A launch vehicle control system fault diagnosis method, system and device based on time-frequency dual-branch graph attention reconstruction

Abstract

The application discloses a launch vehicle control system fault diagnosis method, system and device based on time-frequency double-branch graph attention reconstruction, relates to the aerospace fault diagnosis technical field, and the method comprises the steps of acquiring three-channel real-time power supply voltage time sequence data of a launch vehicle control system; inputting the three-channel real-time power supply voltage time sequence data into a voltage reconstruction model to obtain three-channel reconstructed power supply voltage time sequence data of a current period; determining the residual error of each moment based on the three-channel real-time power supply voltage time sequence data and the three-channel reconstructed power supply voltage time sequence data; determining a dynamic fault threshold based on the residual error of each moment; and determining the fault occurrence of a time sequence controller in the launch vehicle control system at each moment based on the residual error of each moment and the dynamic fault threshold. The application can improve the accuracy and efficiency of launch vehicle control system fault diagnosis.

Description

Technical Field

[0001] This application relates to the field of aerospace fault diagnosis technology, and in particular to a fault diagnosis method, system and device for launch vehicle control systems based on time-frequency bi-branch graph attention reconstruction. Background Technology

[0002] The launch vehicle's control system is the core hub ensuring the success or failure of a flight mission. The timing controller typically employs a three-machine redundancy architecture (i.e., channels 1, 2, and 3 operate simultaneously) to improve system reliability. The stability of the timing controller's power supply voltage is a key indicator of its health; abnormal voltage fluctuations, drift, or inconsistencies between channels often indicate faults such as power module aging, electromagnetic interference, or hardware failure.

[0003] Traditional launch vehicle fault diagnosis mainly relies on manually set fixed thresholds. However, during different stages of rocket refueling, ignition, and takeoff, the load current changes drastically, causing normal transient voltage fluctuations, making the fixed threshold method prone to false alarms. With the development of artificial intelligence technology, fault diagnosis methods based on deep learning have been introduced, but the relevant methods still have the following shortcomings: (1) Most current methods are based on a single time domain analysis, ignoring the frequency domain features contained in the voltage signal, making it difficult to capture the periodic fluctuations and harmonic anomalies of the voltage signal; (2) Current diagnostic models usually treat the three redundant channels as independent inputs, failing to explicitly model the inherent topological correlation of the three-engine redundant system, and failing to make full use of the consistency information between channels to identify minor faults; (3) The voltage data distribution of launch vehicles varies in different flight stages, and the relevant methods lack an adaptive adjustment mechanism, making it difficult to adapt to stable diagnosis under multiple operating conditions.

[0004] Therefore, there is an urgent need for a fault diagnosis method for launch vehicle control systems that can integrate the time-domain and frequency-domain characteristics of voltage signals, explicitly model the topological relationships between redundant channels of the three machines, and have the ability to adapt to operating conditions. Summary of the Invention

[0005] The purpose of this application is to provide a method, system, and device for fault diagnosis of launch vehicle control systems based on time-frequency bi-branch graph attention reconstruction, so as to improve the accuracy and efficiency of fault diagnosis of launch vehicle control systems.

[0006] To achieve the above objectives, this application provides the following solution.

[0007] Firstly, this application provides a fault diagnosis method for a launch vehicle control system based on time-frequency bi-branch graph attention reconstruction, including: The real-time power supply voltage timing data of the three channels of the launch vehicle control system is obtained; the architecture of the launch vehicle control system is a three-machine redundant channel; the real-time power supply voltage timing data of the three channels includes the power supply voltage data of the three channels at multiple consecutive moments in the current time period; The real-time power supply voltage time series data of the three channels is input into the voltage reconstruction model to obtain the reconstructed power supply voltage time series data of the three channels for the current time period. The voltage reconstruction model is obtained by training a time-frequency dual-branch graph attention reconstruction model, which includes a dual-stream encoder, a dynamic topology graph learner, and a feature reconstruction decoder. The dual-stream encoder is built based on a multi-scale time-frequency fusion module, and each of the multi-scale time-frequency fusion modules is built based on a time-domain convolution unit, a frequency-domain spectrum extraction unit, and a time-frequency interactive attention unit. The dynamic topology graph learner is used to construct the dynamic dependency relationship between the three redundant channels. The feature reconstruction decoder is built based on a gated residual reconstruction module, and each of the gated residual reconstruction modules is built based on a channel adaptive weighting subunit and a multilayer perceptron reconstruction subunit. Based on the real-time power supply voltage timing data of the three channels in the current time period and the reconstructed power supply voltage timing data of the three channels in the current time period, the residuals at each moment in the current time period are determined. Based on the residuals at each time point in the current time period, determine the dynamic fault threshold for the current time period; Based on the residuals at each moment in the current time period and the dynamic fault threshold of the current time period, the fault occurrence status of the timing controller in the launch vehicle control system at each moment is determined; the fault occurrence status is either a fault occurred or no fault occurred.

[0008] In one embodiment, the training process of the voltage reconstruction model includes: Obtain a training sample set; the training sample set includes the three-channel historical power supply voltage timing data of the launch vehicle control system under normal operating conditions for historical periods; Construct a time-frequency bi-branch graph attention reconstruction model; Using the training sample set as input, and with the goal of minimizing the reconstruction error between the three-channel reconstructed power supply voltage time series data of the historical period output by the time-frequency bi-branch graph attention reconstruction model and the training sample set, the time-frequency bi-branch graph attention reconstruction model is trained unsupervised to obtain the voltage reconstruction model.

[0009] In one embodiment, the dual-stream encoder comprises M cascaded multi-scale time-frequency fusion modules, each of which includes a temporal convolution unit, a frequency domain spectrum extraction unit, a time-frequency interactive attention unit, and a first adder; the temporal convolution unit includes three parallel dilated convolutional layers, and the frequency domain spectrum extraction unit includes a fast Fourier transform subunit, a first multilayer perceptron, and an inverse fast Fourier transform subunit connected in sequence; The dynamic topology graph learner includes a learnable node embedding layer, a similarity calculation layer, and a graph convolutional network layer connected in sequence. The feature reconstruction decoder includes M-1 matrix concatenation operation layers and M cascaded decoding modules; the M cascaded decoding modules include one first decoding module and M-1 second decoding modules; the first decoding module includes a first sampling layer, a second sampling layer, and a gated residual reconstruction module connected in sequence, and the second decoding module includes a third sampling layer, a second sampling layer, and a gated residual reconstruction module connected in sequence; the first sampling layer is an upsampling mapping layer with a time dimension step size of 2, the second sampling layer is a mapping layer with a time dimension step size of 1, and the third sampling layer is an upsampling mapping layer with a time dimension step size of 2; the gated residual reconstruction module includes a channel adaptive weighting subunit, a multilayer perceptron reconstruction subunit, a sigmoid activation function, a first multiplier, and a second adder; the channel adaptive weighting subunit includes a global average pooling layer, a second multilayer perceptron, and a second multiplier; the multilayer perceptron reconstruction subunit includes a first linear layer and a second linear layer; The first adder in the first multi-scale time-frequency fusion module is connected to the time-domain convolution unit, the frequency-domain spectrum extraction unit, and the first sampling layer in the second multi-scale time-frequency fusion module, respectively. The first adder in the m1-th multi-scale time-frequency fusion module is connected to the time-domain convolution unit and the frequency-domain spectrum extraction unit in the m1+1-th multi-scale time-frequency fusion module, respectively; m1=2,…,M-1, M≥4. The first adder in the m2-th multi-scale time-frequency fusion module and the third sampling layer in the m2-th second decoding module are both connected to the m2-th matrix concatenation operation layer. The m2-th matrix concatenation operation layer is connected to the second sampling layer in the m2-th second decoding module; m2=1,…,M-1. The first adder of the Mth multi-scale time-frequency fusion module is connected to the learnable node embedding layer and the graph convolutional network layer, respectively. The graph convolutional network layer is connected to the third sampling layer in the (M-1)th second decoding module. The second adder in the m3th second decoding module is connected to the third sampling layer in the m3-1th second decoding module; the second adder in the 1st second decoding module is connected to the first sampling layer; m3 = M-1, ..., 2.

