Aircraft fuel control system fault path mining and root cause diagnosis method

Abstract

The invention belongs to the technical field of aircraft fault diagnosis, and particularly relates to an aircraft fuel control system fault path mining and root cause diagnosis method. The method comprises the following steps: providing a fault propagation-oriented fault root cause diagnostic graph model to realize more interpretable fault source positioning, and backtracking from an alarm node to a fault source based on propagation mechanism information; constructing a variable propagation graph of the aircraft fuel control system in combination with correlation coefficient analysis and a Lingam causal algorithm, and then performing correction based on a system simulation structure in a manual inspection mode so as to establish an interpretable data structure; introducing an edge propagation quantization factor and a path scoring criterion integrating nodes, edges and path lengths to measure the possibility of candidate fault propagation paths; and learning an optimal threshold set of each node in the propagation graph structure by adopting a particle swarm optimization algorithm. The method supports path-level transparent backtracking, improves the accuracy and flexibility of the model, and can excavate a plurality of fault propagation paths.

Description

Technical Field

[0001] This invention belongs to the field of aircraft fault diagnosis technology, specifically relating to a method for fault path mining and root cause diagnosis of an aircraft fuel control system. Background Technology

[0002] The fuel control system is a crucial component of aircraft systems, playing a vital role in flight safety and operational efficiency. According to aircraft malfunction statistics, fuel system failures in civil aircraft, such as insufficient fuel, fuel instability, pipeline ruptures, and leaks, account for over 30% of all aircraft malfunctions. Therefore, root cause diagnosis research on aircraft fuel control systems is essential. However, in large industrial systems, components are interconnected, and variables are highly coupled, making it easy for abnormal information to propagate between components, causing even non-faulty components to exhibit anomalies. For example, when a fault occurs, multiple sensors may trigger alarms, leading to alarm flooding. Therefore, this phenomenon can affect the determination of the true cause of the fault using currently used data-driven fault root cause diagnosis classification algorithms. Modeling fault propagation, conducting in-depth quantitative assessments of faults, uncovering fault propagation paths, and locating the true fault source are of great significance in the field of fault root cause diagnosis.

[0003] Currently, fault tracing methods can be mainly divided into three categories: knowledge-based methods, data-driven methods, and hybrid methods. Most knowledge-based methods heavily rely on expert knowledge and do not fully utilize the rich quantitative information contained in system variables, making them inefficient for modeling more complex and large-scale industrial systems. While data-driven methods have achieved high accuracy in some prediction and classification tasks, as black-box models, most lack interpretability and visibility. Furthermore, time-series data from different domains have different characteristics, and the manifestation of causal relationships may vary; therefore, modeling using purely data-driven methods cannot guarantee consistency with the original system mechanism. Combining knowledge and data approaches can improve model interpretability, make fault propagation analysis results more aligned with domain knowledge, and improve the efficiency of fault source localization.

[0004] Currently, in the field of fault propagation, graph structures are either constructed based on pure knowledge or pure data-driven methods, or they directly use prior graph structures provided by public datasets. If we want to extend this to other industrial fields, we may not be able to guarantee applicability and accuracy. In addition, many graph structure methods, although they can implement backtracking logic, often lack in-depth mining and utilization of quantitative information such as nodes and edges in the graph, and lack physically meaningful quantification of the propagation path.

[0005] Related terms:

[0006] Aircraft fault diagnosis: Aircraft fault diagnosis;

[0007] propagation path tracing: Propagation path tracing;

[0008] Linear Non-Gaussian Acyclic Model (Lingam): Linear Non-Gaussian Acyclic Model;

[0009] Particle Swarm Optimization (PSO): Particle Swarm Optimization Algorithm;

[0010] Graph Convolutional Networks (GCN): Graph Convolutional Neural Network;

[0011] Graph Attention Networks (GAT): Graph Attention Neural Network. Summary of the Invention

[0012] The purpose of the present invention is to provide a method for fault path mining and root cause diagnosis of an aircraft fuel control system with high diagnostic accuracy and strong applicability to ensure the safe and efficient operation of the aircraft.

