An intelligent defect diagnosis method for unmanned vehicle hybrid software system based on mechanism-embedded graph neural network

Abstract

The application provides an intelligent defect diagnosis method for a hybrid software system of an unmanned vehicle based on a mechanism-embedded graph neural network, mainly including construction of a hybrid system defect propagation graph, multi-modal node feature representation, design of a mechanism-embedded graph neural network, and defect positioning and tracing output. The application introduces an intelligent diagnosis algorithm based on a graph neural network, combines system prior mechanism and operation data, constructs a structured defect propagation graph, and designs a graph neural network with a mechanism-oriented information propagation mechanism on the basis, so that the modeling and diagnosis of a defect propagation path in a hybrid software system are realized, deep learning and accurate reasoning of defect behaviors are realized, and technical support is provided for reliability evaluation and fault troubleshooting of the software system of the unmanned vehicle. The application can systematically reveal the cross-modal propagation law of defects under a hybrid architecture, construct multi-modal defect diagnosis data resources, and realize accurate diagnosis by using a graph neural network with mechanism and data fusion.

Description

Technical Field

[0001] This invention belongs to the field of intelligent unmanned vehicle software system diagnostic technology, specifically relating to an intelligent defect diagnosis method for unmanned vehicle hybrid software systems based on mechanism-embedded graph neural networks. Background Technology

[0002] As modern warfare evolves towards intelligence, unmanned vehicles (such as unmanned combat vehicles and drone swarms) have become a crucial component of combat systems. The core competitiveness of these unmanned vehicles increasingly relies on the intelligence and reliability of their software systems. Generally, the software system of an unmanned vehicle constitutes a closed-loop control system through multiple modules including sensing, decision-making, and control. Common modules such as perception, decision-making / planning, and execution / control play key roles in the "Observe-Orient-Decide-Act (OODA)" cycle. The software system is the "brain" and "nerve center" of the unmanned vehicle; any software defect can lead to operational failure or even serious consequences. For example, in a simulated exercise, the visual perception module of an unmanned ground vehicle misidentified ground obstacles under poor lighting and smoke interference. This misidentification was adopted by the subsequent decision-making logic, ultimately causing the vehicle to deviate from its intended path and the mission to fail. Similar real-world cases clearly demonstrate that software defect diagnosis is crucial for ensuring the reliability of unmanned vehicles.

[0003] Currently, autonomous vehicles generally adopt a hybrid architecture of "intelligent algorithms + traditional algorithms". Intelligent algorithms (such as deep learning and reinforcement learning algorithms) endow autonomous vehicles with the ability to perceive complex environments and make autonomous decisions; traditional algorithms (such as rule-based control algorithms and classical signal processing algorithms) dominate in control, navigation, and signal processing stages with high deterministic and high reliability requirements. This hybrid system can use intelligent algorithms to process complex environmental information in the perception stage and use traditional algorithms to ensure predictable system responses in the control stage. However, it is precisely because of the deep coupling of these two types of heterogeneous algorithms that the system's defective behaviors exhibit higher complexity.

[0004] Traditional software defect diagnosis techniques mainly include static analysis, dynamic monitoring and analysis, model-based diagnosis, and data-driven methods. Static analysis relies on source code or design documents and can detect programming errors and design inconsistencies, but it struggles to capture defects caused by runtime interactions. Dynamic monitoring and model-based diagnosis methods, through analysis of runtime logs, sensor data, and physical models, can identify obvious failures in the system, but they are insufficient in identifying hidden defects (such as deep logic errors and misjudgments caused by data adversarial mechanisms). In recent years, data-driven methods such as deep learning have been used for fault detection and diagnosis, automatically extracting features from massive amounts of multidimensional data, but they often lack constraints on physical mechanisms and system structure. The advantages and disadvantages of these existing methods can be summarized as follows: 1) Static analysis and rule-based methods: The advantages are clear logic and high verifiability; the disadvantages are that it is difficult to analyze intelligent modules with non-deterministic outputs and cannot capture hidden defects under environmental interference.

[0005] 2) Model-based diagnostics: Relying on accurate system mathematical models and sensor data, it can monitor known failure modes in real time, but it is difficult to deal with new defects outside the design and is sensitive to model uncertainties.

[0006] 3) Data-driven diagnostic methods: These methods have automatic feature learning capabilities and can handle multimodal big data. However, they are highly dependent on training data and are susceptible to sample scarcity and distribution drift. They also lack the utilization of prior mechanisms of the system, resulting in poor interpretability.