[0010] In one embodiment, the real-time power supply voltage timing data of the three channels is input into the voltage reconstruction model to obtain the three-channel reconstructed power supply voltage timing data for the current time period, including: The real-time power supply voltage timing data of the three channels is input into the dual-stream encoder in the voltage reconstruction model to obtain the three-channel time-frequency fusion feature map at M scales; The three-channel time-frequency fusion feature map at the Mth scale is input into the dynamic topology graph learner in the voltage reconstruction model to obtain the graph update features; The graph update features and the three-channel time-frequency fusion feature maps at the first M-1 scales are both input into the feature reconstruction decoder in the voltage reconstruction model to obtain the three-channel reconstructed power supply voltage time series data for the current time period.

[0011] In one embodiment, the three-channel real-time power supply voltage timing data is input into a dual-stream encoder in the voltage reconstruction model to obtain a three-channel time-frequency fusion feature map at M scales, including: The real-time power supply voltage timing data of the three channels is input to the first multi-scale time-frequency fusion module, and then processed by M multi-scale time-frequency fusion modules for feature extraction to obtain the three-channel time-frequency fusion feature map at M scales; wherein, the feature extraction process of the m-th multi-scale time-frequency fusion module in the voltage reconstruction model includes: The temporal convolution unit of the m-th multi-scale time-frequency fusion module performs multi-path parallel dilated convolution on the three-channel time-frequency fusion feature map at the (m-1)-th scale to obtain the temporal waveform features under the m-th multi-receptive field; when m=1, the three-channel time-frequency fusion feature map at the (m-1)-th scale is the real-time power supply voltage timing data of the three channels; m=1, 2, ..., M; The frequency domain spectrum extraction unit of the m-th multi-scale time-frequency fusion module performs fast Fourier transform, multilayer perceptron mapping and inverse fast Fourier transform on the three-channel time-frequency fusion feature map at the (m-1)-th scale in sequence to obtain the m-th global frequency domain distribution feature. The time-frequency interaction attention unit of the m-th multi-scale time-frequency fusion module performs cross-domain attention calculation on the m-th time-domain waveform feature and the m-th global frequency-domain distribution feature to obtain the m-th time-frequency cross feature; The first adder of the m-th multi-scale time-frequency fusion module adds the m-th time-frequency cross feature and the three-channel time-frequency fusion feature map at the (m-1)-th scale to obtain the three-channel time-frequency fusion feature map at the m-th scale.

[0012] In one embodiment, the three-channel time-frequency fusion feature map at the Mth scale is input into the dynamic topology graph learner in the voltage reconstruction model to obtain graph update features, including: The three-channel time-frequency fusion feature map at the Mth scale is input into the learnable node embedding layer in the voltage reconstruction model to obtain the node embedding vectors corresponding to the three channels. The node embedding vectors corresponding to each channel are input into the similarity calculation layer in the voltage reconstruction model to obtain the similarity calculation results. The similarity calculation results are then activated to obtain the dynamic adjacency matrix. The similarity calculation results include the semantic similarity between each pair of channels. The three-channel time-frequency fusion feature map at the Mth scale and the dynamic adjacency matrix are input into the graph convolutional network layer in the voltage reconstruction model to perform graph convolution operation and obtain the graph update feature.

[0013] In one embodiment, the graph update features and the three-channel time-frequency fusion feature maps at the first M-1 scales are both input into the feature reconstruction decoder in the voltage reconstruction model to obtain the three-channel reconstructed power supply voltage time series data for the current time period, including: The updated graph features are input to the (M-1)th second decoding module. Combined with the three-channel time-frequency fusion feature maps at the previous (M-1) scales, the features are sequentially processed through the (M-1)th second decoding modules and one first decoding module in the feature reconstruction decoder to obtain the three-channel reconstructed power supply voltage time series data for the current time period. The feature reconstruction process of the r-th second decoding module in the voltage reconstruction model includes: The third sampling layer of the r-th second decoding module upsamples the (r+1)-th decoding feature map to obtain the r-th sampled feature map; r = 1, 2, ..., M-1; when r = M-1, the (r+1)-th decoding feature map is the updated feature of the graph; The matrix splicing operation layer of the r-th second decoding module splices the r-th sampled feature map and the three-channel time-frequency fusion feature map at the r-th scale to obtain the r-th spliced ​​map; The second sampling layer of the r-th second decoding module performs feature extraction and channel dimension compression on the r-th stitched image in sequence to obtain the r-th integrated feature map; The gated residual reconstruction module of the r-th second decoding module reconstructs the r-th integrated feature map to obtain the r-th decoded feature map; The first sampling layer of the first decoding module upsamples the first decoded feature map to obtain the first sampled feature map; The second sampling layer of the first decoding module performs feature extraction and channel dimension compression to obtain the second sampling feature map; The gated residual reconstruction module of the first decoding module reconstructs the second sampling feature map to obtain the three-channel reconstructed power supply voltage timing data for the current time period.

[0014] In one embodiment, determining the dynamic fault threshold for the current time period based on the residuals at each time point within the current time period includes: Calculate the mean and standard deviation of the residuals for the current time period based on the residuals at each time point in the current time period; Based on the mean and standard deviation of the residuals for the current time period, the dynamic fault threshold for the current time period is determined using the dynamic fault threshold setting formula. The dynamic fault threshold setting formula is as follows: ; in, Dynamic fault threshold; The mean; The preset sensitivity coefficient; The standard deviation is denoted as .