[0013] The method for fault path mining and root cause diagnosis of the aircraft fuel control system provided by the present invention is as follows: First, a fault root cause diagnosis graph model for fault propagation is proposed to achieve more interpretable fault source localization, and based on the propagation mechanism information, trace back from the alarm node to the fault source; combine the correlation coefficient analysis and the Lingam causal algorithm to construct a variable propagation graph of the aircraft fuel control system, and then correct it based on the system simulation structure in an artificial inspection manner to establish an interpretable data structure; second, introduce an edge propagation quantization factor and a path scoring criterion that comprehensively considers nodes, edges, and path lengths to measure the likelihood of candidate fault propagation paths; further, use the particle swarm optimization algorithm to learn the optimal threshold set of each node in the propagation graph structure. The specific steps are as follows:

[0014] Step (1): Construction of the topological structure diagram of the fuel control system

[0015] In this step, abstract the key variable parameters and the relationships between variables of each component in the fuel system, such as fuel pumps, metering valves, differential pressure valves, distributors, etc., into a topological structure diagram, which is represented by a quadruple:

[0016]

[0017] Where:

[0018] V = {v1, v2,..., vn} is a finite non-empty set of nodes, which correspond to key variables of various components in the fuel system, including pressure, flow rate, valve displacement, etc.

[0019] E = {e1, e2, ..., e n} is a finite set of relationships between nodes (variables), also called the set of edges or branches. For a directed edge, it means that the change of the first node variable A affects the change of the tail node variable B.

[0020] S reflects the specific dependencies between variables. e The symbols representing dependencies are: "+" indicates that the tail node variable increases as the head node variable increases and decreases as the head node variable decreases; "-" indicates the opposite, representing a suppressive effect. Variable information is also added to the edges, such as λ. e Defined as the propagation factor of an edge, it is used to reflect the strength of fault information propagation between components.

[0021] ψ={(S v ,f v )|S v ∈{+,0,-}} represents the state information of the variable, where "+", "0", and "-" indicate that the fuel system variable data is in an extremely high, normal, or extremely low state, respectively; quantitative information is added to the node, f v Defined as the degree of fault or deviation of a node, it is used to reflect the likelihood of a fault occurring or propagating in that node variable within the fuel system.

[0022] Correlation coefficient and Lingam algorithm were used respectively. [1] The algorithm calculates and extracts the set of edges between each variable node. The correlation coefficient result includes the corresponding variable node pair and the sign of the edge (positive or negative correlation), but not the direction. Lingam's edge set includes both the direction and sign of the edge. Since correlation relationships are often more generalized, the fusion rule is to take the node pairs common to both sets, while the direction and sign of the edge are determined by the Lingam algorithm. This initially constructed directed graph structure provides a good foundation for manual review, saving time and manpower, deeply integrating domain knowledge, and strengthening the physical support of the graph structure. Subsequently, combined with knowledge of the fuel system structure, key connections in the graph that do not conform to the physical flow or are missing are corrected or supplemented. Through these steps, the final directed graph reflects both data-driven statistics and causal relationships and incorporates the physical priors of the fuel control system, providing a reliable graph structure foundation for subsequent backtracking fault diagnosis algorithms.

[0023] Step (2): Embedding fault propagation graph structure information

[0024] For graph structure information embedding, it is divided into two parts: nodes and edges. Among them:

[0025] (1) The node represents the variable v in the fuel system, including three types of system parameters: pressure, flow rate and displacement. Specifically, there are fuel pressure after pump, fuel pressure before metering valve, fuel flow rate of main oil circuit branch after pump, fuel flow rate of constant pressure valve branch after pump, piston displacement of constant pressure valve, piston displacement of differential pressure valve, etc. In order to analyze the fault conditions of the nodes in the system, the degree of abnormality of the variable data is used as the node value.