[0007] For hybrid software systems, traditional diagnostic methods are increasingly limited. On the one hand, the non-deterministic output of intelligent algorithms (such as the image recognition results of deep neural networks) and the deterministic control logic of traditional algorithms form a complex interface. Defect propagation paths often span multiple components, are hidden, and non-linear, making it difficult to track them using techniques based on single-modality or single-component analysis. On the other hand, due to the scarcity of fault samples in real adversarial environments, data-driven models struggle to obtain sufficient training; while purely static or model-driven methods lack an understanding of the learning biases of intelligent modules, making it difficult to design diagnostic strategies for hybrid architectures. Therefore, existing diagnostic systems for autonomous vehicle hybrid software systems often suffer from "inaccurate diagnosis and incomplete source tracing," urgently requiring new technological means to reveal the patterns of defect propagation and accurately locate the root causes of faults. Summary of the Invention

[0008] (a) Technical problems to be solved This invention proposes an intelligent defect diagnosis method for unmanned vehicle hybrid software systems based on mechanistic embedded graph neural networks. This method addresses the technical challenges of how to quantitatively model and diagnose the complex and hidden defect propagation behavior in hybrid software systems; how to generate and enhance diagnostic data through intelligent algorithms when a large number of real fault samples are lacking; and how to deeply integrate system mechanism knowledge with data-driven learning models to obtain diagnostic results that are both highly accurate and interpretable.

[0009] (II) Technical Solution To address the aforementioned technical problems, this invention proposes an intelligent defect diagnosis method for autonomous vehicle hybrid software systems based on mechanism-embedded graph neural networks. This method includes the following steps: S1. Construct a defect propagation diagram for a hybrid software system. Based on the OODA closed-loop concept, the functional modules of the autonomous vehicle hybrid software system to be diagnosed are divided and analyzed, and the hybrid software system is divided into... N Each key module represents a node; determine the interaction relationships and defect propagation paths between modules, and establish a directed defect propagation topology graph, where nodes represent functional components or subsystems of the system, and edges represent possible defect propagation paths. Define a directed graph ,in express Each module This represents a directed interaction between modules; for each pair of modules... If there is a risk of defect propagation, that is, in Add directed edges to For each edge Assign initial weights , used to indicate from component To Component The probability or intensity of defect propagation, with an initial weight of [value]. ; S2. Multimodal node feature extraction For each module node, multi-source data is collected from the perception, decision-making, and control stages to form a feature vector reflecting the state and behavior. The above features are concatenated or fused to obtain the overall feature vector of each node. ,in Indicates dynamic runtime characteristics, The module's mechanistic characteristics are represented; the features of all nodes form a feature matrix. ; S3. Graph Neural Networks with Embedded Design Mechanisms After obtaining the system topology and node features, a graph neural network is introduced to complete defect diagnosis. The constructed defect propagation graph and multimodal node features are input into the graph neural network to realize intelligent reasoning for defect propagation and localization. S4. Defect Location and Source Tracing Output After several layers of inference in the graph neural network, the final layer of the graph neural network outputs the final representation of each node. Based on this, one or more fully connected layers are used to classify or score nodes for defects, mapping node representations to corresponding defect probabilities or fault scores. Threshold judgment or maximum likelihood decision-making is used to locate faulty nodes in the hybrid system, and the message passing attention coefficients of the graph neural network are analyzed. This allows us to infer the propagation path of defects in the graph, thereby enabling fault tracing.

[0010] Furthermore, the key modules include a perception module, a decision-making module, and a control module; wherein, the perception module includes cameras and radar, the decision-making module includes a task planning module and a path selection module, and the control module includes a navigation controller and an actuator controller.

[0011] Furthermore, for the sensing module, the sensor output is converted into vector features by a feature encoder to obtain the sensing module feature vector. For the decision-making module, statistical features are extracted from system logs and command signals. These features are then fused to generate a node feature vector, which describes the current state and behavior pattern of the decision-making module, thus obtaining the decision-making module feature vector. For the control module, statistical features are extracted from system logs and command signals. These features are then fused to generate a node feature vector, which describes the current state and behavior pattern of the control module, thus obtaining the control module feature vector. .

[0012] Furthermore, in step S2, the prior mechanism features of each key module are extracted and incorporated into the node representation as additional features.

