Liquid rocket engine detection method, system, medium, and apparatus

Abstract

A liquid rocket engine detection method, system, medium and device, in the method, the physical knowledge of the engine dynamics model of the liquid rocket engine is introduced and embedded into the graph neural network as prior information, the consistency of the graph structure relationship and the actual physical relationship of the engine is determined; the topological relationship between the variables is extracted from the engine dynamics model by using the small deviation method, the nonlinear dynamics model of the engine is equivalent converted into a linear state space equation, then the quantitative relationship is converted into a qualitative relationship, and finally a directed graph is constructed; the node connection weight of the directed graph is updated by using a graph learning method based on sparse update, and a sensor posterior matrix is obtained; the graph convolution is performed according to the sensor posterior matrix, and a mixed jump layer is introduced therein, and an updated feature matrix is obtained; a time series convolution module is executed on the feature matrix, a double inflation convolution mechanism is adopted, and finally an engine detection model is obtained to detect the abnormal state of the liquid rocket engine.

Description

Technical Field

[0001] This invention relates to the field of liquid rocket engine testing technology, and in particular to a liquid rocket engine testing method, system, medium, and equipment. Background Technology

[0002] Liquid rocket engines are the core propulsion devices of modern aerospace launch systems. With their high thrust, re-start capability, and precise thrust adjustment, they are widely used in launch vehicles and various spacecraft, becoming the preferred propulsion system for launch vehicles, deep space probes, and manned spacecraft. Liquid rockets operate under extremely harsh conditions and must withstand complex loads. To better ensure the safe operation of the system, anomaly detection in liquid rocket engines is a key technology for ensuring stable operation and reducing the risk of failure. Residual analysis based on surrogate models and observation models can achieve real-time fault detection, identifying anomalies early and preventing escalation and serious losses. Currently, anomaly detection in liquid rocket engines is mainly achieved through data-driven and model-based methods (such as nonlinear Kalman filtering). Data-driven methods rely on a deep understanding of the engine system to establish accurate mathematical models. Faults are detected and diagnosed by comparing the model's predicted output with the actual sensor measurements (i.e., the "residual"). Model-based methods automatically learn characteristics and patterns from historical engine operating data to construct diagnostic models.

[0003] Existing liquid rocket engine fault detection technologies suffer from problems such as large computational load, limited generalization ability, and lack of physical interpretability. Limited by engineering conditions such as data quality and measurement cost, they are unable to meet the high-precision requirements of monitoring tasks for different engine models and cannot be well applied to fault detection technology for liquid rocket engines.

[0004] The information disclosed in the background section is only for enhancing the understanding of the background of this invention, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] This invention provides a liquid rocket engine detection method, system, medium, and equipment. It proposes a new means for accurate detection under different system operating conditions based on graph learning with sparse updates and spatiotemporal graph prediction based on temporal convolution and graph convolution.

[0006] A method for testing a liquid rocket engine includes:

[0007] Step A: Introduce the physical knowledge of the engine dynamics model of the liquid rocket engine and embed it as prior information into the graph neural network to determine the consistency between the graph structure relationship and the actual physical relationship of the engine;

[0008] Step B: Use the small deviation method to extract the topological relationships between variables from the engine dynamics model, convert the nonlinear dynamics model of the engine into a linear state-space equation, then convert the quantitative relationships into qualitative relationships, and finally construct a directed graph;

[0009] Step C: Update the node connection weights of the directed graph using a sparse update-based graph learning method to optimize the prior adjacency matrix. The non-zero elements and zero-valued elements remain unchanged to obtain the sensor posterior matrix. ;

[0010] Step D: Based on the sensor posterior matrix Perform graph convolution, where N is the number of nodes (number of sensors), R is the maximum graph propagation order in the hybrid hop layer, and introduce a hybrid hop layer to obtain the updated feature matrix. ;

[0011] Step E: Perform a temporal convolution module on the feature matrix, using a double dilation convolution mechanism, to obtain the final engine detection model for detecting abnormal states of liquid rocket engines.

