Unmanned aerial vehicle fmeca detectability evaluation method based on mbse-gnn collaborative driving

Abstract

The application relates to the technical field of unmanned aerial vehicle evaluation, and discloses an unmanned aerial vehicle FMECA detectability evaluation method based on MBSE-GNN collaborative driving, which comprises the following steps: extracting unmanned aerial vehicle system source data to construct an initial system heterogeneous topology graph; extracting constraint parameters from a node set, collecting time sequence features in a flight process, and extracting real-time control compensation gain; performing cascade fusion of the constraint parameters and the time sequence features to generate an initial node hidden state vector; inputting into a graph attention network, converting the real-time control compensation gain into aggregated attention weights through a conditional edge feature mapping function, and outputting a confidence probability of a compensation failure state; and calculating a detection penalty term and performing nonlinear coupling calculation with the confidence probability. The application unifies static constraints and dynamic features to construct a complete data basis, and the network uses conditional edge mapping to process abnormal gain, so that the fault feature masking problem is overcome.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) evaluation technology, specifically to a UAV FMECA (Fast-Easy Detection Assessment) method based on MBSE-GNN collaborative drive. Background Technology

[0002] With the continuous improvement of the integration and control precision of UAV systems, system reliability analysis has become a core link in ensuring flight mission safety. In the failure mode, effect and hazard analysis system, detectability, as an indicator of the ease with which a failure can be identified before it leads to catastrophic consequences, directly affects the accuracy of system risk assessment.

[0003] Current detectability assessment and fault monitoring rely on qualitative scoring based on expert experience or threshold determination based on sensor signal residuals. Experts assess the difficulty for the monitoring system to identify a component failure based on the system design architecture and historical failure modes. During flight operation, the flight control system collects feedback signals from physical actuators and sensors, compares them with preset commands to generate state residuals, and when the residuals exceed the set threshold, the monitoring system triggers an early warning and attempts to isolate the damaged component.

[0004] Modern unmanned aerial vehicle (UAV) systems exhibit strong physical and logical coupling. The compensation effect of closed-loop control laws can mask early failure characteristics of physical layer actuators, creating a control masking effect. Traditional evaluation methods lack the ability to comprehensively process multi-source operational data based on the physical connection topology. When facing a distributed redundant architecture, the characteristics of underlying faults are diluted by multi-path parallel gain compensation, and conventional sensor feedback is in a pseudo-stationary state, leading to the risk of missed detections due to monitoring mechanisms relying solely on a single residual threshold. Faced with subtle feature evolution under strong compensation conditions, analysis methods relying on subjective experience and a single residual threshold cannot generate direct quantitative risk warning scores, causing the monitoring platform to fail to accurately trigger the UAV's redundancy backup or mission degradation closed-loop response logic.

[0005] Therefore, the purpose of this invention is to provide a method for evaluating the detectability of unmanned aerial vehicles (UAVs) based on MBSE-GNN collaborative drive, in order to address the shortcomings of existing technologies. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a UAV FMECA detectability assessment method based on MBSE-GNN collaborative drive. This method solves the problems of traditional detectability assessment methods that separate static design attributes from dynamic operational data, and that in a distributed redundant architecture, the closed-loop control compensation effect masks the underlying physical fault characteristics, resulting in missed detections and inaccurate triggering of the closed-loop response logic.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A UAV FMECA detectability assessment method based on MBSE-GNN collaborative drive includes: Extract the source data of the UAV system design, and construct an initial system heterogeneous topology graph based on the source data, which includes a node set and an edge set with heterogeneous edge characteristics. The node set contains multiple nodes. Constraint parameters of each node are extracted from the node set, time series features of the UAV during flight are collected, and real-time control compensation gain is extracted from the time series features. The time series features are mapped to the corresponding nodes in the node set, and the constraint parameters and time series features are cascaded and fused to generate the initial node hidden state vector. The initial node hidden state vector and heterogeneous edge features are input into the graph attention network. The real-time control compensation gain is transformed into the aggregate attention weight in the graph space through the conditional edge feature mapping function. The graph attention network outputs the certain probability that the corresponding node is in the compensation failure state. The detection penalty term is calculated based on the real-time control compensation gain. The detection penalty term is then nonlinearly coupled with the confidence probability to generate a quantified score for the detection measure. Determine the numerical range of the probe quantification score. If the probe quantification score is greater than or equal to the preset safety threshold, trigger the closed-loop response logic.

[0008] This invention addresses the fault masking problem caused by control compensation in complex, highly redundant architectures. It constructs a heterogeneous graph network structure by concatenating the static constraint topology derived from system model engineering with dynamic flight time-series data. This invention introduces a conditional edge mapping mechanism to perform high-dimensional spatial mapping on control feedback edges, extracting real-time control compensation gains and transforming these gains into attention weights for graph node aggregation, thus penetrating the feature dilution barrier generated by the distributed compensation process.

[0009] During the parameter iteration stage of the graph attention network, a probe-sensitive loss function is constructed. The control gain surge criterion and the tracking error stationarity criterion are used to define the control masking period time window of the system. The weight coefficient of the corresponding node in the loss function is increased by the indicator function for the node within the control masking period time window, thereby enhancing the model's extraction strength of physical conflict signals.

[0010] During the detection quantification phase, a detection penalty term is calculated by combining the real-time control compensation gain with a preset maximum controller output gain threshold. Simultaneously, the square of the difference between the numerical value and the confidence probability is obtained as a nonlinear mapping term. The detection penalty term and the nonlinear mapping term are then coupled together. This coupled operation logic transforms the feature evolution under strong compensation conditions into a quantified early warning score, thereby driving the automated takeover of physical redundancy backup and degradation logic.

[0011] Preferably, the steps of extracting source data from the UAV system design and constructing an initial heterogeneous topology graph of the system based on the source data, including a node set and an edge set with heterogeneous edge features, specifically include: reading and parsing a standard exchange format file exported from the system engineering design as source data; extracting component entities from the internal block graph in the standard exchange format file; instantiating the component entities as nodes in a directed heterogeneous graph; and constructing a directed edge set describing the coupling relationship between component entities based on the connector elements defined in the system modeling language model. The edge set is divided into physical conduction edges, signal transmission edges, and control feedback edges. When constructing control feedback edges, the control gain term of the model parameter graph in the standard exchange format file is extracted as a numerical label and bound to the edge set to form heterogeneous edge features.

[0012] Preferably, the step of extracting constraint parameters from each node in the node set specifically includes: extracting multi-dimensional constraint parameters for the nodes in the node set, combining the constraint parameters to construct a six-dimensional initial static feature vector, the six-dimensional initial static feature vector including type parameters, redundancy parameters, observability parameters, sampling frequency parameters, criticality parameters, and noise intensity parameters; extracting type parameters based on the stereotype attributes of the system modeling language, reading redundancy parameters from the instance multiplicity attributes of the internal block diagram, obtaining observability parameters from the reachability constraint labels in the demand diagram or state machine diagram, obtaining sampling frequency parameters by parsing the clock constraint block in the parameter diagram, calculating criticality parameters based on the logical series-parallel relationships in the reliability block diagram, and extracting noise intensity parameters from the tolerance definition of the attribute block.