[0026] Membership degree is an important concept in fuzzy mathematics, used to describe the degree to which an element belongs to a fuzzy set. This invention chooses to use a membership function to measure the abnormal state of variable v. Traditionally, Z-shaped and trapezoidal membership functions are used to characterize node states. This invention replaces the linear segment of the Z-shaped membership function with a more continuous and asymptotic Sigmoid curve, which better reflects the gradual transition between normal and abnormal states of a physical quantity. Furthermore, it has an adjustable kurtosis parameter, allowing control of the transition region based on system characteristics. The improved membership function expression is as follows:

[0027]

[0028] in, The standard Sigmoid function is represented by x, where x is the parameter and k is the slope coefficient, used to measure the steepness of the Sigmoid curve near the threshold, i.e., the model's sensitivity to parameter changes. When k is large, the Sigmoid curve changes more steeply, meaning the membership degree quickly approaches 0 or 1; when k is small, the curve changes more gently. a, b, c, and d represent the low threshold on the negative side, the high threshold on the negative side, the low threshold on the positive side, and the high threshold on the positive side, respectively. During threshold initialization, the "normal range" (measured by the median and robust dispersion) of each measurement point is first estimated using only normal operating condition samples, and this range is used as the initial boundary of the middle interval. Then, the abnormal thresholds at both ends are determined using the statistical distribution information of fault samples on both sides to reflect the deviation magnitude during actual faults. The resulting set of ordered thresholds is used for membership degree calculation and serves as the initial parameters for subsequent fine-tuning.

[0029] (2) Besides node variables, the topological information of edges also affects the fault propagation process in the system. Therefore, the normalized result of edge betweenness is introduced as a propagation factor for the edge, which measures the frequency of information transmission along the edge in the graph structure, thereby reflecting the strength or criticality of fault propagation along the edge. The expression for edge betweenness is shown below:

[0030]

[0031] For a directed graph G = (V, E), where V is the set of nodes and E is the set of directed edges, for any edge e ij ∈E, its betweenness b ij It is defined as the sum of the ratios of the number of shortest paths passing through this edge to the total number of paths. Where σ st σ represents the total number of shortest paths between node pairs (s,t). st (e ij ) indicates that these paths pass through edge e ij The quantity.

[0032] Step (3): Building an interpretable fault backtracking diagnostic model

[0033] Fault backtracking diagnosis is a search process that starts from the downstream node where an anomaly is observed, traces upwards along the reverse topology of the graph, progressively constructing candidate fault propagation paths and pruning and scoring them until a possible fault source node is reached or a stopping condition is triggered. Depth-first search (DFS) is employed in this process. [2] The method (a graph search algorithm) starts from the initial node and searches as deep as possible along a path until it cannot continue. Then it backtracks to the previous node and tries other branches. During this process, pruning conditions are set, that is, the search stops and a different path is taken when an edge that does not conform to the fault propagation rules is encountered. The pruning conditions are as follows: pruning is performed when equation (3) is not satisfied. This principle is the compatible path principle, which believes that when the sign product of the first and last nodes and the edge is positive, it conforms to the consistency law of fault propagation.

[0034] S v (v i )S e (e ij )S v (v j )=+1 (3)

[0035] Among them, v i and v j Let e ​​represent the first and last nodes of an edge, respectively. ij Indicates from v i Point to v j The edge, S v The symbol representing a node, S e Symbols representing edges.

[0036] For the research object of high reliability requirements of aircraft fuel control system, in the process of fault analysis and diagnosis, this invention first constructs a complete set of candidate fault paths through the aforementioned reverse depth-first search process, and on this basis designs a path scoring criterion with clear physical meaning to more accurately locate the real fault source. The path scoring criterion designed in this invention is shown in equation (4):

[0037]

[0038] Where μ(v) is the membership degree of a node, N is the set of nodes in the path, b(u,v) is the propagation factor of a directed edge, E is the set of edges in the path, and P Length K is the length of the path. max This is the preset maximum path length, used for standardizing the length penalty. The first term represents the average membership degree of all nodes in the propagation path, multiplied by the λ factor, reflecting the relative importance of membership information and edge propagation factors. The second term is the average propagation factor of all edges in the path, multiplied by the (1-λ) factor. The third term is the path length penalty. Since fault propagation information weakens with increasing path length, the probability of propagation also decreases. Therefore, the longer the path length, the greater the penalty. Here, γ is the penalty factor. Setting it too large will over-favor shorter paths, while setting it too small may retain redundant and unreasonable paths. A grid search is performed on a proportional number of points to determine the appropriate penalty factor. The above path evaluation criteria are interpretable and can reasonably evaluate different candidate paths to locate the true fault source.

