Carrier rocket pre-launching parameter fault prediction method fused with space-time attention mechanism
By integrating a temporal convolutional network with a spatiotemporal attention mechanism and a fault classification network, and utilizing multi-dimensional pre-launch parameters for launch vehicle fault prediction, this method solves the problems of wasted manpower and resources and safety risks in traditional methods. It achieves efficient and accurate fault early warning without power outages for maintenance, thereby improving the safety and reliability of launch vehicles.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-26
- Publication Date
- 2026-04-10
AI Technical Summary
Existing technologies are insufficient for efficient and comprehensive prediction of launch vehicle failures. Traditional methods involve waste of human and material resources and pose safety risks. Furthermore, traditional periodic maintenance cannot meet the needs of complex rocket structures.
A temporal convolutional network and a fault classification network with spatiotemporal attention mechanism are used to predict faults through multi-dimensional pre-launch parameters. Deep learning methods are used to learn the historical operation of the launch vehicle, and multi-dimensional time series data are combined to provide fault early warning.
It enables fault early warning without power outages for maintenance, reduces costs and risks, provides comprehensive and accurate fault prediction results, and improves the safety and reliability of launch vehicles.
Smart Images
Figure CN121834543A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of pre-launch parameter prediction and analysis of launch vehicles, and specifically relates to a method for analyzing and predicting launch vehicle failures based on multi-dimensional pre-launch parameters. Background Technology
[0002] As the number of launch missions increases, the scale of launch vehicle design and production plans continues to expand, and the importance of quality increases. Damage or breakage of a launch vehicle can not only cause huge economic losses but also pose significant safety hazards.
[0003] For a long time, manual maintenance methods have been used, leading to a significant waste of human and material resources and posing risks to maintenance personnel due to a lack of understanding of the rocket's operational status. Furthermore, the conditions of a launch vehicle during testing, such as operating voltage and temperature, are inconsistent with its dynamic state during operation, affecting the accuracy of diagnostics. Since most launch vehicle damage is sudden, traditional periodic maintenance methods are no longer sufficient to meet the needs of current aerospace development. At the same time, the increasing number and complexity of various models, along with the complex structures of new rockets, renders traditional experience-based methods of launch vehicle maintenance inapplicable. Summary of the Invention
[0004] The purpose of this invention is to overcome the aforementioned shortcomings and provide a method for predicting pre-launch parameter faults in launch vehicles by incorporating a spatiotemporal attention mechanism, thus solving the technical problem of high difficulty in early warning of launch vehicle faults. This invention can provide fault information in advance, which has important guiding significance for the operation and maintenance of launch vehicles.
[0005] To achieve the above-mentioned objectives, the present invention provides the following technical solution:
[0006] A method for predicting pre-launch parameter failures of launch vehicles by incorporating a spatiotemporal attention mechanism includes:
[0007] (1) Obtain multi-dimensional pre-launch parameters of the launch vehicle;
[0008] (2) Standardize the multi-dimensional pre-launch parameters and use the processed data to obtain the training and test sets of the temporal convolutional network;
[0009] (3) Construct a temporal convolutional network for predicting multidimensional pre-fire parameters;
[0010] (4) Input the training and test data samples of the temporal convolutional network into the temporal convolutional network used to predict multi-dimensional pre-shot parameters to obtain the trained temporal convolutional network.
[0011] (5) Define time windows for the multi-dimensional pre-launch parameters obtained in step (1), with each time window corresponding to a fault type label, to obtain the training set and test set of the fault classification network;
[0012] (6) Construct a fault classification network for predicting specific fault types corresponding to multi-dimensional pre-launch parameters;
[0013] (7) Input the training set and test set data samples of the fault classification network into the fault classification network to obtain the trained fault classification network;
[0014] (8) Input the standardized multi-dimensional pre-launch parameters into the trained temporal convolutional network to obtain the multi-dimensional pre-launch parameter prediction values. Input the multi-dimensional pre-launch parameter prediction values into the trained fault classification network to output the fault type.