[0013] Preferably, the step of performing concatenated fusion of constraint parameters and time series features to generate the initial node hidden state vector specifically includes: extracting the dynamic feature vector of the corresponding node at the current time as the time series feature; performing a feature concatenation and concatenation operation on the six-dimensional initial static feature vector and the dynamic feature vector to generate the concatenated feature vector; and inputting the concatenated feature vector into a fully connected layer for mapping processing to transform it into an initial node hidden state vector of a unified dimension.

[0014] Preferably, the step of transforming the real-time control compensation gain into aggregated attention weights in the graph space through the conditional edge feature mapping function specifically includes: when the edges in the edge set belong to physical conduction edges or signal transmission edges, the conditional edge feature mapping function outputs a learnable vector representing the edge type; when the edges in the edge set belong to control feedback edges, the conditional edge feature mapping function reads and concatenates the control gain term in real time, and uses a multilayer perceptron network containing a linear transformation matrix to upgrade the control gain term to a high-dimensional feature space to represent the real-time control compensation gain.

[0015] Preferably, the steps for the graph attention network to output the certainty probability that the corresponding node is in a compensated failure state specifically include: the graph attention network performs multi-layer feature aggregation calculation based on the aggregated attention weights to generate node spatial aggregated features; the node spatial aggregated features at the current time are input into the gated recurrent unit network, and the temporal state update operation is performed in combination with the historical hidden features at the previous time to extract the final spatiotemporal hidden features of the corresponding node; the spatiotemporal hidden features are input into the fully connected classification layer, and the certainty probability is calculated in combination with the logistic regression function.

[0016] Preferably, before inputting the initial node hidden state vector and heterogeneous edge features into the graph attention network, a step of training the graph attention network is also included. Specifically, this includes: extracting time-series features from the historical flight process as a dynamic temporal sliding window; performing a dual-criteria definition of the control masking period within the time window of the dynamic temporal sliding window, which includes the control gain surge criterion and the tracking error stationarity criterion; when the real-time monitoring control gain term satisfies the control gain surge criterion and the monitoring attitude tracking error satisfies the tracking error stationarity criterion, determining that the corresponding node is within the control masking period time window, and setting the value of the indicator function to a value of one; constructing a probe sensitivity loss function, introducing a probe sensitivity penalty term, and increasing the weight of nodes within the control masking period time window in the probe sensitivity loss function through the indicator function; and completing the parameter iteration of the graph attention network based on the probe sensitivity loss function.

[0017] Preferably, the step of nonlinearly coupling the detection penalty term with the confidence probability to generate a detection quantification score specifically includes: calculating the square of the difference between the numerical value and the confidence probability as a nonlinear mapping term to adjust the detection score, and introducing the detection penalty term into the nonlinear mapping term to obtain the detection evaluation index.

[0018] Preferably, the step of nonlinearly coupling the detection penalty term with the confidence probability to generate a detection quantification score further includes: rounding the detection evaluation index to the nearest integer and converting it into a level 1 to 10 evaluation index, and mapping the level 1 to 10 evaluation indexes into detection quantification scores.

[0019] Preferably, the steps for triggering the closed-loop response logic specifically include: reading the system configuration model associated with the compensatory failure state and obtaining the redundancy parameter in the constraint parameters of the corresponding node; if the redundancy parameter is greater than or equal to a value of two, activating the redundancy backup mechanism and blocking the data interaction port of the corresponding physical component, and enabling the backup physical component to take over the system's flight control; if the redundancy parameter is equal to a value of one, activating the task degradation mechanism and forcibly terminating the high-maneuverability flight mission currently being executed by the UAV system, and instructing the UAV system to enter the safe return mode or emergency landing mode.

[0020] This invention provides a method for evaluating the detectability of unmanned aerial vehicles (UAVs) based on MBSE-GNN collaborative driving. It has the following beneficial effects: 1. This invention, through the cascaded fusion steps of setting constraint parameters and time series features, maps the static topology structure constructed based on UAV system source data with the dynamic physical state during flight. The fusion design unifies the system's design boundary attributes and real-time sensing feedback in the graph node space, enabling the network model to comprehensively process multi-source operational data based on the physical connection topology, providing a complete data foundation for the subsequent extraction and evaluation of node hidden state vectors.

[0021] 2. By setting a conditional edge feature mapping function, this invention directly transforms the extracted real-time control compensation gain into the aggregated attention weight of the graph space during the parameter calculation process of the graph attention network. The network structure configuration enables the graph inference calculation to perform high-dimensional spatial mapping specifically for the abnormal gains generated by the control loop. This overcomes the problem that the underlying fault features of the closed-loop control system are masked and diluted due to distributed margin compensation, and improves the accuracy of determining the certain probability that a node is in a compensated failure state.

[0022] 3. This invention sets up a nonlinear coupled calculation logic that includes a detection penalty term, calculates the penalty value based on the real-time control compensation gain, and combines the penalty value with the confidence probability output by graph inference to map it into a detection quantified score. This transforms the weak feature evolution under strong compensation into a direct risk warning indicator, corrects the risk of missed detection caused by relying solely on a single threshold in the traditional method, and can accurately trigger the UAV's redundant backup or mission degradation closed-loop response logic based on the quantified score. Attached Figure Description

[0023] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a flowchart of the method of the present invention; Figure 3 This is a block definition diagram of the unmanned aerial vehicle system of the present invention; Figure 4 This is a block diagram of the internal structure of the unmanned aerial vehicle system of the present invention; Figure 5 This is the core logic of the MBSE-GNN detectivity automatic evaluation method of the present invention; Figure 6 This is a comparison chart of expert scores and system pre-evaluation scores under different modes of the present invention. Detailed Implementation

[0024] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] See attached document Figure 1 This invention provides a UAV FMECA detectability assessment system based on MBSE-GNN collaborative drive, including a topology construction module, a feature fusion module, a graph reasoning module, and a detectability measurement module. The topology construction module connects to the system model engineering tool to extract design parameters and collects UAV flight control sensor data and actuator commands through the airborne communication bus.

[0026] The topology building module has a built-in model-to-graph conversion unit, which transforms the system architecture diagram into a directed heterogeneous graph containing prior attributes. The feature fusion module is connected to the topology building module and receives static design constraint vectors and flight sequence data streams.

[0027] The feature fusion module aligns the static design constraint vector with the flight sequence data stream in the temporal domain and splices them spatially. The spliced ​​data is then mapped to the hidden state features of the initial nodes through an embedding layer. The graph inference module is connected to the feature fusion module and is equipped with a graph attention network.

[0028] The graph reasoning module receives the initial node hidden state features and the heterogeneous edge features containing real-time control gain information. Based on the initial node hidden state features and the heterogeneous edge features, it outputs the certainty probability that each node is in a failed state. The probe quantization module is connected to the graph reasoning module and is used to calculate the probe penalty term.

