Disclosure of Invention
The embodiment of the invention provides a fault diagnosis method and system for a lightweight electric propulsion system, which can solve the problems in the prior art.
In a first aspect of an embodiment of the present invention,
Provided is a fault diagnosis method for a lightweight electric propulsion system, comprising:
Determining a fault causal relation and a component dependency relation based on historical data and structural data of the lightweight electric propulsion system, constructing a fault tree and a directed graph, constructing an initial graph model of the lightweight electric propulsion system based on the fault tree and the directed graph, mapping nodes in the initial graph model into a low-dimensional dense vector space, determining node embedded representation, and generating a fault knowledge graph;
based on the node embedded representation in the fault knowledge graph, searching to obtain a corresponding neighbor embedded representation, extracting node high-level characteristic representation by carrying out convolution operation on the node and the neighbor, determining a fault association relationship, extracting fault importance representation by pooling operation, and constructing a fault association characteristic extraction network;
inputting the real-time monitoring data of the lightweight electric propulsion system into a fault diagnosis model, extracting fault correlation characteristics and importance characteristics corresponding to the real-time monitoring data, carrying out classification decision by combining the fault knowledge graph through a fault classifier, and determining a fault diagnosis result, wherein the fault diagnosis model is constructed by integrating the fault correlation characteristics extraction network and the fault classifier based on the fault knowledge graph.
In an alternative embodiment of the present invention,
Determining a fault causal relationship and a component dependency relationship based on historical data and structural data of the lightweight electric propulsion system, constructing a fault tree and a directed graph, and constructing an initial graph model of the lightweight electric propulsion system based on the fault tree and the directed graph comprises:
Carrying out statistical analysis on the historical data, identifying a fault mode and occurrence frequency, tracing a fault reason, determining a fault causal relationship, taking a system-level fault as a top event, a component-level fault as an intermediate event and a root cause as a basic event, generating logic nodes of a fault tree, determining a logic edge based on the fault causal relationship and a causal strength factor, connecting all the logic nodes, and constructing the fault tree;
Identifying physical components according to the structural data, determining directed graph nodes, constructing a dependency relationship based on interfaces of the physical components and the directions of substance transmission and information interaction, determining directed graph directed edges in combination with dependency strength factors, and determining directed graphs based on the directed graph nodes and the directed graph directed edges;
And integrating the fault tree with the directed graph, and generating an initial graph model of the lightweight electric propulsion system through node merging, edge merging and attribute addition, wherein nodes of the initial graph model comprise fault events and system components, and edges of the initial graph model comprise fault causal relationships and component dependency relationships.
In an alternative embodiment of the present invention,
Mapping the nodes in the initial graph model into a low-dimensional dense vector space, determining node embedded representation, and generating a fault knowledge graph comprises:
determining the edge weight of the initial graph model based on the causal strength factor and the dependent strength factor, calculating a node transition probability matrix, and performing random walk from a starting node based on the transition probability matrix to generate a plurality of random walk sequences;
Based on the random walk sequence, taking each node as a central node, determining context nodes according to a preset window size based on the central nodes, constructing a search sample pair of the central nodes and the context nodes, applying a pre-constructed word vector model, taking the search sample pair as input, and capturing an embedded representation vector of each node in the initial graph model by maximizing the probability of correctly predicting the central nodes and the context nodes;
Taking the embedded expression vector as a new attribute of the node, integrating the new attribute with the original attribute of the node to form comprehensive feature expression of the node, calculating similarity among the nodes based on the embedded expression vector, extracting a semantic correlation node pair, and adding a semantic correlation edge between the semantic correlation node pair; and integrating the comprehensive characteristic representation and the semantic association side of the initial graph model with the original side of the initial graph model to construct a fault knowledge graph of the light-weight electric propulsion system.
In an alternative embodiment of the present invention,
The fault associated feature extraction network comprises:
Constructing a node characteristic matrix based on node embedded expression in the fault knowledge graph, determining an adjacency matrix according to the edge connection relation of the fault knowledge graph, and dividing a plurality of sub-adjacency matrices according to the types of edges;
Inputting the node characteristic matrix and the sub-adjacent matrix into a pre-constructed multi-layer heterogeneous graph convolution network, determining a corresponding convolution kernel according to the type of the node, determining attention weight according to the type of the edge, extracting heterogeneous neighborhood information and associated features of the node layer by layer based on the convolution kernel and the attention weight, and constructing a heterogeneous graph convolution characteristic matrix;
The heterogeneous graph convolution feature matrix calculates attention weights among nodes through a multi-head graph attention mechanism, gathers node neighborhood information based on the attention weights, extracts graph attention feature matrices, calculates self-attention weights of the nodes through a self-attention pooling mechanism, gathers node features, determines graph level feature representations and obtains fault association features.
In an alternative embodiment of the present invention,
The training of the fault diagnosis model comprises the following steps:
Acquiring marked historical fault data, wherein the historical fault data comprises fault association features corresponding to fault scenes and corresponding fault labels;
Initializing model parameters of a fault diagnosis model;
Constructing a fault classification loss function based on the fault classification loss and the graph regularization term;
Inputting the fault correlation characteristics into the fault classifier, and calculating the total loss of a fault diagnosis model through forward propagation based on the fault classification loss function;
And calculating the parameter gradients of the fault correlation characteristic extraction network and the fault classifier through back propagation according to the total loss of the fault diagnosis model, and updating model parameters based on the parameter gradients until a preset model training round is reached, so that training of the fault diagnosis model is completed.
