Aero-engine fault mode analysis method based on trivial causal graph neural network

Abstract

The application provides an aero-engine fault mode analysis method based on a trivial causal graph neural network, and belongs to the technical field of data processing, and specifically comprises the following steps: collecting engine multi-source sensor signals, segmenting the signals through a sliding window, constructing an input graph, an adjacency matrix and a node feature matrix by taking sensor components as nodes and the correlation between sensors as edges; obtaining node feature representation by using a graph neural network encoder, and estimating a soft mask through a node-level and edge-level attention module to untangle the input graph into a causal attention graph and a trivial attention graph; in the training stage, taking the weighted sum of a supervision loss, a uniform distribution constraint loss and a causal intervention loss as a total target, and optimizing model parameters in an end-to-end manner; in the reasoning stage, inputting the graph representation of the causal attention graph into a classifier to output a fault mode. Through the scheme of the application, the confusion effect of non-causal trivial features can be effectively inhibited, and the accuracy and robustness of cross-condition fault recognition are improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for analyzing the failure modes of aero-engines based on a trivial causal graph neural network. Background Technology

[0002] Currently, aero-engine systems are characterized by complex structures, strong component coupling, variable operating conditions, and long fault propagation chains. The complex interactions between various components and sensors mean that aero-engine fault mode identification needs to focus not only on local anomalies in individual sensor signals but also on the temporal relationships, spatial correlations, and potential causal relationships between multiple sensors. However, existing methods have the following shortcomings: (1) Insufficient stability: It is easy to learn trivial features that are bound to the distribution of the training set, which leads to a decrease in performance after changing working conditions, equipment or collection environment.

[0003] (2) Insufficient interpretability: Traditional models can usually only provide classification results, and it is difficult to explain which sensor nodes, which edge relationships or which graph structures are truly related to the causal relationship of the fault.

[0004] (3) Insufficient resistance to confounding: The influence of trivial features as confounding factors is not explicitly modeled, and the stable causal pattern cannot be effectively distinguished from the accidental statistical correlation pattern.

[0005] (4) Insufficient utilization of graph structure: Although existing methods use graph neural networks, they usually only learn a global graph representation and lack a mechanism to decompose the input graph into causal attention graphs and trivial attention graphs.

[0006] (5) Lack of causal intervention in the representation layer: Traditional models generally do not make reverse adjustments or interventions to causal features and trivial features in the representation layer, making it difficult to ensure that the causal representation maintains stable predictive ability under different interference factors.

[0007] It is evident that there is an urgent need for a fault mode analysis method for aero-engines based on trivial causal graph neural networks, which offers high accuracy, stability, and interpretability. Summary of the Invention

[0008] In view of this, embodiments of the present invention provide a method for analyzing the failure modes of aero-engines based on trivial causal graph neural networks, which at least partially solves the problems existing in the prior art.

[0009] This invention provides a method for fault mode analysis of aero-engines based on trivial causal graph neural networks, including: Step 1: Acquire multi-source sensor signals of the aero-engine, wherein the multi-source sensor signals include status monitoring signals collected by multiple sensors distributed in different parts of the engine; Step 2: The multi-source sensor signals are segmented using a sliding window to obtain multiple short-time signal segments; Step 3: Using sensor components as nodes and the relationships between sensors as edges, construct the input graph and the corresponding adjacency matrix and node feature matrix based on short-time signal segments; Step 4: Use a graph neural network encoder to encode the adjacency matrix and the node feature matrix to obtain the node feature representation; Step 5: Based on the node feature representation, estimate the node-level soft mask through the node-level attention module and estimate the edge-level soft mask through the edge-level attention module; Step 6: Dewrap the input graph into a causal attention graph and a trivial attention graph based on the node-level soft mask and the edge-level soft mask. The causal attention graph gathers causal features with stable causal relationships with the fault modes, while the trivial attention graph gathers non-causal trivial features. Step 7: During the training phase, the weighted sum of the supervision loss, uniform distribution constraint loss, and causal intervention loss is used as the overall training objective. The trainable parameters of the graph neural network encoder, node-level attention module, edge-level attention module, and classifier are optimized in an end-to-end manner. Specifically, a fault category supervision loss is applied to the graph representation of the causal attention graph to enable it to learn a stable discriminative representation. A uniform distribution constraint loss is applied to the graph representation of the trivial attention graph to suppress its direct participation in fault category discrimination. In the representation layer, the causal representation of the causal attention graph is combined with multiple trivial feature estimates. The causal intervention loss constrains the classifier to maintain stable prediction under different combinations of trivial features. Step 8: In the inference phase, multi-source sensor signals of the aero-engine under test are collected, a graph representation of the causal attention map is generated and input into the classifier, and the engine failure mode analysis results are output.