[0029] The detection quantification module maps the confidence probability to the detection penalty term to generate a detection score. Based on the detection score, the detection quantification module sends the corresponding control logic trigger command to the flight control system.

[0030] See attached document Figure 2 This invention provides a method for evaluating the detectability of unmanned aerial vehicles (UAVs) based on MBSE-GNN collaborative driving, comprising the following steps: Extract the source data of the UAV system design, construct an initial system heterogeneous topology graph containing node set and edge set based on the stereotype and constraint block definition in the source data, and extract the constraint parameters of physical layer, perception layer and logic layer nodes from the node set.

[0031] The time-series features of the UAV during flight are collected. These features include attitude observations, command assignment weights, and actuator physical feedback. The time-series features are mapped to corresponding nodes based on the connection relationships of the heterogeneous topology graph of the initial system. The execution constraint parameters are cascaded and fused with the time-series features to generate the hidden state vector of the initial node required by the graph neural network.

[0032] The initial node hidden state vector and heterogeneous edge features are input into the graph attention network. The real-time gain variables generated by control loop compensation are transformed into aggregate attention weights in the graph space through a preset conditional edge feature mapping function. The graph attention network outputs the certain probability that the corresponding node in the system is in a compensation failure state.

[0033] Extract the real-time control compensation gain to maintain flight steady state, calculate the detection penalty term based on the real-time control compensation gain, and perform nonlinear coupling calculation between the detection penalty term and the confidence probability to generate a detection quantification score.

[0034] The system determines the numerical range of the detection quantification score. If the detection quantification score is lower than the preset safety threshold, it maintains normal status monitoring and alarm. If the detection quantification score is greater than or equal to the preset safety threshold, it sends a warning command to the flight control bus, triggering distributed redundant link backup or mission degradation logic.

[0035] The process of constructing an initial heterogeneous topology graph of the system, which includes a set of nodes and a set of edges, involves the following steps: reading and parsing a standard exchange format file exported from the system engineering design.

[0036] The standard exchange format file specifically adopts the Extensible Markup Language Metadata Exchange Format (XMI format file). The file contains system modeling language model elements of the UAV architecture. The extraction program reads the component entities in the internal block graph and instantiates the component entities as nodes of the directed heterogeneous graph.

[0037] For methods of parsing and extracting the underlying code of standard exchange format files, those skilled in the art can call existing extensible markup language parsing libraries to write scripts to implement it. The process of traversing document tree nodes and extracting tags is a well-known technology in this field.

[0038] Based on the connector elements defined in the system modeling language model, a set of directed edges describing the coupling relationships between components is constructed. The mathematical structure of the directed heterogeneous graph is as follows: ; In the formula, This represents a directed anisomorphic graph. This represents the collection of extracted functional component nodes. This represents the set of directed edges connecting components.

[0039] To accurately characterize the fault suppression effect generated by the closed-loop control of the UAV system, the directed edge set is divided into three categories of edge attributes with physical semantics. The first category of edge attributes is defined as physical conduction edges.

[0040] Physical transmission edges are mainly used to quantify the attenuation of mechanical vibration or heat source failure between different physical structural components. The second type of edge attribute is defined as signal transmission edge, which describes the transmission characteristics of sensor sampling data on the communication bus, such as delay and bandwidth.

[0041] The communication bus specifically adopts the uORB bus, a micro publish-subscribe bus commonly used in UAV systems, and the third type of edge attribute is defined as a control feedback edge. The control feedback edge is responsible for identifying the control masking effect after a fault occurs.

[0042] When constructing control feedback edges, the reading tool extracts the control gain term from the model parameter graph and binds it to the connection edge as a numerical label. The control gain term represents the flight control system controller's ability to compensate for the output deviation of the physical actuator.

[0043] Based on the definition and classification of the above three types of heterogeneous edge attributes, the topological constraints of the UAV system design are completely transformed into the network structure relationship required for graph neural network calculation. This network structure relationship provides a physical link basis for tracing the subsequent hidden failure path.

[0044] See attached document Figure 3 To further illustrate the above topology construction process, Figure 3 The provided Block Definition Diagram (BDD) shows that the architecture adopts a top-down three-layer combinatorial logical relationship, aiming to provide standardized component entities and prior constraints for the subsequent construction of the heterogeneous topology diagram of the system through a model-driven approach. The top layer in the diagram is the overall block of the UAV system; the middle layer is rigorously divided into a physical layer, a perception layer, and a logical control layer based on physical semantics and signal flow, standardizing the functional positioning of components within the system. In the bottom-level functional nodes, the BDD defines in detail 12 key functional components, including the power unit, core control unit, and state estimator. These components are instantiated into a set of nodes in the initial heterogeneous topology diagram during the algorithm execution phase.

[0045] Combination Figure 4 The provided Internal Block Diagram (IBD) shows that... Figure 4 The provided Internal Block Graph (IBD) details the coupling and interaction relationships among the 12 core functional components of the UAV system, providing a physical link foundation for constructing an initial system heterogeneous topology graph with heterogeneous edge characteristics.

[0046] exist Figure 4In this system, the internal logical interactions are defined into three types of edge attributes with clear physical semantics: physical conduction edges, signal transmission edges, and control feedback edges. Specifically, the connection between the sensor group (including IMU, GPS, and Alt) and the state estimator (EKF2), and the communication between EKF2 and the core control loop (FlightCtrl), constitute the signal transmission edges. The path extending from the power unit (Power) to the power and sensor components is defined as the physical conduction edge, used to quantify the transmission attenuation of energy distribution and environmental disturbances in the physical topology. Crucially, the command output path from the core control loop to the mixer and then to the propulsion system, and the dynamic intervention link between health monitoring (HealthMon) and failsafe, constitute the control feedback edges.

[0047] By extracting the control gain terms on these control feedback edges, the system can obtain the real-time control compensation gain for maintaining flight steady state. This edge feature extraction mechanism based on IBD topology enables the graph attention network to perceive the feature masking effect caused by distributed compensation. The extracted dynamic gain is transformed into the aggregated attention weight of the graph space through the conditional edge feature mapping function, thereby penetrating the signal dilution barrier generated by the compensation process and realizing the accurate calculation of the certainty probability of the fault state.

[0048] After constructing the initial heterogeneous topology graph of the system, the extraction program extracts multi-dimensional constraint parameters for the functional component nodes in the directed heterogeneous graph. These multi-dimensional constraint parameters constitute a six-dimensional initial static feature vector. The specific mathematical expression of the six-dimensional initial static feature vector is defined as follows: ; In the formula, Represents a node The six-dimensional initial static eigenvector; Represents a node Type parameters; Represents a node Redundancy parameters; Represents a node Observability parameters; Represents a node The sampling frequency parameter; Represents a node Keyness parameters; Represents a node Noise intensity parameters, Indicates the transpose operation; It represents a six-dimensional real space.