In an alternative embodiment of the present invention,
Based on the fault classification loss and the graph regularization term, constructing a fault classification loss function includes:
；
Wherein L represents total loss, k represents sample number, C represents failure category number, C represents failure category total number, yk,c represents real label of kth sample belonging to the C-th category, pk,c represents probability of model prediction of kth sample belonging to the C-th category, lambda1 represents weight coefficient of graph regularization term, r represents relationship type number, i represents one node, j represents another node, Hi,rj represents the representation of node i and node j under the r-th relationship type, Xi,rj represents the characteristics of node i and node j under the r-th relationship type, Wi,rj represents the elements of the characteristic transformation matrix of node i and node j under the r-th relationship type, sigmar represents the weight coefficient of the loss term under each relationship type in the regularization term of the adjustment graph, lambda2 represents the weight coefficient of the L2 regularization term, and MF2 represents the L2 regularization term of the model parameter.
In a second aspect of an embodiment of the present invention,
Provided is a fault diagnosis system for a lightweight electric propulsion system, comprising:
The first unit is used for determining a fault causal relationship and a component dependency relationship based on historical data and structural data of the lightweight electric propulsion system, constructing a fault tree and a directed graph, constructing an initial graph model of the lightweight electric propulsion system based on the fault tree and the directed graph, mapping nodes in the initial graph model into a low-dimensional dense vector space, determining node embedded representation and generating a fault knowledge graph;
The second unit is used for searching and obtaining corresponding neighbor embedded representations based on the node embedded representations in the fault knowledge graph, extracting node high-level characteristic representations through convolution operation on the nodes and the neighbors, determining a fault association relationship, extracting fault importance representations through pooling operation, and constructing a fault association characteristic extraction network;
The third unit is used for inputting the real-time monitoring data of the lightweight electric propulsion system into a fault diagnosis model, extracting fault correlation features and importance features corresponding to the real-time monitoring data, carrying out classification decision by combining the fault knowledge graph through a fault classifier, and determining a fault diagnosis result, wherein the fault diagnosis model is constructed by integrating the fault correlation feature extraction network and the fault classifier based on the fault knowledge graph.
In a third aspect of an embodiment of the present invention,
There is provided an electronic device including:
A processor;
A memory for storing processor-executable instructions;
wherein the processor is configured to invoke the instructions stored in the memory to perform the method described previously.
In a fourth aspect of an embodiment of the present invention,
There is provided a computer readable storage medium having stored thereon computer program instructions which, when executed by a processor, implement the method as described above.
In the embodiment of the invention, the propagation path and the influence range of the fault state in the system are reflected through the directed graph, so that basis is provided for fault tracing, isolation and blocking; the fault tree and the directed graph are further abstracted into graph models, and the graph models are mapped into a low-dimensional vector space through a node embedding technology, so that compact and efficient fault knowledge graph representation is formed, and subsequent storage, retrieval and reasoning application are facilitated; by utilizing the constructed fault knowledge graph, intelligent fault diagnosis based on methods such as causal reasoning, path searching and the like can be realized, potential fault risks are predicted through associated information in the graph, and preventive maintenance is guided; by utilizing the graph neural network technology, the embedded representation of the nodes in the fault knowledge graph can be effectively obtained, and the structural similarity and semantic similarity between the nodes are captured; the neighbor nodes of the nodes are obtained through searching, and the embedded representations of the neighbor nodes are obtained, so that the context information of the nodes is enriched, and the richness of the feature representation is improved; the fault classifier utilizes the fault associated features to extract the features extracted by the network, and can quickly and effectively classify and decide the real-time monitoring data by combining the existing fault knowledge graph, thereby improving the accuracy and efficiency of fault classification; the integrated design of the fault diagnosis model can effectively improve the fault diagnosis performance of the light electric propulsion system, reduce the cost and time cost of fault diagnosis and improve the reliability and stability of the system.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
The technical scheme of the invention is described in detail below by specific examples. The following embodiments may be combined with each other, and some embodiments may not be repeated for the same or similar concepts or processes.
Fig. 1 is a flow chart of a fault diagnosis method of a lightweight electric propulsion system according to an embodiment of the present invention, as shown in fig. 1, the method includes:
s101, determining a fault causal relation and a component dependency relation based on historical data and structural data of the lightweight electric propulsion system, constructing a fault tree and a directed graph, constructing an initial graph model of the lightweight electric propulsion system based on the fault tree and the directed graph, mapping nodes in the initial graph model into a low-dimensional dense vector space, determining node embedded representation, and generating a fault knowledge graph;
The historical data specifically refers to various data recorded in the past operation process of the lightweight electric propulsion system, including: sensor data, such as physical quantity records of temperature, pressure, flow, voltage, current and the like; fault logs, recording time, position, fault phenomenon, fault codes and the like of the system; maintenance records, recording maintenance processes, replacement parts, maintenance time and the like after faults occur; performance data such as historical variation of key performance indexes including thrust, specific impulse, efficiency and the like;
The structural data specifically refer to data describing the internal structure and composition of the lightweight electric propulsion system, and the data comprises the following components: the system topology structure reflects the connection relation and the signal flow direction among all the components in the system; a component hierarchy describing component components of the system and hierarchical dependencies thereof; component parameters such as intrinsic property data for the materials, dimensions, nominal operating conditions, etc. of the component.