[0012] In the liquid rocket engine testing method described above, step B includes:

[0013] Step B1: Calculate the partial derivatives of a certain sensor variable in the engine dynamics model with respect to the other sensor variables. Analyze their mutual influence relationships;

[0014] Step B2: Determine the qualitative relationship between the variables. If the partial derivatives... If it is zero, then the sensor variable With independent variable There is no direct qualitative relationship between them; the corresponding elements in the adjacency matrix... If the value is assigned to 0, and the partial derivative is not zero, it indicates that the sensor variable... With independent variable There are dependencies between them, at the corresponding positions in the adjacency matrix. The value is assigned to 1;

[0015] Step B3: Repeat step B2, calculating the partial derivative relationships between all variables in the differential equation and the algebraic equation, and converting them into the corresponding values ​​of the adjacency matrix to obtain the complete adjacency matrix. ;

[0016] Step B4: Based on the adjacency matrix Construct a directed graph where nodes represent sensor variables and the direction of edges is determined by the partial derivative relationship, ensuring that the graph structure characterizes the variable dependencies in the dynamic model.

[0017] In the liquid rocket engine testing method described above, step C includes:

[0018] Step C1: Use the PyTorch framework to generate the global pre-training matrix A2 and the mask matrix respectively. Assign gradients to the pre-training matrix and update the weights in the global pre-training matrix A2 through gradient descent. The mask matrix is ​​not assigned gradients and is only responsible for data preprocessing, so that the input at a specific position in the data is 0.

[0019] Step C2: Sparse update, only update the parameters in matrix A2 whose weights are not 0, "deactivate" all the parameters in matrix A2 whose weights are 0, even if their gradients are 0, "deactivate" the elements at specified positions in matrix A2, indicating that there is no correlation between the corresponding nodes;

[0020] Step C3: Data dimension expansion. The expansion dimension depends on the actual sensor data dimension, and then the data is multiplied by the mask matrix to achieve data "deactivation".

[0021] Step C4: Graph convolution, using Einstein's summation formula to aggregate sensor information;

[0022] Step C5: Gradient descent, update the adjacency matrix, and obtain the sensor posterior matrix. .

[0023] In the liquid rocket engine testing method described above, step D includes:

[0024] Step D1: Using the sensor posterior matrix A pos Adding the self-connect matrix yields ,Right now ,pass get degree matrix ;

[0025] Step D2: Using the activation function , No. Layer weight matrix and the degree matrix ,pass Obtain the updated feature matrix ;

[0026] Step D3: Utilizing hyperparameters and propagation layers For the updated feature matrix Introducing a hybrid skip layer Using parameter matrix As a feature selector When the given graph structure does not require spatial dependencies, the parameter matrix Adjust the values ​​to zero to mask irrelevant information, and obtain the output feature matrix. .

[0027] In the liquid rocket engine testing method described above, step E includes:

[0028] Step E1: Different activation functions are used for the dilated convolution modules. The first dilated convolution layer uses the hyperbolic tangent function (Tanh) as the activation function, and the second dilated convolution layer uses the Sigmoid function to construct a gating mechanism. Finally, the outputs of the two modules are multiplied element-wise to form a gated convolution unit.

[0029] Step E2: Four different sizes of convolution kernels are used to capture sequential patterns in time series data to form an engine fault detection model.

[0030] In the liquid rocket engine detection method described above, four different sizes of convolution kernels are 1×2, 1×3, 1×6, and 1×7.

[0031] In the liquid rocket engine detection method described above, the abnormal detection result is judged by the residual between the engine detection model output and the actual sensor measurement value, and a threshold is set. If the residual exceeds the threshold, it is judged as abnormal.

[0032] A system for implementing the method includes:

[0033] The mechanism graph construction module is configured to introduce the physical knowledge of the engine dynamics model of the liquid rocket engine and embed it as prior information into the graph neural network to determine the consistency between the graph structure relationship and the actual physical relationship of the engine. It uses the small deviation method to extract the topological relationship between variables from the engine dynamics model, and converts the nonlinear dynamics model of the engine into a linear state-space equation. Then, it converts the quantitative relationship into a qualitative relationship and finally constructs a directed graph.

[0034] The graph learning optimization module is configured to update the node connection weights of the directed graph using a sparse update-based graph learning method, thereby optimizing the prior adjacency matrix. The non-zero elements and zero-valued elements remain unchanged to obtain the sensor posterior matrix. ;

[0035] The spatiotemporal feature extraction module is configured to extract features based on the sensor posterior matrix. Perform graph convolution and introduce a hybrid skip layer to obtain the updated feature matrix. The feature matrix is ​​executed by a temporal convolution module, which uses a double dilatation convolution mechanism to obtain the final engine detection model to detect abnormal states of liquid rocket engines.