[0010] According to a specific implementation of the present invention, the correlation between the sensors includes one or more of the following: physical connection relationship, spatial proximity relationship, historical correlation coefficient, mutual information, coherence function value, expert prior relationship, and learnable adjacency matrix; The input graph has a graph structure that is either isomorphic or heteromorphic.

[0011] According to a specific implementation of an embodiment of the present invention, the estimation methods for the node-level soft mask and the edge-level soft mask are as follows: Using a multilayer perceptron to analyze node features Features of splicing nodes at both ends of the edge After transformation, the softmax activation function outputs causal attention scores and trivial attention scores, satisfying... and ; in, Constructing a causal soft mask, This constitutes a trivial soft mask, which is the complement of a causal soft mask.

[0012] According to a specific implementation of an embodiment of the present invention, the causal attention graph is constructed as follows: ; The method for constructing the trivial attention graph is as follows: ; in, It is an adjacency matrix. The node feature matrix, and These are node-level soft masks and edge-level soft masks, respectively. This represents the complement of the soft mask. This indicates an element-wise multiplication operation.

[0013] According to a specific implementation of an embodiment of the present invention, the fault category supervision loss is represented as: ; in, This is a set of graph samples used for model training. The number of samples in the graph sample set. For the graph sample set Any input graph; To based on the input image And the causal attention graph constructed using causal attention soft masks; Input image The corresponding real fault category label vector is a one-hot encoded vector; Total number of fault categories; For classifiers to apply causal attention graphs Output fault category probability vector; superscript Indicates vector transpose; This indicates that the logarithm is performed on each element of the probability vector. ; The uniform distribution constraint loss is represented by KL divergence as follows: ; ; in, To based on the input image And the trivial attention graph constructed from the complementary mask of the causal attention soft mask; For the classifier's attention graph of trivialities Output fault category probability vector; It is a uniformly distributed vector, and all its elements are... ; The Kullback–Leibler divergence measures the difference between the predicted probability distribution of a trivial attention map and a uniform distribution. The first uniformly distributed vector is the... The probability corresponding to each fault category For the prediction probability vector of a trivial attention map, the first... Predicted probability for each fault category .

[0014] According to a specific implementation of an embodiment of the present invention, the expression for the causal intervention loss is: ; ; in, It is an implicit intervention diagram In the classifier The classification results in It is a causal attention graph The expression, It is layered The expression, It is a set of estimates of trivial features present in the training data.

[0015] According to a specific implementation of an embodiment of the present invention, the expression for the overall training objective is: ; in, and It is a constant used to control the degree of untangling and causal intervention.

[0016] According to one specific implementation of the present invention, the graph neural network encoder is any one of a graph convolutional network, a graph attention network, GraphSAGE, a graph Transformer, or a spatiotemporal graph neural network.

[0017] The fault mode analysis scheme for aero-engines based on trivial causal graph neural networks in this embodiment of the invention includes: Step 1, acquiring multi-source sensor signals of the aero-engine, wherein the multi-source sensor signals include status monitoring signals collected by multiple sensors distributed in different parts of the engine; Step 2, segmenting the multi-source sensor signals using a sliding window to obtain multiple short-time signal segments; Step 3, constructing an input graph and corresponding adjacency matrix and node feature matrix based on the short-time signal segments, using sensor components as nodes and the relationships between sensors as edges; Step 4, encoding the adjacency matrix and node feature matrix using a graph neural network encoder to obtain node feature representations; Step 5, estimating node-level soft masks using a node-level attention module and edge-level soft masks based on the node feature representations; Step 6, dewinding the input graph into a causal attention graph and a trivial attention graph based on the node-level soft masks and edge-level soft masks, wherein the causal attention graph and the trivial attention graph are... If the attention map aggregates causal features with a stable causal relationship with the fault mode, the trivial attention map aggregates non-causal trivial features; Step 7, in the training phase, the weighted sum of supervision loss, uniform distribution constraint loss and causal intervention loss is used as the overall training objective. The trainable parameters of the graph neural network encoder, node-level attention module, edge-level attention module and classifier are optimized in an end-to-end manner. Among them, the fault category supervision loss is applied to the graph representation of the causal attention map to enable it to learn a stable discriminative representation. The uniform distribution constraint loss is applied to the graph representation of the trivial attention map to suppress its direct participation in fault category discrimination. In the representation layer, the causal representation of the causal attention map is combined with multiple trivial feature estimates. The causal intervention loss constrains the classifier to maintain stable prediction under different combinations of trivial features; Step 8, in the inference phase, multi-source sensor signals of the aero-engine to be detected are collected, the graph representation of the causal attention map is generated and input into the classifier, and the engine fault mode analysis results are output.