The failure causal relation specifically refers to causal relation that the failure or abnormal state of one component or parameter in the system causes the failure or abnormal state of the other component or parameter; the component dependency relationship specifically means that the normal operation of one component in the system depends on the normal operation of another component, thus reflecting the dependency and influence of the components on functions; the fault tree specifically refers to a top-down logic analysis method, a tree structure is used for representing causal relations among various faults in a system, a root node represents a system-level fault, an intermediate node represents a subsystem or component-level fault, and a leaf node represents a basic fault event; the directed graph specifically refers to a propagation path of a fault or an abnormal state in a system by using a directed graph, a node represents the fault state, a directed edge represents a causal relationship or a dependency relationship between the fault states, and an edge direction represents a fault propagation direction.
In the embodiment, the fault conditions and the mutual influence of the fault conditions of each layer of the system are comprehensively considered, so that the fault occurrence and evolution rules of the system are comprehensively understood and mastered; the propagation path and the influence range of the fault state in the system are reflected through the directed graph, and basis is provided for fault tracing, isolation and blocking; the fault tree and the directed graph are further abstracted into graph models, and the graph models are mapped into a low-dimensional vector space through a node embedding technology, so that compact and efficient fault knowledge graph representation is formed, and subsequent storage, retrieval and reasoning application are facilitated; by utilizing the constructed fault knowledge graph, intelligent fault diagnosis based on methods such as causal reasoning, path searching and the like can be realized, potential fault risks are predicted through associated information in the graph, and preventive maintenance is guided; new fault cases and expert experiences are continuously added into the fault knowledge graph, so that continuous accumulation and optimization of a diagnosis knowledge base are realized, and the capability of the system for coping with new fault types is improved.
In an alternative embodiment, determining a fault causal relationship and a component dependency relationship based on historical data and structural data of the lightweight electric propulsion system, constructing a fault tree and a directed graph, and constructing an initial graph model of the lightweight electric propulsion system based on the fault tree and the directed graph includes:
Carrying out statistical analysis on the historical data, identifying a fault mode and occurrence frequency, tracing a fault reason, determining a fault causal relationship, taking a system-level fault as a top event, a component-level fault as an intermediate event and a root cause as a basic event, generating logic nodes of a fault tree, determining a logic edge based on the fault causal relationship and a causal strength factor, connecting all the logic nodes, and constructing the fault tree;
Identifying physical components according to the structural data, determining directed graph nodes, constructing a dependency relationship based on interfaces of the physical components and the directions of substance transmission and information interaction, determining directed graph directed edges in combination with dependency strength factors, and determining directed graphs based on the directed graph nodes and the directed graph directed edges;
And integrating the fault tree with the directed graph, and generating an initial graph model of the lightweight electric propulsion system through node merging, edge merging and attribute addition, wherein nodes of the initial graph model comprise fault events and system components, and edges of the initial graph model comprise fault causal relationships and component dependency relationships.
In the fault tree, the causal relation between the events is not a simple equal weight relation, but a certain causal intensity difference exists, so the causal intensity factor can be a weight value used for representing the causal relation intensity between the events, and the causal intensity factor can be determined through analysis of historical data or expert experience, so that different influence degrees between the events are considered when the fault tree is constructed;
In the initial graph model of the system, the dependency relationship among the components is important, and reflects the interaction among the components in the system and the importance degree of information transmission, so the dependency strength factor can be a weight value used for representing the strength of the dependency relationship among the components, and the dependency strength factor can be determined through analysis of the system structure data, so that the dependency degree among the components is considered when the initial graph model is constructed;
Firstly, constructing a fault tree, carrying out statistical analysis on historical data of a lightweight electric propulsion system, identifying a main fault mode and occurrence frequency thereof in the system through indexes such as fault frequency, fault influence and the like, tracing the occurrence reason of each fault mode, determining a causal relation chain of faults through causal analysis methods such as root cause analysis, 5why analysis and the like, taking a system-level fault as a top event of the fault tree, taking a component-level fault in the fault causal chain as an intermediate event, taking the lowest fault cause as a basic event, generating logic nodes of the fault tree, connecting directed edges between the logic nodes of the fault tree according to the determined fault causal relation, adding logic gates (such as AND gates, OR gates and the like) on the directed edges to represent the combined logic relation among a plurality of events, quantitatively describing the strength degree of the causal relation, and measuring by indexes such as conditional probability, correlation coefficient and the like to form a complete fault tree representation;
Reconstructing a pointing graph, and identifying all physical components in the system as nodes of the pointing graph according to the structural data of the light electric propulsion system; analyzing the interface relation of each physical component, determining the directions of substance transmission (such as fluid flow, heat conduction and the like) and signal interaction (such as control instructions, feedback information and the like) between the physical components, connecting directed edges between corresponding nodes based on the physical interfaces and the interaction directions, representing the dependency relation between the components, wherein the direction of the edges represents the dependent direction, adding a dependent strength factor on the directed edges, quantitatively describing the tightness degree of the dependency relation, and measuring by indexes such as coupling degree, interconnection degree and the like to form a complete directed graph representation;