[0049] Type parameters are extracted based on the stereotype attributes of the system modeling language. The type parameters distinguish the physical entity categories corresponding to the nodes. The physical entity categories specifically include sensor nodes, actuator nodes, and computing unit nodes. The extraction program maps different physical entity categories to discrete values. The specific implementation of the discrete values ​​adopts one-hot encoding.

[0050] The redundancy parameter is read from the instance multiplicity attribute of the internal block diagram. The redundancy parameter reflects the number of hardware backups of functional components in the UAV system. The instance multiplicity attribute directly corresponds to the dual-redundancy configuration state or triple-redundancy configuration state in the flight control system.

[0051] Obtain observability parameters from the reachability constraint labels in the demand diagram or state machine diagram. The observability parameters describe whether the state of a functional component can be directly measured by a sensor. For hidden mechanical structures without direct sensor monitoring, the observability parameters are assigned a value of zero.

[0052] The sampling frequency parameter is obtained by parsing the clock constraint block in the parameter diagram. The sampling frequency parameter determines the time granularity of data refresh. After normalization calculation, the sampling frequency parameter is mapped into the numerical range of zero to one.

[0053] The criticality parameter is calculated based on the logical series-parallel relationships in the reliability block diagram. The criticality parameter indicates the degree of impact of functional component failure on the overall flight safety of the UAV. For the series nodes of the control core attitude calculation, the criticality parameter is assigned the highest weight of one.

[0054] The noise intensity parameter is extracted from the tolerance definition of the self-attributed block. This parameter characterizes the inherent physical fluctuation level of sensor measurement data or actuator output data, providing a boundary for subsequently defining subtle faults and anomalies. For the numerical standardization of the six-dimensional initial static eigenvector, those skilled in the art can use existing data normalization function libraries to complete the code. Scale scaling of multidimensional data is a well-known technique in this field.

[0055] The system extracts time-series data frames generated during real-time UAV flight. For nodes in the node set, it extracts the dynamic feature vectors of each node at the current moment. The dynamic feature vectors specifically include the attitude angle observations and acceleration observations of sensor nodes and the pulse width modulation output values ​​of actuator nodes. A dynamic and static feature embedding and fusion mechanism is introduced to perform feature concatenation and splicing operations on the initial six-dimensional static feature vector and the dynamic feature vector.

[0056] The concatenated feature vectors are input into a fully connected layer for mapping, transforming them into initial node hidden state vectors of uniform dimension. The preferred value range for this uniform dimension is 64 to 128 dimensions. The mathematical expression for the initial node hidden state vector is: ; In the formula, Represents a node The initial node at level zero hides the state vector; This represents the learnable weight matrix of the fully connected layer; Represents a node The six-dimensional initial static eigenvector; This represents the vector concatenation operator; Represents a node At the present moment The dynamic feature vector; This represents the bias vector of the fully connected layer.

[0057] The feature concatenation operation establishes a numerical correlation between discrete system design constraints and continuous flight physical states. The fully connected layer aligns constraint parameters of different dimensions with time-series features to the same high-dimensional feature space. The hidden state vectors of the initial nodes form the input basis for subsequent graph attention network data inference operations. For the initialization of weight parameters and configuration of bias vectors in the fully connected layer, those skilled in the art can call the built-in functions of existing deep learning frameworks, specifically the PyTorch or TensorFlow frameworks. The dimensionality transformation and matrix multiplication operations of feature vectors are well-known techniques in this field.

[0058] The graph inference module receives initial node hidden state features and heterogeneous edge features containing real-time control gain information. Internally, the graph inference module is configured with a graph attention network. This network processes the received initial node hidden state features and heterogeneous edge features. The attention weights between nodes in the graph attention network are determined not only by the node's own features but also by edge attribute features injected through a pre-defined conditional edge feature mapping function. The mathematical expression for the attention weight calculation mechanism of the graph attention network is as follows: ; In the formula, Indicates the first Layer neighbor nodes For nodes Attention weight coefficient; Represents an exponential function; This represents a linear activation function with leakage correction. This represents the learnable parameter vector in the attention mechanism; Indicates the transpose operation; Indicates the first The learnable weight transformation matrix of the layer; Represents a node In the The hidden state vector of the layer; Representing neighboring nodes In the The hidden state vector of the layer; This represents the vector concatenation operator; This represents the edge attribute conditional mapping function; Indicates from neighboring nodes Pointing to node The edge; Represents the nodes in the graph The set of neighboring nodes; Indicates the traversal index of the neighboring nodes; Representing neighboring nodes In the The hidden state vector of the layer; Indicates from neighboring nodes Pointing to node The edge.

[0059] The edge attribute conditional mapping function adopts a conditional feature injection mechanism. When the edge belongs to a physical conduction edge or a signal transmission edge, the edge attribute conditional mapping function only outputs a learnable vector that represents the edge type. The learnable vector is a randomly initialized low-dimensional floating-point array.

[0060] When an edge belongs to a control feedback edge, the edge attribute condition mapping function reads and concatenates the control gain term in real time. The control gain term represents the flight control system controller's ability to compensate for the output deviation of the physical actuator.

[0061] The edge attribute conditional mapping function maps the control gain term to a feature tensor of the same dimension as the hidden state vector. The mapping operation of the feature tensor is specifically implemented through a multilayer perceptron network containing a linear transformation matrix. The multilayer perceptron network increases the dimension of the one-dimensional control gain term to a high-dimensional feature space.

[0062] By relying on a conditional feature injection mechanism, graph attention networks (GANs) perceive control features that surge due to abnormal compensation in complex network structures. GANs transform these control features into aggregated attention weights in the graph space, thus overcoming the masking effect caused by closed-loop control. For the forward propagation computation code of the activation function in GANs, those skilled in the art can utilize the built-in operators of existing deep learning frameworks, specifically PyTorch or TensorFlow. The gradient differentiation and backpropagation update of the activation function are well-known techniques in this field.

[0063] Graph attention networks obtain attention weight coefficients based on an attention weight calculation mechanism. Then, based on these attention weight coefficients, the graph attention network performs a weighted aggregation operation on the features of neighboring nodes. The mathematical expression for this weighted aggregation operation is: ; In the formula, Represents a node In the The hidden state vector of the layer; Represents a non-linear activation function; Represents the nodes in the graph The set of neighboring nodes; This represents the traversal index in the set of neighboring nodes; Indicates the first Layer neighbor nodes For nodes Attention weight coefficient; Indicates the first The learnable weight transformation matrix of the layer; Representing neighboring nodes In the Hidden state vectors of the layer; nonlinear activation function Specifically, the exponential linear unit activation function is used, which can alleviate the gradient vanishing problem during the training process of deep networks.

[0064] After the multi-layer feature aggregation calculation is completed, the graph attention network outputs the node space aggregation features. The node space aggregation features represent the local topological coupling state of the UAV system. In order to capture the temporal evolution pattern in the UAV flight sequence, the graph inference module is configured with a gated cyclic unit network.