[0015] Furthermore, in step (1), the multi-dimensional pre-launch parameters include power supply voltage, flight data, and various key pressures or temperatures.
[0016] Furthermore, in step (2), the method for standardizing the multi-dimensional pre-launch parameters includes:
[0017] For each feature dimension in the multi-dimensional pre-launch parameters, mean and variance normalization are performed to ensure that the processed multi-dimensional pre-launch parameters conform to a standard normal distribution, i.e., the mean is 0 and the standard deviation is 1; the nearest neighbor substitution method is used to fill in the missing data values.
[0018] Furthermore, in step (3), the input layer of the temporal convolutional network is used to input multi-dimensional pre-fire parameter samples, and the output layer outputs multi-dimensional pre-fire parameter prediction values.
[0019] In step (6), the input layer of the fault classification network is used to input multi-dimensional pre-launch parameter samples and the fault types corresponding to the time windows, the hidden layer is used to learn the relationship between multi-dimensional pre-launch parameters and fault types, and the output layer outputs the fault type prediction results.
[0020] Furthermore, in step (5), when defining the time window for the multi-dimensional pre-launch parameters obtained in step (1), the length of the real-time multi-dimensional pre-launch parameter time window can be different, and the average sampling of a single time window in the training data is 60 times.
[0021] Furthermore, in step (5), the fault type label is set as follows:
[0022] 0 corresponds to no fault, while 1, 2, and 3 correspond to three different types of faults.
[0023] Furthermore, the parameter settings for the fault classification network are as follows:
[0024] Input data length: n is the length of the multi-dimensional pre-launch parameters, corresponding to the n dimensions of the multi-dimensional pre-launch parameters; m is the length of the fault type array, corresponding to the time window length m (units) of the multi-dimensional pre-launch parameters;
[0025] The average length of the time series is 60.
[0026] The number of neurons in the hidden layer is 128.
[0027] Furthermore, for temporal convolutional networks, the input of each sample is the multi-dimensional pre-fire parameters corresponding to a certain time point; for fault classification networks, the input of a single sample is a multi-dimensional pre-fire parameter matrix including the temporal domain for a certain time window.
[0028] Furthermore, the fault classification network uses Inception-v4 as the backbone network and introduces an LSTM network with attention mechanism and a graph attention network GAT.
[0029] LSTM networks are used to learn the temporal correlations of time-series data, and classification accuracy is improved by adding a Squeeze-and-Excitation Block module to the fully convolutional block;
[0030] The Graph Attention Network (GAT) is used to discover spatial correlations between variables; by fusing the features extracted from each branch, it is fed into a softmax classifier to output the event category.
[0031] Furthermore, in the Graph Attention Network (GAT), each node is assumed to have a connection with itself. Self-loops are added to enhance the self-information ratio of nodes to avoid excessive interference from neighboring nodes.
[0032] When the prior knowledge of the connection relationships between variables is not available, the default connection matrix is a fully connected matrix;
[0033] If x (i) With x (j) Neighboring nodes, an attention coefficient e is introduced. ij x represents (i) For x (j) The importance of, namely:
[0034] e ij =a(Wx (i) Wx (j) )
[0035] Where a(·) is the Attention module implemented using a single-layer feedforward network, and W is the weight matrix of the graph model.
[0036] Compared with the prior art, the present invention has at least one of the following advantages:
[0037] (1) Operability: This invention can provide early warning of launch vehicle malfunctions without power outage maintenance, which greatly reduces costs and reduces the operational risks for maintenance personnel. The launch vehicle malfunction early warning system can run automatically and provides easy-to-understand early warning results. It is easy to operate and requires less professional knowledge.
[0038] (2) Comprehensiveness: This invention comprehensively evaluates and predicts the launch vehicle's fault conditions by integrating characteristic data such as the pressure compensation value, the temperature value during pre-cooling, the command feedback value during servo swing, and the key voltage of the control system. Compared with single-parameter fault repair schemes, it is more comprehensive and can provide fault information in advance.