generating an initial graph model, integrating a constructed fault tree with a directed graph, merging nodes representing the same entity (fault event or system component), selectively merging causal relation edges in the fault tree and dependent relation edges in the directed graph according to whether the connected nodes are the same, adding attribute information such as severity, occurrence probability of the fault event, materials, sizes and the like of the components on the merged graph model nodes, merging causal strength factors and dependent strength factors on the merged graph model edges to form comprehensive associated strength weights, further simplifying and optimizing the merged graph model, removing redundant nodes and edges, and adjusting the attributes of the nodes and the edges to obtain the initial graph model of the light electric propulsion system;
In the embodiment, through statistical analysis of historical data, the main failure modes and the occurrence frequencies of the main failure modes in the system can be accurately identified, so that a system manager is helped to know the main risks existing in the system, and preventive and countermeasure measures are purposefully carried out; the cause of each fault mode can be deeply analyzed through the cause and effect analysis method, and a cause and effect relationship chain of the fault is determined, so that a system manager is facilitated to know the root cause of the fault and take corresponding measures for repair and prevention; the system level fault is taken as a top event, the component level fault is taken as an intermediate event, the fault reason is taken as a basic event, the logic nodes of the fault tree are generated, the complete fault tree representation is formed by connecting the logic nodes and adding the directed edges and the logic gates, the causal relationship and the logic association between different fault events in the system can be clearly shown, all physical components in the system are identified according to the system structure data, the interface relationship between the physical components is analyzed, the material transfer and the signal interaction direction is determined, and the dependency relationship between the directed graph representation components is constructed.
In an alternative embodiment, mapping nodes in the initial graph model into a low-dimensional dense vector space, determining a node embedded representation, generating a fault knowledge-graph includes:
determining the edge weight of the initial graph model based on the causal strength factor and the dependent strength factor, calculating a node transition probability matrix, and performing random walk from a starting node based on the transition probability matrix to generate a plurality of random walk sequences;
Based on the random walk sequence, taking each node as a central node, determining context nodes according to a preset window size based on the central nodes, constructing a search sample pair of the central nodes and the context nodes, applying a pre-constructed word vector model, taking the search sample pair as input, and capturing an embedded representation vector of each node in the initial graph model by maximizing the probability of correctly predicting the central nodes and the context nodes;
Taking the embedded expression vector as a new attribute of the node, integrating the new attribute with the original attribute of the node to form comprehensive feature expression of the node, calculating similarity among the nodes based on the embedded expression vector, extracting a semantic correlation node pair, and adding a semantic correlation edge between the semantic correlation node pair; and integrating the comprehensive characteristic representation and the semantic association side of the initial graph model with the original side of the initial graph model to construct a fault knowledge graph of the light-weight electric propulsion system.
The transition probability matrix specifically describes transition probabilities among nodes in the graph structure, for a node i, the ith row of the transition probability matrix represents probability distribution from the node i to other nodes, and the transition probability matrix is usually a sparse matrix, wherein non-zero elements represent transition probabilities among the nodes with connection;
The random walk sequence specifically refers to that a node is randomly selected as an initial node in the graph structure, then random movement is carried out in the graph according to a certain transfer rule until a stop condition is met, the random walk sequence is a node sequence passing through in the process of recording the random walk, and the random transfer process between the nodes in the graph structure can be simulated from different initial nodes by generating a plurality of random walk sequences, so that local information and global information of the graph structure are captured.
For each edge in the initial graph model, according to the type of the connected node (fault event or system component), extracting the corresponding causal intensity factor or dependent intensity factor from the fault tree or the directed graph, normalizing the extracted intensity factor to be between 0 and 1, adding the intensity factor as an edge weight value into the edge attribute of the initial graph model, and calculating a node transition probability matrix of the initial graph model based on the edge weight. Counting the sum of the weights of all the outgoing edges of each node, dividing the weight of each outgoing edge by the sum of the weights to obtain the probability of transferring from the node to the adjacent node, forming one row of a transfer probability matrix, and carrying out normalization processing on the transfer probability matrix to ensure that the sum of the probabilities of each row is 1;
Randomly selecting a node from the initial graph model as an initial node, starting random walk, selecting the next node in a probability mode according to a transition probability matrix of the current node, forming the next step of random walk, repeating the walk until the walk reaches a preset length or meets a specific stopping condition (such as repeatedly accessing the same node, etc.), generating a complete random walk sequence, and circularly generating a plurality of random walk sequences;
For each random walk sequence, traversing each node, taking the current node as a central node, taking the central node as a core, selecting a certain number of nodes before and after the central node as context nodes according to a preset window size (such as 3, 5 and the like), combining the central node and each context node into a exploration sample pair, and repeating the operation on all nodes in all random walk sequences to construct a complete exploration sample pair set;
The method comprises the steps of using a pre-constructed Word vector model, preferably Word2Vec, to explore a sample pair as input, learning an embedded representation vector of each node in an initial graph model, optimizing the Word vector model by maximizing the probability of correctly predicting a central node and a context node, thereby capturing structural and semantic similarity between the nodes, setting proper Word vector dimensionality, training iteration times, learning rate and other super-parameters to balance the quality and the computing efficiency of representation, mapping each node into a low-dimensional dense real-value vector space after training is completed, and reflecting the similarity degree of the nodes in the graph structure by the distance between the vectors.