[0065] The gated recurrent unit network takes the node spatial aggregation features at the current moment as input, and combines them with the historical hidden features of the previous moment to perform a temporal state update operation. After the temporal state update operation, the graph reasoning module extracts the final spatiotemporal hidden features of the nodes.

[0066] The graph reasoning module inputs spatiotemporal hidden features into the fully connected classification layer. The fully connected classification layer, combined with a logistic regression function, calculates the certainty probability that each node is in a compensated failure state. The mathematical expression for calculating the certainty probability is: ; In the formula, Represents a node The certainty of being in a state of compensatory failure; This represents the logistic regression activation function; This represents the weight matrix of the fully connected classification layer; Represents a node go through The final spatiotemporal hidden features after layer spatiotemporal update; This represents the total number of layers in the graph attention network; This represents the bias coefficient of the fully connected classification layer.

[0067] The confidence probability is mapped to a continuous numerical range from zero to one. The higher the confidence probability, the greater the risk of a hidden fault in the physical component corresponding to the node. The confidence probability is output to the downstream detection quantification module for subsequent evaluation calculation.

[0068] For writing the code logic of tensor state update inside the gated recurrent unit network, those skilled in the art can call the built-in network layer interface of existing deep learning frameworks to implement it. Existing deep learning frameworks specifically include the PyTorch framework or the TensorFlow framework. The calculation of backpropagation error of the neural network classification layer is a well-known technology in this field.

[0069] The probe quantization module receives the confidence probability output by the graph reasoning module and simultaneously extracts the historical time series data of the flight control system. The probe quantization module establishes a dynamic time series sliding window on the historical time series data. The dynamic time series sliding window is used to capture the local flight control data stream before the current moment.

[0070] The detection quantization module performs physical boundary delineation on the local flight control data stream. This physical boundary delineation defines the masking period as the time window from the moment of failure to when the flight control system completely loses its steady state. Within this time window, the detection quantization module executes a dual-criteria definition for the control masking period. These dual criteria include the control gain surge criterion and the tracking error stationarity criterion. The control gain surge criterion is used to determine whether the controller's compensatory output exhibits abnormal gain. The mathematical expression for the control gain surge criterion is: ; In the formula, Indicates the difference in control gain; Indicates the current time Real-time monitoring and control gain term; This represents the baseline control gain under normal, fault-free operating conditions. This represents the preset control gain surge threshold. The specific value range is from 0.1 to 0.3.

[0071] The baseline control gain term is calculated using a dynamic timing sliding window. The probe quantization module extracts historical gain data under fault-free conditions within the dynamic timing sliding window. The probe quantization module then performs a smoothed average of the historical gain data to derive the baseline control gain term. The tracking error stationarity criterion is used to determine whether the system track has not yet deviated. The mathematical expression for the tracking error stationarity criterion is: ; In the formula, Represents the vector norm operator; This indicates the monitoring of attitude tracking error; This represents the preset tracking error tolerance threshold. The specific value range is from 0.01 radians to 0.05 radians.

[0072] When the real-time monitoring control gain term meets the control gain surge criterion and the monitoring attitude tracking error meets the tracking error stationarity criterion, the detection quantization module assigns the current masking state indication value to the value one. The masking state indication value equal to the value one indicates that the failure characteristics of the physical components have been temporarily masked by the closed-loop regulation effect of the flight controller.

[0073] The detection quantification module statistically analyzes the masking status indicator value within the dynamic time-series sliding window, and calculates the duration of the control masking period based on the statistical results. The mathematical expression for the duration of the control masking period is as follows: ; In the formula, Indicates the duration of the control cover period; Indicates the length of the dynamic timing sliding window; Dynamic timing sliding window length The value ranges from 50 to 200 sampling periods; This represents the index of discrete time steps within a dynamic time-series sliding window; Indicates the current time; This indicates the sampling time step of the drone's sensors. The value ranges from 10 milliseconds to 50 milliseconds; Indicates at discrete time step The masking status indicator value.

[0074] For the code writing of mean filtering and threshold comparison of arrays within a dynamic time-series sliding window, those skilled in the art can call existing open-source numerical computing extension libraries to implement it. Specifically, the existing open-source numerical computing extension libraries use the NumPy arithmetic library, and the integral accumulation loop logic of discrete time steps is a well-known technology in this field.

[0075] In UAV hidden failure detection missions, traditional binary cross-entropy loss focuses on overall prediction accuracy, but it is difficult to characterize the importance of detection confidence during the fault masking period.

[0076] The probe quantification module constructs a probe-sensitive loss function during the model training phase. The mathematical expression of the probe-sensitive loss function is: ; In the formula, This represents the probe sensitivity loss function. This represents the total number of nodes participating in the training. Represents the traversal index in the node set; Indicates the first The weight coefficients of each node; Indicates the first The true state label of each node; Indicates the first The predicted certainty probability that a node is in a state of compensatory failure; Represents a logarithmic function.

[0077] The detectivity-sensitive loss function introduces a detectivity-sensitive penalty term to compensate for the weight of nodes that fail during the masking period. The mathematical expression for the weight coefficient of a node is: ; In the formula, Indicates the first The weight coefficients of each node; Indicates the detection of sensitive penalty items; Detection of sensitive penalty items The value of is 7; The indicator function takes the value of one when the node is within the control cover-up time window. The specific criterion for determining whether a node is within the control cover-up time window is that the relevant node simultaneously satisfies both the control gain surge criterion and the tracking error stationarity criterion. When the node is not within the control cover-up time window, the indicator function takes the value of zero.

[0078] The probe quantization module increases the weight of the masking period node in the total loss through the indicator function. The weight increase operation enhances the response strength of the model gradient update to the masking period failure feature. The probe quantization module strengthens the model's ability to extract weak physical conflict signals under control gain compensation behavior.

[0079] The detectability-sensitive loss function solves the problem of judgment bias during the masking period of UAV faults, and the weight compensation logic effectively reduces the probability of missed detection when the flight control system is in a strong compensation state.

[0080] For the forward computation and backpropagation code writing of the probe sensitivity loss function, those skilled in the art can call the loss function interface provided by the existing deep learning framework to rewrite the implementation. The existing deep learning frameworks specifically include the PyTorch framework or the TensorFlow framework. The iterative update of network weights based on gradient descent is a well-known technology in this field.

[0081] The detection quantification module constructs an evaluation model that integrates predictive confidence and physical compensation characteristics. The physical logic of the evaluation model lies in the fact that the observability of failure is limited by the compensation depth of the closed-loop control system. Based on the evaluation model, the detection quantification module calculates the final detection evaluation index. The mathematical expression for the detection evaluation index is as follows: ; In the formula, Indicates the detectability assessment index; This represents the rounding function; Indicates the certainty of node failure; This indicates a detection penalty.

[0082] The node failure confidence probability is output by the graph inference module. The node failure confidence probability represents the degree of confidence of the graph inference module in the current failure state of the system. The probe quantification module adjusts the probe score using a nonlinear mapping term.