[0013] Furthermore, in step S3, the graph neural network is divided into a graph coding layer, a message passing layer, and an output layer; among which, The graph coding layer is based on the initial weights. Construct the initial adjacency matrix and with the characteristic matrix Both are included as input; The message passing layer consists of multiple layers of graph neural networks; nodes In the The hidden state of a layer is denoted as Its update formula is expressed as: (1) in, It is a non-linear activation function; Represents a node The set of neighboring nodes; This is the learnable weight matrix for this layer; The attention coefficient for message passing, used to dynamically adjust the attention of slave nodes. To the node Information flow intensity; No. In the layer, nodes The hidden state is For each edge Calculate the raw attention score To measure nodes For nodes In the The influence of the layer; The raw attention score is calculated using the following formula. : (2) Message Passing Attention Coefficient The calculation formula is as follows: (3) in, This represents a vector concatenation operation. For the first The matrix used for feature transformation in the layer. These are the vector parameters used to calculate the attention weights; Mechanism-guided aggregation: Initial weights Incorporating mechanistic knowledge into message passing; designing regularization terms to enable... More inclined to Consistent, or combined during updates : (4) in, This is a weighting factor used to smoothly integrate prior and learned attention.

[0014] Furthermore, in step S4, a linear mapping plus Softmax is used to obtain the failure probability distribution of each node: (5) in, , These are the output layer weights and biases, respectively. Representation Component The probability distribution of being in a fault state or fault mode. The dimension Represents a node The first The probability of a defective pattern or a healthy state.

[0015] (III) Beneficial Effects This invention proposes an intelligent defect diagnosis method for autonomous vehicle hybrid software systems based on mechanism-embedded graph neural networks. The method mainly includes constructing a defect propagation graph of the hybrid system, representing multimodal node features, designing a mechanism-embedded graph neural network, and outputting defect localization and source tracing. This invention introduces an intelligent diagnostic algorithm based on graph neural networks, combining prior system mechanisms and operational data to construct a structured defect propagation graph. Based on this graph, a graph neural network with a mechanism-guided information propagation mechanism is designed to model and diagnose the defect propagation path in the hybrid software system. This enables deep learning and accurate inference of defect behavior, providing technical support for reliability assessment and fault diagnosis of autonomous vehicle software systems.

[0016] This invention systematically reveals the cross-modal propagation patterns of defects in hybrid architectures, constructs multimodal defect diagnostic data resources, and utilizes graph neural networks that fuse mechanisms and data to achieve accurate diagnosis. Compared to existing technologies, this invention has the following main advantages and innovations: 1. Mechanism-embedded graph structure design: For the first time, the operating mechanism of the autonomous vehicle hybrid software system is mapped to the OODA closed-loop structure as a graph model, and prior information such as defect propagation probability is directly embedded in the graph, so as to realize the respect and utilization of physical mechanism in the diagnostic process and improve the reliability and interpretability of diagnostic results.

[0017] 2. Learnable attention propagation mechanism: Through dynamic weighted message passing via the attention mechanism, it can adaptively capture unknown or changing defect propagation patterns. At the same time, combined with prior weights, it ensures that the diagnosis conforms to known constraints, which can overcome the defect of traditional deep models that are prone to "black box" decision-making.

[0018] 3. Multimodal feature fusion: Multiple sensor outputs, command information and mechanism data are fused into node features, enabling the diagnostic model to understand comprehensive information, which is more comprehensive than traditional single-modal methods.

[0019] 4. Precise location and traceable diagnosis: This invention not only provides probabilistic predictions of faulty components, but also explicitly constructs fault propagation paths using attention weights, which helps in root cause analysis and the formulation of subsequent corrective strategies.

[0020] In summary, this invention proposes a mechanism-data collaboration graph neural network fault diagnosis method for autonomous vehicle hybrid software systems. This method can fill the gap in the existing technology for defect diagnosis in scenarios with deep integration of heterogeneous algorithms, and provides new ideas and methods for the reliability analysis of intelligent autonomous vehicles. Attached Figure Description

[0021] Figure 1 This is a flowchart of the intelligent defect diagnosis method for the unmanned vehicle hybrid software system of the present invention. Detailed Implementation

[0022] To make the objectives, contents, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples.

[0023] This embodiment proposes an intelligent defect diagnosis method for autonomous vehicle hybrid software systems based on mechanistic embedded graph neural networks. Its main process is as follows: Figure 1 As shown, the specific steps include the following: S1. Construct a defect propagation diagram for a hybrid software system. Based on the OODA closed-loop concept, the functional modules of the autonomous vehicle hybrid software system to be diagnosed are divided and analyzed, including the perception module, decision-making module, and control module in the hybrid software system. N Each key module represents a node. The perception module includes cameras and radar, the decision-making module includes task planning and path selection, and the control module includes navigation controllers and actuator controllers.

