of the degradation stage of the spacecraft telemetry data currently located comprise the following steps:

S51, setting parameters of a clustering algorithm, and taking feature vector data obtained by fusing time domain, frequency domain and structural features as input;

s52, adopting an incremental DBSCAN clustering algorithm, and dynamically generating clusters under the condition of the nearest neighbor distance and the minimum telemetry data sample number;

s53, counting data distribution conditions of different clusters, and determining corresponding spacecraft degradation stages;

s54, calculating the distance between the newly input characteristic data and each clustering center, and judging the class of the newly input characteristic data so as to identify the current degradation stage;

s55, with continuous input of new data, the clustering center is adjusted through incremental learning, and dynamic clustering and stage recognition are achieved.

Optionally, the dynamically generating clusters using the incremental DBSCAN clustering algorithm, on the condition of the nearest neighbor distance and the minimum number of telemetry data samples, includes the following steps:

s521, DBSCAN clustering is carried out on the initial training data set, and core objects, boundary points and noise points are found;

s522, setting a clustering distance threshold and a minimum telemetry data sample number;

s523, when a new telemetry data sample point is added, calculating the distance between the new telemetry data sample point and the existing telemetry data sample;

s524, if the distance from the new telemetry data sample point to the core object is smaller than the distance threshold, adding the new telemetry data sample point to the category of the corresponding core object;

If the distance from all the core objects is greater than the distance threshold value, but the distance from the boundary point is less than the distance threshold value, adding the new telemetry data sample point into the category of the corresponding boundary point;

s525, if the new telemetry data sample point becomes a new core object, searching a new neighborhood telemetry data sample which accords with a distance threshold value and the minimum telemetry data sample number by taking the new telemetry data sample point as a center to form a new category;

s526, if the new telemetry data sample point does not belong to any existing category, and a new core object is not formed, marking the new telemetry data sample point as an outlier;

s527, repeatedly executing the steps of S521-S526, and repeatedly judging and updating the category each time a new telemetry data sample is added, so as to realize dynamic clustering.

Optionally, the calculating the distance between the newly input feature data and each cluster center and judging the class to which the newly input feature data belongs, so as to identify the current degradation stage comprises the following steps:

s541, calculating Euclidean distance between the new feature vector and each clustering center;

s542, assigning the new feature vector to a category corresponding to the cluster center with the nearest Euclidean distance;

s543, searching a corresponding degradation stage according to the category of the new feature vector through a pre-established category and degradation stage corresponding relation table;

S544, if the Euclidean distance between the new feature vector and all the clustering centers is larger than a preset threshold value, marking the new feature vector as an abnormal point, and failing to judge the degradation stage;

s545, counting the number of new feature vectors of each category in the set time window;

s546, determining a degradation stage corresponding to the most category in the number of the new feature vectors as a current degradation stage;

s547, sliding along with a time window, repeating the steps of S541-S546, and dynamically identifying the current degradation stage;

s548, if a new class appears, updating a corresponding relation table of the class and the degradation stage;

s549, judging the category through the distance between the clustering centers, and realizing the degradation stage identification of the new features.

Optionally, the step of performing weighted fusion on the prediction results of the prediction models of each stage to obtain the residual life prediction result of the spacecraft includes the following steps:

s81, collecting prediction results of the LSTM model at each stage on a new telemetry data sample, wherein the prediction results comprise a predicted life value;

s82, setting the weight of the prediction model of each stage;

s83, carrying out weighted fusion on each telemetry data sample;

s84, repeatedly executing the steps of S81-S83 to obtain a weighted fusion result of the new telemetry data sample on the prediction model of each stage;

S85, calculating the prediction results of all telemetry data samples, and obtaining the average prediction life at the current moment by calculation;

s86, dynamically predicting the residual life of the spacecraft along with the entry of a new telemetry data sample, and dynamically adjusting the weight of the prediction model at each stage according to the prediction effect.

Compared with the prior art, the application has the following beneficial effects:

1. the acquired telemetry data of the spacecraft comprise various information reflecting the working state and environmental condition of the spacecraft, such as position, running state and other parameters, the acquisition of the data provides a basis for subsequent fault prediction and service life estimation, the time sequence data are processed by utilizing a convolutional neural network model, so that time domain features and frequency domain features related to the working performance and possible fault modes of the spacecraft are extracted, the features are helpful for comprehensively understanding the working state and possible changes of the spacecraft, in addition, the structural features among the spacecraft parts can be captured by constructing a topological structure diagram of the spacecraft part connection and using a graph rolling network, so that the inherent interaction and dependency relationship of the spacecraft system can be further understood, and the step is particularly important, because in a complex spacecraft system, the fault of one part can affect the performance of other parts, the influence can be captured through the structural features, and the series of steps are fused with various data processing and feature extraction technologies to comprehensively and deeply understand the working state of the spacecraft, and predict the possible fault modes of the spacecraft, so that the service life prediction accuracy and reliability are improved.