[0018] The beneficial effects of the embodiments of the present invention are as follows: (1) Improve the accuracy of fault diagnosis: By focusing on sensor nodes and edges that have a stable relationship with fault modes through causal attention graphs, the effectiveness of fault classification can be improved.

[0019] (2) Improve generalization ability across operating conditions: By weakening the influence of trivial features and confounding factors, the model no longer relies solely on the accidental statistical correlations in the training data, which is conducive to adapting to different operating conditions, different equipment individuals and different data distributions.

[0020] (3) Enhance model robustness: By intervening in the causal representation layer, the causal representation can maintain stable predictive ability under different combinations of trivial features, thereby reducing the impact of noise, bias and interference features.

[0021] (4) Improve interpretability: Node-level and edge-level attention scores can help locate key sensors, key connections and potential fault propagation paths related to fault modes.

[0022] (5) More suitable for complex aero-engine systems: Graph structure modeling can simultaneously express the spatial correlation, temporal segment information and system coupling relationship of multiple source sensors, which is more in line with the characteristics of complex engine systems compared with single sensor or simple splicing feature methods.

[0023] (6) Facilitates engineering deployment: This method can be implemented based on existing sensor data and conventional computing platforms, and can be used as a fault mode analysis module, condition monitoring module or predictive maintenance auxiliary module in the aero-engine health management system. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart illustrating a method for analyzing the failure modes of an aero-engine based on a trivial causal graph neural network, provided in an embodiment of the present invention. Detailed Implementation

[0026] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0027] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0028] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this invention, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0029] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. The illustrations only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0030] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0031] This invention provides a method for analyzing the failure modes of aero-engines based on a trivial causal graph neural network. This method can be applied to the engine failure detection process in aero-engine system scenarios.

[0032] See Figure 1 This is a flowchart illustrating a method for analyzing the failure modes of an aero-engine based on a trivial causal graph neural network, provided by an embodiment of the present invention. Figure 1 As shown, the method mainly includes the following steps: Step 1: Acquire multi-source sensor signals of the aero-engine, wherein the multi-source sensor signals include status monitoring signals collected by multiple sensors distributed in different parts of the engine; Step 2: The multi-source sensor signals are segmented using a sliding window to obtain multiple short-time signal segments; Step 3: Using sensor components as nodes and the relationships between sensors as edges, construct the input graph and the corresponding adjacency matrix and node feature matrix based on short-time signal segments; Step 4: Use a graph neural network encoder to encode the adjacency matrix and the node feature matrix to obtain the node feature representation; Step 5: Based on the node feature representation, estimate the node-level soft mask through the node-level attention module and estimate the edge-level soft mask through the edge-level attention module; Step 6: Dewrap the input graph into a causal attention graph and a trivial attention graph based on the node-level soft mask and the edge-level soft mask. The causal attention graph gathers causal features with stable causal relationships with the fault modes, while the trivial attention graph gathers non-causal trivial features. Step 7: During the training phase, the weighted sum of the supervision loss, uniform distribution constraint loss, and causal intervention loss is used as the overall training objective. The trainable parameters of the graph neural network encoder, node-level attention module, edge-level attention module, and classifier are optimized in an end-to-end manner. Specifically, a fault category supervision loss is applied to the graph representation of the causal attention graph to enable it to learn a stable discriminative representation. A uniform distribution constraint loss is applied to the graph representation of the trivial attention graph to suppress its direct participation in fault category discrimination. In the representation layer, the causal representation of the causal attention graph is combined with multiple trivial feature estimates. The causal intervention loss constrains the classifier to maintain stable prediction under different combinations of trivial features. Step 8: In the inference phase, multi-source sensor signals of the aero-engine under test are collected, a graph representation of the causal attention map is generated and input into the classifier, and the engine failure mode analysis results are output.