[0039] To further improve the accuracy of fault location and the adaptability of the method, the PSO algorithm is adopted. [3] A global search and fine-tuning of the membership function thresholds for each node is performed. This method has a certain degree of scalability and can achieve good optimization results for different research object systems.

[0040] Step (4): Model training and validation

[0041] During the model training phase, the initial symbolic directed graph of the fuel system is automatically generated based on the correlation coefficient analysis and Lingam causal analysis algorithm, following the method for constructing the system topology diagram in step (1). The threshold for determining the strength of correlation (i.e., the correlation coefficient analysis threshold) is α (generally set to 0.6). When the absolute value of the correlation coefficient between two variables is greater than α, a candidate edge is added to the initial graph. The threshold for retaining causal strength (i.e., the Lingam algorithm threshold) is β (generally set to 0.5). To filter out edges with weak causal relationships, only connections with an absolute value of the causal coefficient greater than β are retained in the initial symbolic directed graph; other edges are considered to have insufficient causal influence and are deleted. Subsequently, key edges are fine-tuned based on expert experience to ensure that the physical flow of the graph is basically consistent with the system mechanism. Based on this graph, a reverse depth-first backtracking strategy is used to search all candidate propagation paths starting from the deepest alarm node, and multi-factor support is calculated for each path—that is, a weighted combination of the improved Sigmoid membership product of each node, the path edge betweenness product, and the path length penalty. The kurtosis parameter of the membership function was set to k=5; subsequently, particle swarm optimization (swarm size 100, iterations 10) was used to globally fine-tune all upper and lower thresholds to maximize the diagnostic accuracy on the training set. Five-fold cross-validation was used to evaluate the model performance on normal and fault mode datasets. To verify the effectiveness of the proposed method, two advanced fault diagnosis graph neural network models were compared with the method proposed in this invention: Graph Convolutional Neural Network (GCN). [4] Graph Attention Neural Network (GAT) [5] The GCN model comprises a two-layer graph convolutional structure, uses the ReLU activation function, and employs global average pooling to aggregate node features. The GAT model adaptively learns relationships between nodes by introducing an attention mechanism; in this experiment, a four-head attention mechanism is used. Both models utilize the cross-entropy loss function and the Adam optimizer, and incorporate an early stopping mechanism. Results show that our proposed method achieves a diagnostic accuracy of 97.1% on the test set, while the average accuracies of GCN and GAT are 93.2% and 93.3%, respectively.

[0042] The main technical features and performance advantages of this invention are as follows:

[0043] (1) A more robust system graph structure is constructed by combining statistical and causal analysis methods. Based on this, the graph structure is corrected according to the system structure, and a multi-factor scoring strategy for the propagation path is designed to support transparent backtracking at the path level.

[0044] (2) It is proposed to use the particle swarm optimization algorithm to adaptively and dynamically adjust the membership threshold of graph nodes, thereby improving the accuracy and flexibility of the model.

[0045] Compared to traditional fault tracing methods based on expert experience or purely data-driven approaches, this invention provides a fault propagation path identification and root cause diagnosis strategy that is interpretable throughout the entire process, improving diagnostic accuracy and enabling the discovery of multiple fault propagation paths. Attached Figure Description

[0046] Figure 1 This is the improved sigmoid membership function.

[0047] Figure 2 This section compares the accuracy of each method in diagnosing root causes of failures.

[0048] Figure 3 This section compares the average accuracy of each method in diagnosing root causes of failures using a small number of training samples.

[0049] Figure 4 Overall technical roadmap of this method. Detailed Implementation

[0050] The present invention is illustrated using a faulty pump with insufficient power as an example.