[0036] The anomaly detection module is used to determine anomalies based on the residual between the engine detection model output and the actual data.

[0037] A computer storage medium including computer instructions that, when run on a computer, cause the computer to perform the method.

[0038] An electronic device, the electronic device comprising:

[0039] Memory, processor, and computer programs stored in memory and executable on the processor, wherein,

[0040] The processor implements the method when executing the program.

[0041] Compared with existing technologies, this invention has the following advantages: It employs graph neural networks to extract the correlation characteristics between sensors and constructs a spatiotemporal graph model that conforms to physical mechanisms, effectively improving the accuracy of fault detection; it fully utilizes the small-deviation method to extract the topological relationships between variables from the dynamic equations, effectively enhancing the generalization ability of the surrogate model and strengthening its interpretability; and it fully utilizes the spatiotemporal graph prediction method combining graph convolution and temporal convolution to effectively capture the spatial and temporal dependencies in the data, thereby effectively modeling complex spatiotemporal dynamic relationships. This not only meets the need for effective fusion of low- and high-level data features but also effectively improves information transmission efficiency and reduces time and economic costs. Attached Figure Description

[0042] Various other advantages and benefits of the present invention will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiments below. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. Furthermore, the same reference numerals denote the same parts throughout the drawings.

[0043] In the attached diagram:

[0044] Figure 1 This is a flowchart of the liquid rocket engine anomaly detection method based on mechanistic knowledge embedded in a graph neural network as described in this invention;

[0045] Figure 2 A flowchart of a graph learning method based on coefficient updates;

[0046] Figure 3 This is a schematic diagram of the spatiotemporal convolution module.

[0047] The present invention will be further explained below with reference to the accompanying drawings and embodiments. Detailed Implementation

[0048] Specific embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While specific embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0049] It should be noted that certain terms are used in the specification and claims to refer to specific components. Those skilled in the art will understand that different terms may be used to refer to the same component. This specification and claims do not distinguish components based on differences in terminology, but rather on differences in function. The terms "comprising" or "including" used throughout the specification and claims are open-ended and should be interpreted as "comprising but not limited to." The following descriptions are preferred embodiments for carrying out the invention; however, these descriptions are for the purpose of understanding the general principles of the specification and are not intended to limit the scope of the invention. The scope of protection of this invention is determined by the appended claims.

[0050] To facilitate understanding of the embodiments of the present invention, further explanations and descriptions will be provided below with reference to the accompanying drawings and specific embodiments. The accompanying drawings do not constitute a limitation on the embodiments of the present invention.

[0051] like Figures 1 to 3 As shown, the liquid rocket engine testing method includes the following steps:

[0052] Step A: Introduce the physical knowledge of the engine dynamics model of the liquid rocket engine and embed it as prior information into the graph neural network to determine the consistency between the graph structure relationship and the actual physical relationship of the engine;

[0053] Step B: Extract the topological relationships between sensor variables from the engine dynamics model using the small deviation method, and then apply this to the nonlinear dynamics model of the engine from Step A. Where t is the time variable, calculated using the Jacobian matrix at the operating point The matrices A and B are then equivalently transformed into linear state-space equations. Then, the quantitative relationships such as partial derivatives and Jacobian matrices in the dynamic model are converted into qualitative relationships of whether there is a dependency between variables, and finally a directed graph is constructed.

[0054] Step C: Update the node connection weights of the directed graph using a sparse update-based graph learning method to optimize the prior adjacency matrix. For non-zero elements, a mask parameterization method is used. A fixed prior adjacency matrix is ​​used as the mask matrix, which is then multiplied element-wise with the learnable weight matrix to form the effective adjacency matrix. Only the weights at non-zero positions in the prior matrix are updated, while zero-valued elements remain unchanged, resulting in the sensor posterior matrix. ;

[0055] Step D: Based on the sensor posterior matrix Graph convolution is performed, and a hybrid skip layer is introduced to obtain feature matrices of different propagation layers. Then, the features of each layer are weighted and fused through the hybrid skip layer. The importance of information from different layers is adaptively adjusted using learnable parameters, thereby integrating shallow local features and deep global dependencies.