[0024] Based on system design and professional knowledge, identify the possible interactions and defect propagation paths between modules, and establish a directed defect propagation topology graph. Nodes in the graph represent functional components or subsystems of the system, and edges represent potential defect propagation paths.

[0025] For example, an error in the output of the perception module may affect the decision-making module, and an error in the decision-making module may be transmitted to the control module. These relationships can be abstracted as edges in a graph.

[0026] Define a directed graph ,in express Each module This represents directed interactions between modules. To reflect cross-modal propagation characteristics, for each pair of modules... If there is a risk of defect propagation, that is, in Add directed edges to By analyzing the system design documents and operating principles, and combining prior knowledge or historical data, each edge... Assign initial weights , used to indicate from component To Component The probability or intensity of defect propagation, with an initial weight of [value]. It can also be obtained through engineering experience or statistical analysis.

[0027] For example, suppose the perception module To the decision module If there is an information flow, then add an edge. Decision module To the control module If there is a decision output, then add an edge. Regarding defect propagation, if data such as the probability of false recognition in visual sensing and the robustness of the algorithm are known, then... Assign a non-zero value.

[0028] The defect propagation topology diagram constructed in this step can reflect the structural characteristics of the interaction between intelligent algorithms and traditional algorithms in the hybrid system and the possible flow of defects, providing basic topology, physical and logical constraints and mechanism information for subsequent diagnostic models.

[0029] S2. Multimodal node feature extraction For each module node, multi-source data is collected from sensing, decision-making, and control processes to form feature vectors reflecting state and behavior. .

[0030] 1) For the perception module, sensor outputs (such as images and radar data) are converted into vector features using a feature encoder (convolutional network, descriptor extraction, etc.). For sensors such as cameras and infrared sensors, feature extraction is performed on the raw data (images, video streams) using a deep neural network (e.g., a pre-trained convolutional neural network) to obtain the perception module feature vector. , indicating time Visual input features; or empirical features (edge ​​detection, histograms, etc.).

[0031] 2) For decision-making modules, statistical features are extracted using system logs, command signals, and other information. These features are then fused to generate node feature vectors, which describe the current state and behavior patterns of the module. For example, in the path planning module, features are extracted from task logs, policy outputs, and internal module state variables. For instance, discrete action commands are encoded into one-dimensional or multi-dimensional vectors, or a sequence feature encoder is used to process historical command sequences to obtain the decision-making module feature vector. .

[0032] 3) For control modules, statistical features are extracted using system logs, command signals, and other information. These features are then fused to generate node feature vectors, which describe the current state and behavior patterns of the module. For example, for speed controllers and servo controllers, the feature vector of the control module is calculated using sampled feedback control signals, actuator states (speed, acceleration), and error values. .

[0033] Furthermore, prior mechanistic features such as functional constraints or model characteristics (e.g., sensor accuracy range, controller response function, etc.) of each module can be extracted from a mechanistic perspective and incorporated as additional features into the node representation. For example, sensor noise model parameters and measurement accuracy range can be added to sensor nodes; control gain, stability limits, and other technical characteristics can be added to controller nodes. The aforementioned mechanistic information can be represented as static features attached to the corresponding nodes.

[0034] The above features are concatenated or fused to obtain the overall feature vector of each node. ,in Indicates dynamic runtime characteristics, This represents the mechanistic characteristics of the module. All node features form a feature matrix. .

[0035] S3. Graph Neural Networks with Embedded Design Mechanisms After obtaining the system topology and node features, a specialized graph neural network is introduced to complete defect diagnosis. The defect propagation graph and multimodal node features constructed above are input into the graph neural network to achieve intelligent reasoning for defect propagation and localization. The graph neural network consists of a graph encoding layer, a message passing layer, and an output layer.

[0036] Graph coding layer: This layer is based on the initial weights. Construct the initial adjacency matrix and with the characteristic matrix These are also used as input. To enhance the utilization of mechanistic knowledge, nodes in the graph can be grouped according to function, and block structure weight matrices can be set, etc.

[0037] Message passing layer: Consists of a multi-layer graph neural network.