[0039] (3) Accuracy: This invention mainly uses deep learning to learn the historical operation of launch vehicles through the network, and uses a large amount of data to drive the training model, achieving relatively accurate prediction results in the simulation data testing stage. Attached Figure Description
[0040] Figure 1 This is an overall flowchart of the pre-launch parameter fault prediction method for launch vehicles that integrates a spatiotemporal attention mechanism, as described in this invention.
[0041] Figure 2 This is a network structure diagram for fault classification in this invention. Detailed Implementation
[0042] The features and advantages of the present invention will become clearer and more apparent from the following detailed description.
[0043] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments. Although various aspects of embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless specifically indicated otherwise.
[0044] Methods for analyzing and predicting launch vehicle faults based on multi-dimensional pre-launch parameters have significant practical implications. Among these, detection technology based on multi-dimensional pre-launch parameter information represents an important development direction for launch vehicle maintenance. Currently used fault detection methods do not fully utilize the interrelationships between pre-launch parameters, making it difficult to provide comprehensive and accurate fault predictions for launch vehicles.
[0045] This invention proposes a launch vehicle fault analysis and prediction method based on multi-dimensional pre-launch parameters. It can comprehensively analyze key information such as pressure values during pressurization, temperature values during pre-cooling, command feedback values during servo oscillation, and critical voltages of the control system to provide early warnings of abnormal launch vehicle faults. The feature matrix extracted from the launch vehicle's pre-launch parameters is referred to as multi-dimensional pre-launch parameters. Continuous parameters constitute time-series data, with 60 time windows. The method of this invention is composed of an integrated prediction network and a classification network.
[0046] Specifically, this invention provides a method for analyzing and predicting launch vehicle failures based on multi-dimensional pre-launch parameters. The feature data stream extracted before launch is referred to as multi-dimensional pre-launch parameters. These time-domain aligned multi-dimensional pre-launch parameters form a matrix-based multi-dimensional time series. The specific implementation steps are as follows:
[0047] Step 1: Collect characteristic data such as pressurization pressure, precooling temperature, servo swing command feedback value, and key voltage of the control system before normal launch and during malfunctions of the launch vehicle, and extract multi-dimensional pre-launch parameters. The extraction time interval is 1 minute, and each extraction corresponds to a time point.
[0048] Step 2: For the multi-dimensional pre-fire parameters extracted in Step 1, after preprocessing, they are standardized to better fit the temporal convolutional network and conform to a normal distribution. After processing, 60% of the time points are selected as the training set and 40% as the test set.
[0049] Step 3: Construct a temporal convolutional network to predict the values of multi-dimensional pre-launch parameters over a future period of time.
[0050] Step 4: Input the training set data, validation set data, and test set data processed in Step 2 into the temporal convolutional network constructed in Step 3. The input of a single sample is the multi-dimensional pre-fire parameters corresponding to a certain time point; the output is the predicted value of the multi-dimensional pre-fire parameters, thus obtaining the trained temporal convolutional network.
[0051] Step 5: For the multi-dimensional pre-launch parameters extracted in Step 1, define time windows, with each time window corresponding to a fault type label, serving as an input sample for the fault classification network. Label 0 corresponds to no fault, while 1, 2, and 3 correspond to the three fault types, respectively. Select 60% of the labeled multi-dimensional pre-launch parameters as the training set and 40% as the test set.
[0052] Step 6: Construct a classification network to predict the specific fault types corresponding to multi-dimensional pre-launch parameters.
[0053] Step 7: Input the processed training and test data from Step 5 into the fault classification network constructed in Step 6. The input for a single sample is a multi-dimensional pre-fire parameter matrix including the time domain within a certain time window; the output is the fault classification result corresponding to the multi-dimensional pre-fire parameter matrix. The trained fault classification network is then obtained.