[0033] In specific implementation, the technical route of the present invention can be summarized as follows: multi-source sensor signal acquisition and sliding window segmentation, signal segment graph structure construction, GNN-based node representation learning, node-level and edge-level soft mask estimation, causal attention graph and trivial attention graph untangling, causal attention graph supervised learning, trivial attention graph uniform distribution constraint, representation layer causal intervention, and fault mode classification output.

[0034] Step 1: Acquire signals from multiple sensor sources on the aero-engine. Different sensor components are distributed in different parts of the engine, forming multiple raw measurement variables.

[0035] Step 2: Use a sliding window to divide the long-running signal into multiple short-term stable signal segments, so that each group of signal segments can be used as input for primary graph structure modeling.

[0036] Step 3: Using sensor components or signal variables as nodes and the correlations, connections, or interactions between sensors as edges, construct the input graph and its adjacency matrix.

[0037] Specifically, the nodes of the graph can be determined based on the actual sensor layout, for example, vibration, temperature, pressure, speed, or other state sensors can be used as nodes respectively; the edges of the graph can be determined based on sensor physical connections, spatial proximity, historical correlation, mutual information, coherence, expert priors, or learnable adjacency matrices. If the engine component topology already exists on site, the component connection relationships and data-driven correlation relationships can also be merged to form an adjacency matrix.

[0038] Optionally, in the graph construction method, edge relationships can be determined by physical topology, sensor installation location, correlation coefficient, mutual information, coherence function, expert rules, or learnable adjacency matrix, or a heterogeneous graph with multiple relationships can be used.

[0039] Step four: Obtain node feature representations using a graph neural network encoder, and estimate the soft mask using node-level and edge-level attention modules.

[0040] Specifically, soft masks can be implemented using multilayer perceptrons, graph attention networks, gating networks, or other differentiable weight estimation modules. Node-level soft masks are used to filter sensor nodes or signal variables related to fault modes, while edge-level soft masks are used to filter connections related to fault propagation or sensor coupling. The complement of the soft masks is used to construct a trivial attention graph, enabling the causal and trivial components to be jointly learned within the same training framework.

[0041] Alternatively, graph neural network encoders can be replaced with graph convolutional networks, graph attention networks, GraphSAGE, graph Transformers, spatiotemporal graph neural networks, or other graph representation learning models.

[0042] Alternatively, alternatives to the soft mask estimation module include: attention scores for nodes and edges can be obtained from multilayer perceptrons, gating networks, attention networks, learnable parameter matrices, or sampled masking networks.

[0043] Step 5: Decompose the input graph into a causal attention graph and a trivial attention graph using a soft mask and its complement, and learn the causal representation and the trivial representation respectively.

[0044] Step 6: Apply real fault label supervision to the causal attention graph to make it focus on stable discriminative features; apply uniform distribution constraint to the trivial attention graph to make it approximate a non-causal trivial pattern.

[0045] Alternatively, alternatives to the trivial attention graph constraint method include: in addition to KL divergence and uniform distribution constraints, entropy maximization, class confusion loss, adversarial loss, or mutual information constraints can also be used to prevent the trivial attention graph from directly carrying stable class discrimination information.

[0046] Step 7: In the representation layer, the inverse adjustment idea is used to pair or combine causal features with various trivial features to form an implicit intervention graph, and the stability of causal representation is enhanced by causal intervention loss.

[0047] Specifically, causal intervention can be performed at the data layer, graph structure layer, or representation layer. Considering the irregularity of aero-engine graph data and the cost of on-site data intervention, this scheme preferably performs implicit causal intervention at the representation layer, that is, combining the causal attention graph representation with different sets of trivial feature estimates, and using intervention loss to constrain the classifier to maintain stable predictions under different combinations of disturbances.

[0048] Alternatively, alternative causal intervention methods include: intervention at the representation layer, graph structure layer, or sample augmentation layer; or combining causal representations with trivial representations from different batches, operating conditions, or equipment sources.

[0049] Step 8: Input the learned causal attention graph representation into the classifier and output the analysis results of the aircraft engine fault category, fault mode, or health status.