[0083] The nonlinear mapping term is represented by the square of the difference between the probability of node failure and the probability of certainty. During the fault masking period, the nonlinear mapping term achieves risk warning by increasing the detectability score. The detection penalty term is used to quantify and control the contribution of the masking effect to the detection difficulty. The mathematical expression for the detection penalty term is: ; In the formula, This indicates the preset sensitivity coefficient; sensitivity coefficient The value range is from 1 to 3; This indicates real-time control of compensatory gain; This indicates the preset maximum output gain threshold of the controller. The value is determined based on the physical limits of the drone actuator, and ranges from 0.8 to 1.

[0084] The aforementioned preset control gain surge threshold, tracking error tolerance threshold, sensitivity coefficient, and controller maximum output gain threshold were all obtained by those skilled in the art through joint tuning based on historical fault-free test flight data and engineering experience of the UAV. The detection quantification module introduces a compensation strength constraint through a detection penalty term. The compensation strength constraint ensures that when the UAV controller is in a strong compensation phase, the detection quantification module determines that the detection difficulty is relatively high, and the detection penalty term meets the criteria requirements of traditional reliability analysis standards that are difficult to identify through conventional observation.

[0085] The probe metric module transforms abstract feature vectors into evaluation metrics ranging from one to ten. When a probe metric is less than or equal to 4, the probe metric module determines that the current node is in a manifest failure state. A manifest failure state indicates that the physical component failure is easily detectable.

[0086] When the detection rate evaluation index is greater than or equal to 7, the detection rate quantification module determines that the current node is in a hidden failure state. The hidden failure state indicates that the fault is masked by the control feedback closed loop and belongs to a high-risk state. The determination result of the hidden failure state is used to trigger downstream redundancy backup or task degradation logic.

[0087] For the scalar operations and conditional branch decision code for detectability evaluation metrics, those skilled in the art can use existing basic programming language control flow statements, specifically Python or C++.

[0088] After acquiring the detectability evaluation index, the detection quantification module generates a status evaluation report. The status evaluation report is sent to the flight control computing unit via the UAV's internal bus. The UAV's internal bus specifically adopts a micro-aircraft communication protocol or an asynchronous object request proxy bus. The flight control computing unit continuously parses the detectability evaluation index in the status evaluation report. When the detectability evaluation index is greater than or equal to the value of seven, the flight control computing unit triggers the closed-loop response logic.

[0089] The closed-loop response logic includes a redundancy backup mechanism and a task degradation mechanism. The flight control computing unit reads the system configuration model associated with the fault node, and the flight control computing unit synchronously obtains the redundancy parameters in the initial feature vector corresponding to the fault node.

[0090] If the redundancy parameter is greater than or equal to two, the flight control computing unit determines that the faulty node has a usable hardware redundancy link. The flight control computing unit activates the redundancy backup mechanism and blocks the data interaction port of the physical component corresponding to the faulty node. The flight control computing unit simultaneously enables the backup physical component to take over the flight control of the system.

[0091] If the redundancy parameter is equal to one, the flight control computing unit determines that the faulty node lacks a hardware redundancy link. The flight control computing unit activates the mission degradation mechanism and forcibly terminates the high-maneuverability flight mission currently being executed by the UAV system. The flight control computing unit instructs the UAV system to enter the safe return mode or emergency landing mode.

[0092] The safe return-to-home mode limits the maximum cruise speed and maximum roll maneuver angle of the UAV system. This operational limitation prevents the UAV system from losing control due to compensatory overload, and the closed-loop triggering mechanism completes the conversion of physical node failure assessment results into aircraft control commands.

[0093] See attached document Figure 5 , Figure 5 The core logic of the MBSE-GNN automatic detection evaluation method demonstrated achieves an automated closed loop from system architecture modeling to detection index quantification through three key stages.

[0094] In the first stage, the system performs MBSE-driven modeling and heterogeneous feature injection. Through an automated model-to-graph (M2G) conversion algorithm, the PX4 flight architecture and SysML system model are deeply analyzed and mapped into a full system topology containing 12 typical heterogeneous nodes. In this stage, six-dimensional static parameters are extracted from the SysML attribute library to construct the initial feature vectors of the nodes. At the same time, three types of heterogeneous edge relationships are defined: physical conduction, signal transmission, and control feedback.

[0095] The second stage focuses on failure detection inference based on graph attention network (GAT). GAT is used to perform deep spatial inference on the failure characteristics of 12 nodes to deal with the control masking effect caused by complex closed-loop control. By learning the weight distribution of heterogeneous edges in failure propagation, especially the feedback edges bound to the real-time control gain term, the network can penetrate the signal dilution barrier generated by the distributed compensation process and keenly capture the abnormal control compensation behavior behind the seemingly pseudo-stationary sensor data.

[0096] In the third stage, the system completes the final quantization through the detectivity score mapping function. It performs nonlinear coupling calculation between the failure significance probability output by GNN and the coverage of observation nodes (or detection penalty term). Based on the quantization mapping mechanism of global attention weight, the system transforms the abstract inference results into 1-10 level detectivity scores that conform to the FMECA standard.

[0097] For the code writing of the flight control computing unit bus message publishing and subscription communication, those skilled in the art can call the existing flight control open source system communication interface. The existing flight control open source system communication interface adopts the PX4 system firmware communication interface. The general input and output port level control of the microcontroller is a well-known technology in the field and will not be described in detail here.

[0098] The computer device includes a processor, memory, and a communication interface, and is configured to execute task instructions related to the aforementioned latent failure detection.

[0099] The processor, memory, and communication interface exchange data and transmit power through an internal system bus. This internal system bus can be either a peripheral component interconnect bus or an internal integrated circuit bus. The internal system bus ensures the communication operation of all physical hardware modules within the computer device. The processor provides control and numerical computation power to the computer device. A processor may include a central processing unit (CPU), a graphics processing unit (GPU), or a dedicated neural network processing unit. The processor calls and executes pre-written computer program code stored in memory.

[0100] When the processor executes the computer program code, it performs the initial graph structure construction operation based on the aforementioned model-based systems engineering and the failure certainty probability prediction operation of the graph reasoning model. When the processor executes the computer program code, it simultaneously completes the dynamic timing sliding window dual-criteria delimitation logic operation and the probe quantification evaluation operation.

[0101] The memory provides storage space for computer program code and temporary computational data streams. Specifically, the memory includes random access memory, read-only memory, or non-volatile solid-state drives. The memory stores the weight parameters of the trained graph inference network and the UAV system configuration model. The communication interface is responsible for data transmission and reception between the computer device and external flight control system nodes. Specifically, the communication interface adopts a controller area network bus transceiver or a universal asynchronous transceiver. The computer device relies on the communication interface to receive historical flight data and send out the final status assessment report.

[0102] This invention also provides a non-volatile computer-readable storage medium storing a set of computer-executable instructions. The set of computer-executable instructions is read and executed by a processor to implement the aforementioned method for detecting hidden failures of unmanned aerial vehicles.