And taking the embedded expression vector learned by each node as a new attribute of the node, integrating the new attribute with the original attribute of the node, carrying out necessary pretreatment such as normalization, discretization and the like on the original attribute, coordinating the original attribute with the embedded expression vector on a numerical scale, splicing or combining the embedded expression vector with the processed original attribute to form a comprehensive feature expression of the node, and reflecting the multidimensional feature of the node.
Based on the embedded expression vector of the nodes, calculating the similarity between each pair of nodes in the initial graph model, preferably, calculating cosine similarity, extracting node pairs with higher semantic relevance as candidate semantic relevance edges according to the similarity, screening and filtering the candidate semantic relevance edges, removing repeated edges with low confidence to obtain a final semantic relevance edge set, adding the semantic relevance edges into the initial graph model, and forming a new edge set with original causal relation edges and dependency relation edges;
Updating the comprehensive characteristic representation of the nodes into the node attribute of the initial graph model, replacing or expanding the original attribute, integrating the newly added semantic association edge with the original edge of the initial graph model to form an expanded edge set, carrying out necessary optimization and adjustment on the integrated graph model, such as removing isolated nodes, combining redundant edges and the like, improving the quality and efficiency of the graph, storing the optimized graph model as a fault knowledge graph of the lightweight electric propulsion system, wherein each node contains the comprehensive characteristic representation, the edges contain causal relationship, dependency relationship and semantic association relationship, designing a proper storage mode and retrieval interface for the knowledge graph, and facilitating subsequent query, reasoning and application.
In the embodiment, the edge weight of the initial graph model is determined based on the calculation of the causal strength factor and the dependent strength factor, so that a node transition probability matrix is constructed, and the information transfer and the association degree between nodes in a simulation system are facilitated; generating a plurality of random walk sequences in a graph model through random walk to help reveal association modes and paths among nodes in the system; constructing an exploration sample pair set, wherein the sample pair combines a central node and surrounding context nodes, thereby facilitating subsequent node embedded representation learning; learning an embedded representation vector of each node in the initial graph model through a pre-trained word vector model, and reflecting the structural similarity and semantic similarity of the nodes in the graph structure; integrating the embedded expression vector of the node with the original attribute to form a comprehensive node characteristic expression, which is helpful for describing the characteristics of the node more comprehensively, extracting node pairs with higher semantic relevance, and adding the node pairs as semantic relevance edges into a graph model to enrich the information of the graph; and updating the comprehensive characteristic representation of the nodes into a graph model, optimizing and adjusting the graph model, and finally constructing a fault knowledge graph of the lightweight electric propulsion system, thereby providing comprehensive support for fault diagnosis and decision of the system.
S102, searching and obtaining corresponding neighbor embedded representations based on node embedded representations in the fault knowledge graph, extracting node high-level feature representations through convolution operation on the nodes and the neighbors, determining a fault association relationship, extracting fault importance representations through pooling operation, and constructing a fault association feature extraction network;
Extracting embedded expression vectors of each node from a fault knowledge graph as characteristic expression of the node, searching and acquiring the embedded expression vectors of neighboring nodes of each node based on the side relation of the knowledge graph, combining the embedded expressions of the node and the neighboring nodes into a local subgraph, applying graph convolution operation to the local subgraph, extracting high-level characteristic expression of the node by aggregating characteristic information of the node and the neighboring nodes, and determining fault association relation between the nodes by similarity calculation, threshold screening and other methods based on the extracted high-level characteristic of the node;
pooling the high-level features of the nodes extracted by convolution, aggregating the feature information of the nodes, and extracting importance representation of the nodes; the fault importance reflects the key degree of the node in the whole fault propagation process, convolution operation and pooling operation are combined into a complete neural network model to form a fault associated feature extraction network, the node embedded expression is used as input, and the high-level associated features and importance features of the node are extracted through multi-layer convolution and pooling;
in the embodiment, the embedded representation of the nodes in the fault knowledge graph can be effectively obtained by utilizing the graph neural network technology, and the structural similarity and semantic similarity between the nodes are captured; the neighbor nodes of the nodes are obtained through searching, and the embedded representations of the neighbor nodes are obtained, so that the context information of the nodes is enriched, and the richness of the feature representation is improved; by utilizing convolution operation, the high-level characteristic representation of the nodes and the neighbor nodes thereof can be effectively extracted, and the association relation and characteristic interaction between the nodes can be captured; by carrying out convolution operation on the nodes and the neighbors, the association relation between faults can be effectively determined, and the mutual influence and propagation paths between the faults are revealed; by pooling, a fault importance representation can be extracted from the high-level features of the nodes, helping to identify and distinguish important fault events.