[0050] Specifically, a fault mode analysis model based on a trivial causal attention graph neural network (GNN) is constructed by a graph neural network encoder, node-level attention modules, edge-level attention modules, and a classifier. In aero-engine systems, complex interactions exist between components and sensors, making fault mode identification challenging. Considering that embedded components and their interactions in industrial systems can be represented as nodes and edges in a graph, graph representation algorithms have become a cutting-edge research direction in fault analysis and diagnosis. As one of the most commonly used graph representation algorithms, graph neural networks (GNNs) primarily follow the principle of "learning participation."

[0051] GNNs extract features from training data and learn the statistical correlation between features and labels, making the participation graph more conducive to accessing non-causal features. Although these non-causal features, as trivial features, can be used for downstream tasks such as prediction and classification, they are extremely unstable and depend on the data distribution characteristics in the training dataset. This reduces the generalization ability of the classifier. Based on existing causal analysis theories, certain biased features in the dataset become confounding factors (trivial features) between causal features and predictions, causing the classifier to learn incorrect correlations.

[0052] Therefore, in order to discover causal relationship patterns among various sensors and reduce the confusion effect of invalid features, this study proposes a graph attention neural network strategy based on trivial causality.

[0053] First, node and edge representations are given by estimating soft masks. Second, causal and perturbation features are obtained from the graph through decoupling. Then, the inverse adjustment parameterization of causal theory is performed, combining each causal feature with multiple perturbation features.

[0054] Engine sensing signals are typically acquired by sensor assemblies distributed across the engine. These different sensor assemblies are located in different positions, thus generating a variety of signals. One original measurement variable. In time... , No. The signal segments generated by each component are: .

[0055] However, due to the long operating time of the engine system, the resulting signal segments span a large range, which is usually difficult to process directly. Therefore, it is necessary to obtain multiple signal segments through a sliding window, which can be represented as: .

[0056] Since signal segments are stable and do not change much in a short period of time, they can be used as input for graph structure modeling.

[0057] The input graph is represented as The vertex is , side is In this system, vertices represent various sensor components in the aero-engine system, and edges represent the correlations between them.

[0058] Adjacency Matrix Used to record details of the entire graph, including if edges exist. ,but ,otherwise .

[0059] Node features can describe signal segments of a component, represented by the symbol . ,in It is the size of the signal segment.

[0060] This represents a graph neural network module, where The matrix represents the features of the nodes.

[0061] Because existing experience loss depends on the distribution characteristics and statistical correlation of the training data, this learning strategy obtains trivial features for prediction but fails to find key causal features.

[0062] To address the aforementioned issues, a trivial causal attention mechanism needs to be applied simultaneously to the input graph to distinguish between causal and trivial features.

[0063] Since causal features are the fundamental features that distinguish different fault topologies, the corresponding labels of the graph representation learned from the causal attention graph should be regarded as fundamental facts.

[0064] The trivial attention graph complements the causal attention graph. The graph representation it learns cannot adequately distinguish fault representations, so its predictions will average across all categories.

[0065] Our goal is to learn these two attention graph representations to obtain causal and trivial features, which can then be applied to fault diagnosis.

[0066] Specifically, we propose a framework for a trivial causal attention graph neural network based on multi-source signal fragment information.

[0067] First, we analyzed the problems in GNN learning from the perspective of causal relationships, and identified trivial features as the confounding factor between causal features and predictions.

[0068] Then, a causal attention graph neural network framework was proposed to reduce the confusion effect and improve the generalization ability of the model.

[0069] The framework consists of three key parts: 1) Estimate the soft mask and provide representations of nodes and edges; 2) Untangling: Obtain the causal attention map and the trivial attention map through two loss functions; 3) Causal intervention: obtain the causal intervention diagram through the inverse gating formula.

[0070] When obtaining the input image Then, the soft mask is represented as , while node features are represented as .

[0071] Given a soft mask Its complement mask can be represented as .

[0072] If the diagram is represented in another form ,in The connections in the graph are recorded, and Represents the characteristics of a node.

[0073] Then, one graph can be split into two graphs: and .

[0074] According to relevant theoretical research, the category of a graph can usually be derived from more fundamental causal features. Therefore, attention graphs that aggregate causal features are defined as causal attention graphs. The corresponding graph is defined as a trivial attention graph. .

[0075] However, in practical applications, attention maps containing real information cannot usually be used directly. Therefore, it is necessary to obtain these two types of attention maps by learning masks: and To achieve causal reverse gating adjustment, we propose a trivial causal attention graph neural network.

[0076] First, an attention module is needed to filter out causal and trivial features. Then, based on the obtained features, causal candidates and trivial candidates are generated.