[0103] The specific physical form of a computer-readable storage medium includes a read-only optical disc, a flash drive, or a portable hard disk drive. The code implementation of the underlying data read / write addressing logic of the storage medium is accomplished by those skilled in the art by calling the existing operating system's underlying application programming interface (API). The calling logic of the existing operating system's underlying API is a well-known technology in this field.

[0104] Specific application examples: To verify the effectiveness of the proposed MBSE-GNN-based UAV FMECA detectability assessment method in solving the design blind spot problem caused by the fault masking effect due to distributed control compensation under complex and highly redundant architecture, this embodiment is based on the zero-shot pre-evaluation application scenario of a novel hexarotor UAV with power redundancy, and is explained in detail in combination with blind evaluation scoring data from human FMECA experts.

[0105] In the application scenario of this embodiment, in order to examine the system's ability to generalize to complex heterogeneous topologies, the original single power node was deconstructed into 6 independent power execution nodes through topology deconstruction, and a system topology diagram containing 17 heterogeneous nodes was constructed. The system extracted the internal block diagram of the new model and found that the control output of the Mixer node (N8) evolved into 6 parallel pulse width modulation (PWM) control edges, and the redundancy parameter of the underlying power system was set to 2.

[0106] Implementation of dual-criteria determination for dynamic control of cover-up period: In the simulated test conditions, a partial failure of a power motor is set (such as a decrease in motor efficiency leading to a 25% reduction in output power).

[0107] Actual conditions: Due to the high redundancy level of the system with a redundancy parameter of 2, the other 5 active control paths quickly perform margin compensation. The feedback from conventional sensors (inertial navigation unit and GPS) is in a pseudo-stationary state, and traditional monitoring systems that rely on residual thresholds are prone to missing alarms.

[0108] System determination: The detection quantization module extracts the real-time monitoring and control gain term at the current moment. The normal reference control gain term was measured within the dynamic timing sliding window. The system calculates the control gain difference based on the control gain surge criterion formula. Due to the calculation The control gain surge threshold is greater than the preset threshold. (This embodiment is set) The surge criterion is then met.

[0109] Simultaneously, the system extracts and monitors attitude tracking errors. radian, satisfying , Less than the preset tracking error tolerance threshold (This embodiment is set) (Radian), to determine if the system meets the stability criterion, the system assigns the current cover status indicator value to the value one, confirming that the UAV system has entered a highly deceptive control cover period.

[0110] Feature dilution calculation and probe metric evaluation implementation: After confirming that the system is in the control masking period, the system performs probe metric calculation to counteract the feature dilution effect caused by distributed compensation.

[0111] Actual conditions: Three senior engineers with more than 5 years of operation and maintenance experience conducted a blind evaluation of the failure condition of the power motor. Based on physical intuition, the expert group believed that a single-point power failure under the redundancy parameter of 2 would produce a noticeable physical residual. The average detection score given was 5.33 points (considered to be of medium detection difficulty and easy to detect).

[0112] Data Analysis and Calculation: When processing topology, the graph reasoning module of this invention outputs the node failure certainty probability because the abnormal characteristics of failed nodes are diluted by the characteristics of multiple normal nodes. If relying solely on traditional confidence prediction, the system will also make incorrect judgments. However, this system introduces a probe penalty term. To quantify the contribution of the control masking effect.

[0113] Set sensitivity coefficient Controller maximum output gain threshold The system is based on the penalty formula. Calculation Subsequently, the detection quantification module uses the evaluation model formula. The internal high-precision calculation yields a value of 10 × (1 0.20) 2 +1.72 = 10 × 0.64 + 1.72 = 8.12 (this value is the MBSE-GAT predicted score before rounding). After rounding, the final detectivity evaluation index is output. .

[0114] Multi-indicator evaluation and closed-loop trigger control implementation: To comprehensively verify the accuracy of the system's assessment under various failure modes, this embodiment simultaneously recorded the system's pre-assessment scores for two other typical failure conditions (decreased elevator performance and abnormal aileron feedback link), and compared them with the blind evaluation data from the expert group. The results were compiled into the following comparison table of the detection assessment results for the new hexacoach UAV: Table 1. Comparison of Detection Assessment Results for New Hexacopter UAVs Comparative conclusions: Combining the table and appendix Figure 6 As can be seen, the two line trends representing the average expert score and the pre-evaluation score are consistent. The bar chart in the figure shows that the absolute errors of the two sets of working conditions are only 0.38 and 0.45, indicating that under normal working conditions without being deeply masked by multiple physical redundancies, the detection quantification score output by this system is highly consistent with the experience judgment of senior experts, verifying the reliability of the system's basic evaluation logic.

[0115] However, in the high-redundancy operating condition region of the power motor failure, the data trend showed a significant difference. The line representing the pre-evaluation score of this system climbed to 8.12 here, far higher than the line representing the average expert score at 5.33, forming an absolute error of 2.79. Combined with the analysis of the control mechanism in the implementation, it can be seen that the huge error is not due to the inaccuracy of the system calculation, but because the redundancy parameter of the power node of the model is 2. The multi-path parallel gain compensation makes the external sensors present a false impression of weak physical residuals, thus affecting the intuition of experts. This system, relying on the attention mechanism inside the graph reasoning module, successfully penetrated the feature dilution effect and determined that the operating condition belongs to a high-risk hidden failure (i.e., design blind spot) that is difficult to detect manually.

[0116] Closed-loop action: After obtaining the high-risk detection rate assessment index of 8.12 (rounded to 8, which is greater than the preset safety threshold of 7), the flight control computing unit reads the power node redundancy parameter and determines that there is a usable hardware redundancy link. It immediately activates the redundancy backup mechanism, blocks the data interaction port of the faulty motor, and reconstructs the mixed control output matrix of the remaining 5 motors. The automated closed-loop command safely eliminates the potential crash hazard caused by the strong control masking period, realizing a complete safety closed loop from accurate identification to timely suppression of the hidden failure of complex redundant UAVs.

[0117] Summary of application examples: In the face of complex and highly redundant architectures, traditional FMECA assessments, which rely on human intuition, are easily misled by the compensatory effects of closed-loop systems, leading to misjudgments. This system not only closely matches expert scores under conventional failure modes but also quantifies the depth of feature masking by distributed compensation. The system breaks through the blind spots of expert experience assessments with zero flight samples and achieves a complete safety closed loop from accurate identification to timely suppression of latent failures of complex and redundant UAVs through automated closed-loop commands.

Claims

1. A UAV FMECA detectability assessment method based on MBSE-GNN collaborative driving, characterized in that, Includes the following steps: Extract source data for UAV system design, and construct an initial system heterogeneous topology graph based on the source data, which includes a node set and an edge set with heterogeneous edge features. The node set contains multiple nodes. Constraint parameters of each node are extracted from the node set, time series features of the UAV during flight are collected, and real-time control compensation gain is extracted from the time series features. The time series features are mapped to the corresponding nodes in the node set, and the constraint parameters and the time series features are cascaded and fused to generate an initial node hidden state vector. The initial node hidden state vector and the heterogeneous edge features are input into the graph attention network. The real-time control compensation gain is transformed into the aggregate attention weight of the graph space through the conditional edge feature mapping function. The graph attention network outputs the certain probability that the node is in the compensation failure state. The detection penalty term is calculated based on the real-time control compensation gain, and the detection penalty term is nonlinearly coupled with the confidence probability to generate a detection quantification score. The numerical range of the probe quantification score is determined. If the probe quantification score is greater than or equal to a preset safety threshold, a closed-loop response logic is triggered.