[0056] When the input graph structure does not require spatial dependencies, the updated feature matrix is ​​obtained by setting the feature selection parameter matrix to zero, thereby masking irrelevant features. ;

[0057] Step E: Perform a temporal convolution module on the feature matrix, using a double dilation convolution mechanism, to obtain the final engine detection model for detecting abnormal states of liquid rocket engines.

[0058] In a preferred embodiment of the liquid rocket engine testing method, step A includes:

[0059] Step A1: Solve the differential equations using the Newton-Raphson iteration method to obtain the physical and mathematical relationships between sensor variables of components such as the combustion chamber, valves, and turbine in the liquid rocket engine;

[0060] Step A2: A graph neural network consists of nodes and edges, where nodes represent sensor variables and the direction of the edges is determined by the sensor variables. With independent variable partial derivatives Decide.

[0061] In a preferred embodiment of the liquid rocket engine testing method, step B includes:

[0062] Step B1: Calculate the partial derivatives of a certain sensor variable in the engine dynamics model with respect to the other sensor variables. ;

[0063] Step B2: Determine the qualitative relationship between the variables. If the partial derivatives... If it is zero, then the sensor variable With independent variable There is no direct qualitative relationship between them; the corresponding elements in the adjacency matrix... If the value is assigned to 0, and the partial derivative is not zero, it indicates that the sensor variable... With independent variable There are dependencies between them, at the corresponding positions in the adjacency matrix. The value is assigned to 1;

[0064] Step B3: Repeat step B2, sequentially calculating the partial derivative relationships between all variables in the differential and algebraic equations of the engine dynamic model, and converting them into the corresponding values ​​of the adjacency matrix to obtain the complete adjacency matrix. ;

[0065] Step B4: Based on the adjacency matrix Construct a directed graph where nodes represent sensor variables and the direction of edges is determined by the partial derivative relationship, ensuring that the graph structure characterizes the variable dependencies in the dynamic model.

[0066] In a preferred embodiment of the liquid rocket engine testing method, step C includes:

[0067] Step C1: Using the PyTorch framework, generate a global pre-training matrix A and a mask matrix, respectively. Assign gradients to the pre-training matrix. The global pre-training matrix A is a learnable weight matrix, randomly initialized and with gradient computation enabled. The mask matrix is ​​a fixed binary matrix indicating which edges need updating during training. During training, the global pre-training matrix A is optimized according to the loss function, while the mask matrix remains unchanged. The weights in the global pre-training matrix A are updated using gradient descent. The mask matrix is ​​not assigned gradients; it only handles data preprocessing, ensuring that the input at specific locations in the data is 0.

[0068] Step C2: Sparse update, only update the parameters with non-zero weights in the global pre-training matrix A, "deactivate" all parameters with zero weights in the global pre-training matrix A, even if their gradients are zero, "deactivate" the elements at specified positions in the global pre-training matrix A, and specify that there is no correlation between the corresponding nodes;

[0069] Step C3: Data dimension expansion. The expansion dimension depends on the actual sensor data dimension, and then the data is multiplied by the mask matrix to achieve data "deactivation".

[0070] Step C4: Graph convolution, using Einstein's summation formula. Aggregate sensor information, among which It is an adjacency matrix. For the features of node j in the l-th layer, Let be the weight matrix of the l-th layer. As the activation function, in this way, the features of node i aggregate information from its connected nodes j, and are then used to update the node features;

[0071] Step C5: Gradient descent, update the adjacency matrix, and obtain the sensor posterior matrix. .

[0072] In a preferred embodiment of the liquid rocket engine testing method, step D includes:

[0073] Step D1: Use the sensor posterior matrix Apos plus the self-connection matrix I to obtain... ,Right now ,pass get degree matrix , where matrix D is composed of The calculated degree matrix;

[0074] Step D2: Using the activation function , No. Layer weight matrix and the degree matrix ,pass Obtain the updated feature matrix ;

[0075] Step D3: Utilizing hyperparameters and propagation layers For the updated feature matrix Introducing a hybrid skip layer Using parameter matrix As a feature selector When the given graph structure does not require spatial dependencies, the parameter matrix Adjust the values ​​to zero to mask irrelevant information, and obtain the output feature matrix. .