[0077] The GNN-based encoder is represented as The input image is represented as The nodes are represented as follows: ; To obtain attention scores, we can approach it from both the node and edge levels. Then, we use two multilayer perceptrons (MLPs): and .

[0078] For nodes and edge We can obtain: ; ; in, In the causal attention graph node Node-level attention score and edge The edge attention score, This is used for trivial attention maps. It is the softmax activation function. This indicates a splicing operation.

[0079] Obviously, , Attention score They were used to construct soft masks. .

[0080] Therefore, the initial causal attention graph and trivial attention graph can be constructed as follows: and .

[0081] To obtain causal and trivial attention maps, representations of the attention maps can be obtained using GNN modules respectively. Finally, the class of the input map is predicted using a readout function and a classifier. To obtain graph-level representations of causal attention maps and trivial attention maps, the graph structure and node features after soft masking can be encoded using a graph neural network encoding module, and the corresponding graph-level feature representations can be obtained through a graph-level readout function. Subsequently, the graph-level feature representations are input into a classifier to obtain the corresponding fault category prediction results. Specifically, this is represented as follows: ; ; ; ; in, This is a causal attention graph. For trivial attention graphs; Input image The adjacency matrix is ​​used to characterize the connection relationships between the sensor nodes of the aero-engine. Input image The node feature matrix is ​​used to characterize the signal segment features corresponding to each sensor node.

[0082] and These are the causal attention soft masks learned for edges and nodes, respectively. Used to perform weighted filtering of edge connections in a graph. Used for weighted filtering of node features. and These are complementary masks for the edge-level causal attention soft mask and the node-level causal attention soft mask, respectively, used to extract trivial features not preserved by the causal attention soft mask.

[0083] This represents the graph neural network encoding operation used for causal attention graphs, whose input is the graph connectivity filtered by causal attention soft masks. and node features ; This represents the graph neural network encoding operation used for trivial attention graphs, whose input is the graph connectivity filtered by complementary masks. and node features .

[0084] This is a graph-level readout function used to aggregate the node representations output by the graph neural network to obtain the graph-level feature representation of the entire graph. Causal attention graph The causal graph-level feature representation obtained after graph neural network encoding and graph-level readout; For ordinary attention graphs The trivial graph-level feature representation obtained after graph neural network encoding and graph-level readout.

[0085] and These represent the classifier mapping functions applied to the causal branch and the trivial branch, respectively; The classifier is based on causal graph-level feature representation The output fault category prediction probability vector, The classifier is based on trivial graph-level features. The output is a probability vector predicting the fault categories. If the total number of fault categories is... ,but and All include The predicted probabilities for each category.

[0086] The purpose of causal attention graphs is to estimate causal features, which can be classified as true labels. Accordingly, the supervised loss on the graph classification problem is defined as: ; Conversely, trivial attention graphs aim to approximate non-causal trivial patterns. Therefore, for all known fault categories, the predictions of trivial attention graphs can be incentivized, and a unified loss can be defined for graph classification problems: ; ; Where KL represents the KL divergence, The distribution is uniform. The two objectives mentioned above are used to distinguish between causal and trivial features. However, related research shows that real-world graph data contains noise, which undoubtedly leads to a greater correlation between the causal part and the label than the correlation between the entire graph and the label. Furthermore, due to the existence of trivial patterns, the causal attention graph obtained through the above decoupling method is unlikely to eventually converge to the complete graph.

[0087] Reverse adjustment can effectively reduce the confusion effect by stratifying the confounding factors and pairing each layer of the target causal attention map with a trivial attention map, thus forming an intervention map.

[0088] Due to the irregularity of graph data, we cannot intervene at the data level, but can only intervene implicitly at the representation level. Therefore, we propose a loss guided by inverse adjustment.

[0089] ; ; This formula is called the causal intervention loss function at the representation level, where It is an implicit intervention diagram In the classifier The classification results. It is a causal attention graph The expression . It is layered The expression . It is a set of estimates of trivial features present in the training data. Due to the shared nature of causal features, it enables intervention maps to make predictions stably across different strata.

[0090] Finally, the overall learning objective of our designed method is as follows: ; in, and These are constants used to control the degree of untangling and causal intervention; they are adjustable hyperparameters.

[0091] The structural parameters of the model are shown in Table 1.