[0015] Secondly, this application provides a fault diagnosis system for a launch vehicle control system based on time-frequency bi-branch graph attention reconstruction, to realize the aforementioned fault diagnosis method for a launch vehicle control system based on time-frequency bi-branch graph attention reconstruction. The launch vehicle control system fault diagnosis system includes: The data acquisition module is used to acquire the real-time power supply voltage timing data of the three channels of the launch vehicle control system; the architecture of the launch vehicle control system is a three-machine redundant channel; the real-time power supply voltage timing data of the three channels includes the power supply voltage data of the three channels at multiple consecutive moments in the current time period; The data reconstruction module is used to input the real-time power supply voltage time-series data of the three channels into the voltage reconstruction model to obtain the reconstructed power supply voltage time-series data of the three channels for the current time period. The voltage reconstruction model is obtained by training a time-frequency dual-branch graph attention reconstruction model, which includes a dual-stream encoder, a dynamic topology graph learner, and a feature reconstruction decoder. The dual-stream encoder is built based on a multi-scale time-frequency fusion module, and each of the multi-scale time-frequency fusion modules is built based on a time-domain convolutional unit, a frequency-domain spectrum extraction unit, and a time-frequency interactive attention unit. The dynamic topology graph learner is used to construct the dynamic dependency relationship between the three redundant channels. The feature reconstruction decoder is built based on a gated residual reconstruction module, and each of the gated residual reconstruction modules is built based on a channel adaptive weighting subunit and a multilayer perceptron reconstruction subunit. The residual calculation module is used to determine the residuals at each moment in the current time period based on the real-time power supply voltage timing data of the three channels in the current time period and the reconstructed power supply voltage timing data of the three channels in the current time period. The threshold determination module is used to determine the dynamic fault threshold for the current time period based on the residuals at each time point in the current time period. The fault diagnosis module is used to determine the fault occurrence status of the timing controller in the launch vehicle control system at each moment based on the residuals at each moment in the current time period and the dynamic fault threshold of the current time period; the fault occurrence status is either a fault has occurred or no fault has occurred.

[0016] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described fault diagnosis method for a launch vehicle control system based on time-frequency bi-branch graph attention reconstruction.

[0017] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application discloses a fault diagnosis method, system, and device for a launch vehicle control system based on time-frequency bibranch graph attention reconstruction. It acquires real-time power supply voltage time-series data of the three channels of the launch vehicle control system and inputs the data into a voltage reconstruction model trained on a time-frequency bibranch graph attention reconstruction model for processing. The dual-stream encoder in the time-frequency bibranch graph attention reconstruction model includes multiple multi-scale time-frequency fusion modules. Each module includes a time-domain convolution unit, a frequency-domain spectrum extraction unit, and a time-frequency interactive attention unit, which can establish complementary feature representations of the voltage signal between time-domain trends and frequency-domain noise, avoiding information omissions from a single perspective. Based on this, the dual-stream encoder can effectively improve feature extraction performance and enhance the ability to capture small signal fluctuations. The feature reconstruction decoder in the time-frequency bibranch graph attention reconstruction model is combined with a dynamic topology graph learner. The dynamic topology graph learner models the dynamic correlation between the three channels, and the feature reconstruction decoder uses a gated residual reconstruction module. The gated mechanism adaptively controls feature flow, effectively avoiding the gradient vanishing problem in long sequence reconstruction. Simultaneously, through adaptive channel weighting, no additional noise interference is introduced. Based on this, the feature reconstruction decoder can accurately reconstruct the voltage waveform under normal operating conditions, thereby identifying faults through the reconstruction residuals, and thus improving the accuracy and efficiency of fault diagnosis in the launch vehicle control system. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A schematic flowchart of a fault diagnosis method for a launch vehicle control system based on time-frequency bi-branch graph attention reconstruction provided in an embodiment of this application; Figure 2 A schematic diagram of the time-frequency bi-branch graph attention reconstruction model when M=4; Figure 3 This is a schematic diagram of the multi-scale time-frequency fusion module structure; Figure 4 This is a schematic diagram of the gated residual reconstruction module. Figure 5 This is a schematic diagram of the dynamic topology graph learner structure; Figure 6 This is a schematic diagram of the voltage reconstruction effect under normal operating conditions; Figure 7 This is a schematic diagram of the fault detection results; Figure 8 This is an application environment diagram of a fault diagnosis method for a launch vehicle control system based on time-frequency bi-branch graph attention reconstruction in one embodiment of this application; Figure 9 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] The purpose of this application is to provide a fault diagnosis method, system and device for launch vehicle control systems based on time-frequency bi-branch graph attention reconstruction, which aims to solve the problems of difficulty in diagnosing minor voltage faults and high false alarm rate in launch vehicle control systems under complex and variable operating conditions with missing fault samples.

[0022] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] In one exemplary embodiment, such as Figure 1 As shown, a fault diagnosis method for a launch vehicle control system based on time-frequency bi-branch graph attention reconstruction is provided. This method is executed by computer equipment, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to a server as an example, and includes the following steps.

[0024] Step 1: Obtain the real-time power supply voltage timing data of the three channels of the launch vehicle control system.

[0025] The launch vehicle control system has a three-machine redundant channel architecture; the real-time power supply voltage timing data of the three channels includes the power supply voltage data of the three channels at multiple consecutive moments in the current period.

[0026] Step 2: Input the real-time power supply voltage timing data of the three channels into the voltage reconstruction model to obtain the reconstructed power supply voltage timing data of the three channels for the current time period.

[0027] The voltage reconstruction model is obtained by training the time-frequency dual-branch graph attention reconstruction model, which includes a two-stream encoder, a dynamic topology graph learner, and a feature reconstruction decoder. The two-stream encoder is built based on a multi-scale time-frequency fusion module, and each multi-scale time-frequency fusion module is built based on a time-domain convolutional unit, a frequency-domain spectrum extraction unit, and a time-frequency interactive attention unit. The dynamic topology graph learner is used to construct the dynamic dependency relationship between the three redundant channels. The feature reconstruction decoder is built based on a gated residual reconstruction module, and each gated residual reconstruction module is built based on a channel adaptive weighting sub-unit and a multilayer perceptron reconstruction sub-unit.

[0028] As an optional implementation, step 2, the training process of the voltage reconstruction model, includes: Obtain the training sample set; the training sample set includes the three-channel historical power supply voltage timing data of the launch vehicle control system under normal operating conditions for historical periods; Construct a time-frequency bi-branch graph attention reconstruction model; Using the training sample set as input, and aiming to minimize the reconstruction error between the three-channel reconstructed power supply voltage time series data of the historical period output by the time-frequency bi-branch graph attention reconstruction model and the training sample set, the time-frequency bi-branch graph attention reconstruction model is trained unsupervised to obtain the voltage reconstruction model.

[0029] Specifically, the training framework for the time-frequency dual-branch graph attention reconstruction model is PyTorch. During the forward propagation phase, the input three-channel historical power supply voltage time-series data is downsampled through multiple multi-scale time-frequency fusion modules in the dual-stream encoder to extract multi-scale high-level time-frequency features, generating three-channel time-frequency fusion feature maps at M scales. Subsequently, the extracted deep features (i.e., the three-channel time-frequency fusion feature map at the Mth scale) are used by a dynamic topology graph learner to model inter-channel dependencies, generating graph update features that incorporate spatial redundancy information. Next, the feature reconstruction decoder recovers the temporal details and spatial resolution of the features through multiple upsampling operations of the second and first decoding modules and skip connections with the multi-scale time-frequency fusion modules, generating high-resolution reconstructed feature maps. Finally, the model outputs three-channel reconstructed power supply voltage time-series data for historical periods.