[0038] node In the The hidden state of a layer is denoted as Its update formula is expressed as: (1) in, It is a non-linear activation function (such as ReLU); Represents a node The set of neighboring nodes; This is the learnable weight matrix for this layer; The attention coefficient for message passing, used to dynamically adjust the attention of slave nodes. To the node Information flow intensity.

[0039] This update includes its own characteristic transformation term. and weighted aggregation items from neighbors By stacking multiple layers, nodes can fuse information from more distant neighbors, capturing the impact of defect propagation across modalities. Introducing an attention mechanism enables the network to adaptively capture defect propagation paths in complex and ever-changing operating environments.

[0040] No. In the layer, nodes The hidden state is (Initially, it can be set) For each edge Calculate the raw attention score To measure nodes For nodes In the The influence of layers. By introducing an attention mechanism, the network can adaptively capture defect propagation paths in complex and ever-changing operating environments.

[0041] The raw attention score is calculated using the following formula. : (2) Message Passing Attention Coefficient The calculation formula is as follows: (3) in, This represents a vector concatenation operation. For the first The matrix used for feature transformation in the layer. These are the vector parameters used to calculate the attention weights.

[0042] By concatenating prior information, initial weights are added to the attention input, thus enabling prior information to participate in the attention calculation. This allows system mechanism information to be directly integrated into the message passing process. Through this mechanism, graph neural networks can utilize data-driven feature associations and be guided by prior mechanistic knowledge during the learning process, thereby achieving efficient diagnosis in hybrid systems with both deterministic and uncertain components.

[0043] Mechanism-guided aggregation: Initial weights Mechanism knowledge is incorporated into message passing. By setting mechanisms, constraints on the attention coefficient can be added during training. For example, a regularization term can be designed to... More inclined to Consistent, or combined during updates : (4) in, The coefficient is used to smoothly fuse prior and learned attention. This strategy ensures that the network takes into account both known mechanisms and data features.

[0044] S4. Defect Location and Source Tracing Output After several layers of inference in the graph neural network, the final layer of the graph neural network outputs the final representation of each node. Building upon this, one or more fully connected layers are used to classify or score nodes for defects, mapping node representations to corresponding defect probabilities or fault scores. For example, a linear mapping combined with Softmax is used to obtain the fault probability distribution for each node: (5) in, , These are the output layer weights and biases, respectively. Representation Component The probability distribution of being in a fault state or fault mode. The dimension Represents a node The first The probability of a defective mode (or healthy state). Depending on the task requirements, the cross-entropy loss function can be used to optimize the network parameters during training.

[0045] Faulty nodes in a hybrid system can be located by using threshold judgment or maximum likelihood decision. Furthermore, the message-passing attention coefficients of graph neural networks can be analyzed. This allows us to infer the propagation path of defects in the graph, thus enabling fault tracing. Since the attention coefficient integrates data characteristics and mechanistic information, its maximum path can be considered as the most probable defect propagation link, which helps engineers further analyze the causes of defects.

[0046] In actual diagnosis, take The category corresponding to the maximum value in the middle is the diagnostic result for that node. For example, if If the corresponding category is "fault", then the decision node is determined. A fault exists. When multiple nodes are identified as faulty, attention weights can be further incorporated. A larger value indicates a decision node The impact is significant; and so on, eventually locating the damaged control node. This provides both fault localization and a clear causal chain of the fault, aiding in subsequent fault eradication and system improvement. The fault propagation chain can be represented as starting from a certain source node... (The root cause of the observed fault) reaches the final fault node after passing through several nodes. path In graph neural networks, this path corresponds to the set of attention edges with a large propagation component, which can be visualized as a directed subgraph with the largest weight. By observing the edges with strong attention, we can understand how the fault gradually spreads from the source along the OODA loop to other modules.

[0047] Through the above technical solutions, the present invention can systematically reveal the cross-modal propagation law of defects under hybrid architecture, construct multimodal defect diagnosis data resources, and achieve accurate diagnosis by using graph neural networks that fuse mechanism and data.