[0076] In a preferred embodiment of the liquid rocket engine testing method, step E includes:

[0077] Step E1: Different activation functions are used for the dilated convolution modules. The first dilated convolution layer uses the hyperbolic tangent function (Tanh) as the activation function, and the second dilated convolution layer uses the Sigmoid function to construct a gating mechanism. Finally, the outputs of the two modules are multiplied element-wise to form a gated convolution unit.

[0078] Step E2: Four different sizes of convolution kernels are used to capture sequential patterns in time series data to form an engine fault detection model.

[0079] In a preferred embodiment of the liquid rocket engine detection method, the four different sizes of convolution kernels are 1×2, 1×3, 1×6, and 1×7.

[0080] In a preferred embodiment of the liquid rocket engine detection method, the abnormal detection result is judged by the residual between the engine detection model output and the actual sensor measurement value, and a threshold is set. If the residual exceeds the threshold, it is judged as abnormal.

[0081] A system for implementing the method includes:

[0082] The mechanism graph construction module is configured to introduce the physical knowledge of the engine dynamics model of the liquid rocket engine and embed it as prior information into the graph neural network to determine the consistency between the graph structure relationship and the actual physical relationship of the engine. It uses the small deviation method to extract the topological relationship between variables from the engine dynamics model, and converts the nonlinear dynamics model of the engine into a linear state-space equation. Then, it converts the quantitative relationship into a qualitative relationship and finally constructs a directed graph.

[0083] The graph learning optimization module is configured to update the node connection weights of the directed graph using a sparse update-based graph learning method, thereby optimizing the prior adjacency matrix. The non-zero elements and zero-valued elements remain unchanged to obtain the sensor posterior matrix. ;

[0084] The spatiotemporal feature extraction module is configured to extract features based on the sensor posterior matrix. Perform graph convolution and introduce a hybrid skip layer to obtain the updated feature matrix. The feature matrix is ​​executed by a temporal convolution module, which uses a double dilatation convolution mechanism to obtain the final engine detection model to detect abnormal states of liquid rocket engines.

[0085] The anomaly detection module is used to determine anomalies based on the residual between the engine detection model output and the actual data.

[0086] A computer storage medium including computer instructions that, when run on a computer, cause the computer to perform the method.

[0087] An electronic device, the electronic device comprising:

[0088] Memory, processor, and computer programs stored in memory and executable on the processor, wherein,

[0089] The processor implements the method when executing the program.

[0090] In one embodiment, the relationships between engine sensors are extracted using a small-bias method, effectively transforming the engine's nonlinear dynamics model into a linear state-space equation. A graph structure conforming to the physical mechanism is then constructed, and a sparse update method is used to train the graph network. Subsequently, a one-dimensional convolutional network (1D-CNN) is combined for spatiotemporal feature extraction. Finally, anomaly detection of liquid rocket engines is achieved based on a combination of graph convolution and temporal convolution. By embedding mechanistic knowledge into a graph neural network for anomaly detection in liquid rocket engines, an accurate prediction model is constructed. This solves the problems of data-driven methods being limited by data quality and generalization ability, and has broad application prospects in the field of liquid rocket engine fault detection.