[0030] During the backpropagation phase, the gradient of each layer's parameters is calculated using the error backpropagation algorithm. The optimizer Adam updates the network weights based on the gradient information, with an initial learning rate of 10. -3The learning rate strategy is linear, the batch size is 32, the number of training epochs is 120, and the random kill rate is 0.1. Since this application employs an unsupervised reconstruction mechanism, the voltage reconstruction model uses the mean squared error loss function to calculate the loss between the reconstructed sequence and the original sequence. Furthermore, the voltage reconstruction model uses a Batch Norm layer to accelerate training and improve generalization ability. The mean squared error loss function is: ; in, This represents the mean squared error loss function value; The total number of historical power supply voltage data for the three channels in the three-channel historical power supply voltage timing data; This represents the loss value of the i-th historical power supply voltage data point in the three-channel historical power supply voltage time series data for a historical period. Specifically, this is reflected in the squared Euclidean distance between the reconstructed value and the true value, i.e. , For the i-th three-channel historical power supply voltage data, This is the i-th three-channel reconstructed power supply voltage timing data for a historical period.

[0031] During training, to evaluate model performance and prevent overfitting, the fault detection rate (Fault Detection Rate) is used. ) and false alarm rate ) as evaluation indicators. Among them, It is used to measure the proportion of all actual fault samples that are correctly diagnosed. Used to calculate the proportion of normal operating condition samples that are incorrectly identified as faults, measuring the robustness of the model's diagnostics: ; ; in, A true positive (the number of faults that were correctly detected); False negatives (the number of unreported faults); False positives (inflated numbers of normal samples); True negative (the number of correctly identified normal samples).

[0032] Finally, training is terminated when the mean squared error loss function value stops decreasing or the maximum number of training iterations is reached. The voltage reconstruction model is then loaded with optimal training weights and evaluated on the test set to verify its diagnostic accuracy for voltage waveforms at different flight stages of the launch vehicle, thereby ensuring its generalization ability.

[0033] Specifically, during the experimental phase, this application used the voltage telemetry sample set of the launch vehicle's three-channel timing controller to train and validate the voltage reconstruction model. During unsupervised training, the model's MSE on the training set eventually stabilized at around 0.0045, indicating that the model has fully grasped the voltage evolution law under normal operating conditions.

[0034] During testing, the voltage reconstruction model performed excellently in the scenario of monitoring the power supply voltage of the timing controller. It reached 99.2%, while The error rate is only 0.8%. Experimental data demonstrates that this model has extremely high sensitivity to nonlinear characteristics in voltage data. Furthermore, Figure 6 The results demonstrate the voltage reconstruction effect under normal operating conditions. The reconstructed voltage sequence closely matches the original voltage sequence, and even in the presence of random disturbances, the two still maintain a very high degree of consistency, verifying the model's strong representation ability for normal operating modes. Figure 7 The real-time dynamic process of fault determination is demonstrated, with the model successfully detecting the injected step-off fault between steps 100 and 130. In the comparison graph, it can be seen that the original voltage sequence shifted significantly upward within the fault region, while the reconstructed voltage sequence, constrained by the normal evolution pattern learned by the model, remained near the normal baseline, thus creating a prediction bias. In the residual graph, the corresponding reconstruction residual rapidly increases at the fault initiation moment and significantly exceeds the preset red adaptive dynamic threshold line. Experimental results show that the voltage reconstruction model proposed in this application has extremely high sensitivity to small offset faults, enabling not only rapid fault determination but also effective avoidance of environmental noise interference through residual analysis, exhibiting superior diagnostic accuracy and effectively supporting the health management tasks of critical spacecraft systems.

[0035] As an optional implementation method, such as Figures 2-5 As shown, the dual-stream encoder is constructed by cascading M multi-scale time-frequency fusion modules. Each multi-scale time-frequency fusion module includes a time-domain convolution unit, a frequency-domain spectrum extraction unit, a time-frequency interactive attention unit, and a first adder. The time-domain convolution unit includes three parallel dilated convolutional layers, and the frequency-domain spectrum extraction unit includes a fast Fourier transform subunit, a first multilayer perceptron, and an inverse fast Fourier transform subunit connected in sequence. The dynamic topology graph learner consists of a learnable node embedding layer, a similarity calculation layer, and a graph convolutional network layer connected in sequence. The feature reconstruction decoder includes M-1 matrix concatenation operation layers and M cascaded decoding modules; the M cascaded decoding modules include one first decoding module and M-1 second decoding modules; the first decoding module includes a first sampling layer, a second sampling layer, and a gated residual reconstruction module connected in sequence, and the second decoding module includes a third sampling layer, a second sampling layer, and a gated residual reconstruction module connected in sequence; the first sampling layer is an upsampling mapping layer with a time dimension step of 2, the second sampling layer is a mapping layer with a time dimension step of 1, and the third sampling layer is an upsampling mapping layer with a time dimension step of 2; the gated residual reconstruction module includes a channel adaptive weighting subunit, a multilayer perceptron reconstruction subunit, a sigmoid activation function, a first multiplier, and a second adder; the channel adaptive weighting subunit includes a global average pooling layer, a second multilayer perceptron, and a second multiplier; the multilayer perceptron reconstruction subunit includes a first linear layer and a second linear layer; The first adder in the first multi-scale time-frequency fusion module is connected to the time-domain convolution unit, the frequency-domain spectrum extraction unit, and the first sampling layer in the second multi-scale time-frequency fusion module, respectively. The first adder in the m1-th multi-scale time-frequency fusion module is connected to the time-domain convolution unit and the frequency-domain spectrum extraction unit in the m1+1-th multi-scale time-frequency fusion module, respectively; m1=2,…,M-1, M≥4. The first adder in the m2-th multi-scale time-frequency fusion module and the third sampling layer in the m2-th second decoding module are both connected to the m2-th matrix concatenation operation layer. The m2-th matrix concatenation operation layer is connected to the second sampling layer in the m2-th second decoding module; m2=1,…,M-1. The first adder of the Mth multi-scale time-frequency fusion module is connected to the learnable node embedding layer and the graph convolutional network layer, respectively. The graph convolutional network layer is connected to the third sampling layer in the (M-1)th second decoding module. The second adder in the m3th second decoding module is connected to the third sampling layer in the m3-1th second decoding module; the second adder in the 1st second decoding module is connected to the first sampling layer; m3 = M-1, ..., 2.

[0036] Specifically, the expansion rates of the three hollow convolutional layers are 1, 2, and 4, respectively.

[0037] As an optional implementation, step 2 includes the following steps.

[0038] Step 21: Input the three-channel real-time power supply voltage timing data into the dual-stream encoder in the voltage reconstruction model to obtain the three-channel time-frequency fusion feature map at M scales.