2. The method has the advantages that the time domain features, the frequency domain features and the structural features are aligned and fused on a time axis to obtain the fused features, the process has the advantages that the time sequence information among the features can be reserved by using a transducer model, the inherent dependency relationship among the feature sequences can be effectively learned and modeled, and in addition, the method can continuously update the parameters of the model by using a cross entropy loss function to perform iterative optimization on the transducer model, so that the prediction capability of the model is improved, the accuracy of a prediction task is also improved, and meanwhile, the method can be used for combining the features into a unified feature sequence by aligning and splicing and fusing different features on the time axis.

3. The method integrates various data processing and feature extraction technologies to comprehensively and deeply understand the working state of the spacecraft, predicts a possible failure mode of the spacecraft, so that the accuracy and the reliability of life prediction are improved, the process particularly emphasizes the extraction and the application of structural features, because in a complex spacecraft system, the failure of one part can influence the performance of other parts, the influence can be captured through the structural features, in addition, the dynamic clustering of the telemetry data of the spacecraft can be realized by adopting a density-based clustering algorithm DBSCAN, so that the current degradation stage is identified, the method has good self-adaptability, the clustering center can be updated along with the arrival of new data, the dynamic monitoring of the degradation process is realized, the deep mining and intelligent processing of the telemetry data of the spacecraft are realized in the process, the operation safety and the efficiency of the spacecraft are improved, and the method has important practical value for the health management and the failure early warning of the spacecraft.

Drawings

The above features, features and advantages of the present invention, as well as the manner of attaining them and method of attaining them, will become more apparent and the invention will be better understood by reference to the following description of embodiments, taken in conjunction with the accompanying drawings. Here shown in schematic diagram:

Fig. 1 is a flowchart of a method of predicting remaining life of a spacecraft according to an embodiment of the invention.

Detailed Description

In order to make the present application solution better understood by those skilled in the art, the following description will clearly and completely describe the technical solution in the embodiments of the present application with reference to the accompanying drawings in the embodiments of the present application, and it is apparent that the described embodiments are only some embodiments of the present application, not all embodiments. All other embodiments, which can be made by one of ordinary skill in the art without undue burden from the present disclosure, are within the scope of the present disclosure.

According to an embodiment of the invention, a method for predicting the residual life of a spacecraft is provided.

The invention will be further described with reference to the accompanying drawings and detailed description, as shown in fig. 1, a method for predicting the residual life of a spacecraft according to an embodiment of the invention, the method for predicting the residual life of the spacecraft comprises the following steps:

s1, acquiring spacecraft telemetry data, wherein the spacecraft telemetry data at least comprises time sequence data of each parameter.

It should be noted that, the telemetry data of the spacecraft refers to state parameter data acquired by various sensors installed on the spacecraft and transmitted to the ground station during the in-orbit flight of the spacecraft. The method can reflect the information of the position, the running state, the working environment and the like of the spacecraft. The spacecraft telemetry system consists of a space section and a ground section, and performs data transmission through radio signals. The space section mainly comprises a sensor, a signal modulation and transmission device; the ground section then includes signal receiving and demodulating equipment.

Spacecraft telemetry data is of a wide variety including orbit telemetry, attitude telemetry, temperature telemetry, voltage telemetry, etc., and records the state of each parameter in a time series. For spacecraft life predictions, data reflecting the state of different components and systems of the spacecraft is required to be acquired. At the same time, there are also deletions and outliers in the telemetry data, requiring pre-processing to improve quality. Acquiring high quality telemetry data is the basis for spacecraft health management and life prediction.

S2, performing convolution operation on time sequence data of each parameter by using a convolution neural network model, and extracting time domain features and frequency domain features.