[0091] Furthermore, this invention extracts qualitative relationships between variables based on the small-deviation method and constructs an initial adjacency matrix, achieving explicit embedding of physical mechanisms and improving model interpretability and prior rationality. By analyzing the partial derivative relationships between variables in the dynamic equations, the physical causal dependence of the engine is transformed into directed edges in a graph structure, ensuring that the initial graph topology conforms to the behavior of the real system. This avoids the spurious connection problem caused by data noise or insufficient samples in traditional data-driven methods, providing high-quality prior knowledge for subsequent learning and significantly enhancing the physical consistency and generalization ability of the model. Data-driven optimization is achieved while preserving physical constraints, preventing model overfitting and structural drift. Gradient updates are performed only on non-zero elements in the prior adjacency matrix, forcing zero elements to remain unchanged, effectively suppressing the introduction of spurious associations that do not exist physically during training. This mechanism achieves "parameter adaptive optimization under structural sparsity constraints," both utilizing data to correct the inaccuracies of prior knowledge and preventing the model from deviating from physical laws, thus improving model robustness and engineering credibility. It enhances the flexibility and selectivity of feature propagation and improves spatial modeling capabilities. By introducing a learnable parameter matrix as a feature selector, the fusion weights of features at different levels in skip connections are dynamically adjusted. When local graph structure information is insufficient or noise exists, the model can automatically reduce spatial dependency weights and instead rely on the original input or high-level semantic features, achieving "on-demand aggregation." This mechanism effectively alleviates the oversmoothing problem in deep GNNs and enhances the model's adaptability to complex operating conditions. It efficiently captures long and short-term temporal dependencies, enhancing the expressive power and selectivity of temporal modeling. A Tanh-Sigmoid dual-branch structure is used to form gated convolutional units, where the Tanh branch extracts temporal features and the Sigmoid branch generates gating signals to control information flow. Combined with multi-scale dilated convolutional kernels (1×2, 1×3, 1×6, 1×7), it can perceive longer time spans without increasing network depth, effectively capturing the dynamic evolution patterns of liquid rocket engine startup, steady state, and shutdown stages, improving the response speed to sudden anomalies. It decouples spatial dependencies from temporal dynamics, achieving efficient and interpretable spatiotemporal feature learning. First, graph convolution is used to model the physical coupling relationship between sensors (spatial dimension), and then temporal convolution is used to capture the dynamic characteristics of each variable evolving over time (temporal dimension). This separate design avoids parameter explosion and training difficulties caused by spatiotemporal joint convolution, while facilitating the analysis of the independent contributions of spatial structure and temporal patterns, enhancing the model's debuggability and engineering applicability. The trained model is used to reconstruct or predict sensor data, calculating the residuals between actual and predicted values. The statistical characteristics of the residuals (such as mean, variance, and peak value) directly reflect the degree to which the system deviates from its normal state. By setting reasonable thresholds, automated alarms can be achieved, suitable for real-time online monitoring scenarios, and meeting the high reliability requirements of aerospace missions.

[0092] Although embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments and application fields described above. The specific embodiments described above are merely illustrative and instructive, and not restrictive. Those skilled in the art can make many other forms based on the guidance of this specification and without departing from the scope of protection of the claims of the present invention, and all of these are within the scope of protection of the present invention.

Claims

1. A method for testing a liquid rocket engine, characterized in that, Includes the following steps: Step A: Introduce the physical knowledge of the engine dynamics model of the liquid rocket engine and embed it as prior information into the graph neural network to determine the consistency between the graph structure relationship and the actual physical relationship of the engine; Step B: Use the small deviation method to extract the topological relationships between variables from the engine dynamics model, convert the nonlinear dynamics model of the engine into a linear state-space equation, then convert the quantitative relationships into qualitative relationships, and finally construct a directed graph; Step C: Update the node connection weights of the directed graph using a sparse update-based graph learning method to optimize the prior adjacency matrix. The non-zero elements and zero-valued elements remain unchanged to obtain the sensor posterior matrix. ; Step D: Based on the sensor posterior matrix Perform graph convolution and introduce a hybrid skip layer to obtain the updated feature matrix. ; Step E: Perform a temporal convolution module on the feature matrix, using a double dilation convolution mechanism, to obtain the final engine detection model for detecting abnormal states of liquid rocket engines.

2. The liquid rocket engine testing method according to claim 1, characterized in that, Preferably, step B includes: Step B1: Calculate the partial derivatives of a certain sensor variable in the engine dynamics model with respect to the other sensor variables. Analyze their mutual influence relationships; Step B2: Determine the qualitative relationship between the variables. If the partial derivatives... If it is zero, then the sensor variable With independent variable There is no direct qualitative relationship between them; the corresponding elements in the adjacency matrix... If the value is assigned to 0, and the partial derivative is not zero, it indicates that the sensor variable... With independent variable There are dependencies between them, at the corresponding positions in the adjacency matrix. The value is assigned to 1; Step B3: Repeat step B2, calculating the partial derivative relationships between all variables in the differential equation and the algebraic equation, and converting them into the corresponding values ​​of the adjacency matrix to obtain the complete adjacency matrix. ; Step B4: Based on the adjacency matrix Construct a directed graph where nodes represent sensor variables and the direction of edges is determined by the partial derivative relationship, ensuring that the graph structure characterizes the variable dependencies in the dynamic model.