In an alternative embodiment, the fault-associated feature extraction network includes:
Constructing a node characteristic matrix based on node embedded expression in the fault knowledge graph, determining an adjacency matrix according to the edge connection relation of the fault knowledge graph, and dividing a plurality of sub-adjacency matrices according to the types of edges;
Inputting the node characteristic matrix and the sub-adjacent matrix into a pre-constructed multi-layer heterogeneous graph convolution network, determining a corresponding convolution kernel according to the type of the node, determining attention weight according to the type of the edge, extracting heterogeneous neighborhood information and associated features of the node layer by layer based on the convolution kernel and the attention weight, and constructing a heterogeneous graph convolution characteristic matrix;
The heterogeneous graph convolution feature matrix calculates attention weights among nodes through a multi-head graph attention mechanism, gathers node neighborhood information based on the attention weights, extracts graph attention feature matrices, calculates self-attention weights of the nodes through a self-attention pooling mechanism, gathers node features, determines graph level feature representations and obtains fault association features.
And extracting the embedded expression vector of each node from the fault knowledge graph to form a node characteristic matrix. Each row of the node characteristic matrix corresponds to one node, each column corresponds to one dimension of the embedded vector, and an adjacent matrix is constructed according to the edge connection relation of the fault knowledge graph. The adjacency matrix is a two-dimensional square matrix, rows and columns respectively represent nodes, matrix elements represent whether edges are connected between the nodes, 1 represents edges, 0 represents no edges, the adjacency matrix is divided into a plurality of sub-adjacency matrices according to types of the edges, such as causal relationship, dependency relationship, semantic association and the like, and each sub-adjacency matrix represents a specific type of edge relationship;
A multi-layer heterogeneous graph convolution network is constructed, the input of the network is a node characteristic matrix and a sub-adjacency matrix, and different convolution kernels are designed according to the types of nodes, such as fault events, system components and the like. Each node type corresponds to a convolution kernel, and is used for extracting the characteristics of the node of the type, and different attention weights are set according to the types of the edges. The attention weight reflects the importance degree of different types of edges in feature propagation, and in the heterogeneous graph convolution layer, corresponding convolution kernels are selected for each node according to the types of the nodes, and corresponding attention weights are selected according to the types of the edges between the nodes and the neighboring nodes; performing convolution operation on the convolution kernel and the node characteristics, performing weighted aggregation on the attention weight and the neighbor node characteristics, and extracting heterogeneous neighborhood information and associated characteristics of the nodes; the multi-layer heterograph convolution network propagates and extracts the characteristics layer by layer to obtain a final heterograph convolution characteristic matrix;
On the basis of the heterogeneous graph convolution feature matrix, introducing a multi-head graph attention mechanism, further extracting the association information among the nodes, dividing the heterogeneous graph convolution feature matrix into a plurality of sub-feature matrices, wherein each sub-feature matrix corresponds to one attention head, calculating the attention weight among the nodes for each attention head, carrying out weighted aggregation on the neighborhood information of each node based on the attention weight to obtain node feature representation under the attention head, and splicing the node feature representations of all the attention heads to form the graph attention feature matrix. The attention feature matrix of the graph fuses node association information of a plurality of angles to provide more comprehensive and accurate feature representation;
On the basis of a graph attention feature matrix, a self-attention pooling mechanism is introduced, global feature representation of a graph level is extracted, attention weights of each node and all other nodes in the graph are calculated through the self-attention mechanism, importance of the nodes in the whole graph is reflected, based on the self-attention weights, the features of all the nodes are weighted and aggregated to obtain feature vectors of the graph level, the feature vectors of the graph level represent global information of the whole fault knowledge graph, and the whole association mode among faults is captured;
and combining the heterogeneous graph convolution feature matrix, the graph annotation force feature matrix and the graph-level feature vector to form a final fault correlation feature representation, wherein the fault correlation feature fusion node has heterogeneous neighborhood information, multi-angle correlation information and global graph-level information, and provides comprehensive and fine fault correlation description.
In the embodiment, according to the embedded expression vector of the nodes in the fault knowledge graph, a node characteristic matrix is constructed, so that characteristic information of each node can be comprehensively represented; dividing the adjacency matrix into a plurality of sub adjacency matrices according to the types of the edges, so that different types of edge relations can be better distinguished, and finer characteristic representation is provided; the multi-layer heterogeneous graph convolution network is constructed, so that the characteristics can be effectively transmitted and extracted, heterogeneous neighborhood information and associated characteristics among nodes are captured, and the characteristic capability of the characteristics is improved; the attention mechanism of the multi-head graph is introduced, the association information among the nodes can be further extracted, the node association information of a plurality of angles is fused, and more comprehensive and accurate characteristic representation is provided; the self-attention pooling mechanism is introduced, global feature representation of the graph level can be extracted, global information of the whole fault knowledge graph is captured, importance of nodes in the whole graph is reflected, the heterogeneous graph convolution feature matrix, the graph annotation force feature matrix and the graph level feature vector are combined to form final fault association feature representation, association relations among faults can be comprehensively and finely described, and support is provided for follow-up fault diagnosis and decision.