[0039] As an optional implementation, step 21 includes the following steps.

[0040] Step 211: Input the real-time power supply voltage timing data of the three channels into the first multi-scale time-frequency fusion module, and then extract features through M multi-scale time-frequency fusion modules in sequence to obtain the three-channel time-frequency fusion feature map at M scales.

[0041] The feature extraction process of the m-th multi-scale time-frequency fusion module in the voltage reconstruction model includes the following steps.

[0042] Step 2111: The temporal convolution unit of the m-th multi-scale time-frequency fusion module performs multi-path parallel dilated convolution on the three-channel time-frequency fusion feature map at the (m-1)-th scale to obtain the temporal waveform features under the m-th multi-receptive field; when m=1, the three-channel time-frequency fusion feature map at the (m-1)-th scale is the three-channel real-time power supply voltage timing data; m=1, 2, ..., M.

[0043] Step 2112: The frequency domain spectrum extraction unit of the m-th multi-scale time-frequency fusion module performs fast Fourier transform, multilayer perceptron mapping and inverse fast Fourier transform on the three-channel time-frequency fusion feature map at the (m-1)-th scale in sequence to obtain the m-th global frequency domain distribution feature.

[0044] Step 2113: The time-frequency interaction attention unit of the m-th multi-scale time-frequency fusion module performs cross-domain attention calculation on the m-th time-domain waveform feature and the m-th global frequency domain distribution feature to obtain the m-th time-frequency cross feature.

[0045] Specifically, the time-frequency interactive attention unit uses time-domain waveform features as query vectors and global frequency-domain distribution features as key and value vectors. It uses a multiplier and a Softmax function to calculate the dot product similarity between the query vector and the key vector, generates a cross-domain attention map, and uses the cross-domain attention map to perform a weighted summation of the value vectors to obtain the time-frequency interactive features.

[0046] Step 2114: The first adder of the m-th multi-scale time-frequency fusion module adds the m-th time-frequency cross feature and the three-channel time-frequency fusion feature map at the (m-1)-th scale to obtain the three-channel time-frequency fusion feature map at the m-th scale.

[0047] Specifically, M multi-scale time-frequency fusion modules are set up. As the number of layers increases, the time dimension of the three-channel time-frequency fusion feature map is halved layer by layer, and the number of channels is doubled layer by layer, thereby effectively capturing the multi-scale spatiotemporal evolution law of voltage data.

[0048] Step 22: Input the three-channel time-frequency fusion feature map at the Mth scale into the dynamic topology graph learner in the voltage reconstruction model to obtain the graph update features.

[0049] As an optional implementation, step 22 includes the following steps.

[0050] Step 221: Input the three-channel time-frequency fusion feature map at the Mth scale into the learnable node embedding layer in the voltage reconstruction model to obtain the node embedding vectors corresponding to the three channels.

[0051] Specifically, the size of the three-channel time-frequency fusion feature map at the Mth scale is [C, 3, L], where C represents the number of feature channels, 3 represents the number of three processing channels (i.e., three channels) of the three-machine redundant system, and L represents the sequence length of the real-time power supply voltage timing data of the three channels.

[0052] Step 222: Input the node embedding vectors corresponding to each channel into the similarity calculation layer in the voltage reconstruction model to obtain the similarity calculation results, and perform activation processing on the similarity calculation results to obtain the dynamic adjacency matrix; the similarity calculation results include: the semantic similarity between pairs of channels.

[0053] Specifically, a dot product operation is performed on the node embedding vectors corresponding to any two channels to obtain the semantic similarity between the two channels. The sigmoid activation function is then used to activate the similarity calculation results, generating a dynamic adjacency matrix that reflects the strength of real-time topological associations.

[0054] Step 223: Input the three-channel time-frequency fusion feature map and dynamic adjacency matrix at the Mth scale into the graph convolutional network layer in the voltage reconstruction model, perform graph convolution operation, and obtain graph update features.

[0055] Specifically, the three-channel time-frequency fusion feature map at the Mth scale, the dynamic adjacency matrix, and the final value of the trainable weight model at the end of training are input into the graph convolutional network layer to perform graph convolution operations, realize information interaction and fusion between channels, and obtain graph updated features.

[0056] During the operation of the voltage reconstruction model, the dynamic adjacency matrix adaptively learns the dependencies between the three redundant channels. Compared to traditional fixed-weight feature aggregation, this application establishes a data-driven dynamic interaction mechanism through a dynamic topology graph learner. When the voltage of a certain channel drifts or becomes abnormal, the corresponding weights in the dynamic adjacency matrix will be adaptively adjusted, enabling the output graph update features to capture consistency deviations between channels, thereby improving the accuracy of identifying minor faults in the launch vehicle control system.

[0057] Step 23: Input the graph update features and the three-channel time-frequency fusion feature maps of the first M-1 scales into the feature reconstruction decoder in the voltage reconstruction model to obtain the three-channel reconstructed power supply voltage time series data for the current time period.

[0058] As an optional implementation, step 23 includes the following steps.

[0059] Step 231: Input the updated graph features into the (M-1)th second decoding module, combine the three-channel time-frequency fusion feature maps of the previous M-1 scales, and sequentially pass through the M-1 second decoding modules and 1 first decoding module in the feature reconstruction decoder to perform feature reconstruction, and obtain the three-channel reconstructed power supply voltage timing data for the current time period.

[0060] The feature reconstruction process of the r-th second decoding module in the voltage reconstruction model includes: Step 2311: The third sampling layer of the r-th second decoding module upsamples the (r+1)-th decoding feature map to obtain the r-th sampled feature map; r=1, 2, ..., M-1; when r=M-1, the (r+1)-th decoding feature map is the graph update feature.

[0061] Specifically, the time length of the sampled feature map is increased by an upsampling mapping with a step size of 2, thereby matching the scale of the three-channel time-frequency fusion feature map.

[0062] Step 2312: The matrix splicing operation layer of the r-th second decoding module splices the r-th sampled feature map and the three-channel time-frequency fusion feature map at the r-th scale to obtain the r-th spliced ​​map.

[0063] Specifically, the sampled feature map is spliced ​​with the three-channel time-frequency fusion feature map to obtain a spliced ​​image, thereby realizing the fusion of the underlying graph semantics and the shallow fine-grained temporal features of the encoder.

[0064] Step 2313: The second sampling layer of the r-th second decoding module performs feature extraction and channel dimension compression on the r-th stitched image in sequence to obtain the r-th integrated feature map.

[0065] Specifically, the second sampling layer sequentially extracts features and compresses channel dimensions on the stitched image to meet the input requirements of the gated residual reconstruction module.

[0066] Step 2314: The gated residual reconstruction module of the r-th second decoding module reconstructs the r-th integrated feature map to obtain the r-th decoded feature map.

[0067] Specifically, the integrated feature map is input into the gated residual reconstruction module for feature reconstruction and nonlinear transformation to obtain the decoded feature map.