3. The liquid rocket engine testing method according to claim 1, characterized in that, Step C includes: Step C1: Use the PyTorch framework to generate a global pre-training matrix A and a mask matrix respectively. Assign gradients to the pre-training matrix and update the weights in the global pre-training matrix A through gradient descent. The mask matrix is ​​not assigned gradients and is only responsible for data preprocessing, so that the input at a specific position in the data is 0. Step C2: Sparse update, only update the parameters with non-zero weights in the global pre-training matrix A, "deactivate" all parameters with zero weights in the global pre-training matrix A, even if their gradients are zero, "deactivate" the elements at specified positions in the global pre-training matrix A, indicating that there is no correlation between the corresponding nodes; Step C3: Data dimension expansion. The expansion dimension depends on the actual sensor data dimension, and then the data is multiplied by the mask matrix to achieve data "deactivation". Step C4: Graph convolution, using Einstein's summation formula to aggregate sensor information; Step C5: Gradient descent, update the global pre-trained matrix, and obtain the sensor posterior matrix. .

4. The liquid rocket engine testing method according to claim 3, characterized in that, Step D includes: Step D1: Use the global pre-trained matrix A plus the self-connection matrix to obtain... ,Right now ,pass get degree matrix ; Step D2: Using the activation function , No. Layer weight matrix and the degree matrix ,pass Obtain the updated feature matrix ; Step D3: Utilizing hyperparameters and propagation layers For the updated feature matrix Introducing a hybrid skip layer Using parameter matrix As a feature selector When the given graph structure does not require spatial dependencies, the parameter matrix Adjust the values ​​to zero to mask irrelevant information, and obtain the output feature matrix. .

5. The liquid rocket engine testing method according to claim 1, characterized in that, Step E includes: Step E1: Different activation functions are used for the dilated convolution modules. The first dilated convolution layer uses the hyperbolic tangent function (Tanh) as the activation function, and the second dilated convolution layer uses the Sigmoid function to construct a gating mechanism. Finally, the outputs of the two modules are multiplied element-wise to form a gated convolution unit. Step E2: The dilated convolution module is used with four different sizes of convolution kernels to capture sequential patterns in time series data, forming an engine fault detection model.

6. The liquid rocket engine testing method according to claim 1, characterized in that, The four different sizes of convolution kernels are 1×2, 1×3, 1×6, and 1×7.

7. The liquid rocket engine testing method according to claim 1, characterized in that, The abnormal detection results are judged by the residual between the engine detection model output and the actual sensor measurement value, and a threshold is set. If the residual exceeds the threshold, it is judged as abnormal.

8. A system for implementing the method of any one of claims 1-7, characterized in that, It includes: The mechanism graph construction module is configured to introduce the physical knowledge of the engine dynamics model of the liquid rocket engine and embed it as prior information into the graph neural network to determine the consistency between the graph structure relationship and the actual physical relationship of the engine. It uses the small deviation method to extract the topological relationship between variables from the engine dynamics model, and converts the nonlinear dynamics model of the engine into a linear state-space equation. Then, it converts the quantitative relationship into a qualitative relationship and finally constructs a directed graph. The graph learning optimization module is configured to update the node connection weights of the directed graph using a sparse update-based graph learning method, thereby optimizing the prior adjacency matrix. The non-zero elements and zero-valued elements remain unchanged to obtain the sensor posterior matrix. ; The spatiotemporal feature extraction module is configured to extract features based on the sensor posterior matrix. Perform graph convolution and introduce a hybrid skip layer to obtain the updated feature matrix. The feature matrix is ​​executed by a temporal convolution module, which uses a double dilatation convolution mechanism to obtain the final engine detection model to detect abnormal states of liquid rocket engines. The anomaly detection module is used to determine anomalies based on the residual between the engine detection model output and the actual data.

9. A computer storage medium, characterized in that, The storage medium includes computer instructions that, when executed on a computer, cause the computer to perform the method as described in any one of claims 1-7.

10. An electronic device, characterized in that, The electronic device includes: Memory, processor, and computer programs stored in memory and executable on the processor, wherein, When the processor executes the program, it implements the method as described in any one of claims 1-7.

Images