Preferably, the convolution neural network model is used for carrying out convolution operation on the time sequence data of each parameter, and the time domain feature and the frequency domain feature are extracted, which comprises the following steps:

It should be noted that, the convolutional neural network can learn the time domain features in the time-series data through the stacking of the convolutional layer and the pooling layer. The convolution kernel of the one-dimensional convolution layer scans time sequence data, the correlation of surrounding time steps is captured, and the superposition of the convolution layers can learn multi-level time domain characteristics. The pooling layer reduces feature dimensions while retaining the primary features. And the time sequence data is subjected to Fourier transformation to obtain the frequency domain information. The frequency domain features can also be learned on the frequency domain using a convolution layer and a pooling layer. Thus, both time domain information and frequency domain information can be learned. The parameter sharing characteristic of the convolutional neural network reduces the number of model parameters, and the translation invariant property of the model parameters makes the model parameters suitable for processing time sequence data. The activation function increases nonlinearity and the pooling layer dimension reduction prevents overfitting. Finally, splicing the time domain and frequency domain characteristics, and performing multidimensional modeling. The convolutional neural network can effectively extract representative time-frequency domain features from complex time-series data, provides information input for subsequent modeling, and is one of important technologies in time-series data analysis.

S3, constructing a spacecraft part connection topological structure diagram, and learning structural features among the parts by adopting a graph convolution network model.

Preferably, the construction of the spacecraft component connection topological structure diagram and the study of the structural features among components by adopting a graph roll-up network model comprise the following steps:

It should be noted that a spacecraft is a complex system that includes numerous components and their interconnection. The components and the connections are abstracted into graph data structures, so that the topological structure of the spacecraft can be visually represented. Wherein nodes represent components, edges represent connections between components, and weights represent the strength of the connections. Based on the topology map, modeling can be performed using a graph rolling network. The method can learn the dependency relationship among the nodes by aggregating the neighbor features of the nodes. The graph convolution network can directly perform convolution operation on the graph structure, does not need to be converted into fixed grid data, and is more suitable for irregular graph data. The graph rolling network can efficiently learn the connection characteristics among spacecraft components, and provides important structural knowledge for subsequent fault prediction and health management. The method can process the image structure data more than the traditional deep learning, and is more automatic and efficient than the manual extraction of the structure features. In conclusion, the part topological graph is constructed and the graph rolling network is applied, so that the internal structural knowledge of the spacecraft can be effectively learned, additional structural features are provided for life prediction tasks, and the prediction effect of the model is improved.

And S4, aligning and fusing the time domain features, the frequency domain features and the structural features on a time axis by adopting a splicing and transferring time sequence model to obtain fusion features.

Preferably, the aligning and fusing the time domain feature, the frequency domain feature and the structural feature on the time axis by adopting a splicing and transforming sequence model to obtain the fused feature includes the following steps:

Preferably, the inputting the feature vector data into the transducer model and learning the dependency relationship between the feature vector data by the self-attention mechanism includes the steps of:

Preferably, the calculation formula for calculating the cross entropy loss function is:

；

In the method, in the process of the invention,as a loss function;

is a mean square error loss function;

is a root mean square error loss function;

is a weight coefficient.

It should be noted that the transducer model (preamble codec predictor) can learn the dependency between input sequences through the encoder-decoder structure. Wherein both the encoder and decoder use a multi-headed attention mechanism, the inherent correlation of sequence data can be efficiently modeled. For complex spacecraft telemetry data, the transducer can learn the characteristics of three aspects of time domain, frequency domain and structure, respectively. And then, by aligning different features on a time axis, splicing and fusing, combining the different features into a unified feature sequence. The self-attention mechanism of the transducer can effectively learn the inherent dependencies of the feature sequence. By means of the encoder-decoder framework, timing information between features can be preserved. And finally, generating a target feature sequence aiming at the prediction task. Compared with simple splicing, the method uses a transducer to perform feature fusion, so that richer feature representation can be obtained. It learns the correlation between different features, which is more helpful for subsequent failure prediction and life estimation. In the whole, the method fuses the multi-source characteristic information and improves the prediction capability of the model.

S5, dynamically clustering the fusion characteristics through a clustering algorithm, and identifying the degradation stage of the spacecraft telemetry data at present.

Preferably, the dynamic clustering of the fusion features by a clustering algorithm and the identification of the degradation stage of the spacecraft telemetry data currently comprises the following steps:

s52, adopting an incremental DBSCAN clustering algorithm (Density-Based Spatial Clustering ofApplications withNoise, a Density-based clustering algorithm), and dynamically generating clusters under the condition of nearest neighbor distance and minimum telemetry data sample number;

Preferably, the dynamically generating clusters using the incremental DBSCAN clustering algorithm, on the condition of the nearest neighbor distance and the minimum telemetry data sample number, includes the following steps:

Preferably, the step of calculating the distance between the newly input feature data and each cluster center and judging the class to which the newly input feature data belongs so as to identify the current degradation stage comprises the following steps:

It should be explained that DBSCAN is a density-based clustering algorithm, and clusters and distribution patterns thereof in data can be identified through cluster analysis. The incremental DBSCAN can realize dynamic clustering, namely, update new data in situ without slave clustering. The mode in the telemetry data of the spacecraft can be found through clustering, and different modes can correspond to different working states or degradation phases of the spacecraft. After new data enter, judging which cluster the new data belongs to, and identifying the current degradation stage. The degradation phase is more adaptive than the preset threshold using cluster analysis. With the arrival of new data, the clustering center can be updated, and the dynamic monitoring of the degradation process is realized. The DBSCAN only needs to set two parameters, is simple to realize, can find clusters with any shape, and is suitable for data with complex distribution. DBSCAN is also less sensitive to parameters and initial values than KMeans et al. In conclusion, the DBSCAN clustering algorithm can be used for effectively finding out the mode in the telemetry data, intelligent monitoring of the degradation process of the spacecraft is achieved, and the effect of fault prediction is improved.

And S6, modeling a degradation stage where the telemetry data of the spacecraft are currently located by adopting an LSTM network to obtain a stage prediction model.

It should be noted that LSTM network (long and short term memory network) is a recurrent neural network capable of learning long-sequence data. The system controls the flow of states through mechanisms such as forget gates, input gates and the like, and can capture long-distance dependency relations in time sequence data. The telemetry data patterns of a spacecraft may be different during different degradation phases. The telemetry data of a specific stage is input into the LSTM network for training, and the characteristic extraction and modeling of the stage can be obtained. Compared with a simple fully-connected network, the LSTM is more suitable for processing time sequences, and can learn the internal rule of time sequence state transition. It is also more robust to sequence length variations than convolutional networks. And finally, integrating LSTM models of each stage, wherein the models of different stages can capture time sequence characteristics under corresponding states, so that the modeling capacity of the whole degradation process is improved, and the prediction effect is enhanced. In conclusion, the LSTM network is used for stage modeling, so that deep learning expression of a dynamic degradation process can be obtained, and fault prediction and health management of the spacecraft are facilitated.

And S7, repeatedly executing the steps S2 to S5 on the newly acquired spacecraft telemetry data, identifying the current degradation stage of the newly acquired spacecraft telemetry data, and inputting the degradation stage into a stage prediction model for online prediction.

It should be noted that online prediction refers to timely predictive analysis of newly acquired real-time telemetry data. The method needs to repeatedly perform the processes of feature extraction, cluster recognition and the like, so as to realize the real-time processing of new data. Compared with offline analysis, online prediction can immediately feed back the latest state and perform real-time monitoring. It needs to be ensured that the individual processing modules are efficient enough to accommodate streaming data. Repeated feature extraction may update the state representation and repeated clustering may identify the most recent degradation stage. The new data is input into the corresponding stage model, so that the service life can be dynamically predicted. Meanwhile, the model can be updated through feedback results, so that incremental learning is realized. The on-line prediction enables the whole system to be more intelligent and automatic. In conclusion, online prediction is a key link of spacecraft health management, dynamic tracking of the whole life cycle can be achieved, risks are fed back in time, and decision efficiency is improved.

Preferably, the step of weighting and fusing the prediction results of the prediction models of each stage to obtain the residual life prediction result of the spacecraft includes the following steps:

s82, setting the weight of the prediction model of each stage;

s83, carrying out weighted fusion on each telemetry data sample;

It should be explained that multi-model fusion is a common integrated learning method. Different single models have advantages, and fusion can utilize the advantages of each model to improve the stability and robustness of prediction. The LSTM models for each degradation stage are weighted fused here. The weighted coeffients may be set based on the verification performance of the model or may be set to the same weight. With the arrival of new data, online prediction can be continuously performed, and the weights coeffients of the models are dynamically adjusted according to the prediction effect, so that self-adaptive optimization is realized. Compared with a single model, the model fusion considers the characteristics under different states, can more comprehensively describe the whole degradation process, and has more reliable prediction results. In summary, multi-model fusion is one of the important methods of integrated learning, can improve the accuracy and the robustness of prediction, and is more suitable for the problem of complex sequence prediction.