[0048] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent defect diagnosis of an unmanned vehicle hybrid software system based on a mechanism-embedded graph neural network, characterized in that, The intelligent defect diagnosis method for the unmanned vehicle hybrid software system includes the following steps: S1. Construct a defect propagation diagram for a hybrid software system. Based on the OODA closed-loop concept, the functional modules of the autonomous vehicle hybrid software system to be diagnosed are divided and analyzed, and the hybrid software system is divided into... N Each key module represents a node; determine the interaction relationships and defect propagation paths between modules, and establish a directed defect propagation topology graph, where nodes represent functional components or subsystems of the system, and edges represent possible defect propagation paths. Define a directed graph ,in express Each module This represents a directed interaction between modules; for each pair of modules... If there is a risk of defect propagation, that is, in Add directed edges to For each edge Assign initial weights , used to indicate from component To Component The probability or intensity of defect propagation, with an initial weight of [value]. ; S2. Multimodal node feature extraction For each module node, multi-source data is collected from the perception, decision-making, and control stages to form a feature vector reflecting the state and behavior. The above features are concatenated or fused to obtain the overall feature vector of each node. ,in Indicates dynamic runtime characteristics, The module's mechanistic characteristics are represented; the features of all nodes form a feature matrix. ; S3. Graph Neural Networks with Embedded Design Mechanisms After obtaining the system topology and node features, a graph neural network is introduced to complete defect diagnosis. The constructed defect propagation graph and multimodal node features are input into the graph neural network to realize intelligent reasoning for defect propagation and localization. S4. Defect Location and Source Tracing Output After several layers of inference in the graph neural network, the final layer of the graph neural network outputs the final representation of each node. Based on this, one or more fully connected layers are used to classify or score nodes for defects, mapping node representations to corresponding defect probabilities or fault scores. Threshold judgment or maximum likelihood decision-making is used to locate faulty nodes in the hybrid system, and the message passing attention coefficients of the graph neural network are analyzed. This allows us to infer the propagation path of defects in the graph, thereby enabling fault tracing.

2. The intelligent defect diagnosis method for unmanned vehicle hybrid software systems based on mechanistic embedded graph neural networks as described in claim 1, characterized in that, The key modules include a perception module, a decision-making module, and a control module; wherein, the perception module includes a camera and a radar, the decision-making module includes a task planning module and a path selection module, and the control module includes a navigation controller and an actuator controller.

3. The intelligent defect diagnosis method for unmanned vehicle hybrid software systems based on mechanistic embedded graph neural networks as described in claim 2, characterized in that, For the sensing module, the sensor output is converted into vector features by a feature encoder to obtain the sensing module feature vector. ; For the decision-making module, statistical features are extracted from system logs and command signals. These features are then fused to generate a node feature vector, which describes the current state and behavior pattern of the decision-making module, thus obtaining the decision-making module feature vector. For the control module, statistical features are extracted from system logs and command signals. These features are then fused to generate a node feature vector, which describes the current state and behavior pattern of the control module, thus obtaining the control module feature vector. .

4. The intelligent defect diagnosis method for unmanned vehicle hybrid software systems based on mechanistic embedded graph neural networks as described in claim 1, characterized in that, In step S2, the prior mechanism features of each key module are extracted and incorporated into the node representation as additional features.

5. The intelligent defect diagnosis method for unmanned vehicle hybrid software systems based on mechanistic embedded graph neural networks as described in claim 1, characterized in that, In step S3, the graph neural network is divided into a graph coding layer, a message passing layer, and an output layer; wherein, The graph coding layer is based on the initial weights. Construct the initial adjacency matrix and with the characteristic matrix Both are included as input; The message passing layer consists of multiple layers of graph neural networks; nodes In the The hidden state of a layer is denoted as Its update formula is expressed as: (1) in, It is a non-linear activation function; Represents a node The set of neighboring nodes; This is the learnable weight matrix for this layer; The attention coefficient for message passing, used to dynamically adjust the attention of slave nodes. To the node Information flow intensity; No. In the layer, nodes The hidden state is For each edge Calculate the raw attention score To measure nodes For nodes In the The influence of the layer; The raw attention score is calculated using the following formula. : (2) Message Passing Attention Coefficient The calculation formula is as follows: (3) in, This represents a vector concatenation operation. For the first The matrix used for feature transformation in the layer. These are the vector parameters used to calculate the attention weights; Mechanism-guided aggregation: Initial weights Incorporating mechanistic knowledge into message passing; designing regularization terms to enable... More inclined to Consistent, or combined during updates : (4) in, This is a weighting factor used to smoothly integrate prior and learned attention.

6. The intelligent defect diagnosis method for unmanned vehicle hybrid software systems based on mechanistic embedded graph neural networks as described in claim 5, characterized in that, In step S4, a linear mapping plus Softmax is used to obtain the failure probability distribution of each node: (5) in, , These are the output layer weights and biases, respectively. Representation Component The probability distribution of being in a fault state or fault mode. The dimension Represents a node The first The probability of a defective pattern or a healthy state.

Images