[0068] Specifically, the reconstruction process of the original input feature map by any gated residual reconstruction module includes: The original input feature map is fed into the Sigmoid activation function for feature transformation to obtain the gate weight signal; The original input feature map is input into the channel adaptive weighting subunit for channel importance recalibration to obtain the weighted calibration feature map; The weighted calibration feature map is input into the multilayer perceptron reconstruction subunit, and linear feature extraction is performed using the first and second linear layers to obtain the preliminary reconstructed feature map. The preliminary reconstructed feature map and the gating weight signal are input into the first multiplier for weighted fusion processing to obtain the gating reconstructed feature map; The gated reconstructed feature map and the original input feature map are fed into the second adder for addition to obtain the decoded feature map.

[0069] Step 2315: The first sampling layer of the first decoding module upsamples the first decoded feature map to obtain the first sampled feature map.

[0070] Step 2316: The second sampling layer of the first decoding module performs feature extraction and channel dimension compression to obtain the second sampling feature map.

[0071] Step 2317: The gated residual reconstruction module of the first decoding module reconstructs the second sampling feature map to obtain the three-channel reconstructed power supply voltage timing data for the current time period.

[0072] Step 3: Based on the real-time power supply voltage timing data of the three channels in the current time period and the reconstructed power supply voltage timing data of the three channels in the current time period, determine the residual at each moment in the current time period.

[0073] Specifically, when M=4, the size of the three-channel time-frequency fusion feature map at the Mth scale is [512, 3, L / 16], where 512 is the number of feature channels in the three-channel time-frequency fusion feature map at the Mth scale, 3 is the number of the three processing channels (i.e., three channels) in the three-machine redundant system, the size of the (M-1)th sampling feature map is [256, 3, L / 8], the size of the (M-1)th stitched map is [512, 3, L / 8], the size of the (M-1)th integrated feature map is [256, 3, L / 8], and the size of the (M-1)th decoded feature map is [128, 3, L / 8]. After the cascading reconstruction of the first M-2 second decoding modules and the first decoding module, the final output size of the three-channel reconstructed power supply voltage timing data is [3, L].

[0074] Step 4: Determine the dynamic fault threshold for the current time period based on the residuals at each time point in the current time period.

[0075] As an optional implementation, step 4 includes the following steps.

[0076] Step 41: Calculate the mean and standard deviation of the residuals for the current time period based on the residuals at each time point in the current time period.

[0077] Step 42: Based on the mean and standard deviation of the residuals for the current time period, determine the dynamic fault threshold for the current time period using the dynamic fault threshold setting formula; the dynamic fault threshold setting formula is: ; in, Dynamic fault threshold; The mean; The preset sensitivity coefficient; The standard deviation is denoted as .

[0078] Step 5: Based on the residuals at each moment in the current time period and the dynamic fault threshold of the current time period, determine the fault occurrence status of the timing controller in the launch vehicle control system at each moment; the fault occurrence status is either a fault occurred or no fault occurred.

[0079] The fault diagnosis method for launch vehicle control systems based on time-frequency bi-branch graph attention reconstruction provided in this application can be applied to, for example... Figure 8 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be set up independently, integrated into server 104, or placed in the cloud or on other servers. Terminal 102 can send the three-channel real-time power supply voltage timing data of the launch vehicle control system to server 104. After receiving the three-channel real-time power supply voltage timing data, server 104 inputs the three-channel real-time power supply voltage timing data into the voltage reconstruction model to obtain the three-channel reconstructed power supply voltage timing data for the current time period. Based on the three-channel real-time power supply voltage timing data and the three-channel reconstructed power supply voltage timing data for the current time period, the residuals at each moment in the current time period are determined. Based on the residuals at each moment in the current time period, the dynamic fault threshold for the current time period is determined. Based on the residuals at each moment in the current time period and the dynamic fault threshold for the current time period, the fault occurrence status of the timing controller in the launch vehicle control system at each moment is determined. Server 104 can feed back the obtained fault occurrence status to terminal 102. Furthermore, in some embodiments, the fault diagnosis method for the launch vehicle control system based on time-frequency dual-branch graph attention reconstruction can also be implemented separately by the server 104 or the terminal 102. For example, the terminal 102 can directly perform fault diagnosis of the launch vehicle control system based on the three-channel real-time power supply voltage timing data, or the server 104 can obtain the three-channel real-time power supply voltage timing data from the data storage system and perform fault diagnosis of the launch vehicle control system based on the three-channel real-time power supply voltage timing data.

[0080] The terminal 102 can be, but is not limited to, various measurement and control desktop computers, high-performance workstations, ruggedized laptops, portable measurement and control terminals, etc. The terminal 102 can be connected to monitoring instruments and telemetry front-end equipment. The server 104 can be implemented using a standalone server or a server cluster composed of multiple servers, or it can be a cloud server.

[0081] In an exemplary embodiment, a fault diagnosis system for a launch vehicle control system based on time-frequency bi-branch graph attention reconstruction is provided to implement the aforementioned fault diagnosis method for a launch vehicle control system based on time-frequency bi-branch graph attention reconstruction. The launch vehicle control system fault diagnosis system includes: The data acquisition module is used to acquire the real-time power supply voltage timing data of the three channels of the launch vehicle control system. The architecture of the launch vehicle control system is a three-machine redundant channel. The real-time power supply voltage timing data of the three channels includes the power supply voltage data of the three channels at multiple consecutive moments in the current time period. The data reconstruction module is used to input the real-time power supply voltage time-series data of the three channels into the voltage reconstruction model to obtain the reconstructed power supply voltage time-series data of the three channels for the current period. The voltage reconstruction model is obtained by training the time-frequency dual-branch graph attention reconstruction model, which includes a dual-stream encoder, a dynamic topology graph learner, and a feature reconstruction decoder. The dual-stream encoder is built based on a multi-scale time-frequency fusion module, and each multi-scale time-frequency fusion module is built based on a time-domain convolutional unit, a frequency-domain spectrum extraction unit, and a time-frequency interactive attention unit. The dynamic topology graph learner is used to construct the dynamic dependency relationship between the three redundant channels. The feature reconstruction decoder is built based on a gated residual reconstruction module, and each gated residual reconstruction module is built based on a channel adaptive weighting subunit and a multilayer perceptron reconstruction subunit. The residual calculation module is used to determine the residuals at each moment in the current time period based on the real-time power supply voltage timing data of the three channels in the current time period and the reconstructed power supply voltage timing data of the three channels in the current time period. The threshold determination module is used to determine the dynamic fault threshold for the current time period based on the residuals at each time point in the current time period. The fault diagnosis module is used to determine the fault occurrence status of the timing controller in the launch vehicle control system at each moment based on the residuals at each moment in the current time period and the dynamic fault threshold of the current time period; the fault occurrence status is either a fault has occurred or no fault has occurred.

[0082] In one exemplary embodiment, a computer device is provided, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a fault diagnosis method for a launch vehicle control system based on time-frequency bi-branch graph attention reconstruction.