[0051] The membership degree of each node is calculated according to formula (1) (the return value includes the sign and degree), and the results are as follows: the fuel pressure after the pump is p1 -0.3211, the fuel pressure before the metering valve is p2 0.0000, the control oil pressure of the metering valve is p3 -0.1220, the fuel pressure after the metering valve is p4 -0.6936, the fuel pressure before the interlocking valve is p5 -0.0517, the pressure after the second stage of the distribution valve is p6 -0.3424, the pressure after the first stage of the distribution valve is p7 -0.3132, the fuel flow rate of the main oil circuit branch after the pump is q1 -0.5838, and the fuel flow rate of the constant pressure valve branch after the pump is... The fuel flow rate q2 is -0.7243, the fuel flow rate q4 after the metering valve is -0.3291, the fuel flow rate q5 of the second-stage main fuel line nozzle of the distribution valve is -0.4399, the fuel flow rate q6 of the auxiliary fuel line nozzle of the distribution valve is -0.3867, the piston displacement x1 of the constant pressure valve is +0.2241, the piston displacement x2 of the differential pressure valve is +0.3708, the piston displacement x3 of the differential pressure mechanism is -0.3392, the metering valve x4 is -0.4254, the piston displacement x5 of the second stage of the distribution valve is -0.4494, and the piston displacement x6 of the first stage of the distribution valve is 0.0000. For the membership degree of each node, nodes with an absolute membership degree > 0 are considered abnormal nodes. Then, for each alarm node, an upward (from downstream to upstream) path search is performed based on DFS backtracking. During the search process, the compatible path pruning rule is executed. The expansion continues only when the state direction of the parent node and the edge symbol jointly satisfy the propagation consistency (see Equation (3)). Otherwise, pruning is performed. In this way, a set of candidate propagation paths that satisfy the propagation characteristics can be constructed. The path score calculation and sorting results of each candidate path are as follows using Equation (4) (where the λ factor is 0.7 and the γ penalty factor is set to 0.001): Candidate path 1 is p1→q1→p4→q6, the corresponding fault source is p1, and the path score is 0.347119244; Candidate path 2 is p1→q1→q5, the corresponding fault source is p1, and the path score is 0.313601572. Candidate path 3 is p1→q1→x5→p6→q5, with the corresponding fault source being p1, and a path score of 0.298735942; candidate path 4 is q4→x5→p6→q5, with the corresponding fault source being q4, and a path score of 0.27284903; candidate path 5 is q4→q5, with the corresponding fault source being q4, and a path score of 0.269060562; candidate path 6 is q4→q6, with the corresponding fault source being q4, and a path score of 0.250423113. Through comparison, the final fault propagation path can be determined to be p1→q1→p4→q6 and the fault source being p1. Furthermore, the path ranking results show that the first three paths all point to the actual fault source.

[0052] To better validate the root cause diagnosis and fault path discovery performance of this model, three typical faults—external leakage, valve jamming, and insufficient pump power—were compared with two advanced fault diagnosis graphical neural network models, GCN and GAT, using five-fold cross-validation. The average accuracies of the proposed method, GCN, and GAT were 97.1%, 93.2%, and 93.3%, respectively. (From the comparison bar...) Figure 2 The improvement in accuracy of the proposed method is clearly visible. The results show that the method of this invention has the highest accuracy, GCN has a relatively stable accuracy, and GAT, due to the introduction of an attention mechanism, has a better diagnostic effect than GCN in most cases, but its diagnostic performance is poor for individual samples.

[0053] Secondly, experiments were designed to validate the model with a small number of training samples. In each training iteration, 90% of the training data for a specific fault label was removed, while the test set remained unchanged. The accuracy was then observed to evaluate the model's diagnostic performance when identifying faults with a small number of training samples. The average accuracies of the proposed method, GCN, and GAT were 95.7%, 86.3%, and 93.6%, respectively. The results indicate that the proposed method is not significantly affected by the limited number of fault samples, while GCN exhibits relatively poor diagnostic performance.

[0054] References

[0055] [1] Shimizu S, Hoyer PO, A,et al.A linear non-Gaussianacyclic model for causal discovery[J].Journalof Machine Learning Research,2006,7(10).

[0056] [2]Tarjan R.Depth-first search and linear graph algorithms[J].SIAMjournalon computing,1972,1(2):146-160.

[0057] [3]Kennedy J, Eberhart R.Particle swarm optimization[C] / / Proceedingsof ICNN'95-international conference on neural networks.ieee,1995,4:1942-1948.

[0058] [4]Kipf T N.Semi-Supervised Classification with Graph ConvolutionalNetworks[J].arXiv preprint arXiv:1609.02907,2016.