[0054] Step 8: After obtaining the multi-dimensional pre-launch parameters of a specified real-time length, preprocess and standardize the multi-dimensional pre-launch parameters according to Step 2, and then input them into the pre-trained temporal convolutional network to obtain the predicted values of the multi-dimensional pre-launch parameters. Input the predicted multi-dimensional pre-launch parameter matrix into the fault classification network described in Step 6 to output the fault type value. If the predicted value is 0, it means that no fault will occur within the predicted time period; if the predicted value is 1 / 2 / 3, it means that type 1 / 2 / 3 faults may occur within the predicted time period, and fault category information is sent for early warning.
[0055] There are three types of early warning faults. If more fault types are added, the network should be retrained.
[0056] The number of sampling times for real-time multi-dimensional pre-launch parameters can vary, with an average of 60 sampling times per time window in the training data.
[0057] Furthermore, step 1 extracts multi-dimensional pre-launch parameters, specifically implemented as follows.
[0058] Pre-launch parameter monitoring status includes the following parameters for each system: power supply voltage, flight data, key pressures, and temperature.
[0059] Furthermore, the standardization described in step 2 is specifically implemented as follows:
[0060] Standardization involves removing the mean and normalizing the variance for each feature dimension, ensuring that the processed multi-dimensional pre-projection parameters conform to a standard normal distribution, i.e., a mean of 0 and a standard deviation of 1. For missing values and outliers generated during data collection and recording, the nearest neighbor substitution method is used. That is, for missing values of the i-th dimension variable at time t... Search for the match in the entire sequence closest observation If the observation value at the next moment Not missing, available Alternative
[0061] Furthermore, the temporal convolutional network in step 3 is implemented as follows:
[0062] Temporal convolutional models consist of an input layer, hidden layers, and an output layer.
[0063] The input layer is used to input multi-dimensional pre-fire parameter samples. After setting seq_length, batch_size, and cell_input, the time series is input into the input layer. The hidden layer is used to learn the relationship between multi-dimensional pre-fire parameters in the time domain and different feature dimensions. The output layer outputs the predicted values of multi-dimensional pre-fire parameters.
[0064] Furthermore, the fault classification network is implemented as follows:
[0065] The input layer is used to input multi-dimensional pre-launch parameter samples and corresponding fault types in the time series. The hidden layer is used to learn the relationship between multi-dimensional pre-launch parameters and fault types. The output layer outputs the fault type prediction results.
[0066] Furthermore, the parameter settings for the fault classification network are as follows:
[0067] Input data length: n is the length of the multi-dimensional pre-launch parameters, corresponding to the n dimensions of the multi-dimensional pre-launch parameters; m is the length of the fault type array, corresponding to the time window length m (units) of the multi-dimensional pre-launch parameters.
[0068] Average length of time series: 60;
[0069] Number of neurons in the hidden layer: 128.
[0070] This invention treats characteristic data streams such as pressurization pressure values before normal launch and during malfunctions, temperature values during pre-cooling, command feedback values during servo swing, and key voltages of the control system as matrix-based time series data. It employs a recurrent neural network to learn the horizontal and vertical patterns between different dimensions of the multi-dimensional time series, thereby achieving time series prediction. Based on the prediction results of each parameter feature, a deep learning network is used to learn the relationship between the multi-dimensional pre-launch parameter feature sequences and fault types, thus achieving the goal of early warning of launch vehicle malfunctions.
[0071] Example:
[0072] The invention will now be further described with reference to the accompanying drawings.
[0073] This invention provides a method for predicting launch vehicle faults based on multi-dimensional pre-launch parameters. The invention treats various parameters of the launch vehicle before normal launch and during fault occurrences, including pressure values during pressurization, temperature values during pre-cooling, command feedback values during servo oscillation, and key voltages of the control system, as matrix-based time-series data. A recurrent neural network is used to learn the horizontal and vertical patterns between different dimensions of the multi-dimensional time series, thereby achieving time series prediction. Based on the prediction results of each parameter feature, a deep learning network is used to learn the relationship between the pre-launch parameter feature sequence and fault types, thus achieving data-driven analysis of launch vehicle faults and achieving early warning of launch vehicle faults. This invention can effectively predict and classify the data trends of pre-launch parameters of launch vehicles, provide early warnings before launch vehicle faults occur, and intelligently assess the types of potential faults, greatly improving the safety and reliability of launch vehicles and effectively identifying faults at an early stage.