[0092] Table 1

[0093] During the training phase, the trainable parameters of the graph neural network encoder, node-level attention module, edge-level attention module, and classifier were optimized in an end-to-end manner, as shown in Table 2.

[0094] Table 2

[0095] The fault mode analysis model based on trivial causal attention graph neural networks accurately captures the complex correlations and causal relationships between different sensors in an aero-engine system by modeling information from multi-source sensor signal segments. This graph-based modeling method can more comprehensively capture the temporal and spatial relationships between sensor data, thus helping to reveal fault modes and anomaly characteristics. Furthermore, the designed method effectively solves the causal confusion problem in conventional graph neural networks by utilizing the difference between causal and trivial features. By learning different graph representations from causal and trivial features, this method can effectively distinguish different fault modes, improving the model's generalization performance and accuracy. This is crucial for engine fault analysis. Simultaneously, the fault mode analysis model based on trivial causal attention graph neural networks can further enhance its performance through causal intervention. By introducing causal intervention, the model can better understand causal relationships and more accurately analyze different fault modes.

[0096] Alternatively, alternative input signal types include: in addition to aero-engine sensor signals, the method can also be extended to multi-source condition monitoring and fault diagnosis scenarios for gas turbines, rotating machinery, pumps, compressors or other complex equipment.

[0097] Alternatively, the classification target could be an output that includes fault category, fault location, fault severity, health status level, abnormal mode, or remaining life-related status label.

[0098] The innovative aspect of this invention lies in: (1) The problem of fault diagnosis of multi-source sensors of aero-engine is modeled as a graph structure causal representation learning problem. Sensor components are used as nodes and the correlation or interaction between sensors are used as edges, so that the model can express the local features of sensors and the relationship between multiple sensors at the same time.

[0099] (2) A trivial causal attention graph neural network framework for aero-engine failure mode analysis is proposed. The non-causal trivial features that ordinary GNNs may learn are explicitly regarded as confounding factors and suppressed through model structure.

[0100] (3) By using node-level and edge-level soft masks, the input graph is unwound into a causal attention graph and a trivial attention graph, so that the model can learn stable causal representations and non-causal trivial representations respectively.

[0101] (4) Design the supervised classification loss of the causal attention graph and the uniform distribution loss of the trivial attention graph, so that the causal representation can undertake the fault discrimination function, while the trivial representation is constrained to the interference mode that does not directly distinguish the categories.

[0102] (5) Introduce a representation layer reverse adjustment / causal intervention mechanism to combine causal features with multiple trivial features to form an implicit intervention graph, thereby reducing the confounding effect of trivial features on fault prediction results.

[0103] (6) Unify the untangling loss, causal supervision loss and trivial feature intervention loss into the same training objective, so that the model can be trained end-to-end and output the failure mode analysis results.

[0104] The aero-engine fault mode analysis method based on trivial causal graph neural networks provided in this embodiment constructs graph structure data of multi-source sensor signals of aero-engines using sensor components as nodes and inter-sensor relationships as edges, providing a structured input that preserves the system coupling relationship for subsequent causal analysis. By using node-level and edge-level soft masks, the input graph is dewrapped into causal attention graphs and trivial attention graphs, enabling the model to explicitly distinguish causal features with stable causal relationships to fault modes from non-causal trivial features that only exist in the training data distribution during training. By applying a fault category supervision loss to the causal attention graph to focus on stable discriminative representations, and simultaneously applying a uniform distribution constraint loss to the trivial attention graph to suppress its direct participation in fault discrimination, the method effectively solves the problems of traditional graph neural networks. This addresses the issue of decreased generalization caused by the indiscriminate use of all statistical correlations in network architectures. By combining causal features with multiple trivial features and intervening in causality at the representation layer, the classifier can maintain stable predictions under different combinations of trivial features, thus significantly reducing the interference of trivial features and confounding factors on classification decisions during the training phase. Consequently, during the inference phase, only the graph representation of the causal attention map needs to be input into the classifier, which can maintain high accuracy and robustness in fault mode recognition even under application conditions that vary across operating conditions, equipment, or data distribution. At the same time, node-level and edge-level attention scores can also help locate key sensors and potential fault propagation paths related to fault modes, providing a technical solution for aero-engine health management that combines stability, accuracy, and interpretability.

[0105] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof.