S103, inputting real-time monitoring data of the lightweight electric propulsion system into a fault diagnosis model, extracting fault correlation features and importance features corresponding to the real-time monitoring data, carrying out classification decision by combining the fault knowledge graph through a fault classifier, and determining a fault diagnosis result, wherein the fault diagnosis model is constructed by integrating the fault correlation feature extraction network and the fault classifier based on the fault knowledge graph.
And simultaneously, calculating the importance of each feature in the real-time monitoring data through a specific algorithm or model, and better understanding the influence factors of the real-time monitoring data.
Establishing a fault classifier, wherein the task of the classifier is to classify and diagnose real-time monitoring data according to the extracted characteristics, a fault diagnosis model uses a constructed fault knowledge graph as background knowledge, the fault diagnosis model is integrated into a classification decision process, the classification decision is carried out according to the extracted characteristics of the real-time monitoring data and the node information and association relation in the fault knowledge graph, and the fault diagnosis result of the real-time monitoring data is determined based on the output of the fault classifier and the information in the fault knowledge graph, wherein the fault diagnosis result possibly comprises information on the aspects of fault type, fault grade, fault cause and the like;
in the embodiment, the fault classifier utilizes the fault correlation characteristics to extract the characteristics extracted by the network, and can quickly and effectively classify and decide the real-time monitoring data by combining the existing fault knowledge graph, thereby improving the accuracy and efficiency of fault classification; the fault knowledge graph is used as background knowledge, provides abundant context information and knowledge support for the fault diagnosis model, can help the model to better understand fault characteristics in real-time monitoring data, and makes more reasonable classification decisions; the integrated design of the fault diagnosis model can effectively improve the fault diagnosis performance of the light electric propulsion system, reduce the cost and time cost of fault diagnosis and improve the reliability and stability of the system.
In an alternative embodiment, the training of the fault diagnosis model comprises:
Acquiring marked historical fault data, wherein the historical fault data comprises fault association features corresponding to fault scenes and corresponding fault labels;
Initializing model parameters of a fault diagnosis model;
Constructing a fault classification loss function based on the fault classification loss and the graph regularization term;
Inputting the fault correlation characteristics into the fault classifier, and calculating the total loss of a fault diagnosis model through forward propagation based on the fault classification loss function;
And calculating the parameter gradients of the fault correlation characteristic extraction network and the fault classifier through back propagation according to the total loss of the fault diagnosis model, and updating model parameters based on the parameter gradients until a preset model training round is reached, so that training of the fault diagnosis model is completed.
From sources such as historical fault records, expert knowledge base, and the like, the marked historical fault data are collected and arranged, and the historical fault data comprise two parts of contents: fault associated features corresponding to the fault scene and corresponding fault labels; the fault association features are extracted from the fault knowledge graph through a fault association feature extraction network and represent association modes and important information of a fault scene; the fault labels are the results of classifying and diagnosing fault scenes, such as fault types, severity and the like; dividing the collected historical fault data into a training set, a verification set and a test set for training, tuning and evaluating the model;
The fault diagnosis model consists of a fault associated feature extraction network and a fault classifier, wherein parameters of the fault associated feature extraction network are initialized, the parameters comprise convolution kernel weights, weight matrixes, attention vectors, parameters of the enhanced combination weights and attention heads represented by graph levels of different composition convolution layers, the parameters of the fault classifier are initialized, and the parameters comprise weight matrixes, bias vectors and the like of the full-connection layers;
The fault classification loss function consists of two parts: fault classification loss and graph regularization term, wherein the fault classification loss measures the difference between a fault label predicted by the model and a real label; the graph regularization term is used to constrain the smoothness of the model on the graph structure, encouraging connected nodes to have similar feature representations; the fault classification loss and the graph regularization term are weighted and summed to obtain a final fault classification loss function;
Inputting fault associated features in historical fault data into a fault diagnosis model, extracting high-level feature representation by the fault associated feature extraction network, inputting the extracted feature representation into a fault classifier, obtaining a prediction result of a fault label through a full connection layer and an activation function, substituting the predicted fault label and a real label into a fault classification loss function, and calculating total loss of the model under current parameters;
Calculating the gradient of the model parameters of the loss function through a back propagation algorithm according to the total loss calculated by forward propagation, and updating the model parameters according to gradient information by utilizing a gradient descent optimization algorithm, preferably a random gradient descent algorithm; in the process of parameter updating, proper super parameters such as learning rate, regularization coefficient and the like are required to be set so as to control the updating speed and stability, after one parameter updating is completed, the next iteration is carried out, and forward and backward propagation is repeatedly carried out;
Evaluating the performance of the model on the verification set every time a certain number of iterations are completed, and optionally evaluating indexes such as fault classification accuracy, F1 value and the like; according to the performance on the verification set, adjusting the super parameters of the model; and until the preset model training round is reached, if the performance index meets the preset threshold in advance, the training iteration can be ended in advance, and after the training is finished, the final performance of the model is evaluated on the test set so as to verify the actual diagnosis capability of the model.