[0074] The feature matrix extracted before the launch of the carrier rocket is referred to as the multi-dimensional pre-launch parameter. The continuous multi-dimensional parameters constitute time series data, with 60 time windows. This method is composed of a prediction network and a classification network integrated and connected.
[0075] like Figure 1 As shown, a multispectral-based power equipment fault early warning method is described, and the specific implementation steps are as follows.
[0076] Step 1: Collect characteristic data such as pressurization pressure, pre-cooling temperature, servo swing command feedback value, and key voltages of the control system before normal launch and during periods of malfunction. Extract multi-dimensional pre-launch parameters at 1-minute intervals, with each extraction corresponding to a specific time point. Taking temperature as an example, the same applies to other parameters. Extract multi-dimensional pre-launch parameters at 1-minute intervals, with each extraction corresponding to a specific time point.
[0077] For the entire generated multidimensional pre-launch parameter matrix, the above n dimensions are aligned in the time domain, and the multidimensional pre-launch parameter prediction problem is regarded as a multivariate regression problem.
[0078] Step 2: Preprocess all extracted multi-dimensional pre-projection parameters, using the nearest neighbor substitution method to fill in missing data values. To better fit the temporal convolutional network and improve the speed and accuracy of gradient descent, the multi-dimensional pre-projection parameters are standardized to fit a normal distribution. Standardization involves removing the mean and normalizing the variance for each feature dimension, ensuring that the processed multi-dimensional pre-projection parameters conform to a standard normal distribution, i.e., a mean of 0 and a standard deviation of 1.
[0079] Step 3: Construct a temporal convolutional network to predict the multi-dimensional pre-firing parameters of insulators over a future period. This network consists of an input layer, hidden layers, and an output layer. The input layer takes the multi-dimensional pre-firing parameter samples as input. After setting seq_length, batch_size, and cell_input, the time series data is input into the input layer. The hidden layer learns the relationship between the multi-dimensional pre-firing parameters in the temporal domain and different feature dimensions. The output layer outputs the predicted values of the multi-dimensional pre-firing parameters.
[0080] Given an N-dimensional time series Where the i-th ∈ [1, N] dimension sequence Output sequence Where the i-th dimension sequence is T x and T y These are the lengths of the input and output time series, respectively. Note that... It only depends on the state before t, and not on... Irrelevant.
[0081] Temporal convolutional networks have a longer memory capacity than recurrent networks and avoid the gradient explosion and vanishing problems, making them more suitable for sequence modeling tasks that require long historical records. Furthermore, temporal convolutional networks can increase their receptive field by stacking more dilated convolutional layers, using larger dilation factors, or using larger convolutional kernels, making them easier to transfer to other domains.
[0082] Multidimensional variable sequence prediction networks consist of one-dimensional fully convolutional layers with the same input and output lengths. Their characteristics are as follows: ① Dilated convolution: Temporal convolutional networks utilize the powerful properties of convolution to extract features across time steps; there can be a certain distance between elements in the input sequence. ② Causal convolution: For the output value at time t, its input can only be time t or times prior to t. ③ To avoid gradient vanishing, temporal convolutional networks employ residual block structures, making their structure more generalizable. The parameters that need to be set in a temporal convolutional network are as follows.
[0083] Input data length: m, corresponding to the time window length of the multi-dimensional pre-launch parameters;
[0084] The number of hidden layer neurons is 128. The model has sufficient complexity to ensure learning ability, and 128 is 2 to the power of 7, which is beneficial for model training.
[0085] Step 4: Input the training set data processed in Step 2 into the temporal convolutional network set in Step 3. The input of a single sample is the multi-dimensional pre-fire parameter corresponding to a certain time point; the output is the predicted value of the multi-dimensional pre-fire parameter, thus obtaining the trained temporal convolutional network.