[0059] [5] P,Cucurull G,Casanova A,et al.Graph attention networks[J].arXiv preprint arXiv:1710.10903,2017。

Claims

1. A method for fault path discovery and root cause diagnosis of an aircraft fuel control system, characterized in that, This includes constructing a fault root cause diagnosis graph model oriented towards fault propagation to achieve more interpretable fault source localization, tracing back from alarm nodes to fault sources based on propagation mechanism information; constructing a variable propagation graph of the aircraft fuel control system by combining correlation coefficient analysis and the Lingam causal algorithm, and then refining it based on the system simulation structure through manual verification to establish an interpretable data structure; introducing an edge propagation quantification factor and a path scoring criterion that integrates nodes, edges, and path lengths to measure the probability of candidate fault propagation paths; and using a particle swarm optimization algorithm to learn the optimal threshold set for each node in the propagation graph structure. The specific steps are as follows: Step (1): Construction of the fuel control system topology diagram: The key variable parameters and relationships between the various components of the fuel system, including the fuel pump, metering valve, differential pressure valve, and distributor, are abstracted into a topology diagram, which is represented by quadruples: in: V = {v1, v2, ..., v} n } is a finite non-empty set of nodes, where each node represents a key variable of a component in the fuel system, specifically including pressure, flow rate, and valve displacement. E = {e1, e2, ..., e n } is a finite set of relationships between nodes, also called the set of edges or branches. For a directed edge, it means that the change of the first node variable A affects the change of the tail node variable B. S reflects the specific dependencies between variables. e The symbols representing dependencies are: "+" indicates that the tail node variable increases as the head node variable increases and decreases as the head node variable decreases; "-" indicates the opposite, representing a suppressive effect; variable information is also added to the edges, such as λ. e Defined as the propagation factor of an edge, it is used to reflect the strength of fault information propagation between components; ψ={(S v ,f v )|S v ∈{+,0,-}} represents the state information of the variable, where "+", "0", and "-" indicate that the fuel system variable data belongs to the ultra-high, normal, and ultra-low states, respectively; quantitative information is added to the node, f v Defined as the degree of fault or deviation of a node, it is used to reflect the probability of a fault occurring or propagating in that node variable in the fuel system; The correlation coefficient and the Lingam algorithm were used to calculate and extract the edge sets between each variable node. The correlation coefficient result includes the corresponding variable node pair and the sign of the edge, but not the direction of the edge. The Lingam edge set includes the direction and sign of the edge. The fusion rule is to take the node pairs common to both sets, while the direction and sign of the edge are determined by the Lingam algorithm. Subsequently, combined with the knowledge of the fuel system structure, key connections in the graph that do not conform to the physical flow or are missing are corrected or supplemented. The final directed graph reflects both data-driven statistics and causal relationships and incorporates the physical priors of the fuel control system, providing a reliable graph structure foundation for subsequent backtracking fault diagnosis algorithms. Step (2): Embedding fault propagation graph structure information: It is divided into two parts: nodes and edges, where: (1) For a node, it represents the variable v in the fuel system, including three types of system parameters: pressure, flow rate and displacement. Specifically, there are fuel pressure after pump, fuel pressure before metering valve, fuel flow rate of main oil circuit branch after pump, fuel flow rate of constant pressure valve branch after pump, piston displacement of constant pressure valve, piston displacement of differential pressure valve, etc. In order to analyze the fault conditions of nodes in the system, the degree of abnormality of variable data is used as node value. Specifically, a Z-type membership function is used to measure the abnormal state of variable v. The linear segment of the Z-type membership function is replaced with a sigmoid curve, which is more continuous and asymptotic, to better reflect the gradual transition between normal and abnormal states of the physical quantity. It also has an adjustable kurtosis parameter, allowing control of the transition region based on system characteristics. The improved membership function expression is as follows: in, The standard sigmoid function is represented by k, which is the slope coefficient used to measure the steepness of the sigmoid curve near the threshold, i.e., the sensitivity of the model to parameter changes; a, b, c, and d represent the low threshold on the negative side, the high threshold on the negative side, the low threshold on the positive side, and the high threshold on the positive side, respectively. In the initialization of these thresholds, the "normal range" of each measuring point is first estimated using normal operating condition samples, and this range is used as the initial boundary of the middle