In summary, by means of the above technical solution of the present invention, the acquired telemetry data of the spacecraft include various information reflecting the working states and environmental conditions of the spacecraft, such as position, running states, etc., the acquisition of these data provides a basis for subsequent failure prediction and life estimation, and the time series data are processed by using the convolutional neural network model, so as to extract time domain features and frequency domain features related to the working performance and possible failure modes of the spacecraft, these features are helpful for comprehensively understanding the working states of the spacecraft and possible changes thereof, in addition, by constructing the connection topology of spacecraft components and using the graph-convolution network, the structural features between spacecraft components can be captured, thus the inherent interaction and dependency relationship of the spacecraft system are further understood, which is particularly important, because in the complex spacecraft system, the failure of one component may affect the performance of other components, and this effect may be captured by the structural features, in general terms, this series of steps fuses various data processing and feature extraction techniques, so as to comprehensively and deeply understand the working states of the spacecraft, predict the possible failure modes thereof, thereby improving the reliability of the spacecraft, and the reliability of the present invention, and the time domain features can be input into the model, and the time domain features of the invention, and the time domain features of the time domain and the forward model can be fused, thereby the three-domain features can be input and the model, and the deep learning of the structural features can be better and the model, meanwhile, by aligning different features on a time axis and splicing and fusing the features, the features can be combined into a unified feature sequence, the method is superior to simple splicing, because the method can retain time sequence information among the features, and effectively learn the inherent dependency relationship of the feature sequence through a self-attention mechanism of a transducer, in addition, the method can continuously update the parameters of the transducer model by using a cross entropy loss function, thereby not only enhancing the prediction capability of the model, but also improving the precision of a prediction task, and in general, the method can provide richer feature representation by fusing multi-source feature information, brings great convenience to the subsequent fault prediction and service life estimation task, and further remarkably improves the prediction capability and precision of the model; the method integrates various data processing and feature extraction technologies to comprehensively and deeply understand the working state of the spacecraft, predicts a possible failure mode of the spacecraft, so that the accuracy and the reliability of life prediction are improved, the process particularly emphasizes the extraction and the application of structural features, because in a complex spacecraft system, the failure of one part can influence the performance of other parts, the influence can be captured through the structural features, in addition, the dynamic clustering of the telemetry data of the spacecraft can be realized by adopting a density-based clustering algorithm DBSCAN, so that the current degradation stage is identified, the method has good self-adaptability, the clustering center can be updated along with the arrival of new data, the dynamic monitoring of the degradation process is realized, the deep mining and intelligent processing of the telemetry data of the spacecraft are realized in the process, the operation safety and the efficiency of the spacecraft are improved, and the method has important practical value for the health management and the failure early warning of the spacecraft.

Although the invention has been described with respect to the preferred embodiments, the embodiments are for illustrative purposes only and are not intended to limit the invention, as those skilled in the art will appreciate that various modifications can be made without departing from the spirit and scope of the invention as defined by the appended claims.

Claims

1. The method for predicting the residual life of the spacecraft is characterized by comprising the following steps of:

s1, acquiring spacecraft telemetry data, wherein the spacecraft telemetry data at least comprises time sequence data of each parameter;

s3, constructing a spacecraft part connection topological structure diagram, and learning structural features among parts by adopting a graph convolution network model, wherein the method comprises the following specific steps of:

s34, learning structural features among spacecraft components by using a graph roll-up network model;

s5, dynamically clustering the fusion feature vectors through a clustering algorithm, and identifying the current degradation stage of the telemetry data of the spacecraft;

2. The method for predicting the residual life of a spacecraft according to claim 1, wherein the step of performing convolution operation on time series data of each parameter by using a convolution neural network model, and extracting time domain features and frequency domain features comprises the steps of:

3. The method for predicting the remaining life of a spacecraft of claim 1, wherein step S4 comprises:

；

4. A method of predicting remaining life of a spacecraft as claimed in claim 3, wherein said inputting feature vectors into a transducer model and learning the dependency between feature vector data by a self-attention mechanism comprises the steps of:

5. The method for predicting the remaining life of a spacecraft of claim 4, wherein the calculation formula for calculating the cross entropy loss function is:

；

in the method, in the process of the invention,as a loss function;

is a mean square error loss function;

is a root mean square error loss function;

is a weight coefficient。

6. The method for predicting the residual life of a spacecraft according to claim 5, wherein the step of dynamically clustering the fusion features by a clustering algorithm and identifying the degradation stage in which the telemetry data of the spacecraft is currently located comprises the steps of:

7. The method for predicting the remaining life of a spacecraft of claim 6, wherein said dynamically generating clusters using an incremental DBSCAN clustering algorithm, subject to a nearest neighbor distance and a minimum number of telemetry data samples, comprises the steps of:

8. The method for predicting the remaining life of a spacecraft of claim 6, wherein said identifying the current degradation stage by calculating the distance between the newly inputted feature data and each cluster center and determining the class to which the newly inputted feature data belongs comprises the steps of:

9. The method for predicting the residual life of a spacecraft according to claim 1, wherein the step of weighting and fusing the prediction results of the prediction models of each stage to obtain the residual life prediction result of the spacecraft comprises the following steps:

s82, setting the weight of the prediction model of each stage;

s83, carrying out weighted fusion on each telemetry data sample;