[0086] Step 5: For the multi-dimensional pre-launch parameters extracted in Step 1, time windows are defined according to historical operation data. Each time window corresponds to a fault type label, which serves as an input sample for the fault classification network. Label 0 corresponds to no fault, and labels 1, 2, and 3 correspond to the three types of faults, respectively. 60% of the labeled multi-dimensional pre-launch parameters are selected as the training set, and 40% as the test set.
[0087] Step 6: Construct a classification network to predict specific fault types corresponding to multi-dimensional pre-launch parameters, such as... Figure 2 As shown, it includes an input layer, a hidden layer, and an output layer. The input layer is used to input multi-dimensional pre-launch parameter samples and the corresponding fault types in the time series. The hidden layer is used to learn the relationship between multi-dimensional pre-launch parameters and fault types. The output layer outputs the fault type prediction results.
[0088] Given a multivariate time series If the input x is the output of the preceding prediction network, i.e., a multivariate time series containing both known and predicted values, then F = T. x +T y To extract the spatiotemporal dependencies of multivariate sequences, this invention proposes a deep neural network that integrates multidimensional spatiotemporal information. This network uses Inception-v4 as its backbone, introduces an LSTM network with an attention mechanism to learn the temporal correlations of time-series data, and a Graph Attention Network (GAT) to mine the spatial correlations between variables. The features extracted from each branch are then fused and fed into a softmax classifier to output the event category.
[0089] A multivariate attention model is incorporated into a multivariate LSTM fully convolutional network to learn long-term dependencies in time series. Furthermore, classification accuracy is further improved by adding a Squeeze-and-Excitation Block module to the fully convolutional blocks. Considering that not all feature maps have the same impact on subsequent layers, this adaptive recalibration of feature maps can be viewed as a form of self-attention learned from the output feature maps of previous layers.
[0090] The spatial feature extraction branch utilizes the GAT network to transform the input multivariate time series into higher-dimensional features. By assigning different weights to different nodes in the neighborhood, it can more flexibly learn the structural features of the graph, i.e., the relationships between nodes. The input is a multivariate time series. Each dimension variable is treated as a node in a graph structure, with N nodes. The output is... This involves mapping input variables of length F to feature vectors of length d. Note that in graph models, each node is assumed to be connected to itself. Adding self-loops enhances the self-information of nodes, thus avoiding excessive interference from neighboring nodes and achieving the goal of accurately uncovering the relationships between different variables. Furthermore, when prior knowledge of the connections between variables is unavailable, the connection matrix is assumed to be a fully connected matrix. If x... (i) With x (j) Neighboring nodes, an attention coefficient e is introduced. ij x represents (i) For x (j) The importance of, namely:
[0091] e ij =a(Wx (i) Wx (j) )
[0092] Here, a(·) is the Attention module implemented using a single-layer feedforward network, and W is the weight matrix of the graph model. This branch first parameterizes the weight vector using the softmax function, and then obtains the feature vector of each graph node through LeakyReLU (a function) non-linear activation.
[0093] The spatial feature extraction network branch learns the numerical features of associated variables and can adaptively adjust the attention coefficients between variables when the multivariate relationships change dynamically, thus enhancing the model's generalization ability under different conditions. Considering both the temporal variation patterns during the event occurrence process and the relationships between multidimensional variables, an attention mechanism is introduced to accurately capture the spatiotemporal information carried in multivariate time series. This invention proposes an early fault detection algorithm that combines sequence prediction and sequence recognition, capable of outputting recognition results before the fault is fully observed.
[0094] The parameter settings are as follows:
[0095] Input data length: n is the length of the multi-dimensional pre-launch parameters, corresponding to the n dimensions of the multi-dimensional pre-launch parameters; m is the length of the fault type array, corresponding to the time window length m (units) of the multi-dimensional pre-launch parameters.