interval; then, the abnormal thresholds at both ends are determined using the statistical distribution information of fault samples on both sides to reflect the deviation magnitude when there is a real fault. Finally, a set of ordered thresholds is obtained for membership calculation and serves as the initial parameters for subsequent fine-tuning. (2) For edges, the normalized result of the edge betweenness is introduced as the propagation factor of the edge, which is used to measure the frequency of information transmission by the edge in the graph structure, thereby reflecting the strength or criticality of the fault propagation along the edge; the expression of the edge betweenness is shown in the following formula: For a directed graph G = (V, E), where V is the set of nodes and E is the set of directed edges, for any edge e ij ∈E, its betweenness b ij Defined as the sum of the ratios of the number of shortest paths passing through this edge to the total number of paths; where σ st σ represents the total number of shortest paths between node pairs (s,t). st (e ij ) indicates that these paths pass through edge e ij Quantity; Step (3): Explainable fault retrospective diagnosis: Fault backtracking diagnosis is a search process, that is, starting from the downstream node where the anomaly is observed, tracing upwards along the reverse topology of the graph, gradually constructing candidate fault propagation paths and pruning and scoring them until reaching the possible fault source node or triggering the stopping condition; specifically, the depth-first search method is adopted, starting from the starting node, searching as deep as possible along a path until it can no longer continue, then backtracking to the previous node and trying other branches; in this process, pruning conditions are set, that is, when encountering an edge that does not conform to the fault propagation rules, the search is stopped and the path is changed; the pruning conditions are as follows, when equation (3) is not satisfied, pruning is performed; this principle is the compatible path principle, which believes that when the sign product of the first and last nodes and the edge is positive, it conforms to the consistency law of fault propagation; S v (in i )S e (e ij )S v (v j )=+1 (3) Among them, v i and v j Let e ​​represent the first and last nodes of an edge, respectively. ij Indicates from v i Point to v j The edge, S v The symbol representing a node, S e Symbols representing edges; By constructing a complete set of candidate fault paths through search, a path scoring criterion with physical meaning is designed to more accurately locate the real fault source; the path scoring criterion is shown in equation (4): Where μ(v) is the membership degree of a node, N is the set of nodes in the path, b(u,v) is the propagation factor of a directed edge, E is the set of edges in the path, and P Length K is the length of the path. max The first term represents the preset maximum path length, used for standardizing the length penalty; the second term represents the average membership degree of all nodes in the propagation path, multiplied by the λ factor, reflecting the relative importance of membership degree information and edge propagation factors; the third term is the average propagation factor of all edges in the path, multiplied by the (1-λ) factor; the fourth term is the path length penalty term. Since fault propagation information weakens as the path length increases, the probability of propagation also decreases relatively. Therefore, the longer the path length, the greater the penalty applied; where γ is the penalty factor. Step (4): Model training and validation During the model training phase, the initial symbolic directed graph of the fuel system is automatically generated based on the correlation coefficient analysis and Lingam causal analysis algorithm, following the construction method of the system topology diagram in step (1). The threshold for determining the strength of correlation, i.e. the correlation coefficient analysis threshold, is α. When the absolute value of the correlation coefficient between two variables is greater than α, a candidate edge is added to the initial graph. The threshold for retaining causal strength, i.e. the Lingam algorithm threshold, is β. Only connections with an absolute value of causal coefficient greater than β are retained in the initial symbolic directed graph, and the remaining edges are considered to have insufficient causal influence and are deleted. Subsequently, the key edges are fine-tuned in combination with expert experience to ensure that the physical flow of the graph is basically consistent with the system mechanism. Based on this graph, the reverse depth-first backtracking strategy is used to start from the deepest alarm node, search all candidate propagation paths, and calculate the multi-factor support for each path—that is, the weighted combination of the improved Sigmoid membership product of each node, the path edge betweenness product, and the path length penalty. Then, particle swarm optimization is used to globally fine-tune all upper and lower limit thresholds to maximize the diagnostic accuracy on the training set.

2. The method according to claim 1, characterized in that, In step (3), the penalty factor γ is determined by taking a number of points in equal proportion and performing a grid search.

3. The method according to claim 1, characterized in that, In step (3), the pso algorithm is used to perform a global search and fine-tuning of the membership function threshold of each node.

Images