In the embodiment, the constructed fault classification loss function can combine the fault classification loss and the graph regularization term, and calculate the total loss of the fault diagnosis model through forward propagation, so that the model can classify and diagnose the fault more accurately; the parameter gradients of the network and the fault classifier are extracted through counter-propagation calculation fault correlation characteristics, and model parameters are updated based on the parameter gradients, so that iterative optimization of the model parameters is realized, the model can be gradually converged to an optimal solution, and generalization capability and accuracy of the model are improved; the model parameters are updated by adopting a back propagation algorithm, so that the historical fault data can be effectively utilized for training, the training efficiency and speed are improved, and meanwhile, the accuracy and stability of the model are ensured.
In an alternative embodiment, constructing the fault classification loss function based on the fault classification loss and the graph regularization term includes:
；
Wherein L represents total loss, k represents sample number, C represents failure category number, C represents failure category total number, yk,c represents real label of kth sample belonging to the C-th category, pk,c represents probability of model prediction of kth sample belonging to the C-th category, lambda1 represents weight coefficient of graph regularization term, r represents relationship type number, i represents one node, j represents another node, Hi,rj represents the representation of node i and node j under the r-th relationship type, Xi,rj represents the characteristics of node i and node j under the r-th relationship type, Wi,rj represents the elements of the characteristic transformation matrix of node i and node j under the r-th relationship type, sigmar represents the weight coefficient of the loss term under each relationship type in the regularization term of the adjustment graph, lambda2 represents the weight coefficient of the L2 regularization term, and MF2 represents the L2 regularization term of the model parameter.
According to the formula, the first term represents a cross entropy loss function and is used for measuring the difference between the fault type predicted by the model and the real label, and the classification loss of the model can be obtained by summing the prediction probabilities of all samples and all types and taking the negative logarithm of the sum; the second term is a graph regularization term, which is used for considering the relation between nodes in the fault knowledge graph, the second term comprises the difference between the representation corresponding to each node and the characteristics, and the difference is measured by a square error, wherein the relation type r represents one relation between the node i and the node j, Hi,rj is the representation of the node i and the node j, Xi,rj is the characteristics of the node i and the node j, and Wi,rj is an element of a characteristic conversion matrix; the third term is an L2 regularization term, which is used for preventing model parameters from excessively fitting historical data, carrying out square summation on the model parameters, multiplying the model parameters by a weight coefficient lambda2 so as to control the influence of the regularization term on total loss, and the purpose of the whole loss function is to minimize the total loss, and balance classification loss, graph regularization term and L2 regularization term so as to ensure that the model can accurately classify faults and utilize information in a fault knowledge graph in the training process;
in the embodiment, through the cross entropy loss function, the model can learn classification tasks of different fault categories at the same time, so that generalization capability and classification accuracy of the model are improved; the relation among the nodes in the fault knowledge graph is considered by the graph regularization term, so that the model can enhance the classification performance by utilizing the relation information among the nodes, the model captures the relation pattern and the mutual influence among faults through the graph regularization term, the L2 regularization term can effectively control the complexity of the model, the historical data is prevented from being excessively fitted by the model in the training process, the generalization capability of the model is improved, and the model is better represented on unseen data; the feature and the graph structure are comprehensively considered, so that the robustness and generalization capability of the model are improved, and the model can be better adapted to complex fault scenes.
Fig. 2 is a schematic structural diagram of a fault diagnosis system of a lightweight electric propulsion system according to an embodiment of the present invention, as shown in fig. 2, the system includes:
The first unit is used for determining a fault causal relationship and a component dependency relationship based on historical data and structural data of the lightweight electric propulsion system, constructing a fault tree and a directed graph, constructing an initial graph model of the lightweight electric propulsion system based on the fault tree and the directed graph, mapping nodes in the initial graph model into a low-dimensional dense vector space, determining node embedded representation and generating a fault knowledge graph;
The second unit is used for searching and obtaining corresponding neighbor embedded representations based on the node embedded representations in the fault knowledge graph, extracting node high-level characteristic representations through convolution operation on the nodes and the neighbors, determining a fault association relationship, extracting fault importance representations through pooling operation, and constructing a fault association characteristic extraction network;
The third unit is used for inputting the real-time monitoring data of the lightweight electric propulsion system into a fault diagnosis model, extracting fault correlation features and importance features corresponding to the real-time monitoring data, carrying out classification decision by combining the fault knowledge graph through a fault classifier, and determining a fault diagnosis result, wherein the fault diagnosis model is constructed by integrating the fault correlation feature extraction network and the fault classifier based on the fault knowledge graph.
In a third aspect of an embodiment of the present invention,
There is provided an electronic device including:
A processor;
A memory for storing processor-executable instructions;
wherein the processor is configured to invoke the instructions stored in the memory to perform the method described previously.
In a fourth aspect of an embodiment of the present invention,
There is provided a computer readable storage medium having stored thereon computer program instructions which, when executed by a processor, implement the method as described above.
The present invention may be a method, apparatus, system, and/or computer program product. The computer program product may include a computer readable storage medium having computer readable program instructions embodied thereon for performing various aspects of the present invention.
Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present invention, and not for limiting the same; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some or all of the technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit of the invention.