[0096] Average length of time series: 60;
[0097] Number of neurons in the hidden layer: 128.
[0098] Step 7: Input the training set data and test set data processed in Step 5 into the fault classification network constructed in Step 6. The input of a single sample is a multi-dimensional pre-fire parameter matrix including the time domain within a certain time window; the output is the fault classification result corresponding to the multi-dimensional pre-fire parameter matrix; and the trained fault classification network is obtained.
[0099] Step 8: After obtaining the multi-dimensional pre-launch parameters of a specified real-time length, preprocess and standardize the multi-dimensional pre-launch parameters according to Step 2, and then input them into the pre-trained temporal convolutional network to obtain the predicted values of the multi-dimensional pre-launch parameters. Input the predicted multi-dimensional pre-launch parameter matrix into the fault classification network described in Step 6 to output the fault type value. If the predicted value is 0, it means that no fault will occur within the predicted time period; if the predicted value is 1 / 2 / 3, it means that type 1 / 2 / 3 faults may occur within the predicted time period, and fault category information is sent for early warning.
[0100] The key innovation of this invention lies in treating the multi-dimensional pre-launch parameter feature data stream of a launch vehicle before normal launch and during malfunctions as matrix-based time series data. A temporal convolutional network is used to learn the horizontal and vertical patterns between different dimensions of the multi-dimensional time series, thereby achieving time series prediction. Based on the prediction results of the multi-dimensional parameter features, a deep learning network is used to learn the relationship between the multi-dimensional feature sequence and the type of malfunction, thus achieving the goal of early warning of launch vehicle malfunctions.
[0101] The standardization method described in this invention can be replaced by linear normalization and nonlinear normalization such as log, exponential, and tangent, all with the aim of better fitting the model and improving the gradient descent speed and model accuracy.
[0102] The early warning system based on this invention offers the following advantages:
[0103] 1. Operability. The system can provide early warnings of launch vehicle malfunctions without requiring power outages for maintenance, significantly reducing costs and lowering operational risks for maintenance personnel. The launch vehicle malfunction early warning system can operate automatically and provides easy-to-understand warning results; it is easy to operate and requires minimal specialized knowledge.
[0104] 2. Comprehensiveness. This invention comprehensively evaluates and predicts launch vehicle malfunctions by integrating characteristic data such as pressurization pressure, pre-cooling temperature, servo swing command feedback, and key voltages of the control system. Compared with single-parameter fault repair solutions, it is more comprehensive and can provide fault information in advance.
[0105] 3. Accuracy. This invention primarily employs deep learning methods, using a network to learn from the historical operational data of launch vehicles. The model is trained with a large amount of data, achieving relatively accurate prediction results during the simulation data testing phase.
[0106] The present invention has been described in detail above with reference to specific embodiments and exemplary examples; however, these descriptions should not be construed as limiting the present invention. Those skilled in the art will understand that various equivalent substitutions, modifications, or improvements can be made to the technical solutions and embodiments of the present invention without departing from the spirit and scope of the invention, and all such modifications and improvements fall within the scope of the present invention. The scope of protection of the present invention is defined by the appended claims.
[0107] The contents not described in detail in this specification are common knowledge to those skilled in the art.
Claims
1. A launch vehicle pre-flight parameter fault prediction method based on spatio-temporal attention mechanism, characterized in that, The method comprises the following steps: (1) acquiring multi-dimensional pre-launch parameters of a launch vehicle; (2) standardizing the multi-dimensional pre-launch parameters, and obtaining a training set and a test set of a time convolution network by using the processed data; (3) constructing the time convolution network for predicting the multi-dimensional pre-launch parameters; (4) inputting the training set and the test set of the time convolution network into the time convolution network for predicting the multi-dimensional pre-launch parameters to obtain a trained time convolution network; (5) delimiting a time window for the multi-dimensional pre-launch parameters acquired in step (1), each time window corresponding to a fault type label, and obtaining a training set and a test set of a fault classification network; (6) constructing the fault classification network for predicting the specific fault type corresponding to the multi-dimensional pre-launch parameters; (7) inputting the training set and the test set of the fault classification network into the fault classification network to obtain a trained fault classification network; (8) inputting the standardized multi-dimensional pre-launch parameters into the trained time convolution network to obtain a multi-dimensional pre-launch parameter prediction value, and inputting the multi-dimensional pre-launch parameter prediction value into the trained fault classification network to output a fault type.