[0083] In one exemplary embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements a fault diagnosis method for a launch vehicle control system based on time-frequency bi-branch graph attention reconstruction.

[0084] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements a fault diagnosis method for a launch vehicle control system based on time-frequency bi-branch graph attention reconstruction.

[0085] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 9 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and databases. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a fault diagnosis method for a launch vehicle control system based on time-frequency bi-branch graph attention reconstruction.

[0086] Those skilled in the art will understand that Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0087] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0088] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0089] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0090] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0091] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A fault diagnosis method for a launch vehicle control system based on time-frequency bi-branch graph attention reconstruction, characterized in that, The fault diagnosis method for the launch vehicle control system includes: The real-time power supply voltage timing data of the three channels of the launch vehicle control system is obtained; the architecture of the launch vehicle control system is a three-machine redundant channel; the real-time power supply voltage timing data of the three channels includes the power supply voltage data of the three channels at multiple consecutive moments in the current time period; The real-time power supply voltage time series data of the three channels is input into the voltage reconstruction model to obtain the reconstructed power supply voltage time series data of the three channels for the current time period. The voltage reconstruction model is obtained by training a time-frequency dual-branch graph attention reconstruction model, which includes a dual-stream encoder, a dynamic topology graph learner, and a feature reconstruction decoder. The dual-stream encoder is built based on a multi-scale time-frequency fusion module, and each of the multi-scale time-frequency fusion modules is built based on a time-domain convolution unit, a frequency-domain spectrum extraction unit, and a time-frequency interactive attention unit. The dynamic topology graph learner is used to construct the dynamic dependency relationship between the three redundant channels. The feature reconstruction decoder is built based on a gated residual reconstruction module, and each of the gated residual reconstruction modules is built based on a channel adaptive weighting subunit and a multilayer perceptron reconstruction subunit. Based on the real-time power supply voltage timing data of the three channels in the current time period and the reconstructed power supply voltage timing data of the three channels in the current time period, the residuals at each moment in the current time period are determined. Based on the residuals at each time point in the current time period, determine the dynamic fault threshold for the current time period; Based on the residuals at each moment in the current time period and the dynamic fault threshold of the current time period, the fault occurrence status of the timing controller in the launch vehicle control system at each moment is determined; the fault occurrence status is either a fault occurred or no fault occurred.

2. The fault diagnosis method for launch vehicle control systems based on time-frequency bi-branch graph attention reconstruction according to claim 1, characterized in that, The training process of the voltage reconstruction model includes: Obtain a training sample set; the training sample set includes three-channel historical power supply voltage timing data of the launch vehicle control system under normal operating conditions for multiple historical periods; Construct a time-frequency bi-branch graph attention reconstruction model; Using the training sample set as input, and with the goal of minimizing the reconstruction error between the three-channel reconstructed power supply voltage time series data of the historical period output by the time-frequency bi-branch graph attention reconstruction model and the training sample set, the time-frequency bi-branch graph attention reconstruction model is trained unsupervised to obtain the voltage reconstruction model.

3. The fault diagnosis method for launch vehicle control systems based on time-frequency bi-branch graph attention reconstruction according to claim 1, characterized in that, The dual-stream encoder comprises M cascaded multi-scale time-frequency fusion modules. Each multi-scale time-frequency fusion module includes a time-domain convolution unit, a frequency-domain spectrum extraction unit, a time-frequency interactive attention unit, and a first adder. The time-domain convolution unit includes three parallel dilated convolutional layers. The frequency-domain spectrum extraction unit includes a fast Fourier transform subunit, a first multilayer perceptron, and an inverse fast Fourier transform subunit connected in sequence. The dynamic topology graph learner includes a learnable node embedding layer, a similarity calculation layer, and a graph convolutional network layer connected in sequence. The feature reconstruction decoder includes M-1 matrix concatenation operation layers and M cascaded decoding modules; the M cascaded decoding modules include one first decoding module and M-1 second decoding modules; the first decoding module includes a first sampling layer, a second sampling layer, and a gated residual reconstruction module connected in sequence, and the second decoding module includes a third sampling layer, a second sampling layer, and a gated residual reconstruction module connected in sequence; the first sampling layer is an upsampling mapping layer with a time dimension step size of 2, the second sampling layer is a mapping layer with a time dimension step size of 1, and the third sampling layer is an upsampling mapping layer with a time dimension step size of 2; the gated residual reconstruction module includes a channel adaptive weighting subunit, a multilayer perceptron reconstruction subunit, a sigmoid activation function, a first multiplier, and a second adder; the channel adaptive weighting subunit includes a global average pooling layer, a second multilayer perceptron, and a second multiplier; the multilayer perceptron reconstruction subunit includes a first linear layer and a second linear layer; The first adder in the first multi-scale time-frequency fusion module is connected to the time-domain convolution unit, the frequency-domain spectrum extraction unit, and the first sampling layer in the second multi-scale time-frequency fusion module, respectively. The first adder in the m1-th multi-scale time-frequency fusion module is connected to the time-domain convolution unit and the frequency-domain spectrum extraction unit in the m1+1-th multi-scale time-frequency fusion module, respectively; m1=2,…,M-1, M≥4. The first adder in the m2-th multi-scale time-frequency fusion module and the third sampling layer in the m2-th second decoding module are both connected to the m2-th matrix concatenation operation layer. The m2-th matrix concatenation operation layer is connected to the second sampling layer in the m2-th second decoding module; m2=1,…,M-1. The first adder of the Mth multi-scale time-frequency fusion module is connected to the learnable node embedding layer and the graph convolutional network layer, respectively. The graph convolutional network layer is connected to the third sampling layer in the (M-1)th second decoding module. The second adder in the m3th second decoding module is connected to the third sampling layer in the m3-1th second decoding module; the second adder in the 1st second decoding module is connected to the first sampling layer; m3 = M-1, ..., 2.

4. The fault diagnosis method for launch vehicle control systems based on time-frequency bi-branch graph attention reconstruction according to claim 3, characterized in that, The real-time power supply voltage timing data of the three channels is input into the voltage reconstruction model to obtain the three-channel reconstructed power supply voltage timing data for the current time period, including: The real-time power supply voltage timing data of the three channels is input into the dual-stream encoder in the voltage reconstruction model to obtain the three-channel time-frequency fusion feature map at M scales; The three-channel time-frequency fusion feature map at the Mth scale is input into the dynamic topology graph learner in the voltage reconstruction model to obtain the graph update features; The graph update features and the three-channel time-frequency fusion feature maps at the first M-1 scales are both input into the feature reconstruction decoder in the voltage reconstruction model to obtain the three-channel reconstructed power supply voltage time series data for the current time period.