2. The UAV FMECA detectability assessment method based on MBSE-GNN collaborative driving as described in claim 1, characterized in that, The steps of extracting source data for UAV system design and constructing an initial heterogeneous topology graph of the system based on the source data, including a set of nodes and a set of edges with heterogeneous edge features, specifically include: Read and parse the standard exchange format file exported from the system engineering design as the source data, extract the component entities of the internal block diagram in the standard exchange format file, instantiate the component entities as the nodes of the initial system heterogeneous topology graph, and all the nodes constitute the node set; Based on the connector elements defined in the system modeling language model, a set of directed edges describing the coupling relationship between the component entities is constructed as the edge set, which is divided into physical conduction edges, signal transmission edges, and control feedback edges. When constructing the control feedback edge, the control gain term of the model parameter graph in the standard exchange format file is extracted as a numerical label and bound to the control feedback edge to form the heterogeneous edge feature.

3. The UAV FMECA detectability assessment method based on MBSE-GNN collaborative driving as described in claim 1, characterized in that, The steps for extracting the constraint parameters of each node from the node set specifically include: For the nodes in the node set, extract multi-dimensional constraint parameters, and combine the constraint parameters to construct a six-dimensional initial static feature vector. The six-dimensional initial static feature vector includes type parameter, redundancy parameter, observability parameter, sampling frequency parameter, criticality parameter, and noise intensity parameter. The type parameter is extracted from the source data based on the stereotype attributes of the system modeling language; the redundancy parameter is read from the instance multiplicity attribute of the internal block diagram contained in the source data; the observability parameter is obtained from the reachability constraint label of the demand diagram or state machine diagram contained in the source data; the sampling frequency parameter is obtained by parsing the clock constraint block of the parameter diagram contained in the source data; the criticality parameter is calculated based on the logical serial-parallel relationship in the reliability block diagram contained in the source data; and the noise intensity parameter is extracted from the tolerance definition of the attribute block contained in the source data.

4. The UAV FMECA detectability assessment method based on MBSE-GNN collaborative driving as described in claim 3, characterized in that, The specific steps of performing the concatenated fusion of the constraint parameters and the time series features to generate the initial node hidden state vector include: Extract the dynamic feature vector corresponding to the node at the current time from the time series features, and perform feature concatenation and splicing operation on the six-dimensional initial static feature vector and the dynamic feature vector to generate the spliced ​​feature vector; The concatenated feature vector is input into a fully connected layer for mapping processing, and transformed into the initial node hidden state vector of a unified dimension.

5. The UAV FMECA detectability assessment method based on MBSE-GNN collaborative driving as described in claim 2, characterized in that, The specific steps of transforming the real-time control compensation gain into aggregate attention weights in the graph space using the conditional edge feature mapping function include: When the edges in the edge set belong to physical conduction edges or signal transmission edges, the conditional edge feature mapping function outputs a learnable vector representing the edge type; When an edge in the edge set belongs to the control feedback edge, the conditional edge feature mapping function reads and concatenates the control gain term in real time. The control gain term is then upgraded to a high-dimensional feature space as a control feature through a multilayer perceptron network containing a linear transformation matrix, representing the real-time control compensation gain. The graph attention network perceives the control feature and transforms it into the aggregated attention weights in the graph space.

6. The UAV FMECA detectability assessment method based on MBSE-GNN collaborative driving as described in claim 1, characterized in that, The steps for the graph attention network to output the certain probability that the node is in a compensatory failure state specifically include: The graph attention network performs multi-layer feature aggregation calculations based on the aggregated attention weights to generate node space aggregated features. The node spatial aggregation features at the current moment are input into the gated recurrent unit network, and the temporal state update operation is performed by combining the historical hidden features of the previous moment to extract the final spatiotemporal hidden features of the corresponding node. The spatiotemporal hidden features are input into a fully connected classification layer, and the confidence probability is calculated by combining the logistic regression function.

7. The UAV FMECA detectability assessment method based on MBSE-GNN collaborative driving as described in claim 1, characterized in that, Before the step of inputting the initial node hidden state vector and the heterogeneous edge features into the graph attention network, the method further includes a step of training the graph attention network, specifically including: A dynamic time-series sliding window is established on the time-series data of historical flight processes. The dynamic time-series sliding window is used to capture local flight control data streams. Within the time window of the dynamic time-series sliding window, the dual criteria for control cover period are defined. The dual criteria for control cover period include the control gain surge criterion and the tracking error stationarity criterion. When the real-time monitoring control gain term extracted from the local flight control data stream satisfies the control gain surge criterion, and the extracted monitoring attitude tracking error satisfies the tracking error stationarity criterion, it is determined that the corresponding node is within the control cover period time window, and the value of the indicator function is set to the value of one. A probe sensitivity loss function is constructed, and a probe sensitivity penalty term is introduced. The weight of the node in the probe sensitivity loss function within the control masking period time window is increased through the indicator function. The parameters of the graph attention network are iterated based on the probe sensitivity loss function.

8. The UAV FMECA detectability assessment method based on MBSE-GNN collaborative driving as described in claim 1, characterized in that, The step of non-linearly coupling the detection penalty term with the confidence probability to generate a detection quantification score specifically includes: The square of the difference between the numerical value and the certainty probability is calculated as a nonlinear mapping term to adjust the detectivity score. The detection penalty term is then introduced into the nonlinear mapping term to obtain the detectivity evaluation index.

9. The UAV FMECA detectability assessment method based on MBSE-GNN collaborative driving as described in claim 8, characterized in that, The step of non-linearly coupling the detection penalty term with the confidence probability to generate a detection quantification score further includes: The detectability evaluation index is rounded to the nearest integer and then converted into a level 1 to 10 evaluation index. The level 1 to 10 evaluation indexes are then mapped to the detectability quantitative score.

10. The UAV FMECA detectability assessment method based on MBSE-GNN collaborative driving as described in claim 1, characterized in that, The steps for triggering the closed-loop response logic specifically include: Read the system configuration model associated with the compensation failure state and obtain the redundancy parameter in the constraint parameters corresponding to the node; If the redundancy parameter is greater than or equal to two, the redundancy backup mechanism is activated and the data interaction port of the corresponding physical component is blocked, and the backup physical component is enabled to take over the flight control of the system. If the redundancy parameter is equal to one, the mission degradation mechanism is activated and the high-maneuverability flight mission currently being executed by the UAV system is forcibly terminated, and the UAV system is instructed to enter the safe return mode or emergency landing mode.

Images