2. The launch vehicle pre-flight parameter fault prediction method of claim 1, wherein, In step (1), the multi-dimensional pre-launch parameters include power supply voltage, flight data, and each key pressure or each key temperature.
3. The launch vehicle pre-flight parameter fault prediction method of claim 1, wherein, In step (2), the method for standardizing the multi-dimensional pre-launch parameters comprises: performing mean removal and variance normalization on each feature dimension in the multi-dimensional pre-launch parameters, so that the processed multi-dimensional pre-launch parameters conform to a standard normal distribution, i.e., the mean is 0 and the standard deviation is 1; and adopting a nearest neighbor replacement method to fill in missing data values.
4. The launch vehicle pre-flight parameter fault prediction method of claim 1, wherein, In step (3), the input layer of the time convolution network is used to input multi-dimensional pre-launch parameter samples, and the output layer outputs multi-dimensional pre-launch parameter prediction values. In step (6), the input layer of the fault classification network is used to input multi-dimensional pre-launch parameter samples and the fault type corresponding to the corresponding time window, the hidden layer is used to learn the relationship between the multi-dimensional pre-launch parameters and the fault type, and the output layer outputs a fault type prediction result.
5. The launch vehicle pre-flight parameter fault prediction method of claim 1, wherein, In step (5), when delimiting the time window for the multi-dimensional pre-launch parameters acquired in step (1), the real-time multi-dimensional pre-launch parameter time window length can be different, and the average sampling of a single time window in the training data is 60 times.
6. The launch vehicle pre-flight parameter fault prediction method of claim 1, wherein, In step (5), the fault type label is set as follows: 0 corresponds to no fault, and 1, 2 and 3 correspond to three different faults respectively.
7. The launch vehicle pre-flight parameter fault prediction method of claim 1, wherein, The parameter settings of the fault classification network are as follows: The input data length is the length n of the multi-dimensional pre-launch parameters, corresponding to the n dimensions of the multi-dimensional pre-launch parameters; the fault type array length m corresponds to the time window length m of the multi-dimensional pre-launch parameters; The average length of the time series is 60; The number of hidden layer neurons is 128.
8. The launch vehicle pre-launch parameter fault prediction method of claim 1, wherein, For the time convolution network, the input of a sample is the multi-dimensional pre-launch parameter corresponding to a time point; for the fault classification network, the input of a single sample is a multi-dimensional pre-launch parameter matrix including the time domain.
9. The launch vehicle pre-launch parameter fault prediction method of claim 1, wherein, The fault classification network takes Inception-v4 as a backbone network, and introduces an LSTM network with an attention mechanism and a graph attention network GAT. The LSTM network is used to learn the time domain correlation of time series data, and the classification accuracy is improved by adding a Squeeze-and-Excitation Block module on the full convolution block. The graph attention network GAT is used to mine the spatial correlation between variables; by fusing the features extracted from each branch, the event category is output by a softmax classifier.
10. The launch vehicle pre-launch parameter fault prediction method of claim 9, wherein, In the graph attention network GAT, each node has a connection relationship with itself by default, and the proportion of self-information of the node is increased by adding a self-loop to avoid excessive interference from adjacent nodes. In the case where the prior connection relationship between variables cannot be obtained, the default connection matrix is a full connection matrix. If x (i) With x (j) Neighboring nodes, an attention coefficient e is introduced. ij x represents (i) For x (j) The importance of, namely: e ij = a(Wx (i) ,Wx (j) ) Where a(·) is an Attention module implemented by a single-layer feedforward network, and W is a graph model weight matrix.