5. The fault diagnosis method for launch vehicle control systems based on time-frequency bi-branch graph attention reconstruction according to claim 4, characterized in that, The three-channel real-time power supply voltage timing data is input into the dual-stream encoder in the voltage reconstruction model to obtain three-channel time-frequency fusion feature maps at M scales, including: The real-time power supply voltage timing data of the three channels is input to the first multi-scale time-frequency fusion module, and then processed by M multi-scale time-frequency fusion modules for feature extraction to obtain the three-channel time-frequency fusion feature map at M scales; wherein, the feature extraction process of the m-th multi-scale time-frequency fusion module in the voltage reconstruction model includes: The temporal convolution unit of the m-th multi-scale time-frequency fusion module performs multi-path parallel dilated convolution on the three-channel time-frequency fusion feature map at the (m-1)-th scale to obtain the temporal waveform features under the m-th multi-receptive field; when m=1, the three-channel time-frequency fusion feature map at the (m-1)-th scale is the real-time power supply voltage timing data of the three channels; m=1, 2, ..., M; The frequency domain spectrum extraction unit of the m-th multi-scale time-frequency fusion module performs fast Fourier transform, multilayer perceptron mapping and inverse fast Fourier transform on the three-channel time-frequency fusion feature map at the (m-1)-th scale in sequence to obtain the m-th global frequency domain distribution feature. The time-frequency interaction attention unit of the m-th multi-scale time-frequency fusion module performs cross-domain attention calculation on the m-th time-domain waveform feature and the m-th global frequency-domain distribution feature to obtain the m-th time-frequency cross feature; The first adder of the m-th multi-scale time-frequency fusion module adds the m-th time-frequency cross feature and the three-channel time-frequency fusion feature map at the (m-1)-th scale to obtain the three-channel time-frequency fusion feature map at the m-th scale.

6. The fault diagnosis method for launch vehicle control systems based on time-frequency bi-branch graph attention reconstruction according to claim 4, characterized in that, The three-channel time-frequency fusion feature map at the Mth scale is input into the dynamic topology graph learner in the voltage reconstruction model to obtain graph update features, including: The three-channel time-frequency fusion feature map at the Mth scale is input into the learnable node embedding layer in the voltage reconstruction model to obtain the node embedding vectors corresponding to the three channels. The node embedding vectors corresponding to each channel are input into the similarity calculation layer in the voltage reconstruction model to obtain the similarity calculation results. The similarity calculation results are then activated to obtain the dynamic adjacency matrix. The similarity calculation results include the semantic similarity between each pair of channels. The three-channel time-frequency fusion feature map at the Mth scale and the dynamic adjacency matrix are input into the graph convolutional network layer in the voltage reconstruction model to perform graph convolution operation and obtain the graph update feature.

7. The fault diagnosis method for launch vehicle control systems based on time-frequency bi-branch graph attention reconstruction according to claim 4, characterized in that, The updated graph features and the three-channel time-frequency fusion feature maps at the first M-1 scales are both input into the feature reconstruction decoder in the voltage reconstruction model to obtain the three-channel reconstructed supply voltage time series data for the current time period, including: The updated graph features are input to the (M-1)th second decoding module. Combined with the three-channel time-frequency fusion feature maps at the previous (M-1) scales, the features are sequentially processed through the (M-1)th second decoding modules and one first decoding module in the feature reconstruction decoder to obtain the three-channel reconstructed power supply voltage time series data for the current time period. The feature reconstruction process of the r-th second decoding module in the voltage reconstruction model includes: The third sampling layer of the r-th second decoding module upsamples the (r+1)-th decoding feature map to obtain the r-th sampled feature map; r = 1, 2, ..., M-1; when r = M-1, the (r+1)-th decoding feature map is the updated feature of the graph; The matrix splicing operation layer of the r-th second decoding module splices the r-th sampled feature map and the three-channel time-frequency fusion feature map at the r-th scale to obtain the r-th spliced ​​map; The second sampling layer of the r-th second decoding module performs feature extraction and channel dimension compression on the r-th stitched image in sequence to obtain the r-th integrated feature map; The gated residual reconstruction module of the r-th second decoding module reconstructs the r-th integrated feature map to obtain the r-th decoded feature map; The first sampling layer of the first decoding module upsamples the first decoded feature map to obtain the first sampled feature map; The second sampling layer of the first decoding module performs feature extraction and channel dimension compression to obtain the second sampling feature map; The gated residual reconstruction module of the first decoding module reconstructs the second sampling feature map to obtain the three-channel reconstructed power supply voltage timing data for the current time period.

8. The fault diagnosis method for launch vehicle control systems based on time-frequency bi-branch graph attention reconstruction according to claim 1, characterized in that, Based on the residuals at each time point in the current time period, the dynamic fault threshold for the current time period is determined, including: Calculate the mean and standard deviation of the residuals for the current time period based on the residuals at each time point in the current time period; Based on the mean and standard deviation of the residuals for the current time period, the dynamic fault threshold for the current time period is determined using the dynamic fault threshold setting formula. The dynamic fault threshold setting formula is as follows: ； in, Dynamic fault threshold; The mean; Preset sensitivity coefficient; The standard deviation is denoted as .

9. A fault diagnosis system for a launch vehicle control system based on time-frequency bi-branch graph attention reconstruction, to implement the fault diagnosis method for a launch vehicle control system based on time-frequency bi-branch graph attention reconstruction as described in any one of claims 1-8, characterized in that, The launch vehicle control system fault diagnosis system includes: The data acquisition module is used to acquire the real-time power supply voltage timing data of the three channels of the launch vehicle control system; the architecture of the launch vehicle control system is a three-machine redundant channel; the real-time power supply voltage timing data of the three channels includes the power supply voltage data of the three channels at multiple consecutive moments in the current time period; The data reconstruction module is used to input the real-time power supply voltage time-series data of the three channels into the voltage reconstruction model to obtain the reconstructed power supply voltage time-series data of the three channels for the current time period. The voltage reconstruction model is obtained by training a time-frequency dual-branch graph attention reconstruction model, which includes a dual-stream encoder, a dynamic topology graph learner, and a feature reconstruction decoder. The dual-stream encoder is built based on a multi-scale time-frequency fusion module, and each of the multi-scale time-frequency fusion modules is built based on a time-domain convolutional unit, a frequency-domain spectrum extraction unit, and a time-frequency interactive attention unit. The dynamic topology graph learner is used to construct the dynamic dependency relationship between the three redundant channels. The feature reconstruction decoder is built based on a gated residual reconstruction module, and each of the gated residual reconstruction modules is built based on a channel adaptive weighting subunit and a multilayer perceptron reconstruction subunit. The residual calculation module is used to determine the residuals at each moment in the current time period based on the real-time power supply voltage timing data of the three channels in the current time period and the reconstructed power supply voltage timing data of the three channels in the current time period. The threshold determination module is used to determine the dynamic fault threshold for the current time period based on the residuals at each time point in the current time period. The fault diagnosis module is used to determine the fault occurrence status of the timing controller in the launch vehicle control system at each moment based on the residuals at each moment in the current time period and the dynamic fault threshold of the current time period; the fault occurrence status is either a fault has occurred or no fault has occurred.

10. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the fault diagnosis method for a launch vehicle control system based on time-frequency bi-branch graph attention reconstruction as described in any one of claims 1-8.

Images