[0106] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for aero-engine fault mode analysis based on a trivial causal graph neural network, characterized in that, include: Step 1: Acquire multi-source sensor signals of the aero-engine, wherein the multi-source sensor signals include status monitoring signals collected by multiple sensors distributed in different parts of the engine; Step 2: The multi-source sensor signals are segmented using a sliding window to obtain multiple short-time signal segments; Step 3: Using sensor components as nodes and the relationships between sensors as edges, construct the input graph and the corresponding adjacency matrix and node feature matrix based on short-time signal segments; Step 4: Use a graph neural network encoder to encode the adjacency matrix and the node feature matrix to obtain the node feature representation; Step 5: Based on the node feature representation, estimate the node-level soft mask through the node-level attention module and estimate the edge-level soft mask through the edge-level attention module; Step 6: Dewrap the input graph into a causal attention graph and a trivial attention graph based on the node-level soft mask and the edge-level soft mask. The causal attention graph gathers causal features with stable causal relationships with the fault modes, while the trivial attention graph gathers non-causal trivial features. Step 7: During the training phase, the weighted sum of the supervision loss, uniform distribution constraint loss, and causal intervention loss is used as the overall training objective. The trainable parameters of the graph neural network encoder, node-level attention module, edge-level attention module, and classifier are optimized in an end-to-end manner. Specifically, a fault category supervision loss is applied to the graph representation of the causal attention graph to enable it to learn a stable discriminative representation. A uniform distribution constraint loss is applied to the graph representation of the trivial attention graph to suppress its direct participation in fault category discrimination. In the representation layer, the causal representation of the causal attention graph is combined with multiple trivial feature estimates. The causal intervention loss constrains the classifier to maintain stable prediction under different combinations of trivial features. Step 8: In the inference phase, multi-source sensor signals of the aero-engine under test are collected, a graph representation of the causal attention map is generated and input into the classifier, and the engine failure mode analysis results are output.

2. The method of claim 1, wherein, The correlation between the sensors includes one or more of the following: physical connection relationship, spatial proximity relationship, historical correlation coefficient, mutual information, coherence function value, expert prior relationship, and learnable adjacency matrix; The input graph has a graph structure that is either isomorphic or heteromorphic.

3. The method of claim 1, wherein, The estimation methods for the node-level soft mask and the edge-level soft mask are as follows: Using a multilayer perceptron to analyze node features Features of splicing nodes at both ends of the edge After transformation, the softmax activation function outputs causal attention scores and trivial attention scores, satisfying... and ; wherein, constitute a causal soft mask, constitute a trivial soft mask, the trivial soft mask being a complement of the causal soft mask.

4. The method according to claim 3, characterized in that, The causal attention graph is constructed as follows: The method for constructing the trivial attention graph is as follows: in, It is an adjacency matrix. The node feature matrix, and These are node-level soft masks and edge-level soft masks, respectively. This represents the complement of the soft mask. This indicates an element-wise multiplication operation.

5. The method according to claim 4, characterized in that, The fault category supervision loss is represented as follows: in, This is a set of graph samples used for model training. The number of samples in the graph sample set; For the graph sample set Any input graph in the dataset; To based on the input image And the causal attention graph constructed using causal attention soft masks; Input image The corresponding real fault category label vector is a one-hot encoded vector; Total number of fault categories; For classifiers to apply causal attention graphs Output fault category probability vector; superscript Indicates vector transpose; This indicates that the logarithm is performed on each element of the probability vector. ; The uniform distribution constraint loss is represented by KL divergence as follows: in, To based on the input image And the trivial attention graph constructed from the complementary mask of the causal attention soft mask; For the classifier's attention graph of trivialities Output fault category probability vector; It is a uniformly distributed vector, and all its elements are... ; The Kullback–Leibler divergence measures the difference between the predicted probability distribution of a trivial attention map and a uniform distribution. The first uniformly distributed vector is the... The probability corresponding to each fault category For the prediction probability vector of a trivial attention map, the first... Predicted probability for each fault category .

6. The method according to claim 5, characterized in that, The expression for the causal intervention loss is: in, It is an implicit intervention diagram In the classifier The classification results in It is a causal attention graph The expression, It is layered The expression, It is a set of estimates of trivial features present in the training data.

7. The method according to claim 6, characterized in that, The expression for the overall training objective is: in, and It is a constant used to control the degree of untangling and causal intervention.

8. The method according to claim 1, characterized in that, The graph neural network encoder is any one of graph convolutional network, graph attention network, GraphSAGE, graph Transformer, or spatiotemporal graph neural network.

Images